program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios16>(tensor<fp16, [1, 3, 224, 224]> pixel_values, tensor<fp16, [1, 1, 196, 64]> rope_cos, tensor<fp16, [1, 1, 196, 64]> rope_sin) {
            tensor<bool, []> var_15 = const()[name = tensor<string, []>("op_15"), val = tensor<bool, []>(true)];
            tensor<int32, []> var_17 = const()[name = tensor<string, []>("op_17"), val = tensor<int32, []>(1)];
            tensor<int32, []> var_20 = const()[name = tensor<string, []>("op_20"), val = tensor<int32, []>(2)];
            tensor<int32, []> var_21 = const()[name = tensor<string, []>("op_21"), val = tensor<int32, []>(-1)];
            tensor<string, []> var_57_pad_type_0 = const()[name = tensor<string, []>("op_57_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_57_strides_0 = const()[name = tensor<string, []>("op_57_strides_0"), val = tensor<int32, [2]>([16, 16])];
            tensor<int32, [4]> var_57_pad_0 = const()[name = tensor<string, []>("op_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_57_dilations_0 = const()[name = tensor<string, []>("op_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_57_groups_0 = const()[name = tensor<string, []>("op_57_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 3, 16, 16]> model_patch_embed_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_patch_embed_proj_weight_to_fp16"), val = tensor<fp16, [768, 3, 16, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [768]> model_patch_embed_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_patch_embed_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1179776)))];
            tensor<fp16, [1, 768, 14, 14]> var_57_cast_fp16 = conv(bias = model_patch_embed_proj_bias_to_fp16, dilations = var_57_dilations_0, groups = var_57_groups_0, pad = var_57_pad_0, pad_type = var_57_pad_type_0, strides = var_57_strides_0, weight = model_patch_embed_proj_weight_to_fp16, x = pixel_values)[name = tensor<string, []>("op_57_cast_fp16")];
            tensor<int32, [3]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [3]>([1, 768, 196])];
            tensor<fp16, [1, 768, 196]> var_58_cast_fp16 = reshape(shape = concat_0, x = var_57_cast_fp16)[name = tensor<string, []>("op_58_cast_fp16")];
            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<bool, []> x_3_interleave_0 = const()[name = tensor<string, []>("x_3_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 768]> model_cls_token_to_fp16 = const()[name = tensor<string, []>("model_cls_token_to_fp16"), val = tensor<fp16, [1, 1, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1181376)))];
            tensor<fp16, [1, 196, 768]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_58_cast_fp16)[name = tensor<string, []>("transpose_516")];
            tensor<fp16, [1, 197, 768]> x_3_cast_fp16 = concat(axis = var_17, interleave = x_3_interleave_0, values = (model_cls_token_to_fp16, x_1_cast_fp16))[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<fp16, [1, 197, 768]> model_pos_embed_to_fp16 = const()[name = tensor<string, []>("model_pos_embed_to_fp16"), val = tensor<fp16, [1, 197, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1182976)))];
            tensor<fp16, [1, 197, 768]> input_1_cast_fp16 = add(x = x_3_cast_fp16, y = model_pos_embed_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [1]> x_5_axes_0 = const()[name = tensor<string, []>("x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1485632)))];
            tensor<fp16, [768]> model_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1487232)))];
            tensor<fp16, []> var_6_to_fp16 = const()[name = tensor<string, []>("op_6_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 197, 768]> x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = model_blocks_0_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_0_norm1_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
            tensor<int32, [4]> var_198 = const()[name = tensor<string, []>("op_198"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_7_cast_fp16 = reshape(shape = var_198, x = x_5_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<int32, [4]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_206_pad_type_0 = const()[name = tensor<string, []>("op_206_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_206_strides_0 = const()[name = tensor<string, []>("op_206_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_206_pad_0 = const()[name = tensor<string, []>("op_206_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_206_dilations_0 = const()[name = tensor<string, []>("op_206_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_206_groups_0 = const()[name = tensor<string, []>("op_206_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1488832)))];
            tensor<fp16, [1, 768, 1, 197]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_515")];
            tensor<fp16, [1, 64, 1, 197]> var_206_cast_fp16 = conv(dilations = var_206_dilations_0, groups = var_206_groups_0, pad = var_206_pad_0, pad_type = var_206_pad_type_0, strides = var_206_strides_0, weight = model_blocks_0_attn_q_projs_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_206_cast_fp16")];
            tensor<int32, [1]> var_207_axes_0 = const()[name = tensor<string, []>("op_207_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_207_cast_fp16 = squeeze(axes = var_207_axes_0, x = var_206_cast_fp16)[name = tensor<string, []>("op_207_cast_fp16")];
            tensor<int32, [3]> q_1_perm_0 = const()[name = tensor<string, []>("q_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_214_pad_type_0 = const()[name = tensor<string, []>("op_214_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_214_strides_0 = const()[name = tensor<string, []>("op_214_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_214_pad_0 = const()[name = tensor<string, []>("op_214_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_214_dilations_0 = const()[name = tensor<string, []>("op_214_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_214_groups_0 = const()[name = tensor<string, []>("op_214_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1587200)))];
            tensor<fp16, [1, 64, 1, 197]> var_214_cast_fp16 = conv(dilations = var_214_dilations_0, groups = var_214_groups_0, pad = var_214_pad_0, pad_type = var_214_pad_type_0, strides = var_214_strides_0, weight = model_blocks_0_attn_k_projs_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_214_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = tensor<string, []>("op_215_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_215_cast_fp16 = squeeze(axes = var_215_axes_0, x = var_214_cast_fp16)[name = tensor<string, []>("op_215_cast_fp16")];
            tensor<int32, [3]> k_1_perm_0 = const()[name = tensor<string, []>("k_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_222_pad_type_0 = const()[name = tensor<string, []>("op_222_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_222_strides_0 = const()[name = tensor<string, []>("op_222_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_222_pad_0 = const()[name = tensor<string, []>("op_222_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_222_dilations_0 = const()[name = tensor<string, []>("op_222_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_222_groups_0 = const()[name = tensor<string, []>("op_222_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1685568)))];
            tensor<fp16, [1, 64, 1, 197]> var_222_cast_fp16 = conv(dilations = var_222_dilations_0, groups = var_222_groups_0, pad = var_222_pad_0, pad_type = var_222_pad_type_0, strides = var_222_strides_0, weight = model_blocks_0_attn_v_projs_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")];
            tensor<int32, [1]> var_223_axes_0 = const()[name = tensor<string, []>("op_223_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_223_cast_fp16 = squeeze(axes = var_223_axes_0, x = var_222_cast_fp16)[name = tensor<string, []>("op_223_cast_fp16")];
            tensor<int32, [3]> v_1_perm_0 = const()[name = tensor<string, []>("v_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1783936)))];
            tensor<fp16, [1, 197, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_207_cast_fp16)[name = tensor<string, []>("transpose_514")];
            tensor<fp16, [1, 197, 64]> q_3_cast_fp16 = add(x = q_1_cast_fp16, y = model_blocks_0_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_3_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1784128)))];
            tensor<fp16, [1, 197, 64]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_223_cast_fp16)[name = tensor<string, []>("transpose_512")];
            tensor<fp16, [1, 197, 64]> v_3_cast_fp16 = add(x = v_1_cast_fp16, y = model_blocks_0_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
            tensor<int32, [1]> q_5_axes_0 = const()[name = tensor<string, []>("q_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_5_cast_fp16 = expand_dims(axes = q_5_axes_0, x = q_3_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")];
            tensor<int32, [1]> k_3_axes_0 = const()[name = tensor<string, []>("k_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_215_cast_fp16)[name = tensor<string, []>("transpose_513")];
            tensor<fp16, [1, 1, 197, 64]> k_3_cast_fp16 = expand_dims(axes = k_3_axes_0, x = k_1_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
            tensor<int32, [1]> v_5_axes_0 = const()[name = tensor<string, []>("v_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_5_cast_fp16 = expand_dims(axes = v_5_axes_0, x = v_3_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
            tensor<int32, [4]> var_232_begin_0 = const()[name = tensor<string, []>("op_232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_232_end_0 = const()[name = tensor<string, []>("op_232_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_232_end_mask_0 = const()[name = tensor<string, []>("op_232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_232_cast_fp16 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = q_5_cast_fp16)[name = tensor<string, []>("op_232_cast_fp16")];
            tensor<int32, [4]> var_236_begin_0 = const()[name = tensor<string, []>("op_236_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_236_end_0 = const()[name = tensor<string, []>("op_236_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_236_end_mask_0 = const()[name = tensor<string, []>("op_236_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_236_cast_fp16")];
            tensor<int32, [4]> var_240_begin_0 = const()[name = tensor<string, []>("op_240_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_240_end_0 = const()[name = tensor<string, []>("op_240_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_240_end_mask_0 = const()[name = tensor<string, []>("op_240_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_240_cast_fp16 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = q_5_cast_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
            tensor<int32, [4]> var_244_begin_0 = const()[name = tensor<string, []>("op_244_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_244_end_0 = const()[name = tensor<string, []>("op_244_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_244_end_mask_0 = const()[name = tensor<string, []>("op_244_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_244_cast_fp16 = slice_by_index(begin = var_244_begin_0, end = var_244_end_0, end_mask = var_244_end_mask_0, x = k_3_cast_fp16)[name = tensor<string, []>("op_244_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_246_cast_fp16 = mul(x = var_240_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_246_cast_fp16")];
            tensor<int32, [5]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_11_cast_fp16 = reshape(shape = var_254, x = var_240_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [2]> var_256_split_sizes_0 = const()[name = tensor<string, []>("op_256_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_256_axis_0 = const()[name = tensor<string, []>("op_256_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_256_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_256_cast_fp16_1 = split(axis = var_256_axis_0, split_sizes = var_256_split_sizes_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_256_cast_fp16")];
            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = tensor<string, []>("squeeze_0_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_256_cast_fp16_0)[name = tensor<string, []>("squeeze_0_cast_fp16")];
            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = tensor<string, []>("squeeze_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_256_cast_fp16_1)[name = tensor<string, []>("squeeze_1_cast_fp16")];
            tensor<fp16, []> const_7_promoted_to_fp16 = const()[name = tensor<string, []>("const_7_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_259_cast_fp16 = mul(x = squeeze_1_cast_fp16, y = const_7_promoted_to_fp16)[name = tensor<string, []>("op_259_cast_fp16")];
            tensor<int32, []> x_13_axis_0 = const()[name = tensor<string, []>("x_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_13_cast_fp16 = stack(axis = x_13_axis_0, values = (var_259_cast_fp16, squeeze_0_cast_fp16))[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_266_cast_fp16 = reshape(shape = var_265, x = x_13_cast_fp16)[name = tensor<string, []>("op_266_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_267_cast_fp16 = mul(x = var_266_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_267_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_1_cast_fp16 = add(x = var_246_cast_fp16, y = var_267_cast_fp16)[name = tensor<string, []>("q_patches_1_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_269_cast_fp16 = mul(x = var_244_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_269_cast_fp16")];
            tensor<int32, [5]> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_17_cast_fp16 = reshape(shape = var_277, x = var_244_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<int32, [2]> var_279_split_sizes_0 = const()[name = tensor<string, []>("op_279_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_279_axis_0 = const()[name = tensor<string, []>("op_279_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_279_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_279_cast_fp16_1 = split(axis = var_279_axis_0, split_sizes = var_279_split_sizes_0, x = x_17_cast_fp16)[name = tensor<string, []>("op_279_cast_fp16")];
            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = tensor<string, []>("squeeze_2_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_279_cast_fp16_0)[name = tensor<string, []>("squeeze_2_cast_fp16")];
            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = tensor<string, []>("squeeze_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_279_cast_fp16_1)[name = tensor<string, []>("squeeze_3_cast_fp16")];
            tensor<fp16, []> const_15_promoted_to_fp16 = const()[name = tensor<string, []>("const_15_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_282_cast_fp16 = mul(x = squeeze_3_cast_fp16, y = const_15_promoted_to_fp16)[name = tensor<string, []>("op_282_cast_fp16")];
            tensor<int32, []> x_19_axis_0 = const()[name = tensor<string, []>("x_19_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_19_cast_fp16 = stack(axis = x_19_axis_0, values = (var_282_cast_fp16, squeeze_2_cast_fp16))[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<int32, [4]> var_288 = const()[name = tensor<string, []>("op_288"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_289_cast_fp16 = reshape(shape = var_288, x = x_19_cast_fp16)[name = tensor<string, []>("op_289_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_290_cast_fp16 = mul(x = var_289_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_290_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_1_cast_fp16 = add(x = var_269_cast_fp16, y = var_290_cast_fp16)[name = tensor<string, []>("k_patches_1_cast_fp16")];
            tensor<bool, []> var_293_interleave_0 = const()[name = tensor<string, []>("op_293_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_293_cast_fp16 = concat(axis = var_20, interleave = var_293_interleave_0, values = (var_232_cast_fp16, q_patches_1_cast_fp16))[name = tensor<string, []>("op_293_cast_fp16")];
            tensor<bool, []> var_296_interleave_0 = const()[name = tensor<string, []>("op_296_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_296_cast_fp16 = concat(axis = var_20, interleave = var_296_interleave_0, values = (var_236_cast_fp16, k_patches_1_cast_fp16))[name = tensor<string, []>("op_296_cast_fp16")];
            tensor<fp16, []> var_298_to_fp16 = const()[name = tensor<string, []>("op_298_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_9_cast_fp16 = mul(x = var_293_cast_fp16, y = var_298_to_fp16)[name = tensor<string, []>("q_9_cast_fp16")];
            tensor<bool, []> attn_1_transpose_x_1 = const()[name = tensor<string, []>("attn_1_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_1_transpose_y_1 = const()[name = tensor<string, []>("attn_1_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_1, transpose_y = attn_1_transpose_y_1, x = q_9_cast_fp16, y = var_296_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_3_cast_fp16 = softmax(axis = var_21, x = attn_1_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
            tensor<bool, []> var_303_transpose_x_0 = const()[name = tensor<string, []>("op_303_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_303_transpose_y_0 = const()[name = tensor<string, []>("op_303_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_303_cast_fp16 = matmul(transpose_x = var_303_transpose_x_0, transpose_y = var_303_transpose_y_0, x = attn_3_cast_fp16, y = v_5_cast_fp16)[name = tensor<string, []>("op_303_cast_fp16")];
            tensor<int32, [1]> var_304_axes_0 = const()[name = tensor<string, []>("op_304_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_304_cast_fp16 = squeeze(axes = var_304_axes_0, x = var_303_cast_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
            tensor<string, []> var_310_pad_type_0 = const()[name = tensor<string, []>("op_310_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_310_strides_0 = const()[name = tensor<string, []>("op_310_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_310_pad_0 = const()[name = tensor<string, []>("op_310_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_310_dilations_0 = const()[name = tensor<string, []>("op_310_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_310_groups_0 = const()[name = tensor<string, []>("op_310_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1784320)))];
            tensor<fp16, [1, 64, 1, 197]> var_310_cast_fp16 = conv(dilations = var_310_dilations_0, groups = var_310_groups_0, pad = var_310_pad_0, pad_type = var_310_pad_type_0, strides = var_310_strides_0, weight = model_blocks_0_attn_q_projs_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_310_cast_fp16")];
            tensor<int32, [1]> var_311_axes_0 = const()[name = tensor<string, []>("op_311_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_311_cast_fp16 = squeeze(axes = var_311_axes_0, x = var_310_cast_fp16)[name = tensor<string, []>("op_311_cast_fp16")];
            tensor<int32, [3]> q_11_perm_0 = const()[name = tensor<string, []>("q_11_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_318_pad_type_0 = const()[name = tensor<string, []>("op_318_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_318_strides_0 = const()[name = tensor<string, []>("op_318_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_318_pad_0 = const()[name = tensor<string, []>("op_318_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_318_dilations_0 = const()[name = tensor<string, []>("op_318_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_318_groups_0 = const()[name = tensor<string, []>("op_318_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1882688)))];
            tensor<fp16, [1, 64, 1, 197]> var_318_cast_fp16 = conv(dilations = var_318_dilations_0, groups = var_318_groups_0, pad = var_318_pad_0, pad_type = var_318_pad_type_0, strides = var_318_strides_0, weight = model_blocks_0_attn_k_projs_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_318_cast_fp16")];
            tensor<int32, [1]> var_319_axes_0 = const()[name = tensor<string, []>("op_319_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_319_cast_fp16 = squeeze(axes = var_319_axes_0, x = var_318_cast_fp16)[name = tensor<string, []>("op_319_cast_fp16")];
            tensor<int32, [3]> k_7_perm_0 = const()[name = tensor<string, []>("k_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_326_pad_type_0 = const()[name = tensor<string, []>("op_326_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_326_strides_0 = const()[name = tensor<string, []>("op_326_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_326_pad_0 = const()[name = tensor<string, []>("op_326_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_326_dilations_0 = const()[name = tensor<string, []>("op_326_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_326_groups_0 = const()[name = tensor<string, []>("op_326_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1981056)))];
            tensor<fp16, [1, 64, 1, 197]> var_326_cast_fp16 = conv(dilations = var_326_dilations_0, groups = var_326_groups_0, pad = var_326_pad_0, pad_type = var_326_pad_type_0, strides = var_326_strides_0, weight = model_blocks_0_attn_v_projs_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
            tensor<int32, [1]> var_327_axes_0 = const()[name = tensor<string, []>("op_327_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_327_cast_fp16 = squeeze(axes = var_327_axes_0, x = var_326_cast_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
            tensor<int32, [3]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2079424)))];
            tensor<fp16, [1, 197, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_311_cast_fp16)[name = tensor<string, []>("transpose_511")];
            tensor<fp16, [1, 197, 64]> q_13_cast_fp16 = add(x = q_11_cast_fp16, y = model_blocks_0_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2079616)))];
            tensor<fp16, [1, 197, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_327_cast_fp16)[name = tensor<string, []>("transpose_509")];
            tensor<fp16, [1, 197, 64]> v_9_cast_fp16 = add(x = v_7_cast_fp16, y = model_blocks_0_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
            tensor<int32, [1]> q_15_axes_0 = const()[name = tensor<string, []>("q_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_15_cast_fp16 = expand_dims(axes = q_15_axes_0, x = q_13_cast_fp16)[name = tensor<string, []>("q_15_cast_fp16")];
            tensor<int32, [1]> k_9_axes_0 = const()[name = tensor<string, []>("k_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_319_cast_fp16)[name = tensor<string, []>("transpose_510")];
            tensor<fp16, [1, 1, 197, 64]> k_9_cast_fp16 = expand_dims(axes = k_9_axes_0, x = k_7_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
            tensor<int32, [1]> v_11_axes_0 = const()[name = tensor<string, []>("v_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_11_cast_fp16 = expand_dims(axes = v_11_axes_0, x = v_9_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
            tensor<int32, [4]> var_336_begin_0 = const()[name = tensor<string, []>("op_336_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_336_end_0 = const()[name = tensor<string, []>("op_336_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_336_end_mask_0 = const()[name = tensor<string, []>("op_336_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_336_cast_fp16 = slice_by_index(begin = var_336_begin_0, end = var_336_end_0, end_mask = var_336_end_mask_0, x = q_15_cast_fp16)[name = tensor<string, []>("op_336_cast_fp16")];
            tensor<int32, [4]> var_340_begin_0 = const()[name = tensor<string, []>("op_340_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_340_end_0 = const()[name = tensor<string, []>("op_340_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_340_end_mask_0 = const()[name = tensor<string, []>("op_340_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, x = k_9_cast_fp16)[name = tensor<string, []>("op_340_cast_fp16")];
            tensor<int32, [4]> var_344_begin_0 = const()[name = tensor<string, []>("op_344_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_344_end_0 = const()[name = tensor<string, []>("op_344_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_344_end_mask_0 = const()[name = tensor<string, []>("op_344_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_344_cast_fp16 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = q_15_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")];
            tensor<int32, [4]> var_348_begin_0 = const()[name = tensor<string, []>("op_348_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_348_end_0 = const()[name = tensor<string, []>("op_348_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_348_end_mask_0 = const()[name = tensor<string, []>("op_348_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_348_cast_fp16 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, x = k_9_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_350_cast_fp16 = mul(x = var_344_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_350_cast_fp16")];
            tensor<int32, [5]> var_358 = const()[name = tensor<string, []>("op_358"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_23_cast_fp16 = reshape(shape = var_358, x = var_344_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<int32, [2]> var_360_split_sizes_0 = const()[name = tensor<string, []>("op_360_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_360_axis_0 = const()[name = tensor<string, []>("op_360_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_360_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_360_cast_fp16_1 = split(axis = var_360_axis_0, split_sizes = var_360_split_sizes_0, x = x_23_cast_fp16)[name = tensor<string, []>("op_360_cast_fp16")];
            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = tensor<string, []>("squeeze_4_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_360_cast_fp16_0)[name = tensor<string, []>("squeeze_4_cast_fp16")];
            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = tensor<string, []>("squeeze_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_360_cast_fp16_1)[name = tensor<string, []>("squeeze_5_cast_fp16")];
            tensor<fp16, []> const_23_promoted_to_fp16 = const()[name = tensor<string, []>("const_23_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_363_cast_fp16 = mul(x = squeeze_5_cast_fp16, y = const_23_promoted_to_fp16)[name = tensor<string, []>("op_363_cast_fp16")];
            tensor<int32, []> x_25_axis_0 = const()[name = tensor<string, []>("x_25_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_25_cast_fp16 = stack(axis = x_25_axis_0, values = (var_363_cast_fp16, squeeze_4_cast_fp16))[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<int32, [4]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_370_cast_fp16 = reshape(shape = var_369, x = x_25_cast_fp16)[name = tensor<string, []>("op_370_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_371_cast_fp16 = mul(x = var_370_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_371_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_3_cast_fp16 = add(x = var_350_cast_fp16, y = var_371_cast_fp16)[name = tensor<string, []>("q_patches_3_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_373_cast_fp16 = mul(x = var_348_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_373_cast_fp16")];
            tensor<int32, [5]> var_381 = const()[name = tensor<string, []>("op_381"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_29_cast_fp16 = reshape(shape = var_381, x = var_348_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<int32, [2]> var_383_split_sizes_0 = const()[name = tensor<string, []>("op_383_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_383_axis_0 = const()[name = tensor<string, []>("op_383_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_383_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_383_cast_fp16_1 = split(axis = var_383_axis_0, split_sizes = var_383_split_sizes_0, x = x_29_cast_fp16)[name = tensor<string, []>("op_383_cast_fp16")];
            tensor<int32, [1]> squeeze_6_axes_0 = const()[name = tensor<string, []>("squeeze_6_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_383_cast_fp16_0)[name = tensor<string, []>("squeeze_6_cast_fp16")];
            tensor<int32, [1]> squeeze_7_axes_0 = const()[name = tensor<string, []>("squeeze_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_383_cast_fp16_1)[name = tensor<string, []>("squeeze_7_cast_fp16")];
            tensor<fp16, []> const_31_promoted_to_fp16 = const()[name = tensor<string, []>("const_31_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_386_cast_fp16 = mul(x = squeeze_7_cast_fp16, y = const_31_promoted_to_fp16)[name = tensor<string, []>("op_386_cast_fp16")];
            tensor<int32, []> x_31_axis_0 = const()[name = tensor<string, []>("x_31_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_31_cast_fp16 = stack(axis = x_31_axis_0, values = (var_386_cast_fp16, squeeze_6_cast_fp16))[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<int32, [4]> var_392 = const()[name = tensor<string, []>("op_392"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_393_cast_fp16 = reshape(shape = var_392, x = x_31_cast_fp16)[name = tensor<string, []>("op_393_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_394_cast_fp16 = mul(x = var_393_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_394_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_3_cast_fp16 = add(x = var_373_cast_fp16, y = var_394_cast_fp16)[name = tensor<string, []>("k_patches_3_cast_fp16")];
            tensor<bool, []> var_397_interleave_0 = const()[name = tensor<string, []>("op_397_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_397_cast_fp16 = concat(axis = var_20, interleave = var_397_interleave_0, values = (var_336_cast_fp16, q_patches_3_cast_fp16))[name = tensor<string, []>("op_397_cast_fp16")];
            tensor<bool, []> var_400_interleave_0 = const()[name = tensor<string, []>("op_400_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_400_cast_fp16 = concat(axis = var_20, interleave = var_400_interleave_0, values = (var_340_cast_fp16, k_patches_3_cast_fp16))[name = tensor<string, []>("op_400_cast_fp16")];
            tensor<fp16, []> var_402_to_fp16 = const()[name = tensor<string, []>("op_402_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_19_cast_fp16 = mul(x = var_397_cast_fp16, y = var_402_to_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<bool, []> attn_5_transpose_x_1 = const()[name = tensor<string, []>("attn_5_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_5_transpose_y_1 = const()[name = tensor<string, []>("attn_5_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_1, transpose_y = attn_5_transpose_y_1, x = q_19_cast_fp16, y = var_400_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_7_cast_fp16 = softmax(axis = var_21, x = attn_5_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
            tensor<bool, []> var_407_transpose_x_0 = const()[name = tensor<string, []>("op_407_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_407_transpose_y_0 = const()[name = tensor<string, []>("op_407_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_407_cast_fp16 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = attn_7_cast_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("op_407_cast_fp16")];
            tensor<int32, [1]> var_408_axes_0 = const()[name = tensor<string, []>("op_408_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_408_cast_fp16 = squeeze(axes = var_408_axes_0, x = var_407_cast_fp16)[name = tensor<string, []>("op_408_cast_fp16")];
            tensor<string, []> var_414_pad_type_0 = const()[name = tensor<string, []>("op_414_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_414_strides_0 = const()[name = tensor<string, []>("op_414_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_414_pad_0 = const()[name = tensor<string, []>("op_414_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_414_dilations_0 = const()[name = tensor<string, []>("op_414_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_414_groups_0 = const()[name = tensor<string, []>("op_414_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2079808)))];
            tensor<fp16, [1, 64, 1, 197]> var_414_cast_fp16 = conv(dilations = var_414_dilations_0, groups = var_414_groups_0, pad = var_414_pad_0, pad_type = var_414_pad_type_0, strides = var_414_strides_0, weight = model_blocks_0_attn_q_projs_2_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
            tensor<int32, [1]> var_415_axes_0 = const()[name = tensor<string, []>("op_415_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_415_cast_fp16 = squeeze(axes = var_415_axes_0, x = var_414_cast_fp16)[name = tensor<string, []>("op_415_cast_fp16")];
            tensor<int32, [3]> q_21_perm_0 = const()[name = tensor<string, []>("q_21_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_422_pad_type_0 = const()[name = tensor<string, []>("op_422_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_422_strides_0 = const()[name = tensor<string, []>("op_422_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_422_pad_0 = const()[name = tensor<string, []>("op_422_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_422_dilations_0 = const()[name = tensor<string, []>("op_422_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_422_groups_0 = const()[name = tensor<string, []>("op_422_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2178176)))];
            tensor<fp16, [1, 64, 1, 197]> var_422_cast_fp16 = conv(dilations = var_422_dilations_0, groups = var_422_groups_0, pad = var_422_pad_0, pad_type = var_422_pad_type_0, strides = var_422_strides_0, weight = model_blocks_0_attn_k_projs_2_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_422_cast_fp16")];
            tensor<int32, [1]> var_423_axes_0 = const()[name = tensor<string, []>("op_423_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_423_cast_fp16 = squeeze(axes = var_423_axes_0, x = var_422_cast_fp16)[name = tensor<string, []>("op_423_cast_fp16")];
            tensor<int32, [3]> k_13_perm_0 = const()[name = tensor<string, []>("k_13_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_430_pad_type_0 = const()[name = tensor<string, []>("op_430_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_430_strides_0 = const()[name = tensor<string, []>("op_430_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_430_pad_0 = const()[name = tensor<string, []>("op_430_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_430_dilations_0 = const()[name = tensor<string, []>("op_430_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_430_groups_0 = const()[name = tensor<string, []>("op_430_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2276544)))];
            tensor<fp16, [1, 64, 1, 197]> var_430_cast_fp16 = conv(dilations = var_430_dilations_0, groups = var_430_groups_0, pad = var_430_pad_0, pad_type = var_430_pad_type_0, strides = var_430_strides_0, weight = model_blocks_0_attn_v_projs_2_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_430_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = tensor<string, []>("op_431_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_431_cast_fp16 = squeeze(axes = var_431_axes_0, x = var_430_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
            tensor<int32, [3]> v_13_perm_0 = const()[name = tensor<string, []>("v_13_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2374912)))];
            tensor<fp16, [1, 197, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_415_cast_fp16)[name = tensor<string, []>("transpose_508")];
            tensor<fp16, [1, 197, 64]> q_23_cast_fp16 = add(x = q_21_cast_fp16, y = model_blocks_0_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_23_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2375104)))];
            tensor<fp16, [1, 197, 64]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_431_cast_fp16)[name = tensor<string, []>("transpose_506")];
            tensor<fp16, [1, 197, 64]> v_15_cast_fp16 = add(x = v_13_cast_fp16, y = model_blocks_0_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
            tensor<int32, [1]> q_25_axes_0 = const()[name = tensor<string, []>("q_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_25_cast_fp16 = expand_dims(axes = q_25_axes_0, x = q_23_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
            tensor<int32, [1]> k_15_axes_0 = const()[name = tensor<string, []>("k_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_423_cast_fp16)[name = tensor<string, []>("transpose_507")];
            tensor<fp16, [1, 1, 197, 64]> k_15_cast_fp16 = expand_dims(axes = k_15_axes_0, x = k_13_cast_fp16)[name = tensor<string, []>("k_15_cast_fp16")];
            tensor<int32, [1]> v_17_axes_0 = const()[name = tensor<string, []>("v_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_17_cast_fp16 = expand_dims(axes = v_17_axes_0, x = v_15_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
            tensor<int32, [4]> var_440_begin_0 = const()[name = tensor<string, []>("op_440_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_440_end_0 = const()[name = tensor<string, []>("op_440_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_440_end_mask_0 = const()[name = tensor<string, []>("op_440_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = q_25_cast_fp16)[name = tensor<string, []>("op_440_cast_fp16")];
            tensor<int32, [4]> var_444_begin_0 = const()[name = tensor<string, []>("op_444_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_444_end_0 = const()[name = tensor<string, []>("op_444_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_444_end_mask_0 = const()[name = tensor<string, []>("op_444_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_444_cast_fp16 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = k_15_cast_fp16)[name = tensor<string, []>("op_444_cast_fp16")];
            tensor<int32, [4]> var_448_begin_0 = const()[name = tensor<string, []>("op_448_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_448_end_0 = const()[name = tensor<string, []>("op_448_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_448_end_mask_0 = const()[name = tensor<string, []>("op_448_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_448_cast_fp16 = slice_by_index(begin = var_448_begin_0, end = var_448_end_0, end_mask = var_448_end_mask_0, x = q_25_cast_fp16)[name = tensor<string, []>("op_448_cast_fp16")];
            tensor<int32, [4]> var_452_begin_0 = const()[name = tensor<string, []>("op_452_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_452_end_0 = const()[name = tensor<string, []>("op_452_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_452_end_mask_0 = const()[name = tensor<string, []>("op_452_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_452_cast_fp16 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = k_15_cast_fp16)[name = tensor<string, []>("op_452_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_454_cast_fp16 = mul(x = var_448_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_454_cast_fp16")];
            tensor<int32, [5]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_35_cast_fp16 = reshape(shape = var_462, x = var_448_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<int32, [2]> var_464_split_sizes_0 = const()[name = tensor<string, []>("op_464_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_464_axis_0 = const()[name = tensor<string, []>("op_464_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_464_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_464_cast_fp16_1 = split(axis = var_464_axis_0, split_sizes = var_464_split_sizes_0, x = x_35_cast_fp16)[name = tensor<string, []>("op_464_cast_fp16")];
            tensor<int32, [1]> squeeze_8_axes_0 = const()[name = tensor<string, []>("squeeze_8_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_464_cast_fp16_0)[name = tensor<string, []>("squeeze_8_cast_fp16")];
            tensor<int32, [1]> squeeze_9_axes_0 = const()[name = tensor<string, []>("squeeze_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_464_cast_fp16_1)[name = tensor<string, []>("squeeze_9_cast_fp16")];
            tensor<fp16, []> const_39_promoted_to_fp16 = const()[name = tensor<string, []>("const_39_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_467_cast_fp16 = mul(x = squeeze_9_cast_fp16, y = const_39_promoted_to_fp16)[name = tensor<string, []>("op_467_cast_fp16")];
            tensor<int32, []> x_37_axis_0 = const()[name = tensor<string, []>("x_37_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_37_cast_fp16 = stack(axis = x_37_axis_0, values = (var_467_cast_fp16, squeeze_8_cast_fp16))[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [4]> var_473 = const()[name = tensor<string, []>("op_473"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_474_cast_fp16 = reshape(shape = var_473, x = x_37_cast_fp16)[name = tensor<string, []>("op_474_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_475_cast_fp16 = mul(x = var_474_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_475_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_5_cast_fp16 = add(x = var_454_cast_fp16, y = var_475_cast_fp16)[name = tensor<string, []>("q_patches_5_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_477_cast_fp16 = mul(x = var_452_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_477_cast_fp16")];
            tensor<int32, [5]> var_485 = const()[name = tensor<string, []>("op_485"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_41_cast_fp16 = reshape(shape = var_485, x = var_452_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
            tensor<int32, [2]> var_487_split_sizes_0 = const()[name = tensor<string, []>("op_487_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_487_axis_0 = const()[name = tensor<string, []>("op_487_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_487_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_487_cast_fp16_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = x_41_cast_fp16)[name = tensor<string, []>("op_487_cast_fp16")];
            tensor<int32, [1]> squeeze_10_axes_0 = const()[name = tensor<string, []>("squeeze_10_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_487_cast_fp16_0)[name = tensor<string, []>("squeeze_10_cast_fp16")];
            tensor<int32, [1]> squeeze_11_axes_0 = const()[name = tensor<string, []>("squeeze_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_487_cast_fp16_1)[name = tensor<string, []>("squeeze_11_cast_fp16")];
            tensor<fp16, []> const_47_promoted_to_fp16 = const()[name = tensor<string, []>("const_47_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_490_cast_fp16 = mul(x = squeeze_11_cast_fp16, y = const_47_promoted_to_fp16)[name = tensor<string, []>("op_490_cast_fp16")];
            tensor<int32, []> x_43_axis_0 = const()[name = tensor<string, []>("x_43_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_43_cast_fp16 = stack(axis = x_43_axis_0, values = (var_490_cast_fp16, squeeze_10_cast_fp16))[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<int32, [4]> var_496 = const()[name = tensor<string, []>("op_496"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_497_cast_fp16 = reshape(shape = var_496, x = x_43_cast_fp16)[name = tensor<string, []>("op_497_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_498_cast_fp16 = mul(x = var_497_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_498_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_498_cast_fp16)[name = tensor<string, []>("k_patches_5_cast_fp16")];
            tensor<bool, []> var_501_interleave_0 = const()[name = tensor<string, []>("op_501_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_501_cast_fp16 = concat(axis = var_20, interleave = var_501_interleave_0, values = (var_440_cast_fp16, q_patches_5_cast_fp16))[name = tensor<string, []>("op_501_cast_fp16")];
            tensor<bool, []> var_504_interleave_0 = const()[name = tensor<string, []>("op_504_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_504_cast_fp16 = concat(axis = var_20, interleave = var_504_interleave_0, values = (var_444_cast_fp16, k_patches_5_cast_fp16))[name = tensor<string, []>("op_504_cast_fp16")];
            tensor<fp16, []> var_506_to_fp16 = const()[name = tensor<string, []>("op_506_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_29_cast_fp16 = mul(x = var_501_cast_fp16, y = var_506_to_fp16)[name = tensor<string, []>("q_29_cast_fp16")];
            tensor<bool, []> attn_9_transpose_x_1 = const()[name = tensor<string, []>("attn_9_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_9_transpose_y_1 = const()[name = tensor<string, []>("attn_9_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_1, transpose_y = attn_9_transpose_y_1, x = q_29_cast_fp16, y = var_504_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_11_cast_fp16 = softmax(axis = var_21, x = attn_9_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
            tensor<bool, []> var_511_transpose_x_0 = const()[name = tensor<string, []>("op_511_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_511_transpose_y_0 = const()[name = tensor<string, []>("op_511_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_511_cast_fp16 = matmul(transpose_x = var_511_transpose_x_0, transpose_y = var_511_transpose_y_0, x = attn_11_cast_fp16, y = v_17_cast_fp16)[name = tensor<string, []>("op_511_cast_fp16")];
            tensor<int32, [1]> var_512_axes_0 = const()[name = tensor<string, []>("op_512_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_512_cast_fp16 = squeeze(axes = var_512_axes_0, x = var_511_cast_fp16)[name = tensor<string, []>("op_512_cast_fp16")];
            tensor<string, []> var_518_pad_type_0 = const()[name = tensor<string, []>("op_518_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_518_strides_0 = const()[name = tensor<string, []>("op_518_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_518_pad_0 = const()[name = tensor<string, []>("op_518_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_518_dilations_0 = const()[name = tensor<string, []>("op_518_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_518_groups_0 = const()[name = tensor<string, []>("op_518_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2375296)))];
            tensor<fp16, [1, 64, 1, 197]> var_518_cast_fp16 = conv(dilations = var_518_dilations_0, groups = var_518_groups_0, pad = var_518_pad_0, pad_type = var_518_pad_type_0, strides = var_518_strides_0, weight = model_blocks_0_attn_q_projs_3_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_518_cast_fp16")];
            tensor<int32, [1]> var_519_axes_0 = const()[name = tensor<string, []>("op_519_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_519_cast_fp16 = squeeze(axes = var_519_axes_0, x = var_518_cast_fp16)[name = tensor<string, []>("op_519_cast_fp16")];
            tensor<int32, [3]> q_31_perm_0 = const()[name = tensor<string, []>("q_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_526_pad_type_0 = const()[name = tensor<string, []>("op_526_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_526_strides_0 = const()[name = tensor<string, []>("op_526_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_526_pad_0 = const()[name = tensor<string, []>("op_526_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_526_dilations_0 = const()[name = tensor<string, []>("op_526_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_526_groups_0 = const()[name = tensor<string, []>("op_526_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2473664)))];
            tensor<fp16, [1, 64, 1, 197]> var_526_cast_fp16 = conv(dilations = var_526_dilations_0, groups = var_526_groups_0, pad = var_526_pad_0, pad_type = var_526_pad_type_0, strides = var_526_strides_0, weight = model_blocks_0_attn_k_projs_3_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_526_cast_fp16")];
            tensor<int32, [1]> var_527_axes_0 = const()[name = tensor<string, []>("op_527_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_527_cast_fp16 = squeeze(axes = var_527_axes_0, x = var_526_cast_fp16)[name = tensor<string, []>("op_527_cast_fp16")];
            tensor<int32, [3]> k_19_perm_0 = const()[name = tensor<string, []>("k_19_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_534_pad_type_0 = const()[name = tensor<string, []>("op_534_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_534_strides_0 = const()[name = tensor<string, []>("op_534_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_534_pad_0 = const()[name = tensor<string, []>("op_534_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_534_dilations_0 = const()[name = tensor<string, []>("op_534_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_534_groups_0 = const()[name = tensor<string, []>("op_534_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2572032)))];
            tensor<fp16, [1, 64, 1, 197]> var_534_cast_fp16 = conv(dilations = var_534_dilations_0, groups = var_534_groups_0, pad = var_534_pad_0, pad_type = var_534_pad_type_0, strides = var_534_strides_0, weight = model_blocks_0_attn_v_projs_3_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_534_cast_fp16")];
            tensor<int32, [1]> var_535_axes_0 = const()[name = tensor<string, []>("op_535_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_535_cast_fp16 = squeeze(axes = var_535_axes_0, x = var_534_cast_fp16)[name = tensor<string, []>("op_535_cast_fp16")];
            tensor<int32, [3]> v_19_perm_0 = const()[name = tensor<string, []>("v_19_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2670400)))];
            tensor<fp16, [1, 197, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_519_cast_fp16)[name = tensor<string, []>("transpose_505")];
            tensor<fp16, [1, 197, 64]> q_33_cast_fp16 = add(x = q_31_cast_fp16, y = model_blocks_0_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_33_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2670592)))];
            tensor<fp16, [1, 197, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_535_cast_fp16)[name = tensor<string, []>("transpose_503")];
            tensor<fp16, [1, 197, 64]> v_21_cast_fp16 = add(x = v_19_cast_fp16, y = model_blocks_0_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
            tensor<int32, [1]> q_35_axes_0 = const()[name = tensor<string, []>("q_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_35_cast_fp16 = expand_dims(axes = q_35_axes_0, x = q_33_cast_fp16)[name = tensor<string, []>("q_35_cast_fp16")];
            tensor<int32, [1]> k_21_axes_0 = const()[name = tensor<string, []>("k_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_527_cast_fp16)[name = tensor<string, []>("transpose_504")];
            tensor<fp16, [1, 1, 197, 64]> k_21_cast_fp16 = expand_dims(axes = k_21_axes_0, x = k_19_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
            tensor<int32, [1]> v_23_axes_0 = const()[name = tensor<string, []>("v_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_23_cast_fp16 = expand_dims(axes = v_23_axes_0, x = v_21_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
            tensor<int32, [4]> var_544_begin_0 = const()[name = tensor<string, []>("op_544_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_544_end_0 = const()[name = tensor<string, []>("op_544_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_544_end_mask_0 = const()[name = tensor<string, []>("op_544_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, x = q_35_cast_fp16)[name = tensor<string, []>("op_544_cast_fp16")];
            tensor<int32, [4]> var_548_begin_0 = const()[name = tensor<string, []>("op_548_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_548_end_0 = const()[name = tensor<string, []>("op_548_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_548_end_mask_0 = const()[name = tensor<string, []>("op_548_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_548_cast_fp16 = slice_by_index(begin = var_548_begin_0, end = var_548_end_0, end_mask = var_548_end_mask_0, x = k_21_cast_fp16)[name = tensor<string, []>("op_548_cast_fp16")];
            tensor<int32, [4]> var_552_begin_0 = const()[name = tensor<string, []>("op_552_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_552_end_0 = const()[name = tensor<string, []>("op_552_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_552_end_mask_0 = const()[name = tensor<string, []>("op_552_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, x = q_35_cast_fp16)[name = tensor<string, []>("op_552_cast_fp16")];
            tensor<int32, [4]> var_556_begin_0 = const()[name = tensor<string, []>("op_556_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_556_end_0 = const()[name = tensor<string, []>("op_556_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_556_end_mask_0 = const()[name = tensor<string, []>("op_556_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_556_cast_fp16 = slice_by_index(begin = var_556_begin_0, end = var_556_end_0, end_mask = var_556_end_mask_0, x = k_21_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_558_cast_fp16 = mul(x = var_552_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_558_cast_fp16")];
            tensor<int32, [5]> var_566 = const()[name = tensor<string, []>("op_566"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_47_cast_fp16 = reshape(shape = var_566, x = var_552_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<int32, [2]> var_568_split_sizes_0 = const()[name = tensor<string, []>("op_568_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_568_axis_0 = const()[name = tensor<string, []>("op_568_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_568_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_568_cast_fp16_1 = split(axis = var_568_axis_0, split_sizes = var_568_split_sizes_0, x = x_47_cast_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
            tensor<int32, [1]> squeeze_12_axes_0 = const()[name = tensor<string, []>("squeeze_12_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_568_cast_fp16_0)[name = tensor<string, []>("squeeze_12_cast_fp16")];
            tensor<int32, [1]> squeeze_13_axes_0 = const()[name = tensor<string, []>("squeeze_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_568_cast_fp16_1)[name = tensor<string, []>("squeeze_13_cast_fp16")];
            tensor<fp16, []> const_55_promoted_to_fp16 = const()[name = tensor<string, []>("const_55_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_571_cast_fp16 = mul(x = squeeze_13_cast_fp16, y = const_55_promoted_to_fp16)[name = tensor<string, []>("op_571_cast_fp16")];
            tensor<int32, []> x_49_axis_0 = const()[name = tensor<string, []>("x_49_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_49_cast_fp16 = stack(axis = x_49_axis_0, values = (var_571_cast_fp16, squeeze_12_cast_fp16))[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<int32, [4]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_578_cast_fp16 = reshape(shape = var_577, x = x_49_cast_fp16)[name = tensor<string, []>("op_578_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_579_cast_fp16 = mul(x = var_578_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_579_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_7_cast_fp16 = add(x = var_558_cast_fp16, y = var_579_cast_fp16)[name = tensor<string, []>("q_patches_7_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_581_cast_fp16 = mul(x = var_556_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_581_cast_fp16")];
            tensor<int32, [5]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_53_cast_fp16 = reshape(shape = var_589, x = var_556_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
            tensor<int32, [2]> var_591_split_sizes_0 = const()[name = tensor<string, []>("op_591_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_591_axis_0 = const()[name = tensor<string, []>("op_591_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_591_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_591_cast_fp16_1 = split(axis = var_591_axis_0, split_sizes = var_591_split_sizes_0, x = x_53_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
            tensor<int32, [1]> squeeze_14_axes_0 = const()[name = tensor<string, []>("squeeze_14_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_591_cast_fp16_0)[name = tensor<string, []>("squeeze_14_cast_fp16")];
            tensor<int32, [1]> squeeze_15_axes_0 = const()[name = tensor<string, []>("squeeze_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_591_cast_fp16_1)[name = tensor<string, []>("squeeze_15_cast_fp16")];
            tensor<fp16, []> const_63_promoted_to_fp16 = const()[name = tensor<string, []>("const_63_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_594_cast_fp16 = mul(x = squeeze_15_cast_fp16, y = const_63_promoted_to_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
            tensor<int32, []> x_55_axis_0 = const()[name = tensor<string, []>("x_55_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_55_cast_fp16 = stack(axis = x_55_axis_0, values = (var_594_cast_fp16, squeeze_14_cast_fp16))[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<int32, [4]> var_600 = const()[name = tensor<string, []>("op_600"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_601_cast_fp16 = reshape(shape = var_600, x = x_55_cast_fp16)[name = tensor<string, []>("op_601_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_602_cast_fp16 = mul(x = var_601_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_602_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_7_cast_fp16 = add(x = var_581_cast_fp16, y = var_602_cast_fp16)[name = tensor<string, []>("k_patches_7_cast_fp16")];
            tensor<bool, []> var_605_interleave_0 = const()[name = tensor<string, []>("op_605_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_605_cast_fp16 = concat(axis = var_20, interleave = var_605_interleave_0, values = (var_544_cast_fp16, q_patches_7_cast_fp16))[name = tensor<string, []>("op_605_cast_fp16")];
            tensor<bool, []> var_608_interleave_0 = const()[name = tensor<string, []>("op_608_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_608_cast_fp16 = concat(axis = var_20, interleave = var_608_interleave_0, values = (var_548_cast_fp16, k_patches_7_cast_fp16))[name = tensor<string, []>("op_608_cast_fp16")];
            tensor<fp16, []> var_610_to_fp16 = const()[name = tensor<string, []>("op_610_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_39_cast_fp16 = mul(x = var_605_cast_fp16, y = var_610_to_fp16)[name = tensor<string, []>("q_39_cast_fp16")];
            tensor<bool, []> attn_13_transpose_x_1 = const()[name = tensor<string, []>("attn_13_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_13_transpose_y_1 = const()[name = tensor<string, []>("attn_13_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_1, transpose_y = attn_13_transpose_y_1, x = q_39_cast_fp16, y = var_608_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_15_cast_fp16 = softmax(axis = var_21, x = attn_13_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
            tensor<bool, []> var_615_transpose_x_0 = const()[name = tensor<string, []>("op_615_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_615_transpose_y_0 = const()[name = tensor<string, []>("op_615_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_615_cast_fp16 = matmul(transpose_x = var_615_transpose_x_0, transpose_y = var_615_transpose_y_0, x = attn_15_cast_fp16, y = v_23_cast_fp16)[name = tensor<string, []>("op_615_cast_fp16")];
            tensor<int32, [1]> var_616_axes_0 = const()[name = tensor<string, []>("op_616_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_616_cast_fp16 = squeeze(axes = var_616_axes_0, x = var_615_cast_fp16)[name = tensor<string, []>("op_616_cast_fp16")];
            tensor<string, []> var_622_pad_type_0 = const()[name = tensor<string, []>("op_622_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_622_strides_0 = const()[name = tensor<string, []>("op_622_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_622_pad_0 = const()[name = tensor<string, []>("op_622_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_622_dilations_0 = const()[name = tensor<string, []>("op_622_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_622_groups_0 = const()[name = tensor<string, []>("op_622_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2670784)))];
            tensor<fp16, [1, 64, 1, 197]> var_622_cast_fp16 = conv(dilations = var_622_dilations_0, groups = var_622_groups_0, pad = var_622_pad_0, pad_type = var_622_pad_type_0, strides = var_622_strides_0, weight = model_blocks_0_attn_q_projs_4_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_622_cast_fp16")];
            tensor<int32, [1]> var_623_axes_0 = const()[name = tensor<string, []>("op_623_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_623_cast_fp16 = squeeze(axes = var_623_axes_0, x = var_622_cast_fp16)[name = tensor<string, []>("op_623_cast_fp16")];
            tensor<int32, [3]> q_41_perm_0 = const()[name = tensor<string, []>("q_41_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_630_pad_type_0 = const()[name = tensor<string, []>("op_630_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_630_strides_0 = const()[name = tensor<string, []>("op_630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_630_pad_0 = const()[name = tensor<string, []>("op_630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_630_dilations_0 = const()[name = tensor<string, []>("op_630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_630_groups_0 = const()[name = tensor<string, []>("op_630_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2769152)))];
            tensor<fp16, [1, 64, 1, 197]> var_630_cast_fp16 = conv(dilations = var_630_dilations_0, groups = var_630_groups_0, pad = var_630_pad_0, pad_type = var_630_pad_type_0, strides = var_630_strides_0, weight = model_blocks_0_attn_k_projs_4_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_630_cast_fp16")];
            tensor<int32, [1]> var_631_axes_0 = const()[name = tensor<string, []>("op_631_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_631_cast_fp16 = squeeze(axes = var_631_axes_0, x = var_630_cast_fp16)[name = tensor<string, []>("op_631_cast_fp16")];
            tensor<int32, [3]> k_25_perm_0 = const()[name = tensor<string, []>("k_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_638_pad_type_0 = const()[name = tensor<string, []>("op_638_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_638_strides_0 = const()[name = tensor<string, []>("op_638_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_638_pad_0 = const()[name = tensor<string, []>("op_638_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_638_dilations_0 = const()[name = tensor<string, []>("op_638_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_638_groups_0 = const()[name = tensor<string, []>("op_638_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2867520)))];
            tensor<fp16, [1, 64, 1, 197]> var_638_cast_fp16 = conv(dilations = var_638_dilations_0, groups = var_638_groups_0, pad = var_638_pad_0, pad_type = var_638_pad_type_0, strides = var_638_strides_0, weight = model_blocks_0_attn_v_projs_4_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_638_cast_fp16")];
            tensor<int32, [1]> var_639_axes_0 = const()[name = tensor<string, []>("op_639_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_639_cast_fp16 = squeeze(axes = var_639_axes_0, x = var_638_cast_fp16)[name = tensor<string, []>("op_639_cast_fp16")];
            tensor<int32, [3]> v_25_perm_0 = const()[name = tensor<string, []>("v_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2965888)))];
            tensor<fp16, [1, 197, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_623_cast_fp16)[name = tensor<string, []>("transpose_502")];
            tensor<fp16, [1, 197, 64]> q_43_cast_fp16 = add(x = q_41_cast_fp16, y = model_blocks_0_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2966080)))];
            tensor<fp16, [1, 197, 64]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_639_cast_fp16)[name = tensor<string, []>("transpose_500")];
            tensor<fp16, [1, 197, 64]> v_27_cast_fp16 = add(x = v_25_cast_fp16, y = model_blocks_0_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
            tensor<int32, [1]> q_45_axes_0 = const()[name = tensor<string, []>("q_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_45_cast_fp16 = expand_dims(axes = q_45_axes_0, x = q_43_cast_fp16)[name = tensor<string, []>("q_45_cast_fp16")];
            tensor<int32, [1]> k_27_axes_0 = const()[name = tensor<string, []>("k_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_631_cast_fp16)[name = tensor<string, []>("transpose_501")];
            tensor<fp16, [1, 1, 197, 64]> k_27_cast_fp16 = expand_dims(axes = k_27_axes_0, x = k_25_cast_fp16)[name = tensor<string, []>("k_27_cast_fp16")];
            tensor<int32, [1]> v_29_axes_0 = const()[name = tensor<string, []>("v_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_29_cast_fp16 = expand_dims(axes = v_29_axes_0, x = v_27_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
            tensor<int32, [4]> var_648_begin_0 = const()[name = tensor<string, []>("op_648_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_648_end_0 = const()[name = tensor<string, []>("op_648_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_648_end_mask_0 = const()[name = tensor<string, []>("op_648_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_648_cast_fp16 = slice_by_index(begin = var_648_begin_0, end = var_648_end_0, end_mask = var_648_end_mask_0, x = q_45_cast_fp16)[name = tensor<string, []>("op_648_cast_fp16")];
            tensor<int32, [4]> var_652_begin_0 = const()[name = tensor<string, []>("op_652_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_652_end_0 = const()[name = tensor<string, []>("op_652_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_652_end_mask_0 = const()[name = tensor<string, []>("op_652_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_652_cast_fp16 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = k_27_cast_fp16)[name = tensor<string, []>("op_652_cast_fp16")];
            tensor<int32, [4]> var_656_begin_0 = const()[name = tensor<string, []>("op_656_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_656_end_0 = const()[name = tensor<string, []>("op_656_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_656_end_mask_0 = const()[name = tensor<string, []>("op_656_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_656_cast_fp16 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = q_45_cast_fp16)[name = tensor<string, []>("op_656_cast_fp16")];
            tensor<int32, [4]> var_660_begin_0 = const()[name = tensor<string, []>("op_660_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_660_end_0 = const()[name = tensor<string, []>("op_660_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_660_end_mask_0 = const()[name = tensor<string, []>("op_660_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_660_cast_fp16 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = k_27_cast_fp16)[name = tensor<string, []>("op_660_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_662_cast_fp16 = mul(x = var_656_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_662_cast_fp16")];
            tensor<int32, [5]> var_670 = const()[name = tensor<string, []>("op_670"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_59_cast_fp16 = reshape(shape = var_670, x = var_656_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<int32, [2]> var_672_split_sizes_0 = const()[name = tensor<string, []>("op_672_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_672_axis_0 = const()[name = tensor<string, []>("op_672_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_672_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_672_cast_fp16_1 = split(axis = var_672_axis_0, split_sizes = var_672_split_sizes_0, x = x_59_cast_fp16)[name = tensor<string, []>("op_672_cast_fp16")];
            tensor<int32, [1]> squeeze_16_axes_0 = const()[name = tensor<string, []>("squeeze_16_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_672_cast_fp16_0)[name = tensor<string, []>("squeeze_16_cast_fp16")];
            tensor<int32, [1]> squeeze_17_axes_0 = const()[name = tensor<string, []>("squeeze_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_672_cast_fp16_1)[name = tensor<string, []>("squeeze_17_cast_fp16")];
            tensor<fp16, []> const_71_promoted_to_fp16 = const()[name = tensor<string, []>("const_71_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_675_cast_fp16 = mul(x = squeeze_17_cast_fp16, y = const_71_promoted_to_fp16)[name = tensor<string, []>("op_675_cast_fp16")];
            tensor<int32, []> x_61_axis_0 = const()[name = tensor<string, []>("x_61_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_61_cast_fp16 = stack(axis = x_61_axis_0, values = (var_675_cast_fp16, squeeze_16_cast_fp16))[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<int32, [4]> var_681 = const()[name = tensor<string, []>("op_681"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_682_cast_fp16 = reshape(shape = var_681, x = x_61_cast_fp16)[name = tensor<string, []>("op_682_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_683_cast_fp16 = mul(x = var_682_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_683_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_9_cast_fp16 = add(x = var_662_cast_fp16, y = var_683_cast_fp16)[name = tensor<string, []>("q_patches_9_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_685_cast_fp16 = mul(x = var_660_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_685_cast_fp16")];
            tensor<int32, [5]> var_693 = const()[name = tensor<string, []>("op_693"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_65_cast_fp16 = reshape(shape = var_693, x = var_660_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<int32, [2]> var_695_split_sizes_0 = const()[name = tensor<string, []>("op_695_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_695_axis_0 = const()[name = tensor<string, []>("op_695_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_695_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_695_cast_fp16_1 = split(axis = var_695_axis_0, split_sizes = var_695_split_sizes_0, x = x_65_cast_fp16)[name = tensor<string, []>("op_695_cast_fp16")];
            tensor<int32, [1]> squeeze_18_axes_0 = const()[name = tensor<string, []>("squeeze_18_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_695_cast_fp16_0)[name = tensor<string, []>("squeeze_18_cast_fp16")];
            tensor<int32, [1]> squeeze_19_axes_0 = const()[name = tensor<string, []>("squeeze_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_695_cast_fp16_1)[name = tensor<string, []>("squeeze_19_cast_fp16")];
            tensor<fp16, []> const_79_promoted_to_fp16 = const()[name = tensor<string, []>("const_79_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_698_cast_fp16 = mul(x = squeeze_19_cast_fp16, y = const_79_promoted_to_fp16)[name = tensor<string, []>("op_698_cast_fp16")];
            tensor<int32, []> x_67_axis_0 = const()[name = tensor<string, []>("x_67_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_67_cast_fp16 = stack(axis = x_67_axis_0, values = (var_698_cast_fp16, squeeze_18_cast_fp16))[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<int32, [4]> var_704 = const()[name = tensor<string, []>("op_704"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_705_cast_fp16 = reshape(shape = var_704, x = x_67_cast_fp16)[name = tensor<string, []>("op_705_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_706_cast_fp16 = mul(x = var_705_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_706_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_9_cast_fp16 = add(x = var_685_cast_fp16, y = var_706_cast_fp16)[name = tensor<string, []>("k_patches_9_cast_fp16")];
            tensor<bool, []> var_709_interleave_0 = const()[name = tensor<string, []>("op_709_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_709_cast_fp16 = concat(axis = var_20, interleave = var_709_interleave_0, values = (var_648_cast_fp16, q_patches_9_cast_fp16))[name = tensor<string, []>("op_709_cast_fp16")];
            tensor<bool, []> var_712_interleave_0 = const()[name = tensor<string, []>("op_712_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_712_cast_fp16 = concat(axis = var_20, interleave = var_712_interleave_0, values = (var_652_cast_fp16, k_patches_9_cast_fp16))[name = tensor<string, []>("op_712_cast_fp16")];
            tensor<fp16, []> var_714_to_fp16 = const()[name = tensor<string, []>("op_714_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_49_cast_fp16 = mul(x = var_709_cast_fp16, y = var_714_to_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
            tensor<bool, []> attn_17_transpose_x_1 = const()[name = tensor<string, []>("attn_17_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_17_transpose_y_1 = const()[name = tensor<string, []>("attn_17_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_1, transpose_y = attn_17_transpose_y_1, x = q_49_cast_fp16, y = var_712_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_19_cast_fp16 = softmax(axis = var_21, x = attn_17_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
            tensor<bool, []> var_719_transpose_x_0 = const()[name = tensor<string, []>("op_719_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_719_transpose_y_0 = const()[name = tensor<string, []>("op_719_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_719_cast_fp16 = matmul(transpose_x = var_719_transpose_x_0, transpose_y = var_719_transpose_y_0, x = attn_19_cast_fp16, y = v_29_cast_fp16)[name = tensor<string, []>("op_719_cast_fp16")];
            tensor<int32, [1]> var_720_axes_0 = const()[name = tensor<string, []>("op_720_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_720_cast_fp16 = squeeze(axes = var_720_axes_0, x = var_719_cast_fp16)[name = tensor<string, []>("op_720_cast_fp16")];
            tensor<string, []> var_726_pad_type_0 = const()[name = tensor<string, []>("op_726_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_726_strides_0 = const()[name = tensor<string, []>("op_726_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_726_pad_0 = const()[name = tensor<string, []>("op_726_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_726_dilations_0 = const()[name = tensor<string, []>("op_726_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_726_groups_0 = const()[name = tensor<string, []>("op_726_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2966272)))];
            tensor<fp16, [1, 64, 1, 197]> var_726_cast_fp16 = conv(dilations = var_726_dilations_0, groups = var_726_groups_0, pad = var_726_pad_0, pad_type = var_726_pad_type_0, strides = var_726_strides_0, weight = model_blocks_0_attn_q_projs_5_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_726_cast_fp16")];
            tensor<int32, [1]> var_727_axes_0 = const()[name = tensor<string, []>("op_727_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_727_cast_fp16 = squeeze(axes = var_727_axes_0, x = var_726_cast_fp16)[name = tensor<string, []>("op_727_cast_fp16")];
            tensor<int32, [3]> q_51_perm_0 = const()[name = tensor<string, []>("q_51_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_734_pad_type_0 = const()[name = tensor<string, []>("op_734_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_734_strides_0 = const()[name = tensor<string, []>("op_734_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_734_pad_0 = const()[name = tensor<string, []>("op_734_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_734_dilations_0 = const()[name = tensor<string, []>("op_734_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_734_groups_0 = const()[name = tensor<string, []>("op_734_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3064640)))];
            tensor<fp16, [1, 64, 1, 197]> var_734_cast_fp16 = conv(dilations = var_734_dilations_0, groups = var_734_groups_0, pad = var_734_pad_0, pad_type = var_734_pad_type_0, strides = var_734_strides_0, weight = model_blocks_0_attn_k_projs_5_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_734_cast_fp16")];
            tensor<int32, [1]> var_735_axes_0 = const()[name = tensor<string, []>("op_735_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_735_cast_fp16 = squeeze(axes = var_735_axes_0, x = var_734_cast_fp16)[name = tensor<string, []>("op_735_cast_fp16")];
            tensor<int32, [3]> k_31_perm_0 = const()[name = tensor<string, []>("k_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_742_pad_type_0 = const()[name = tensor<string, []>("op_742_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_742_strides_0 = const()[name = tensor<string, []>("op_742_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_742_pad_0 = const()[name = tensor<string, []>("op_742_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_742_dilations_0 = const()[name = tensor<string, []>("op_742_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_742_groups_0 = const()[name = tensor<string, []>("op_742_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3163008)))];
            tensor<fp16, [1, 64, 1, 197]> var_742_cast_fp16 = conv(dilations = var_742_dilations_0, groups = var_742_groups_0, pad = var_742_pad_0, pad_type = var_742_pad_type_0, strides = var_742_strides_0, weight = model_blocks_0_attn_v_projs_5_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_742_cast_fp16")];
            tensor<int32, [1]> var_743_axes_0 = const()[name = tensor<string, []>("op_743_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_743_cast_fp16 = squeeze(axes = var_743_axes_0, x = var_742_cast_fp16)[name = tensor<string, []>("op_743_cast_fp16")];
            tensor<int32, [3]> v_31_perm_0 = const()[name = tensor<string, []>("v_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3261376)))];
            tensor<fp16, [1, 197, 64]> q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = var_727_cast_fp16)[name = tensor<string, []>("transpose_499")];
            tensor<fp16, [1, 197, 64]> q_53_cast_fp16 = add(x = q_51_cast_fp16, y = model_blocks_0_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_53_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3261568)))];
            tensor<fp16, [1, 197, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_743_cast_fp16)[name = tensor<string, []>("transpose_497")];
            tensor<fp16, [1, 197, 64]> v_33_cast_fp16 = add(x = v_31_cast_fp16, y = model_blocks_0_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_33_cast_fp16")];
            tensor<int32, [1]> q_55_axes_0 = const()[name = tensor<string, []>("q_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_55_cast_fp16 = expand_dims(axes = q_55_axes_0, x = q_53_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
            tensor<int32, [1]> k_33_axes_0 = const()[name = tensor<string, []>("k_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_735_cast_fp16)[name = tensor<string, []>("transpose_498")];
            tensor<fp16, [1, 1, 197, 64]> k_33_cast_fp16 = expand_dims(axes = k_33_axes_0, x = k_31_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
            tensor<int32, [1]> v_35_axes_0 = const()[name = tensor<string, []>("v_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_35_cast_fp16 = expand_dims(axes = v_35_axes_0, x = v_33_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")];
            tensor<int32, [4]> var_752_begin_0 = const()[name = tensor<string, []>("op_752_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_752_end_0 = const()[name = tensor<string, []>("op_752_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_752_end_mask_0 = const()[name = tensor<string, []>("op_752_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_752_cast_fp16 = slice_by_index(begin = var_752_begin_0, end = var_752_end_0, end_mask = var_752_end_mask_0, x = q_55_cast_fp16)[name = tensor<string, []>("op_752_cast_fp16")];
            tensor<int32, [4]> var_756_begin_0 = const()[name = tensor<string, []>("op_756_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_756_end_0 = const()[name = tensor<string, []>("op_756_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_756_end_mask_0 = const()[name = tensor<string, []>("op_756_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = k_33_cast_fp16)[name = tensor<string, []>("op_756_cast_fp16")];
            tensor<int32, [4]> var_760_begin_0 = const()[name = tensor<string, []>("op_760_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_760_end_0 = const()[name = tensor<string, []>("op_760_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_760_end_mask_0 = const()[name = tensor<string, []>("op_760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_760_cast_fp16 = slice_by_index(begin = var_760_begin_0, end = var_760_end_0, end_mask = var_760_end_mask_0, x = q_55_cast_fp16)[name = tensor<string, []>("op_760_cast_fp16")];
            tensor<int32, [4]> var_764_begin_0 = const()[name = tensor<string, []>("op_764_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_764_end_0 = const()[name = tensor<string, []>("op_764_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_764_end_mask_0 = const()[name = tensor<string, []>("op_764_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_764_cast_fp16 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, x = k_33_cast_fp16)[name = tensor<string, []>("op_764_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_766_cast_fp16 = mul(x = var_760_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_766_cast_fp16")];
            tensor<int32, [5]> var_774 = const()[name = tensor<string, []>("op_774"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_71_cast_fp16 = reshape(shape = var_774, x = var_760_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<int32, [2]> var_776_split_sizes_0 = const()[name = tensor<string, []>("op_776_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_776_axis_0 = const()[name = tensor<string, []>("op_776_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_776_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_776_cast_fp16_1 = split(axis = var_776_axis_0, split_sizes = var_776_split_sizes_0, x = x_71_cast_fp16)[name = tensor<string, []>("op_776_cast_fp16")];
            tensor<int32, [1]> squeeze_20_axes_0 = const()[name = tensor<string, []>("squeeze_20_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_776_cast_fp16_0)[name = tensor<string, []>("squeeze_20_cast_fp16")];
            tensor<int32, [1]> squeeze_21_axes_0 = const()[name = tensor<string, []>("squeeze_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_776_cast_fp16_1)[name = tensor<string, []>("squeeze_21_cast_fp16")];
            tensor<fp16, []> const_87_promoted_to_fp16 = const()[name = tensor<string, []>("const_87_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_779_cast_fp16 = mul(x = squeeze_21_cast_fp16, y = const_87_promoted_to_fp16)[name = tensor<string, []>("op_779_cast_fp16")];
            tensor<int32, []> x_73_axis_0 = const()[name = tensor<string, []>("x_73_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_73_cast_fp16 = stack(axis = x_73_axis_0, values = (var_779_cast_fp16, squeeze_20_cast_fp16))[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<int32, [4]> var_785 = const()[name = tensor<string, []>("op_785"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_786_cast_fp16 = reshape(shape = var_785, x = x_73_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_787_cast_fp16 = mul(x = var_786_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_787_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_11_cast_fp16 = add(x = var_766_cast_fp16, y = var_787_cast_fp16)[name = tensor<string, []>("q_patches_11_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_789_cast_fp16 = mul(x = var_764_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_789_cast_fp16")];
            tensor<int32, [5]> var_797 = const()[name = tensor<string, []>("op_797"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_77_cast_fp16 = reshape(shape = var_797, x = var_764_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<int32, [2]> var_799_split_sizes_0 = const()[name = tensor<string, []>("op_799_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_799_axis_0 = const()[name = tensor<string, []>("op_799_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_799_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_799_cast_fp16_1 = split(axis = var_799_axis_0, split_sizes = var_799_split_sizes_0, x = x_77_cast_fp16)[name = tensor<string, []>("op_799_cast_fp16")];
            tensor<int32, [1]> squeeze_22_axes_0 = const()[name = tensor<string, []>("squeeze_22_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_799_cast_fp16_0)[name = tensor<string, []>("squeeze_22_cast_fp16")];
            tensor<int32, [1]> squeeze_23_axes_0 = const()[name = tensor<string, []>("squeeze_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_799_cast_fp16_1)[name = tensor<string, []>("squeeze_23_cast_fp16")];
            tensor<fp16, []> const_95_promoted_to_fp16 = const()[name = tensor<string, []>("const_95_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_802_cast_fp16 = mul(x = squeeze_23_cast_fp16, y = const_95_promoted_to_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
            tensor<int32, []> x_79_axis_0 = const()[name = tensor<string, []>("x_79_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_79_cast_fp16 = stack(axis = x_79_axis_0, values = (var_802_cast_fp16, squeeze_22_cast_fp16))[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<int32, [4]> var_808 = const()[name = tensor<string, []>("op_808"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_809_cast_fp16 = reshape(shape = var_808, x = x_79_cast_fp16)[name = tensor<string, []>("op_809_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_810_cast_fp16 = mul(x = var_809_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_810_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_11_cast_fp16 = add(x = var_789_cast_fp16, y = var_810_cast_fp16)[name = tensor<string, []>("k_patches_11_cast_fp16")];
            tensor<bool, []> var_813_interleave_0 = const()[name = tensor<string, []>("op_813_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_813_cast_fp16 = concat(axis = var_20, interleave = var_813_interleave_0, values = (var_752_cast_fp16, q_patches_11_cast_fp16))[name = tensor<string, []>("op_813_cast_fp16")];
            tensor<bool, []> var_816_interleave_0 = const()[name = tensor<string, []>("op_816_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_816_cast_fp16 = concat(axis = var_20, interleave = var_816_interleave_0, values = (var_756_cast_fp16, k_patches_11_cast_fp16))[name = tensor<string, []>("op_816_cast_fp16")];
            tensor<fp16, []> var_818_to_fp16 = const()[name = tensor<string, []>("op_818_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_59_cast_fp16 = mul(x = var_813_cast_fp16, y = var_818_to_fp16)[name = tensor<string, []>("q_59_cast_fp16")];
            tensor<bool, []> attn_21_transpose_x_1 = const()[name = tensor<string, []>("attn_21_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_21_transpose_y_1 = const()[name = tensor<string, []>("attn_21_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_1, transpose_y = attn_21_transpose_y_1, x = q_59_cast_fp16, y = var_816_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_23_cast_fp16 = softmax(axis = var_21, x = attn_21_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
            tensor<bool, []> var_823_transpose_x_0 = const()[name = tensor<string, []>("op_823_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_823_transpose_y_0 = const()[name = tensor<string, []>("op_823_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_823_cast_fp16 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = attn_23_cast_fp16, y = v_35_cast_fp16)[name = tensor<string, []>("op_823_cast_fp16")];
            tensor<int32, [1]> var_824_axes_0 = const()[name = tensor<string, []>("op_824_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_824_cast_fp16 = squeeze(axes = var_824_axes_0, x = var_823_cast_fp16)[name = tensor<string, []>("op_824_cast_fp16")];
            tensor<string, []> var_830_pad_type_0 = const()[name = tensor<string, []>("op_830_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_830_strides_0 = const()[name = tensor<string, []>("op_830_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_830_pad_0 = const()[name = tensor<string, []>("op_830_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_830_dilations_0 = const()[name = tensor<string, []>("op_830_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_830_groups_0 = const()[name = tensor<string, []>("op_830_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3261760)))];
            tensor<fp16, [1, 64, 1, 197]> var_830_cast_fp16 = conv(dilations = var_830_dilations_0, groups = var_830_groups_0, pad = var_830_pad_0, pad_type = var_830_pad_type_0, strides = var_830_strides_0, weight = model_blocks_0_attn_q_projs_6_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_830_cast_fp16")];
            tensor<int32, [1]> var_831_axes_0 = const()[name = tensor<string, []>("op_831_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_831_cast_fp16 = squeeze(axes = var_831_axes_0, x = var_830_cast_fp16)[name = tensor<string, []>("op_831_cast_fp16")];
            tensor<int32, [3]> q_61_perm_0 = const()[name = tensor<string, []>("q_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_838_pad_type_0 = const()[name = tensor<string, []>("op_838_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_838_strides_0 = const()[name = tensor<string, []>("op_838_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_838_pad_0 = const()[name = tensor<string, []>("op_838_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_838_dilations_0 = const()[name = tensor<string, []>("op_838_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_838_groups_0 = const()[name = tensor<string, []>("op_838_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3360128)))];
            tensor<fp16, [1, 64, 1, 197]> var_838_cast_fp16 = conv(dilations = var_838_dilations_0, groups = var_838_groups_0, pad = var_838_pad_0, pad_type = var_838_pad_type_0, strides = var_838_strides_0, weight = model_blocks_0_attn_k_projs_6_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_838_cast_fp16")];
            tensor<int32, [1]> var_839_axes_0 = const()[name = tensor<string, []>("op_839_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_839_cast_fp16 = squeeze(axes = var_839_axes_0, x = var_838_cast_fp16)[name = tensor<string, []>("op_839_cast_fp16")];
            tensor<int32, [3]> k_37_perm_0 = const()[name = tensor<string, []>("k_37_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_846_pad_type_0 = const()[name = tensor<string, []>("op_846_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_846_strides_0 = const()[name = tensor<string, []>("op_846_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_846_pad_0 = const()[name = tensor<string, []>("op_846_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_846_dilations_0 = const()[name = tensor<string, []>("op_846_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_846_groups_0 = const()[name = tensor<string, []>("op_846_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3458496)))];
            tensor<fp16, [1, 64, 1, 197]> var_846_cast_fp16 = conv(dilations = var_846_dilations_0, groups = var_846_groups_0, pad = var_846_pad_0, pad_type = var_846_pad_type_0, strides = var_846_strides_0, weight = model_blocks_0_attn_v_projs_6_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_846_cast_fp16")];
            tensor<int32, [1]> var_847_axes_0 = const()[name = tensor<string, []>("op_847_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_847_cast_fp16 = squeeze(axes = var_847_axes_0, x = var_846_cast_fp16)[name = tensor<string, []>("op_847_cast_fp16")];
            tensor<int32, [3]> v_37_perm_0 = const()[name = tensor<string, []>("v_37_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3556864)))];
            tensor<fp16, [1, 197, 64]> q_61_cast_fp16 = transpose(perm = q_61_perm_0, x = var_831_cast_fp16)[name = tensor<string, []>("transpose_496")];
            tensor<fp16, [1, 197, 64]> q_63_cast_fp16 = add(x = q_61_cast_fp16, y = model_blocks_0_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_63_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3557056)))];
            tensor<fp16, [1, 197, 64]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_847_cast_fp16)[name = tensor<string, []>("transpose_494")];
            tensor<fp16, [1, 197, 64]> v_39_cast_fp16 = add(x = v_37_cast_fp16, y = model_blocks_0_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_39_cast_fp16")];
            tensor<int32, [1]> q_65_axes_0 = const()[name = tensor<string, []>("q_65_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_65_cast_fp16 = expand_dims(axes = q_65_axes_0, x = q_63_cast_fp16)[name = tensor<string, []>("q_65_cast_fp16")];
            tensor<int32, [1]> k_39_axes_0 = const()[name = tensor<string, []>("k_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_839_cast_fp16)[name = tensor<string, []>("transpose_495")];
            tensor<fp16, [1, 1, 197, 64]> k_39_cast_fp16 = expand_dims(axes = k_39_axes_0, x = k_37_cast_fp16)[name = tensor<string, []>("k_39_cast_fp16")];
            tensor<int32, [1]> v_41_axes_0 = const()[name = tensor<string, []>("v_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_41_cast_fp16 = expand_dims(axes = v_41_axes_0, x = v_39_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")];
            tensor<int32, [4]> var_856_begin_0 = const()[name = tensor<string, []>("op_856_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_856_end_0 = const()[name = tensor<string, []>("op_856_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_856_end_mask_0 = const()[name = tensor<string, []>("op_856_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_856_cast_fp16 = slice_by_index(begin = var_856_begin_0, end = var_856_end_0, end_mask = var_856_end_mask_0, x = q_65_cast_fp16)[name = tensor<string, []>("op_856_cast_fp16")];
            tensor<int32, [4]> var_860_begin_0 = const()[name = tensor<string, []>("op_860_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_860_end_0 = const()[name = tensor<string, []>("op_860_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_860_end_mask_0 = const()[name = tensor<string, []>("op_860_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_860_cast_fp16 = slice_by_index(begin = var_860_begin_0, end = var_860_end_0, end_mask = var_860_end_mask_0, x = k_39_cast_fp16)[name = tensor<string, []>("op_860_cast_fp16")];
            tensor<int32, [4]> var_864_begin_0 = const()[name = tensor<string, []>("op_864_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_864_end_0 = const()[name = tensor<string, []>("op_864_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_864_end_mask_0 = const()[name = tensor<string, []>("op_864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_864_cast_fp16 = slice_by_index(begin = var_864_begin_0, end = var_864_end_0, end_mask = var_864_end_mask_0, x = q_65_cast_fp16)[name = tensor<string, []>("op_864_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = tensor<string, []>("op_868_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = tensor<string, []>("op_868_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = tensor<string, []>("op_868_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = k_39_cast_fp16)[name = tensor<string, []>("op_868_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_870_cast_fp16 = mul(x = var_864_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_870_cast_fp16")];
            tensor<int32, [5]> var_878 = const()[name = tensor<string, []>("op_878"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_83_cast_fp16 = reshape(shape = var_878, x = var_864_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<int32, [2]> var_880_split_sizes_0 = const()[name = tensor<string, []>("op_880_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_880_axis_0 = const()[name = tensor<string, []>("op_880_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_880_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_880_cast_fp16_1 = split(axis = var_880_axis_0, split_sizes = var_880_split_sizes_0, x = x_83_cast_fp16)[name = tensor<string, []>("op_880_cast_fp16")];
            tensor<int32, [1]> squeeze_24_axes_0 = const()[name = tensor<string, []>("squeeze_24_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_880_cast_fp16_0)[name = tensor<string, []>("squeeze_24_cast_fp16")];
            tensor<int32, [1]> squeeze_25_axes_0 = const()[name = tensor<string, []>("squeeze_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_880_cast_fp16_1)[name = tensor<string, []>("squeeze_25_cast_fp16")];
            tensor<fp16, []> const_103_promoted_to_fp16 = const()[name = tensor<string, []>("const_103_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_883_cast_fp16 = mul(x = squeeze_25_cast_fp16, y = const_103_promoted_to_fp16)[name = tensor<string, []>("op_883_cast_fp16")];
            tensor<int32, []> x_85_axis_0 = const()[name = tensor<string, []>("x_85_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_85_cast_fp16 = stack(axis = x_85_axis_0, values = (var_883_cast_fp16, squeeze_24_cast_fp16))[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<int32, [4]> var_889 = const()[name = tensor<string, []>("op_889"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_890_cast_fp16 = reshape(shape = var_889, x = x_85_cast_fp16)[name = tensor<string, []>("op_890_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_891_cast_fp16 = mul(x = var_890_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_891_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_13_cast_fp16 = add(x = var_870_cast_fp16, y = var_891_cast_fp16)[name = tensor<string, []>("q_patches_13_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_893_cast_fp16 = mul(x = var_868_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_893_cast_fp16")];
            tensor<int32, [5]> var_901 = const()[name = tensor<string, []>("op_901"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_89_cast_fp16 = reshape(shape = var_901, x = var_868_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<int32, [2]> var_903_split_sizes_0 = const()[name = tensor<string, []>("op_903_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_903_axis_0 = const()[name = tensor<string, []>("op_903_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_903_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_903_cast_fp16_1 = split(axis = var_903_axis_0, split_sizes = var_903_split_sizes_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_903_cast_fp16")];
            tensor<int32, [1]> squeeze_26_axes_0 = const()[name = tensor<string, []>("squeeze_26_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_903_cast_fp16_0)[name = tensor<string, []>("squeeze_26_cast_fp16")];
            tensor<int32, [1]> squeeze_27_axes_0 = const()[name = tensor<string, []>("squeeze_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_903_cast_fp16_1)[name = tensor<string, []>("squeeze_27_cast_fp16")];
            tensor<fp16, []> const_111_promoted_to_fp16 = const()[name = tensor<string, []>("const_111_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_906_cast_fp16 = mul(x = squeeze_27_cast_fp16, y = const_111_promoted_to_fp16)[name = tensor<string, []>("op_906_cast_fp16")];
            tensor<int32, []> x_91_axis_0 = const()[name = tensor<string, []>("x_91_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_91_cast_fp16 = stack(axis = x_91_axis_0, values = (var_906_cast_fp16, squeeze_26_cast_fp16))[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<int32, [4]> var_912 = const()[name = tensor<string, []>("op_912"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_913_cast_fp16 = reshape(shape = var_912, x = x_91_cast_fp16)[name = tensor<string, []>("op_913_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_914_cast_fp16 = mul(x = var_913_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_914_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_13_cast_fp16 = add(x = var_893_cast_fp16, y = var_914_cast_fp16)[name = tensor<string, []>("k_patches_13_cast_fp16")];
            tensor<bool, []> var_917_interleave_0 = const()[name = tensor<string, []>("op_917_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_917_cast_fp16 = concat(axis = var_20, interleave = var_917_interleave_0, values = (var_856_cast_fp16, q_patches_13_cast_fp16))[name = tensor<string, []>("op_917_cast_fp16")];
            tensor<bool, []> var_920_interleave_0 = const()[name = tensor<string, []>("op_920_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_920_cast_fp16 = concat(axis = var_20, interleave = var_920_interleave_0, values = (var_860_cast_fp16, k_patches_13_cast_fp16))[name = tensor<string, []>("op_920_cast_fp16")];
            tensor<fp16, []> var_922_to_fp16 = const()[name = tensor<string, []>("op_922_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_69_cast_fp16 = mul(x = var_917_cast_fp16, y = var_922_to_fp16)[name = tensor<string, []>("q_69_cast_fp16")];
            tensor<bool, []> attn_25_transpose_x_1 = const()[name = tensor<string, []>("attn_25_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_25_transpose_y_1 = const()[name = tensor<string, []>("attn_25_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_1, transpose_y = attn_25_transpose_y_1, x = q_69_cast_fp16, y = var_920_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_27_cast_fp16 = softmax(axis = var_21, x = attn_25_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
            tensor<bool, []> var_927_transpose_x_0 = const()[name = tensor<string, []>("op_927_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_927_transpose_y_0 = const()[name = tensor<string, []>("op_927_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_927_cast_fp16 = matmul(transpose_x = var_927_transpose_x_0, transpose_y = var_927_transpose_y_0, x = attn_27_cast_fp16, y = v_41_cast_fp16)[name = tensor<string, []>("op_927_cast_fp16")];
            tensor<int32, [1]> var_928_axes_0 = const()[name = tensor<string, []>("op_928_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_928_cast_fp16 = squeeze(axes = var_928_axes_0, x = var_927_cast_fp16)[name = tensor<string, []>("op_928_cast_fp16")];
            tensor<string, []> var_934_pad_type_0 = const()[name = tensor<string, []>("op_934_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_934_strides_0 = const()[name = tensor<string, []>("op_934_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_934_pad_0 = const()[name = tensor<string, []>("op_934_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_934_dilations_0 = const()[name = tensor<string, []>("op_934_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_934_groups_0 = const()[name = tensor<string, []>("op_934_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3557248)))];
            tensor<fp16, [1, 64, 1, 197]> var_934_cast_fp16 = conv(dilations = var_934_dilations_0, groups = var_934_groups_0, pad = var_934_pad_0, pad_type = var_934_pad_type_0, strides = var_934_strides_0, weight = model_blocks_0_attn_q_projs_7_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_934_cast_fp16")];
            tensor<int32, [1]> var_935_axes_0 = const()[name = tensor<string, []>("op_935_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_935_cast_fp16 = squeeze(axes = var_935_axes_0, x = var_934_cast_fp16)[name = tensor<string, []>("op_935_cast_fp16")];
            tensor<int32, [3]> q_71_perm_0 = const()[name = tensor<string, []>("q_71_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_942_pad_type_0 = const()[name = tensor<string, []>("op_942_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_942_strides_0 = const()[name = tensor<string, []>("op_942_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_942_pad_0 = const()[name = tensor<string, []>("op_942_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_942_dilations_0 = const()[name = tensor<string, []>("op_942_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_942_groups_0 = const()[name = tensor<string, []>("op_942_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3655616)))];
            tensor<fp16, [1, 64, 1, 197]> var_942_cast_fp16 = conv(dilations = var_942_dilations_0, groups = var_942_groups_0, pad = var_942_pad_0, pad_type = var_942_pad_type_0, strides = var_942_strides_0, weight = model_blocks_0_attn_k_projs_7_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_942_cast_fp16")];
            tensor<int32, [1]> var_943_axes_0 = const()[name = tensor<string, []>("op_943_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_943_cast_fp16 = squeeze(axes = var_943_axes_0, x = var_942_cast_fp16)[name = tensor<string, []>("op_943_cast_fp16")];
            tensor<int32, [3]> k_43_perm_0 = const()[name = tensor<string, []>("k_43_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_950_pad_type_0 = const()[name = tensor<string, []>("op_950_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_950_strides_0 = const()[name = tensor<string, []>("op_950_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_950_pad_0 = const()[name = tensor<string, []>("op_950_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_950_dilations_0 = const()[name = tensor<string, []>("op_950_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_950_groups_0 = const()[name = tensor<string, []>("op_950_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3753984)))];
            tensor<fp16, [1, 64, 1, 197]> var_950_cast_fp16 = conv(dilations = var_950_dilations_0, groups = var_950_groups_0, pad = var_950_pad_0, pad_type = var_950_pad_type_0, strides = var_950_strides_0, weight = model_blocks_0_attn_v_projs_7_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
            tensor<int32, [1]> var_951_axes_0 = const()[name = tensor<string, []>("op_951_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_951_cast_fp16 = squeeze(axes = var_951_axes_0, x = var_950_cast_fp16)[name = tensor<string, []>("op_951_cast_fp16")];
            tensor<int32, [3]> v_43_perm_0 = const()[name = tensor<string, []>("v_43_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3852352)))];
            tensor<fp16, [1, 197, 64]> q_71_cast_fp16 = transpose(perm = q_71_perm_0, x = var_935_cast_fp16)[name = tensor<string, []>("transpose_493")];
            tensor<fp16, [1, 197, 64]> q_73_cast_fp16 = add(x = q_71_cast_fp16, y = model_blocks_0_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3852544)))];
            tensor<fp16, [1, 197, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_951_cast_fp16)[name = tensor<string, []>("transpose_491")];
            tensor<fp16, [1, 197, 64]> v_45_cast_fp16 = add(x = v_43_cast_fp16, y = model_blocks_0_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_45_cast_fp16")];
            tensor<int32, [1]> q_75_axes_0 = const()[name = tensor<string, []>("q_75_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_75_cast_fp16 = expand_dims(axes = q_75_axes_0, x = q_73_cast_fp16)[name = tensor<string, []>("q_75_cast_fp16")];
            tensor<int32, [1]> k_45_axes_0 = const()[name = tensor<string, []>("k_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_943_cast_fp16)[name = tensor<string, []>("transpose_492")];
            tensor<fp16, [1, 1, 197, 64]> k_45_cast_fp16 = expand_dims(axes = k_45_axes_0, x = k_43_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
            tensor<int32, [1]> v_47_axes_0 = const()[name = tensor<string, []>("v_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_47_cast_fp16 = expand_dims(axes = v_47_axes_0, x = v_45_cast_fp16)[name = tensor<string, []>("v_47_cast_fp16")];
            tensor<int32, [4]> var_960_begin_0 = const()[name = tensor<string, []>("op_960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_960_end_0 = const()[name = tensor<string, []>("op_960_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_960_end_mask_0 = const()[name = tensor<string, []>("op_960_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_960_cast_fp16 = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = q_75_cast_fp16)[name = tensor<string, []>("op_960_cast_fp16")];
            tensor<int32, [4]> var_964_begin_0 = const()[name = tensor<string, []>("op_964_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_964_end_0 = const()[name = tensor<string, []>("op_964_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_964_end_mask_0 = const()[name = tensor<string, []>("op_964_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_964_cast_fp16 = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = k_45_cast_fp16)[name = tensor<string, []>("op_964_cast_fp16")];
            tensor<int32, [4]> var_968_begin_0 = const()[name = tensor<string, []>("op_968_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_968_end_0 = const()[name = tensor<string, []>("op_968_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_968_end_mask_0 = const()[name = tensor<string, []>("op_968_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, x = q_75_cast_fp16)[name = tensor<string, []>("op_968_cast_fp16")];
            tensor<int32, [4]> var_972_begin_0 = const()[name = tensor<string, []>("op_972_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_972_end_0 = const()[name = tensor<string, []>("op_972_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_972_end_mask_0 = const()[name = tensor<string, []>("op_972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_972_cast_fp16 = slice_by_index(begin = var_972_begin_0, end = var_972_end_0, end_mask = var_972_end_mask_0, x = k_45_cast_fp16)[name = tensor<string, []>("op_972_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_974_cast_fp16 = mul(x = var_968_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_974_cast_fp16")];
            tensor<int32, [5]> var_982 = const()[name = tensor<string, []>("op_982"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_95_cast_fp16 = reshape(shape = var_982, x = var_968_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
            tensor<int32, [2]> var_984_split_sizes_0 = const()[name = tensor<string, []>("op_984_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_984_axis_0 = const()[name = tensor<string, []>("op_984_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_984_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_984_cast_fp16_1 = split(axis = var_984_axis_0, split_sizes = var_984_split_sizes_0, x = x_95_cast_fp16)[name = tensor<string, []>("op_984_cast_fp16")];
            tensor<int32, [1]> squeeze_28_axes_0 = const()[name = tensor<string, []>("squeeze_28_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_984_cast_fp16_0)[name = tensor<string, []>("squeeze_28_cast_fp16")];
            tensor<int32, [1]> squeeze_29_axes_0 = const()[name = tensor<string, []>("squeeze_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_984_cast_fp16_1)[name = tensor<string, []>("squeeze_29_cast_fp16")];
            tensor<fp16, []> const_119_promoted_to_fp16 = const()[name = tensor<string, []>("const_119_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_987_cast_fp16 = mul(x = squeeze_29_cast_fp16, y = const_119_promoted_to_fp16)[name = tensor<string, []>("op_987_cast_fp16")];
            tensor<int32, []> x_97_axis_0 = const()[name = tensor<string, []>("x_97_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_97_cast_fp16 = stack(axis = x_97_axis_0, values = (var_987_cast_fp16, squeeze_28_cast_fp16))[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<int32, [4]> var_993 = const()[name = tensor<string, []>("op_993"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_994_cast_fp16 = reshape(shape = var_993, x = x_97_cast_fp16)[name = tensor<string, []>("op_994_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_995_cast_fp16 = mul(x = var_994_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_995_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_15_cast_fp16 = add(x = var_974_cast_fp16, y = var_995_cast_fp16)[name = tensor<string, []>("q_patches_15_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_997_cast_fp16 = mul(x = var_972_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_997_cast_fp16")];
            tensor<int32, [5]> var_1005 = const()[name = tensor<string, []>("op_1005"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_101_cast_fp16 = reshape(shape = var_1005, x = var_972_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<int32, [2]> var_1007_split_sizes_0 = const()[name = tensor<string, []>("op_1007_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1007_axis_0 = const()[name = tensor<string, []>("op_1007_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1007_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1007_cast_fp16_1 = split(axis = var_1007_axis_0, split_sizes = var_1007_split_sizes_0, x = x_101_cast_fp16)[name = tensor<string, []>("op_1007_cast_fp16")];
            tensor<int32, [1]> squeeze_30_axes_0 = const()[name = tensor<string, []>("squeeze_30_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_1007_cast_fp16_0)[name = tensor<string, []>("squeeze_30_cast_fp16")];
            tensor<int32, [1]> squeeze_31_axes_0 = const()[name = tensor<string, []>("squeeze_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_1007_cast_fp16_1)[name = tensor<string, []>("squeeze_31_cast_fp16")];
            tensor<fp16, []> const_127_promoted_to_fp16 = const()[name = tensor<string, []>("const_127_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1010_cast_fp16 = mul(x = squeeze_31_cast_fp16, y = const_127_promoted_to_fp16)[name = tensor<string, []>("op_1010_cast_fp16")];
            tensor<int32, []> x_103_axis_0 = const()[name = tensor<string, []>("x_103_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_103_cast_fp16 = stack(axis = x_103_axis_0, values = (var_1010_cast_fp16, squeeze_30_cast_fp16))[name = tensor<string, []>("x_103_cast_fp16")];
            tensor<int32, [4]> var_1016 = const()[name = tensor<string, []>("op_1016"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1017_cast_fp16 = reshape(shape = var_1016, x = x_103_cast_fp16)[name = tensor<string, []>("op_1017_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1018_cast_fp16 = mul(x = var_1017_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1018_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_15_cast_fp16 = add(x = var_997_cast_fp16, y = var_1018_cast_fp16)[name = tensor<string, []>("k_patches_15_cast_fp16")];
            tensor<bool, []> var_1021_interleave_0 = const()[name = tensor<string, []>("op_1021_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1021_cast_fp16 = concat(axis = var_20, interleave = var_1021_interleave_0, values = (var_960_cast_fp16, q_patches_15_cast_fp16))[name = tensor<string, []>("op_1021_cast_fp16")];
            tensor<bool, []> var_1024_interleave_0 = const()[name = tensor<string, []>("op_1024_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1024_cast_fp16 = concat(axis = var_20, interleave = var_1024_interleave_0, values = (var_964_cast_fp16, k_patches_15_cast_fp16))[name = tensor<string, []>("op_1024_cast_fp16")];
            tensor<fp16, []> var_1026_to_fp16 = const()[name = tensor<string, []>("op_1026_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_79_cast_fp16 = mul(x = var_1021_cast_fp16, y = var_1026_to_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
            tensor<bool, []> attn_29_transpose_x_1 = const()[name = tensor<string, []>("attn_29_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_29_transpose_y_1 = const()[name = tensor<string, []>("attn_29_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_1, transpose_y = attn_29_transpose_y_1, x = q_79_cast_fp16, y = var_1024_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_31_cast_fp16 = softmax(axis = var_21, x = attn_29_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
            tensor<bool, []> var_1031_transpose_x_0 = const()[name = tensor<string, []>("op_1031_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1031_transpose_y_0 = const()[name = tensor<string, []>("op_1031_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_0, transpose_y = var_1031_transpose_y_0, x = attn_31_cast_fp16, y = v_47_cast_fp16)[name = tensor<string, []>("op_1031_cast_fp16")];
            tensor<int32, [1]> var_1032_axes_0 = const()[name = tensor<string, []>("op_1032_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1032_cast_fp16 = squeeze(axes = var_1032_axes_0, x = var_1031_cast_fp16)[name = tensor<string, []>("op_1032_cast_fp16")];
            tensor<string, []> var_1038_pad_type_0 = const()[name = tensor<string, []>("op_1038_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1038_strides_0 = const()[name = tensor<string, []>("op_1038_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1038_pad_0 = const()[name = tensor<string, []>("op_1038_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1038_dilations_0 = const()[name = tensor<string, []>("op_1038_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1038_groups_0 = const()[name = tensor<string, []>("op_1038_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3852736)))];
            tensor<fp16, [1, 64, 1, 197]> var_1038_cast_fp16 = conv(dilations = var_1038_dilations_0, groups = var_1038_groups_0, pad = var_1038_pad_0, pad_type = var_1038_pad_type_0, strides = var_1038_strides_0, weight = model_blocks_0_attn_q_projs_8_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1038_cast_fp16")];
            tensor<int32, [1]> var_1039_axes_0 = const()[name = tensor<string, []>("op_1039_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1039_cast_fp16 = squeeze(axes = var_1039_axes_0, x = var_1038_cast_fp16)[name = tensor<string, []>("op_1039_cast_fp16")];
            tensor<int32, [3]> q_81_perm_0 = const()[name = tensor<string, []>("q_81_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1046_pad_type_0 = const()[name = tensor<string, []>("op_1046_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1046_strides_0 = const()[name = tensor<string, []>("op_1046_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1046_pad_0 = const()[name = tensor<string, []>("op_1046_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1046_dilations_0 = const()[name = tensor<string, []>("op_1046_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1046_groups_0 = const()[name = tensor<string, []>("op_1046_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3951104)))];
            tensor<fp16, [1, 64, 1, 197]> var_1046_cast_fp16 = conv(dilations = var_1046_dilations_0, groups = var_1046_groups_0, pad = var_1046_pad_0, pad_type = var_1046_pad_type_0, strides = var_1046_strides_0, weight = model_blocks_0_attn_k_projs_8_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1046_cast_fp16")];
            tensor<int32, [1]> var_1047_axes_0 = const()[name = tensor<string, []>("op_1047_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1047_cast_fp16 = squeeze(axes = var_1047_axes_0, x = var_1046_cast_fp16)[name = tensor<string, []>("op_1047_cast_fp16")];
            tensor<int32, [3]> k_49_perm_0 = const()[name = tensor<string, []>("k_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1054_pad_type_0 = const()[name = tensor<string, []>("op_1054_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1054_strides_0 = const()[name = tensor<string, []>("op_1054_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1054_pad_0 = const()[name = tensor<string, []>("op_1054_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1054_dilations_0 = const()[name = tensor<string, []>("op_1054_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1054_groups_0 = const()[name = tensor<string, []>("op_1054_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4049472)))];
            tensor<fp16, [1, 64, 1, 197]> var_1054_cast_fp16 = conv(dilations = var_1054_dilations_0, groups = var_1054_groups_0, pad = var_1054_pad_0, pad_type = var_1054_pad_type_0, strides = var_1054_strides_0, weight = model_blocks_0_attn_v_projs_8_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1054_cast_fp16")];
            tensor<int32, [1]> var_1055_axes_0 = const()[name = tensor<string, []>("op_1055_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1055_cast_fp16 = squeeze(axes = var_1055_axes_0, x = var_1054_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")];
            tensor<int32, [3]> v_49_perm_0 = const()[name = tensor<string, []>("v_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4147840)))];
            tensor<fp16, [1, 197, 64]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = var_1039_cast_fp16)[name = tensor<string, []>("transpose_490")];
            tensor<fp16, [1, 197, 64]> q_83_cast_fp16 = add(x = q_81_cast_fp16, y = model_blocks_0_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_83_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4148032)))];
            tensor<fp16, [1, 197, 64]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_1055_cast_fp16)[name = tensor<string, []>("transpose_488")];
            tensor<fp16, [1, 197, 64]> v_51_cast_fp16 = add(x = v_49_cast_fp16, y = model_blocks_0_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_51_cast_fp16")];
            tensor<int32, [1]> q_85_axes_0 = const()[name = tensor<string, []>("q_85_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_85_cast_fp16 = expand_dims(axes = q_85_axes_0, x = q_83_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
            tensor<int32, [1]> k_51_axes_0 = const()[name = tensor<string, []>("k_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = var_1047_cast_fp16)[name = tensor<string, []>("transpose_489")];
            tensor<fp16, [1, 1, 197, 64]> k_51_cast_fp16 = expand_dims(axes = k_51_axes_0, x = k_49_cast_fp16)[name = tensor<string, []>("k_51_cast_fp16")];
            tensor<int32, [1]> v_53_axes_0 = const()[name = tensor<string, []>("v_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_53_cast_fp16 = expand_dims(axes = v_53_axes_0, x = v_51_cast_fp16)[name = tensor<string, []>("v_53_cast_fp16")];
            tensor<int32, [4]> var_1064_begin_0 = const()[name = tensor<string, []>("op_1064_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1064_end_0 = const()[name = tensor<string, []>("op_1064_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1064_end_mask_0 = const()[name = tensor<string, []>("op_1064_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1064_cast_fp16 = slice_by_index(begin = var_1064_begin_0, end = var_1064_end_0, end_mask = var_1064_end_mask_0, x = q_85_cast_fp16)[name = tensor<string, []>("op_1064_cast_fp16")];
            tensor<int32, [4]> var_1068_begin_0 = const()[name = tensor<string, []>("op_1068_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1068_end_0 = const()[name = tensor<string, []>("op_1068_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1068_end_mask_0 = const()[name = tensor<string, []>("op_1068_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1068_cast_fp16 = slice_by_index(begin = var_1068_begin_0, end = var_1068_end_0, end_mask = var_1068_end_mask_0, x = k_51_cast_fp16)[name = tensor<string, []>("op_1068_cast_fp16")];
            tensor<int32, [4]> var_1072_begin_0 = const()[name = tensor<string, []>("op_1072_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1072_end_0 = const()[name = tensor<string, []>("op_1072_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1072_end_mask_0 = const()[name = tensor<string, []>("op_1072_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1072_cast_fp16 = slice_by_index(begin = var_1072_begin_0, end = var_1072_end_0, end_mask = var_1072_end_mask_0, x = q_85_cast_fp16)[name = tensor<string, []>("op_1072_cast_fp16")];
            tensor<int32, [4]> var_1076_begin_0 = const()[name = tensor<string, []>("op_1076_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1076_end_0 = const()[name = tensor<string, []>("op_1076_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1076_end_mask_0 = const()[name = tensor<string, []>("op_1076_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1076_cast_fp16 = slice_by_index(begin = var_1076_begin_0, end = var_1076_end_0, end_mask = var_1076_end_mask_0, x = k_51_cast_fp16)[name = tensor<string, []>("op_1076_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1078_cast_fp16 = mul(x = var_1072_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1078_cast_fp16")];
            tensor<int32, [5]> var_1086 = const()[name = tensor<string, []>("op_1086"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_107_cast_fp16 = reshape(shape = var_1086, x = var_1072_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
            tensor<int32, [2]> var_1088_split_sizes_0 = const()[name = tensor<string, []>("op_1088_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1088_axis_0 = const()[name = tensor<string, []>("op_1088_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1088_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1088_cast_fp16_1 = split(axis = var_1088_axis_0, split_sizes = var_1088_split_sizes_0, x = x_107_cast_fp16)[name = tensor<string, []>("op_1088_cast_fp16")];
            tensor<int32, [1]> squeeze_32_axes_0 = const()[name = tensor<string, []>("squeeze_32_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_1088_cast_fp16_0)[name = tensor<string, []>("squeeze_32_cast_fp16")];
            tensor<int32, [1]> squeeze_33_axes_0 = const()[name = tensor<string, []>("squeeze_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_1088_cast_fp16_1)[name = tensor<string, []>("squeeze_33_cast_fp16")];
            tensor<fp16, []> const_135_promoted_to_fp16 = const()[name = tensor<string, []>("const_135_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1091_cast_fp16 = mul(x = squeeze_33_cast_fp16, y = const_135_promoted_to_fp16)[name = tensor<string, []>("op_1091_cast_fp16")];
            tensor<int32, []> x_109_axis_0 = const()[name = tensor<string, []>("x_109_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_109_cast_fp16 = stack(axis = x_109_axis_0, values = (var_1091_cast_fp16, squeeze_32_cast_fp16))[name = tensor<string, []>("x_109_cast_fp16")];
            tensor<int32, [4]> var_1097 = const()[name = tensor<string, []>("op_1097"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1098_cast_fp16 = reshape(shape = var_1097, x = x_109_cast_fp16)[name = tensor<string, []>("op_1098_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1099_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_17_cast_fp16 = add(x = var_1078_cast_fp16, y = var_1099_cast_fp16)[name = tensor<string, []>("q_patches_17_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1101_cast_fp16 = mul(x = var_1076_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1101_cast_fp16")];
            tensor<int32, [5]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_113_cast_fp16 = reshape(shape = var_1109, x = var_1076_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
            tensor<int32, [2]> var_1111_split_sizes_0 = const()[name = tensor<string, []>("op_1111_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1111_axis_0 = const()[name = tensor<string, []>("op_1111_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1111_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1111_cast_fp16_1 = split(axis = var_1111_axis_0, split_sizes = var_1111_split_sizes_0, x = x_113_cast_fp16)[name = tensor<string, []>("op_1111_cast_fp16")];
            tensor<int32, [1]> squeeze_34_axes_0 = const()[name = tensor<string, []>("squeeze_34_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_1111_cast_fp16_0)[name = tensor<string, []>("squeeze_34_cast_fp16")];
            tensor<int32, [1]> squeeze_35_axes_0 = const()[name = tensor<string, []>("squeeze_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_1111_cast_fp16_1)[name = tensor<string, []>("squeeze_35_cast_fp16")];
            tensor<fp16, []> const_143_promoted_to_fp16 = const()[name = tensor<string, []>("const_143_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1114_cast_fp16 = mul(x = squeeze_35_cast_fp16, y = const_143_promoted_to_fp16)[name = tensor<string, []>("op_1114_cast_fp16")];
            tensor<int32, []> x_115_axis_0 = const()[name = tensor<string, []>("x_115_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_115_cast_fp16 = stack(axis = x_115_axis_0, values = (var_1114_cast_fp16, squeeze_34_cast_fp16))[name = tensor<string, []>("x_115_cast_fp16")];
            tensor<int32, [4]> var_1120 = const()[name = tensor<string, []>("op_1120"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1121_cast_fp16 = reshape(shape = var_1120, x = x_115_cast_fp16)[name = tensor<string, []>("op_1121_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1122_cast_fp16 = mul(x = var_1121_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1122_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_17_cast_fp16 = add(x = var_1101_cast_fp16, y = var_1122_cast_fp16)[name = tensor<string, []>("k_patches_17_cast_fp16")];
            tensor<bool, []> var_1125_interleave_0 = const()[name = tensor<string, []>("op_1125_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1125_cast_fp16 = concat(axis = var_20, interleave = var_1125_interleave_0, values = (var_1064_cast_fp16, q_patches_17_cast_fp16))[name = tensor<string, []>("op_1125_cast_fp16")];
            tensor<bool, []> var_1128_interleave_0 = const()[name = tensor<string, []>("op_1128_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1128_cast_fp16 = concat(axis = var_20, interleave = var_1128_interleave_0, values = (var_1068_cast_fp16, k_patches_17_cast_fp16))[name = tensor<string, []>("op_1128_cast_fp16")];
            tensor<fp16, []> var_1130_to_fp16 = const()[name = tensor<string, []>("op_1130_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_89_cast_fp16 = mul(x = var_1125_cast_fp16, y = var_1130_to_fp16)[name = tensor<string, []>("q_89_cast_fp16")];
            tensor<bool, []> attn_33_transpose_x_1 = const()[name = tensor<string, []>("attn_33_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_33_transpose_y_1 = const()[name = tensor<string, []>("attn_33_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_1, transpose_y = attn_33_transpose_y_1, x = q_89_cast_fp16, y = var_1128_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_35_cast_fp16 = softmax(axis = var_21, x = attn_33_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
            tensor<bool, []> var_1135_transpose_x_0 = const()[name = tensor<string, []>("op_1135_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1135_transpose_y_0 = const()[name = tensor<string, []>("op_1135_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1135_cast_fp16 = matmul(transpose_x = var_1135_transpose_x_0, transpose_y = var_1135_transpose_y_0, x = attn_35_cast_fp16, y = v_53_cast_fp16)[name = tensor<string, []>("op_1135_cast_fp16")];
            tensor<int32, [1]> var_1136_axes_0 = const()[name = tensor<string, []>("op_1136_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1136_cast_fp16 = squeeze(axes = var_1136_axes_0, x = var_1135_cast_fp16)[name = tensor<string, []>("op_1136_cast_fp16")];
            tensor<string, []> var_1142_pad_type_0 = const()[name = tensor<string, []>("op_1142_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1142_strides_0 = const()[name = tensor<string, []>("op_1142_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1142_pad_0 = const()[name = tensor<string, []>("op_1142_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1142_dilations_0 = const()[name = tensor<string, []>("op_1142_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1142_groups_0 = const()[name = tensor<string, []>("op_1142_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4148224)))];
            tensor<fp16, [1, 64, 1, 197]> var_1142_cast_fp16 = conv(dilations = var_1142_dilations_0, groups = var_1142_groups_0, pad = var_1142_pad_0, pad_type = var_1142_pad_type_0, strides = var_1142_strides_0, weight = model_blocks_0_attn_q_projs_9_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1142_cast_fp16")];
            tensor<int32, [1]> var_1143_axes_0 = const()[name = tensor<string, []>("op_1143_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1143_cast_fp16 = squeeze(axes = var_1143_axes_0, x = var_1142_cast_fp16)[name = tensor<string, []>("op_1143_cast_fp16")];
            tensor<int32, [3]> q_91_perm_0 = const()[name = tensor<string, []>("q_91_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1150_pad_type_0 = const()[name = tensor<string, []>("op_1150_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1150_strides_0 = const()[name = tensor<string, []>("op_1150_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1150_pad_0 = const()[name = tensor<string, []>("op_1150_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1150_dilations_0 = const()[name = tensor<string, []>("op_1150_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1150_groups_0 = const()[name = tensor<string, []>("op_1150_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4246592)))];
            tensor<fp16, [1, 64, 1, 197]> var_1150_cast_fp16 = conv(dilations = var_1150_dilations_0, groups = var_1150_groups_0, pad = var_1150_pad_0, pad_type = var_1150_pad_type_0, strides = var_1150_strides_0, weight = model_blocks_0_attn_k_projs_9_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1150_cast_fp16")];
            tensor<int32, [1]> var_1151_axes_0 = const()[name = tensor<string, []>("op_1151_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1151_cast_fp16 = squeeze(axes = var_1151_axes_0, x = var_1150_cast_fp16)[name = tensor<string, []>("op_1151_cast_fp16")];
            tensor<int32, [3]> k_55_perm_0 = const()[name = tensor<string, []>("k_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1158_pad_type_0 = const()[name = tensor<string, []>("op_1158_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1158_strides_0 = const()[name = tensor<string, []>("op_1158_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1158_pad_0 = const()[name = tensor<string, []>("op_1158_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1158_dilations_0 = const()[name = tensor<string, []>("op_1158_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1158_groups_0 = const()[name = tensor<string, []>("op_1158_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4344960)))];
            tensor<fp16, [1, 64, 1, 197]> var_1158_cast_fp16 = conv(dilations = var_1158_dilations_0, groups = var_1158_groups_0, pad = var_1158_pad_0, pad_type = var_1158_pad_type_0, strides = var_1158_strides_0, weight = model_blocks_0_attn_v_projs_9_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1158_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = tensor<string, []>("op_1159_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1159_cast_fp16 = squeeze(axes = var_1159_axes_0, x = var_1158_cast_fp16)[name = tensor<string, []>("op_1159_cast_fp16")];
            tensor<int32, [3]> v_55_perm_0 = const()[name = tensor<string, []>("v_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4443328)))];
            tensor<fp16, [1, 197, 64]> q_91_cast_fp16 = transpose(perm = q_91_perm_0, x = var_1143_cast_fp16)[name = tensor<string, []>("transpose_487")];
            tensor<fp16, [1, 197, 64]> q_93_cast_fp16 = add(x = q_91_cast_fp16, y = model_blocks_0_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_93_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4443520)))];
            tensor<fp16, [1, 197, 64]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = var_1159_cast_fp16)[name = tensor<string, []>("transpose_485")];
            tensor<fp16, [1, 197, 64]> v_57_cast_fp16 = add(x = v_55_cast_fp16, y = model_blocks_0_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_57_cast_fp16")];
            tensor<int32, [1]> q_95_axes_0 = const()[name = tensor<string, []>("q_95_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_95_cast_fp16 = expand_dims(axes = q_95_axes_0, x = q_93_cast_fp16)[name = tensor<string, []>("q_95_cast_fp16")];
            tensor<int32, [1]> k_57_axes_0 = const()[name = tensor<string, []>("k_57_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_55_cast_fp16 = transpose(perm = k_55_perm_0, x = var_1151_cast_fp16)[name = tensor<string, []>("transpose_486")];
            tensor<fp16, [1, 1, 197, 64]> k_57_cast_fp16 = expand_dims(axes = k_57_axes_0, x = k_55_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
            tensor<int32, [1]> v_59_axes_0 = const()[name = tensor<string, []>("v_59_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_59_cast_fp16 = expand_dims(axes = v_59_axes_0, x = v_57_cast_fp16)[name = tensor<string, []>("v_59_cast_fp16")];
            tensor<int32, [4]> var_1168_begin_0 = const()[name = tensor<string, []>("op_1168_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1168_end_0 = const()[name = tensor<string, []>("op_1168_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1168_end_mask_0 = const()[name = tensor<string, []>("op_1168_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, x = q_95_cast_fp16)[name = tensor<string, []>("op_1168_cast_fp16")];
            tensor<int32, [4]> var_1172_begin_0 = const()[name = tensor<string, []>("op_1172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1172_end_0 = const()[name = tensor<string, []>("op_1172_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1172_end_mask_0 = const()[name = tensor<string, []>("op_1172_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1172_cast_fp16 = slice_by_index(begin = var_1172_begin_0, end = var_1172_end_0, end_mask = var_1172_end_mask_0, x = k_57_cast_fp16)[name = tensor<string, []>("op_1172_cast_fp16")];
            tensor<int32, [4]> var_1176_begin_0 = const()[name = tensor<string, []>("op_1176_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1176_end_0 = const()[name = tensor<string, []>("op_1176_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1176_end_mask_0 = const()[name = tensor<string, []>("op_1176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1176_cast_fp16 = slice_by_index(begin = var_1176_begin_0, end = var_1176_end_0, end_mask = var_1176_end_mask_0, x = q_95_cast_fp16)[name = tensor<string, []>("op_1176_cast_fp16")];
            tensor<int32, [4]> var_1180_begin_0 = const()[name = tensor<string, []>("op_1180_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1180_end_0 = const()[name = tensor<string, []>("op_1180_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1180_end_mask_0 = const()[name = tensor<string, []>("op_1180_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1180_cast_fp16 = slice_by_index(begin = var_1180_begin_0, end = var_1180_end_0, end_mask = var_1180_end_mask_0, x = k_57_cast_fp16)[name = tensor<string, []>("op_1180_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1182_cast_fp16 = mul(x = var_1176_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1182_cast_fp16")];
            tensor<int32, [5]> var_1190 = const()[name = tensor<string, []>("op_1190"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_119_cast_fp16 = reshape(shape = var_1190, x = var_1176_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
            tensor<int32, [2]> var_1192_split_sizes_0 = const()[name = tensor<string, []>("op_1192_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1192_axis_0 = const()[name = tensor<string, []>("op_1192_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = var_1192_split_sizes_0, x = x_119_cast_fp16)[name = tensor<string, []>("op_1192_cast_fp16")];
            tensor<int32, [1]> squeeze_36_axes_0 = const()[name = tensor<string, []>("squeeze_36_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_36_cast_fp16 = squeeze(axes = squeeze_36_axes_0, x = var_1192_cast_fp16_0)[name = tensor<string, []>("squeeze_36_cast_fp16")];
            tensor<int32, [1]> squeeze_37_axes_0 = const()[name = tensor<string, []>("squeeze_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_37_cast_fp16 = squeeze(axes = squeeze_37_axes_0, x = var_1192_cast_fp16_1)[name = tensor<string, []>("squeeze_37_cast_fp16")];
            tensor<fp16, []> const_151_promoted_to_fp16 = const()[name = tensor<string, []>("const_151_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1195_cast_fp16 = mul(x = squeeze_37_cast_fp16, y = const_151_promoted_to_fp16)[name = tensor<string, []>("op_1195_cast_fp16")];
            tensor<int32, []> x_121_axis_0 = const()[name = tensor<string, []>("x_121_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_121_cast_fp16 = stack(axis = x_121_axis_0, values = (var_1195_cast_fp16, squeeze_36_cast_fp16))[name = tensor<string, []>("x_121_cast_fp16")];
            tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1202_cast_fp16 = reshape(shape = var_1201, x = x_121_cast_fp16)[name = tensor<string, []>("op_1202_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1203_cast_fp16 = mul(x = var_1202_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1203_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_19_cast_fp16 = add(x = var_1182_cast_fp16, y = var_1203_cast_fp16)[name = tensor<string, []>("q_patches_19_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1205_cast_fp16 = mul(x = var_1180_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1205_cast_fp16")];
            tensor<int32, [5]> var_1213 = const()[name = tensor<string, []>("op_1213"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_125_cast_fp16 = reshape(shape = var_1213, x = var_1180_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
            tensor<int32, [2]> var_1215_split_sizes_0 = const()[name = tensor<string, []>("op_1215_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1215_axis_0 = const()[name = tensor<string, []>("op_1215_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1215_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1215_cast_fp16_1 = split(axis = var_1215_axis_0, split_sizes = var_1215_split_sizes_0, x = x_125_cast_fp16)[name = tensor<string, []>("op_1215_cast_fp16")];
            tensor<int32, [1]> squeeze_38_axes_0 = const()[name = tensor<string, []>("squeeze_38_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_38_cast_fp16 = squeeze(axes = squeeze_38_axes_0, x = var_1215_cast_fp16_0)[name = tensor<string, []>("squeeze_38_cast_fp16")];
            tensor<int32, [1]> squeeze_39_axes_0 = const()[name = tensor<string, []>("squeeze_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_39_cast_fp16 = squeeze(axes = squeeze_39_axes_0, x = var_1215_cast_fp16_1)[name = tensor<string, []>("squeeze_39_cast_fp16")];
            tensor<fp16, []> const_159_promoted_to_fp16 = const()[name = tensor<string, []>("const_159_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1218_cast_fp16 = mul(x = squeeze_39_cast_fp16, y = const_159_promoted_to_fp16)[name = tensor<string, []>("op_1218_cast_fp16")];
            tensor<int32, []> x_127_axis_0 = const()[name = tensor<string, []>("x_127_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_127_cast_fp16 = stack(axis = x_127_axis_0, values = (var_1218_cast_fp16, squeeze_38_cast_fp16))[name = tensor<string, []>("x_127_cast_fp16")];
            tensor<int32, [4]> var_1224 = const()[name = tensor<string, []>("op_1224"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1225_cast_fp16 = reshape(shape = var_1224, x = x_127_cast_fp16)[name = tensor<string, []>("op_1225_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1226_cast_fp16 = mul(x = var_1225_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1226_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_19_cast_fp16 = add(x = var_1205_cast_fp16, y = var_1226_cast_fp16)[name = tensor<string, []>("k_patches_19_cast_fp16")];
            tensor<bool, []> var_1229_interleave_0 = const()[name = tensor<string, []>("op_1229_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1229_cast_fp16 = concat(axis = var_20, interleave = var_1229_interleave_0, values = (var_1168_cast_fp16, q_patches_19_cast_fp16))[name = tensor<string, []>("op_1229_cast_fp16")];
            tensor<bool, []> var_1232_interleave_0 = const()[name = tensor<string, []>("op_1232_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1232_cast_fp16 = concat(axis = var_20, interleave = var_1232_interleave_0, values = (var_1172_cast_fp16, k_patches_19_cast_fp16))[name = tensor<string, []>("op_1232_cast_fp16")];
            tensor<fp16, []> var_1234_to_fp16 = const()[name = tensor<string, []>("op_1234_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_99_cast_fp16 = mul(x = var_1229_cast_fp16, y = var_1234_to_fp16)[name = tensor<string, []>("q_99_cast_fp16")];
            tensor<bool, []> attn_37_transpose_x_1 = const()[name = tensor<string, []>("attn_37_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_37_transpose_y_1 = const()[name = tensor<string, []>("attn_37_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_1, transpose_y = attn_37_transpose_y_1, x = q_99_cast_fp16, y = var_1232_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_39_cast_fp16 = softmax(axis = var_21, x = attn_37_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
            tensor<bool, []> var_1239_transpose_x_0 = const()[name = tensor<string, []>("op_1239_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1239_transpose_y_0 = const()[name = tensor<string, []>("op_1239_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1239_cast_fp16 = matmul(transpose_x = var_1239_transpose_x_0, transpose_y = var_1239_transpose_y_0, x = attn_39_cast_fp16, y = v_59_cast_fp16)[name = tensor<string, []>("op_1239_cast_fp16")];
            tensor<int32, [1]> var_1240_axes_0 = const()[name = tensor<string, []>("op_1240_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1240_cast_fp16 = squeeze(axes = var_1240_axes_0, x = var_1239_cast_fp16)[name = tensor<string, []>("op_1240_cast_fp16")];
            tensor<string, []> var_1246_pad_type_0 = const()[name = tensor<string, []>("op_1246_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1246_strides_0 = const()[name = tensor<string, []>("op_1246_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1246_pad_0 = const()[name = tensor<string, []>("op_1246_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1246_dilations_0 = const()[name = tensor<string, []>("op_1246_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1246_groups_0 = const()[name = tensor<string, []>("op_1246_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4443712)))];
            tensor<fp16, [1, 64, 1, 197]> var_1246_cast_fp16 = conv(dilations = var_1246_dilations_0, groups = var_1246_groups_0, pad = var_1246_pad_0, pad_type = var_1246_pad_type_0, strides = var_1246_strides_0, weight = model_blocks_0_attn_q_projs_10_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1246_cast_fp16")];
            tensor<int32, [1]> var_1247_axes_0 = const()[name = tensor<string, []>("op_1247_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1247_cast_fp16 = squeeze(axes = var_1247_axes_0, x = var_1246_cast_fp16)[name = tensor<string, []>("op_1247_cast_fp16")];
            tensor<int32, [3]> q_101_perm_0 = const()[name = tensor<string, []>("q_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1254_pad_type_0 = const()[name = tensor<string, []>("op_1254_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1254_strides_0 = const()[name = tensor<string, []>("op_1254_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1254_pad_0 = const()[name = tensor<string, []>("op_1254_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1254_dilations_0 = const()[name = tensor<string, []>("op_1254_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1254_groups_0 = const()[name = tensor<string, []>("op_1254_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4542080)))];
            tensor<fp16, [1, 64, 1, 197]> var_1254_cast_fp16 = conv(dilations = var_1254_dilations_0, groups = var_1254_groups_0, pad = var_1254_pad_0, pad_type = var_1254_pad_type_0, strides = var_1254_strides_0, weight = model_blocks_0_attn_k_projs_10_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1254_cast_fp16")];
            tensor<int32, [1]> var_1255_axes_0 = const()[name = tensor<string, []>("op_1255_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1255_cast_fp16 = squeeze(axes = var_1255_axes_0, x = var_1254_cast_fp16)[name = tensor<string, []>("op_1255_cast_fp16")];
            tensor<int32, [3]> k_61_perm_0 = const()[name = tensor<string, []>("k_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1262_pad_type_0 = const()[name = tensor<string, []>("op_1262_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1262_strides_0 = const()[name = tensor<string, []>("op_1262_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1262_pad_0 = const()[name = tensor<string, []>("op_1262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1262_dilations_0 = const()[name = tensor<string, []>("op_1262_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1262_groups_0 = const()[name = tensor<string, []>("op_1262_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4640448)))];
            tensor<fp16, [1, 64, 1, 197]> var_1262_cast_fp16 = conv(dilations = var_1262_dilations_0, groups = var_1262_groups_0, pad = var_1262_pad_0, pad_type = var_1262_pad_type_0, strides = var_1262_strides_0, weight = model_blocks_0_attn_v_projs_10_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1262_cast_fp16")];
            tensor<int32, [1]> var_1263_axes_0 = const()[name = tensor<string, []>("op_1263_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1263_cast_fp16 = squeeze(axes = var_1263_axes_0, x = var_1262_cast_fp16)[name = tensor<string, []>("op_1263_cast_fp16")];
            tensor<int32, [3]> v_61_perm_0 = const()[name = tensor<string, []>("v_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4738816)))];
            tensor<fp16, [1, 197, 64]> q_101_cast_fp16 = transpose(perm = q_101_perm_0, x = var_1247_cast_fp16)[name = tensor<string, []>("transpose_484")];
            tensor<fp16, [1, 197, 64]> q_103_cast_fp16 = add(x = q_101_cast_fp16, y = model_blocks_0_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_103_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739008)))];
            tensor<fp16, [1, 197, 64]> v_61_cast_fp16 = transpose(perm = v_61_perm_0, x = var_1263_cast_fp16)[name = tensor<string, []>("transpose_482")];
            tensor<fp16, [1, 197, 64]> v_63_cast_fp16 = add(x = v_61_cast_fp16, y = model_blocks_0_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_63_cast_fp16")];
            tensor<int32, [1]> q_105_axes_0 = const()[name = tensor<string, []>("q_105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_105_cast_fp16 = expand_dims(axes = q_105_axes_0, x = q_103_cast_fp16)[name = tensor<string, []>("q_105_cast_fp16")];
            tensor<int32, [1]> k_63_axes_0 = const()[name = tensor<string, []>("k_63_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_61_cast_fp16 = transpose(perm = k_61_perm_0, x = var_1255_cast_fp16)[name = tensor<string, []>("transpose_483")];
            tensor<fp16, [1, 1, 197, 64]> k_63_cast_fp16 = expand_dims(axes = k_63_axes_0, x = k_61_cast_fp16)[name = tensor<string, []>("k_63_cast_fp16")];
            tensor<int32, [1]> v_65_axes_0 = const()[name = tensor<string, []>("v_65_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_65_cast_fp16 = expand_dims(axes = v_65_axes_0, x = v_63_cast_fp16)[name = tensor<string, []>("v_65_cast_fp16")];
            tensor<int32, [4]> var_1272_begin_0 = const()[name = tensor<string, []>("op_1272_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1272_end_0 = const()[name = tensor<string, []>("op_1272_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1272_end_mask_0 = const()[name = tensor<string, []>("op_1272_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, x = q_105_cast_fp16)[name = tensor<string, []>("op_1272_cast_fp16")];
            tensor<int32, [4]> var_1276_begin_0 = const()[name = tensor<string, []>("op_1276_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1276_end_0 = const()[name = tensor<string, []>("op_1276_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1276_end_mask_0 = const()[name = tensor<string, []>("op_1276_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1276_cast_fp16 = slice_by_index(begin = var_1276_begin_0, end = var_1276_end_0, end_mask = var_1276_end_mask_0, x = k_63_cast_fp16)[name = tensor<string, []>("op_1276_cast_fp16")];
            tensor<int32, [4]> var_1280_begin_0 = const()[name = tensor<string, []>("op_1280_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1280_end_0 = const()[name = tensor<string, []>("op_1280_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1280_end_mask_0 = const()[name = tensor<string, []>("op_1280_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1280_cast_fp16 = slice_by_index(begin = var_1280_begin_0, end = var_1280_end_0, end_mask = var_1280_end_mask_0, x = q_105_cast_fp16)[name = tensor<string, []>("op_1280_cast_fp16")];
            tensor<int32, [4]> var_1284_begin_0 = const()[name = tensor<string, []>("op_1284_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1284_end_0 = const()[name = tensor<string, []>("op_1284_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1284_end_mask_0 = const()[name = tensor<string, []>("op_1284_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1284_cast_fp16 = slice_by_index(begin = var_1284_begin_0, end = var_1284_end_0, end_mask = var_1284_end_mask_0, x = k_63_cast_fp16)[name = tensor<string, []>("op_1284_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1286_cast_fp16 = mul(x = var_1280_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1286_cast_fp16")];
            tensor<int32, [5]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_131_cast_fp16 = reshape(shape = var_1294, x = var_1280_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
            tensor<int32, [2]> var_1296_split_sizes_0 = const()[name = tensor<string, []>("op_1296_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1296_axis_0 = const()[name = tensor<string, []>("op_1296_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1296_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1296_cast_fp16_1 = split(axis = var_1296_axis_0, split_sizes = var_1296_split_sizes_0, x = x_131_cast_fp16)[name = tensor<string, []>("op_1296_cast_fp16")];
            tensor<int32, [1]> squeeze_40_axes_0 = const()[name = tensor<string, []>("squeeze_40_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_40_cast_fp16 = squeeze(axes = squeeze_40_axes_0, x = var_1296_cast_fp16_0)[name = tensor<string, []>("squeeze_40_cast_fp16")];
            tensor<int32, [1]> squeeze_41_axes_0 = const()[name = tensor<string, []>("squeeze_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_41_cast_fp16 = squeeze(axes = squeeze_41_axes_0, x = var_1296_cast_fp16_1)[name = tensor<string, []>("squeeze_41_cast_fp16")];
            tensor<fp16, []> const_167_promoted_to_fp16 = const()[name = tensor<string, []>("const_167_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1299_cast_fp16 = mul(x = squeeze_41_cast_fp16, y = const_167_promoted_to_fp16)[name = tensor<string, []>("op_1299_cast_fp16")];
            tensor<int32, []> x_133_axis_0 = const()[name = tensor<string, []>("x_133_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_133_cast_fp16 = stack(axis = x_133_axis_0, values = (var_1299_cast_fp16, squeeze_40_cast_fp16))[name = tensor<string, []>("x_133_cast_fp16")];
            tensor<int32, [4]> var_1305 = const()[name = tensor<string, []>("op_1305"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1306_cast_fp16 = reshape(shape = var_1305, x = x_133_cast_fp16)[name = tensor<string, []>("op_1306_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1307_cast_fp16 = mul(x = var_1306_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1307_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_21_cast_fp16 = add(x = var_1286_cast_fp16, y = var_1307_cast_fp16)[name = tensor<string, []>("q_patches_21_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1309_cast_fp16 = mul(x = var_1284_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1309_cast_fp16")];
            tensor<int32, [5]> var_1317 = const()[name = tensor<string, []>("op_1317"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_137_cast_fp16 = reshape(shape = var_1317, x = var_1284_cast_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
            tensor<int32, [2]> var_1319_split_sizes_0 = const()[name = tensor<string, []>("op_1319_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1319_axis_0 = const()[name = tensor<string, []>("op_1319_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1319_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1319_cast_fp16_1 = split(axis = var_1319_axis_0, split_sizes = var_1319_split_sizes_0, x = x_137_cast_fp16)[name = tensor<string, []>("op_1319_cast_fp16")];
            tensor<int32, [1]> squeeze_42_axes_0 = const()[name = tensor<string, []>("squeeze_42_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_42_cast_fp16 = squeeze(axes = squeeze_42_axes_0, x = var_1319_cast_fp16_0)[name = tensor<string, []>("squeeze_42_cast_fp16")];
            tensor<int32, [1]> squeeze_43_axes_0 = const()[name = tensor<string, []>("squeeze_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_43_cast_fp16 = squeeze(axes = squeeze_43_axes_0, x = var_1319_cast_fp16_1)[name = tensor<string, []>("squeeze_43_cast_fp16")];
            tensor<fp16, []> const_175_promoted_to_fp16 = const()[name = tensor<string, []>("const_175_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1322_cast_fp16 = mul(x = squeeze_43_cast_fp16, y = const_175_promoted_to_fp16)[name = tensor<string, []>("op_1322_cast_fp16")];
            tensor<int32, []> x_139_axis_0 = const()[name = tensor<string, []>("x_139_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_139_cast_fp16 = stack(axis = x_139_axis_0, values = (var_1322_cast_fp16, squeeze_42_cast_fp16))[name = tensor<string, []>("x_139_cast_fp16")];
            tensor<int32, [4]> var_1328 = const()[name = tensor<string, []>("op_1328"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1329_cast_fp16 = reshape(shape = var_1328, x = x_139_cast_fp16)[name = tensor<string, []>("op_1329_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1330_cast_fp16 = mul(x = var_1329_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1330_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_21_cast_fp16 = add(x = var_1309_cast_fp16, y = var_1330_cast_fp16)[name = tensor<string, []>("k_patches_21_cast_fp16")];
            tensor<bool, []> var_1333_interleave_0 = const()[name = tensor<string, []>("op_1333_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1333_cast_fp16 = concat(axis = var_20, interleave = var_1333_interleave_0, values = (var_1272_cast_fp16, q_patches_21_cast_fp16))[name = tensor<string, []>("op_1333_cast_fp16")];
            tensor<bool, []> var_1336_interleave_0 = const()[name = tensor<string, []>("op_1336_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1336_cast_fp16 = concat(axis = var_20, interleave = var_1336_interleave_0, values = (var_1276_cast_fp16, k_patches_21_cast_fp16))[name = tensor<string, []>("op_1336_cast_fp16")];
            tensor<fp16, []> var_1338_to_fp16 = const()[name = tensor<string, []>("op_1338_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_109_cast_fp16 = mul(x = var_1333_cast_fp16, y = var_1338_to_fp16)[name = tensor<string, []>("q_109_cast_fp16")];
            tensor<bool, []> attn_41_transpose_x_1 = const()[name = tensor<string, []>("attn_41_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_41_transpose_y_1 = const()[name = tensor<string, []>("attn_41_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_1, transpose_y = attn_41_transpose_y_1, x = q_109_cast_fp16, y = var_1336_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_43_cast_fp16 = softmax(axis = var_21, x = attn_41_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
            tensor<bool, []> var_1343_transpose_x_0 = const()[name = tensor<string, []>("op_1343_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1343_transpose_y_0 = const()[name = tensor<string, []>("op_1343_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1343_cast_fp16 = matmul(transpose_x = var_1343_transpose_x_0, transpose_y = var_1343_transpose_y_0, x = attn_43_cast_fp16, y = v_65_cast_fp16)[name = tensor<string, []>("op_1343_cast_fp16")];
            tensor<int32, [1]> var_1344_axes_0 = const()[name = tensor<string, []>("op_1344_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1344_cast_fp16 = squeeze(axes = var_1344_axes_0, x = var_1343_cast_fp16)[name = tensor<string, []>("op_1344_cast_fp16")];
            tensor<string, []> var_1350_pad_type_0 = const()[name = tensor<string, []>("op_1350_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1350_strides_0 = const()[name = tensor<string, []>("op_1350_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1350_pad_0 = const()[name = tensor<string, []>("op_1350_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1350_dilations_0 = const()[name = tensor<string, []>("op_1350_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1350_groups_0 = const()[name = tensor<string, []>("op_1350_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739200)))];
            tensor<fp16, [1, 64, 1, 197]> var_1350_cast_fp16 = conv(dilations = var_1350_dilations_0, groups = var_1350_groups_0, pad = var_1350_pad_0, pad_type = var_1350_pad_type_0, strides = var_1350_strides_0, weight = model_blocks_0_attn_q_projs_11_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1350_cast_fp16")];
            tensor<int32, [1]> var_1351_axes_0 = const()[name = tensor<string, []>("op_1351_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1351_cast_fp16 = squeeze(axes = var_1351_axes_0, x = var_1350_cast_fp16)[name = tensor<string, []>("op_1351_cast_fp16")];
            tensor<int32, [3]> q_111_perm_0 = const()[name = tensor<string, []>("q_111_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1358_pad_type_0 = const()[name = tensor<string, []>("op_1358_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1358_strides_0 = const()[name = tensor<string, []>("op_1358_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1358_pad_0 = const()[name = tensor<string, []>("op_1358_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1358_dilations_0 = const()[name = tensor<string, []>("op_1358_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1358_groups_0 = const()[name = tensor<string, []>("op_1358_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4837568)))];
            tensor<fp16, [1, 64, 1, 197]> var_1358_cast_fp16 = conv(dilations = var_1358_dilations_0, groups = var_1358_groups_0, pad = var_1358_pad_0, pad_type = var_1358_pad_type_0, strides = var_1358_strides_0, weight = model_blocks_0_attn_k_projs_11_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1358_cast_fp16")];
            tensor<int32, [1]> var_1359_axes_0 = const()[name = tensor<string, []>("op_1359_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1359_cast_fp16 = squeeze(axes = var_1359_axes_0, x = var_1358_cast_fp16)[name = tensor<string, []>("op_1359_cast_fp16")];
            tensor<int32, [3]> k_67_perm_0 = const()[name = tensor<string, []>("k_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1366_pad_type_0 = const()[name = tensor<string, []>("op_1366_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1366_strides_0 = const()[name = tensor<string, []>("op_1366_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1366_pad_0 = const()[name = tensor<string, []>("op_1366_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1366_dilations_0 = const()[name = tensor<string, []>("op_1366_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1366_groups_0 = const()[name = tensor<string, []>("op_1366_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_0_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4935936)))];
            tensor<fp16, [1, 64, 1, 197]> var_1366_cast_fp16 = conv(dilations = var_1366_dilations_0, groups = var_1366_groups_0, pad = var_1366_pad_0, pad_type = var_1366_pad_type_0, strides = var_1366_strides_0, weight = model_blocks_0_attn_v_projs_11_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_1366_cast_fp16")];
            tensor<int32, [1]> var_1367_axes_0 = const()[name = tensor<string, []>("op_1367_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1367_cast_fp16 = squeeze(axes = var_1367_axes_0, x = var_1366_cast_fp16)[name = tensor<string, []>("op_1367_cast_fp16")];
            tensor<int32, [3]> v_67_perm_0 = const()[name = tensor<string, []>("v_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_0_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5034304)))];
            tensor<fp16, [1, 197, 64]> q_111_cast_fp16 = transpose(perm = q_111_perm_0, x = var_1351_cast_fp16)[name = tensor<string, []>("transpose_481")];
            tensor<fp16, [1, 197, 64]> q_113_cast_fp16 = add(x = q_111_cast_fp16, y = model_blocks_0_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_113_cast_fp16")];
            tensor<fp16, [64]> model_blocks_0_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5034496)))];
            tensor<fp16, [1, 197, 64]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = var_1367_cast_fp16)[name = tensor<string, []>("transpose_479")];
            tensor<fp16, [1, 197, 64]> v_69_cast_fp16 = add(x = v_67_cast_fp16, y = model_blocks_0_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_69_cast_fp16")];
            tensor<int32, [1]> q_115_axes_0 = const()[name = tensor<string, []>("q_115_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_115_cast_fp16 = expand_dims(axes = q_115_axes_0, x = q_113_cast_fp16)[name = tensor<string, []>("q_115_cast_fp16")];
            tensor<int32, [1]> k_69_axes_0 = const()[name = tensor<string, []>("k_69_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_67_cast_fp16 = transpose(perm = k_67_perm_0, x = var_1359_cast_fp16)[name = tensor<string, []>("transpose_480")];
            tensor<fp16, [1, 1, 197, 64]> k_69_cast_fp16 = expand_dims(axes = k_69_axes_0, x = k_67_cast_fp16)[name = tensor<string, []>("k_69_cast_fp16")];
            tensor<int32, [1]> v_71_axes_0 = const()[name = tensor<string, []>("v_71_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_71_cast_fp16 = expand_dims(axes = v_71_axes_0, x = v_69_cast_fp16)[name = tensor<string, []>("v_71_cast_fp16")];
            tensor<int32, [4]> var_1376_begin_0 = const()[name = tensor<string, []>("op_1376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1376_end_0 = const()[name = tensor<string, []>("op_1376_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1376_end_mask_0 = const()[name = tensor<string, []>("op_1376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1376_cast_fp16 = slice_by_index(begin = var_1376_begin_0, end = var_1376_end_0, end_mask = var_1376_end_mask_0, x = q_115_cast_fp16)[name = tensor<string, []>("op_1376_cast_fp16")];
            tensor<int32, [4]> var_1380_begin_0 = const()[name = tensor<string, []>("op_1380_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1380_end_0 = const()[name = tensor<string, []>("op_1380_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1380_end_mask_0 = const()[name = tensor<string, []>("op_1380_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = k_69_cast_fp16)[name = tensor<string, []>("op_1380_cast_fp16")];
            tensor<int32, [4]> var_1384_begin_0 = const()[name = tensor<string, []>("op_1384_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1384_end_0 = const()[name = tensor<string, []>("op_1384_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1384_end_mask_0 = const()[name = tensor<string, []>("op_1384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1384_cast_fp16 = slice_by_index(begin = var_1384_begin_0, end = var_1384_end_0, end_mask = var_1384_end_mask_0, x = q_115_cast_fp16)[name = tensor<string, []>("op_1384_cast_fp16")];
            tensor<int32, [4]> var_1388_begin_0 = const()[name = tensor<string, []>("op_1388_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1388_end_0 = const()[name = tensor<string, []>("op_1388_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1388_end_mask_0 = const()[name = tensor<string, []>("op_1388_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1388_cast_fp16 = slice_by_index(begin = var_1388_begin_0, end = var_1388_end_0, end_mask = var_1388_end_mask_0, x = k_69_cast_fp16)[name = tensor<string, []>("op_1388_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1390_cast_fp16 = mul(x = var_1384_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1390_cast_fp16")];
            tensor<int32, [5]> var_1398 = const()[name = tensor<string, []>("op_1398"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_143_cast_fp16 = reshape(shape = var_1398, x = var_1384_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
            tensor<int32, [2]> var_1400_split_sizes_0 = const()[name = tensor<string, []>("op_1400_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1400_axis_0 = const()[name = tensor<string, []>("op_1400_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1400_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1400_cast_fp16_1 = split(axis = var_1400_axis_0, split_sizes = var_1400_split_sizes_0, x = x_143_cast_fp16)[name = tensor<string, []>("op_1400_cast_fp16")];
            tensor<int32, [1]> squeeze_44_axes_0 = const()[name = tensor<string, []>("squeeze_44_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_44_cast_fp16 = squeeze(axes = squeeze_44_axes_0, x = var_1400_cast_fp16_0)[name = tensor<string, []>("squeeze_44_cast_fp16")];
            tensor<int32, [1]> squeeze_45_axes_0 = const()[name = tensor<string, []>("squeeze_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_45_cast_fp16 = squeeze(axes = squeeze_45_axes_0, x = var_1400_cast_fp16_1)[name = tensor<string, []>("squeeze_45_cast_fp16")];
            tensor<fp16, []> const_183_promoted_to_fp16 = const()[name = tensor<string, []>("const_183_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1403_cast_fp16 = mul(x = squeeze_45_cast_fp16, y = const_183_promoted_to_fp16)[name = tensor<string, []>("op_1403_cast_fp16")];
            tensor<int32, []> x_145_axis_0 = const()[name = tensor<string, []>("x_145_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_145_cast_fp16 = stack(axis = x_145_axis_0, values = (var_1403_cast_fp16, squeeze_44_cast_fp16))[name = tensor<string, []>("x_145_cast_fp16")];
            tensor<int32, [4]> var_1409 = const()[name = tensor<string, []>("op_1409"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1410_cast_fp16 = reshape(shape = var_1409, x = x_145_cast_fp16)[name = tensor<string, []>("op_1410_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1411_cast_fp16 = mul(x = var_1410_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1411_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_23_cast_fp16 = add(x = var_1390_cast_fp16, y = var_1411_cast_fp16)[name = tensor<string, []>("q_patches_23_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1413_cast_fp16 = mul(x = var_1388_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1413_cast_fp16")];
            tensor<int32, [5]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_149_cast_fp16 = reshape(shape = var_1421, x = var_1388_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
            tensor<int32, [2]> var_1423_split_sizes_0 = const()[name = tensor<string, []>("op_1423_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1423_axis_0 = const()[name = tensor<string, []>("op_1423_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1423_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1423_cast_fp16_1 = split(axis = var_1423_axis_0, split_sizes = var_1423_split_sizes_0, x = x_149_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
            tensor<int32, [1]> squeeze_46_axes_0 = const()[name = tensor<string, []>("squeeze_46_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_46_cast_fp16 = squeeze(axes = squeeze_46_axes_0, x = var_1423_cast_fp16_0)[name = tensor<string, []>("squeeze_46_cast_fp16")];
            tensor<int32, [1]> squeeze_47_axes_0 = const()[name = tensor<string, []>("squeeze_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_47_cast_fp16 = squeeze(axes = squeeze_47_axes_0, x = var_1423_cast_fp16_1)[name = tensor<string, []>("squeeze_47_cast_fp16")];
            tensor<fp16, []> const_191_promoted_to_fp16 = const()[name = tensor<string, []>("const_191_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1426_cast_fp16 = mul(x = squeeze_47_cast_fp16, y = const_191_promoted_to_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<int32, []> x_151_axis_0 = const()[name = tensor<string, []>("x_151_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_151_cast_fp16 = stack(axis = x_151_axis_0, values = (var_1426_cast_fp16, squeeze_46_cast_fp16))[name = tensor<string, []>("x_151_cast_fp16")];
            tensor<int32, [4]> var_1432 = const()[name = tensor<string, []>("op_1432"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1433_cast_fp16 = reshape(shape = var_1432, x = x_151_cast_fp16)[name = tensor<string, []>("op_1433_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1434_cast_fp16 = mul(x = var_1433_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1434_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_23_cast_fp16 = add(x = var_1413_cast_fp16, y = var_1434_cast_fp16)[name = tensor<string, []>("k_patches_23_cast_fp16")];
            tensor<bool, []> var_1437_interleave_0 = const()[name = tensor<string, []>("op_1437_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1437_cast_fp16 = concat(axis = var_20, interleave = var_1437_interleave_0, values = (var_1376_cast_fp16, q_patches_23_cast_fp16))[name = tensor<string, []>("op_1437_cast_fp16")];
            tensor<bool, []> var_1440_interleave_0 = const()[name = tensor<string, []>("op_1440_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1440_cast_fp16 = concat(axis = var_20, interleave = var_1440_interleave_0, values = (var_1380_cast_fp16, k_patches_23_cast_fp16))[name = tensor<string, []>("op_1440_cast_fp16")];
            tensor<fp16, []> var_1442_to_fp16 = const()[name = tensor<string, []>("op_1442_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_119_cast_fp16 = mul(x = var_1437_cast_fp16, y = var_1442_to_fp16)[name = tensor<string, []>("q_119_cast_fp16")];
            tensor<bool, []> attn_45_transpose_x_1 = const()[name = tensor<string, []>("attn_45_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_45_transpose_y_1 = const()[name = tensor<string, []>("attn_45_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_1, transpose_y = attn_45_transpose_y_1, x = q_119_cast_fp16, y = var_1440_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_47_cast_fp16 = softmax(axis = var_21, x = attn_45_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
            tensor<bool, []> var_1447_transpose_x_0 = const()[name = tensor<string, []>("op_1447_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1447_transpose_y_0 = const()[name = tensor<string, []>("op_1447_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1447_cast_fp16 = matmul(transpose_x = var_1447_transpose_x_0, transpose_y = var_1447_transpose_y_0, x = attn_47_cast_fp16, y = v_71_cast_fp16)[name = tensor<string, []>("op_1447_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_1_cast_fp16 = squeeze(axes = out_1_axes_0, x = var_1447_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<bool, []> input_5_interleave_0 = const()[name = tensor<string, []>("input_5_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_5_cast_fp16 = concat(axis = var_21, interleave = input_5_interleave_0, values = (var_304_cast_fp16, var_408_cast_fp16, var_512_cast_fp16, var_616_cast_fp16, var_720_cast_fp16, var_824_cast_fp16, var_928_cast_fp16, var_1032_cast_fp16, var_1136_cast_fp16, var_1240_cast_fp16, var_1344_cast_fp16, out_1_cast_fp16))[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_0_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5034688)))];
            tensor<fp16, [768]> model_blocks_0_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5036288)))];
            tensor<fp16, [1, 197, 768]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = model_blocks_0_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_0_attn_inner_attn_ln_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
            tensor<int32, [4]> var_1455 = const()[name = tensor<string, []>("op_1455"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_155_cast_fp16 = reshape(shape = var_1455, x = x_153_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
            tensor<int32, [4]> input_7_perm_0 = const()[name = tensor<string, []>("input_7_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_157_pad_type_0 = const()[name = tensor<string, []>("x_157_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_157_strides_0 = const()[name = tensor<string, []>("x_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_157_pad_0 = const()[name = tensor<string, []>("x_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_157_dilations_0 = const()[name = tensor<string, []>("x_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_157_groups_0 = const()[name = tensor<string, []>("x_157_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_0_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5037888)))];
            tensor<fp16, [768]> model_blocks_0_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6217600)))];
            tensor<fp16, [1, 768, 1, 197]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_478")];
            tensor<fp16, [1, 768, 1, 197]> x_157_cast_fp16 = conv(bias = model_blocks_0_attn_proj_bias_to_fp16, dilations = x_157_dilations_0, groups = x_157_groups_0, pad = x_157_pad_0, pad_type = x_157_pad_type_0, strides = x_157_strides_0, weight = model_blocks_0_attn_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("x_157_cast_fp16")];
            tensor<int32, [4]> x_159_perm_0 = const()[name = tensor<string, []>("x_159_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_1466 = const()[name = tensor<string, []>("op_1466"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_159_cast_fp16 = transpose(perm = x_159_perm_0, x = x_157_cast_fp16)[name = tensor<string, []>("transpose_477")];
            tensor<fp16, [1, 197, 768]> var_1467_cast_fp16 = reshape(shape = var_1466, x = x_159_cast_fp16)[name = tensor<string, []>("op_1467_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_9_cast_fp16 = add(x = input_1_cast_fp16, y = var_1467_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [1]> x_161_axes_0 = const()[name = tensor<string, []>("x_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6219200)))];
            tensor<fp16, [768]> model_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6220800)))];
            tensor<fp16, [1, 197, 768]> x_161_cast_fp16 = layer_norm(axes = x_161_axes_0, beta = model_blocks_0_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_0_norm2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("x_161_cast_fp16")];
            tensor<int32, [4]> var_1479 = const()[name = tensor<string, []>("op_1479"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_163_cast_fp16 = reshape(shape = var_1479, x = x_161_cast_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
            tensor<int32, [4]> input_11_perm_0 = const()[name = tensor<string, []>("input_11_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_0_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6222400)))];
            tensor<fp16, [2048]> model_blocks_0_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9368192)))];
            tensor<fp16, [1, 768, 1, 197]> input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = x_163_cast_fp16)[name = tensor<string, []>("transpose_476")];
            tensor<fp16, [1, 2048, 1, 197]> input_13_cast_fp16 = conv(bias = model_blocks_0_mlp_w1_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = model_blocks_0_mlp_w1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<string, []> x2_49_pad_type_0 = const()[name = tensor<string, []>("x2_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_49_strides_0 = const()[name = tensor<string, []>("x2_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_49_pad_0 = const()[name = tensor<string, []>("x2_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_49_dilations_0 = const()[name = tensor<string, []>("x2_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_49_groups_0 = const()[name = tensor<string, []>("x2_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_0_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9372352)))];
            tensor<fp16, [2048]> model_blocks_0_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12518144)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_49_cast_fp16 = conv(bias = model_blocks_0_mlp_w2_bias_to_fp16, dilations = x2_49_dilations_0, groups = x2_49_groups_0, pad = x2_49_pad_0, pad_type = x2_49_pad_type_0, strides = x2_49_strides_0, weight = model_blocks_0_mlp_w2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("x2_49_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_1496_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor<string, []>("op_1496_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_1_cast_fp16 = mul(x = var_1496_cast_fp16, y = x2_49_cast_fp16)[name = tensor<string, []>("hidden_1_cast_fp16")];
            tensor<int32, [4]> hidden_3_perm_0 = const()[name = tensor<string, []>("hidden_3_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_1499 = const()[name = tensor<string, []>("op_1499"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_3_cast_fp16 = transpose(perm = hidden_3_perm_0, x = hidden_1_cast_fp16)[name = tensor<string, []>("transpose_475")];
            tensor<fp16, [1, 197, 2048]> input_15_cast_fp16 = reshape(shape = var_1499, x = hidden_3_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [1]> hidden_5_axes_0 = const()[name = tensor<string, []>("hidden_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_0_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12522304)))];
            tensor<fp16, [2048]> model_blocks_0_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12526464)))];
            tensor<fp16, [1, 197, 2048]> hidden_5_cast_fp16 = layer_norm(axes = hidden_5_axes_0, beta = model_blocks_0_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_0_mlp_ffn_ln_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("hidden_5_cast_fp16")];
            tensor<int32, [4]> var_1505 = const()[name = tensor<string, []>("op_1505"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_7_cast_fp16 = reshape(shape = var_1505, x = hidden_5_cast_fp16)[name = tensor<string, []>("hidden_7_cast_fp16")];
            tensor<int32, [4]> input_17_perm_0 = const()[name = tensor<string, []>("input_17_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_165_pad_type_0 = const()[name = tensor<string, []>("x_165_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_165_strides_0 = const()[name = tensor<string, []>("x_165_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_165_dilations_0 = const()[name = tensor<string, []>("x_165_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_165_groups_0 = const()[name = tensor<string, []>("x_165_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_0_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12530624)))];
            tensor<fp16, [768]> model_blocks_0_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_0_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15676416)))];
            tensor<fp16, [1, 2048, 1, 197]> input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = hidden_7_cast_fp16)[name = tensor<string, []>("transpose_474")];
            tensor<fp16, [1, 768, 1, 197]> x_165_cast_fp16 = conv(bias = model_blocks_0_mlp_w3_bias_to_fp16, dilations = x_165_dilations_0, groups = x_165_groups_0, pad = x_165_pad_0, pad_type = x_165_pad_type_0, strides = x_165_strides_0, weight = model_blocks_0_mlp_w3_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = tensor<string, []>("x_167_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_1516 = const()[name = tensor<string, []>("op_1516"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = x_165_cast_fp16)[name = tensor<string, []>("transpose_473")];
            tensor<fp16, [1, 197, 768]> var_1517_cast_fp16 = reshape(shape = var_1516, x = x_167_cast_fp16)[name = tensor<string, []>("op_1517_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_19_cast_fp16 = add(x = input_9_cast_fp16, y = var_1517_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<int32, [1]> x_169_axes_0 = const()[name = tensor<string, []>("x_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15678016)))];
            tensor<fp16, [768]> model_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15679616)))];
            tensor<fp16, [1, 197, 768]> x_169_cast_fp16 = layer_norm(axes = x_169_axes_0, beta = model_blocks_1_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_1_norm1_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
            tensor<int32, [4]> var_1651 = const()[name = tensor<string, []>("op_1651"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_171_cast_fp16 = reshape(shape = var_1651, x = x_169_cast_fp16)[name = tensor<string, []>("x_171_cast_fp16")];
            tensor<int32, [4]> input_21_perm_0 = const()[name = tensor<string, []>("input_21_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_1659_pad_type_0 = const()[name = tensor<string, []>("op_1659_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1659_strides_0 = const()[name = tensor<string, []>("op_1659_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1659_pad_0 = const()[name = tensor<string, []>("op_1659_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1659_dilations_0 = const()[name = tensor<string, []>("op_1659_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1659_groups_0 = const()[name = tensor<string, []>("op_1659_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15681216)))];
            tensor<fp16, [1, 768, 1, 197]> input_21_cast_fp16 = transpose(perm = input_21_perm_0, x = x_171_cast_fp16)[name = tensor<string, []>("transpose_472")];
            tensor<fp16, [1, 64, 1, 197]> var_1659_cast_fp16 = conv(dilations = var_1659_dilations_0, groups = var_1659_groups_0, pad = var_1659_pad_0, pad_type = var_1659_pad_type_0, strides = var_1659_strides_0, weight = model_blocks_1_attn_q_projs_0_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1659_cast_fp16")];
            tensor<int32, [1]> var_1660_axes_0 = const()[name = tensor<string, []>("op_1660_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1660_cast_fp16 = squeeze(axes = var_1660_axes_0, x = var_1659_cast_fp16)[name = tensor<string, []>("op_1660_cast_fp16")];
            tensor<int32, [3]> q_121_perm_0 = const()[name = tensor<string, []>("q_121_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1667_pad_type_0 = const()[name = tensor<string, []>("op_1667_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1667_strides_0 = const()[name = tensor<string, []>("op_1667_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1667_pad_0 = const()[name = tensor<string, []>("op_1667_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1667_dilations_0 = const()[name = tensor<string, []>("op_1667_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1667_groups_0 = const()[name = tensor<string, []>("op_1667_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15779584)))];
            tensor<fp16, [1, 64, 1, 197]> var_1667_cast_fp16 = conv(dilations = var_1667_dilations_0, groups = var_1667_groups_0, pad = var_1667_pad_0, pad_type = var_1667_pad_type_0, strides = var_1667_strides_0, weight = model_blocks_1_attn_k_projs_0_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1667_cast_fp16")];
            tensor<int32, [1]> var_1668_axes_0 = const()[name = tensor<string, []>("op_1668_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1668_cast_fp16 = squeeze(axes = var_1668_axes_0, x = var_1667_cast_fp16)[name = tensor<string, []>("op_1668_cast_fp16")];
            tensor<int32, [3]> k_73_perm_0 = const()[name = tensor<string, []>("k_73_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1675_pad_type_0 = const()[name = tensor<string, []>("op_1675_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1675_strides_0 = const()[name = tensor<string, []>("op_1675_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1675_pad_0 = const()[name = tensor<string, []>("op_1675_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1675_dilations_0 = const()[name = tensor<string, []>("op_1675_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1675_groups_0 = const()[name = tensor<string, []>("op_1675_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15877952)))];
            tensor<fp16, [1, 64, 1, 197]> var_1675_cast_fp16 = conv(dilations = var_1675_dilations_0, groups = var_1675_groups_0, pad = var_1675_pad_0, pad_type = var_1675_pad_type_0, strides = var_1675_strides_0, weight = model_blocks_1_attn_v_projs_0_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1675_cast_fp16")];
            tensor<int32, [1]> var_1676_axes_0 = const()[name = tensor<string, []>("op_1676_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1676_cast_fp16 = squeeze(axes = var_1676_axes_0, x = var_1675_cast_fp16)[name = tensor<string, []>("op_1676_cast_fp16")];
            tensor<int32, [3]> v_73_perm_0 = const()[name = tensor<string, []>("v_73_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15976320)))];
            tensor<fp16, [1, 197, 64]> q_121_cast_fp16 = transpose(perm = q_121_perm_0, x = var_1660_cast_fp16)[name = tensor<string, []>("transpose_471")];
            tensor<fp16, [1, 197, 64]> q_123_cast_fp16 = add(x = q_121_cast_fp16, y = model_blocks_1_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_123_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15976512)))];
            tensor<fp16, [1, 197, 64]> v_73_cast_fp16 = transpose(perm = v_73_perm_0, x = var_1676_cast_fp16)[name = tensor<string, []>("transpose_469")];
            tensor<fp16, [1, 197, 64]> v_75_cast_fp16 = add(x = v_73_cast_fp16, y = model_blocks_1_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_75_cast_fp16")];
            tensor<int32, [1]> q_125_axes_0 = const()[name = tensor<string, []>("q_125_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_125_cast_fp16 = expand_dims(axes = q_125_axes_0, x = q_123_cast_fp16)[name = tensor<string, []>("q_125_cast_fp16")];
            tensor<int32, [1]> k_75_axes_0 = const()[name = tensor<string, []>("k_75_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_73_cast_fp16 = transpose(perm = k_73_perm_0, x = var_1668_cast_fp16)[name = tensor<string, []>("transpose_470")];
            tensor<fp16, [1, 1, 197, 64]> k_75_cast_fp16 = expand_dims(axes = k_75_axes_0, x = k_73_cast_fp16)[name = tensor<string, []>("k_75_cast_fp16")];
            tensor<int32, [1]> v_77_axes_0 = const()[name = tensor<string, []>("v_77_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_77_cast_fp16 = expand_dims(axes = v_77_axes_0, x = v_75_cast_fp16)[name = tensor<string, []>("v_77_cast_fp16")];
            tensor<int32, [4]> var_1685_begin_0 = const()[name = tensor<string, []>("op_1685_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1685_end_0 = const()[name = tensor<string, []>("op_1685_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1685_end_mask_0 = const()[name = tensor<string, []>("op_1685_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1685_cast_fp16 = slice_by_index(begin = var_1685_begin_0, end = var_1685_end_0, end_mask = var_1685_end_mask_0, x = q_125_cast_fp16)[name = tensor<string, []>("op_1685_cast_fp16")];
            tensor<int32, [4]> var_1689_begin_0 = const()[name = tensor<string, []>("op_1689_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1689_end_0 = const()[name = tensor<string, []>("op_1689_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1689_end_mask_0 = const()[name = tensor<string, []>("op_1689_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1689_cast_fp16 = slice_by_index(begin = var_1689_begin_0, end = var_1689_end_0, end_mask = var_1689_end_mask_0, x = k_75_cast_fp16)[name = tensor<string, []>("op_1689_cast_fp16")];
            tensor<int32, [4]> var_1693_begin_0 = const()[name = tensor<string, []>("op_1693_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1693_end_0 = const()[name = tensor<string, []>("op_1693_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1693_end_mask_0 = const()[name = tensor<string, []>("op_1693_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1693_cast_fp16 = slice_by_index(begin = var_1693_begin_0, end = var_1693_end_0, end_mask = var_1693_end_mask_0, x = q_125_cast_fp16)[name = tensor<string, []>("op_1693_cast_fp16")];
            tensor<int32, [4]> var_1697_begin_0 = const()[name = tensor<string, []>("op_1697_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1697_end_0 = const()[name = tensor<string, []>("op_1697_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1697_end_mask_0 = const()[name = tensor<string, []>("op_1697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1697_cast_fp16 = slice_by_index(begin = var_1697_begin_0, end = var_1697_end_0, end_mask = var_1697_end_mask_0, x = k_75_cast_fp16)[name = tensor<string, []>("op_1697_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1699_cast_fp16 = mul(x = var_1693_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1699_cast_fp16")];
            tensor<int32, [5]> var_1707 = const()[name = tensor<string, []>("op_1707"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_175_cast_fp16 = reshape(shape = var_1707, x = var_1693_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
            tensor<int32, [2]> var_1709_split_sizes_0 = const()[name = tensor<string, []>("op_1709_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1709_axis_0 = const()[name = tensor<string, []>("op_1709_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1709_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1709_cast_fp16_1 = split(axis = var_1709_axis_0, split_sizes = var_1709_split_sizes_0, x = x_175_cast_fp16)[name = tensor<string, []>("op_1709_cast_fp16")];
            tensor<int32, [1]> squeeze_48_axes_0 = const()[name = tensor<string, []>("squeeze_48_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_48_cast_fp16 = squeeze(axes = squeeze_48_axes_0, x = var_1709_cast_fp16_0)[name = tensor<string, []>("squeeze_48_cast_fp16")];
            tensor<int32, [1]> squeeze_49_axes_0 = const()[name = tensor<string, []>("squeeze_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_49_cast_fp16 = squeeze(axes = squeeze_49_axes_0, x = var_1709_cast_fp16_1)[name = tensor<string, []>("squeeze_49_cast_fp16")];
            tensor<fp16, []> const_203_promoted_to_fp16 = const()[name = tensor<string, []>("const_203_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1712_cast_fp16 = mul(x = squeeze_49_cast_fp16, y = const_203_promoted_to_fp16)[name = tensor<string, []>("op_1712_cast_fp16")];
            tensor<int32, []> x_177_axis_0 = const()[name = tensor<string, []>("x_177_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_177_cast_fp16 = stack(axis = x_177_axis_0, values = (var_1712_cast_fp16, squeeze_48_cast_fp16))[name = tensor<string, []>("x_177_cast_fp16")];
            tensor<int32, [4]> var_1718 = const()[name = tensor<string, []>("op_1718"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1719_cast_fp16 = reshape(shape = var_1718, x = x_177_cast_fp16)[name = tensor<string, []>("op_1719_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1720_cast_fp16 = mul(x = var_1719_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1720_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_25_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1720_cast_fp16)[name = tensor<string, []>("q_patches_25_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1722_cast_fp16 = mul(x = var_1697_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1722_cast_fp16")];
            tensor<int32, [5]> var_1730 = const()[name = tensor<string, []>("op_1730"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_181_cast_fp16 = reshape(shape = var_1730, x = var_1697_cast_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = tensor<string, []>("op_1732_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1732_axis_0 = const()[name = tensor<string, []>("op_1732_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1732_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_181_cast_fp16)[name = tensor<string, []>("op_1732_cast_fp16")];
            tensor<int32, [1]> squeeze_50_axes_0 = const()[name = tensor<string, []>("squeeze_50_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_50_cast_fp16 = squeeze(axes = squeeze_50_axes_0, x = var_1732_cast_fp16_0)[name = tensor<string, []>("squeeze_50_cast_fp16")];
            tensor<int32, [1]> squeeze_51_axes_0 = const()[name = tensor<string, []>("squeeze_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_51_cast_fp16 = squeeze(axes = squeeze_51_axes_0, x = var_1732_cast_fp16_1)[name = tensor<string, []>("squeeze_51_cast_fp16")];
            tensor<fp16, []> const_211_promoted_to_fp16 = const()[name = tensor<string, []>("const_211_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1735_cast_fp16 = mul(x = squeeze_51_cast_fp16, y = const_211_promoted_to_fp16)[name = tensor<string, []>("op_1735_cast_fp16")];
            tensor<int32, []> x_183_axis_0 = const()[name = tensor<string, []>("x_183_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_183_cast_fp16 = stack(axis = x_183_axis_0, values = (var_1735_cast_fp16, squeeze_50_cast_fp16))[name = tensor<string, []>("x_183_cast_fp16")];
            tensor<int32, [4]> var_1741 = const()[name = tensor<string, []>("op_1741"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1742_cast_fp16 = reshape(shape = var_1741, x = x_183_cast_fp16)[name = tensor<string, []>("op_1742_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1743_cast_fp16 = mul(x = var_1742_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1743_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_25_cast_fp16 = add(x = var_1722_cast_fp16, y = var_1743_cast_fp16)[name = tensor<string, []>("k_patches_25_cast_fp16")];
            tensor<bool, []> var_1746_interleave_0 = const()[name = tensor<string, []>("op_1746_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1746_cast_fp16 = concat(axis = var_20, interleave = var_1746_interleave_0, values = (var_1685_cast_fp16, q_patches_25_cast_fp16))[name = tensor<string, []>("op_1746_cast_fp16")];
            tensor<bool, []> var_1749_interleave_0 = const()[name = tensor<string, []>("op_1749_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1749_cast_fp16 = concat(axis = var_20, interleave = var_1749_interleave_0, values = (var_1689_cast_fp16, k_patches_25_cast_fp16))[name = tensor<string, []>("op_1749_cast_fp16")];
            tensor<fp16, []> var_1751_to_fp16 = const()[name = tensor<string, []>("op_1751_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_129_cast_fp16 = mul(x = var_1746_cast_fp16, y = var_1751_to_fp16)[name = tensor<string, []>("q_129_cast_fp16")];
            tensor<bool, []> attn_49_transpose_x_1 = const()[name = tensor<string, []>("attn_49_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_49_transpose_y_1 = const()[name = tensor<string, []>("attn_49_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_1, transpose_y = attn_49_transpose_y_1, x = q_129_cast_fp16, y = var_1749_cast_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_51_cast_fp16 = softmax(axis = var_21, x = attn_49_cast_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
            tensor<bool, []> var_1756_transpose_x_0 = const()[name = tensor<string, []>("op_1756_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1756_transpose_y_0 = const()[name = tensor<string, []>("op_1756_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1756_cast_fp16 = matmul(transpose_x = var_1756_transpose_x_0, transpose_y = var_1756_transpose_y_0, x = attn_51_cast_fp16, y = v_77_cast_fp16)[name = tensor<string, []>("op_1756_cast_fp16")];
            tensor<int32, [1]> var_1757_axes_0 = const()[name = tensor<string, []>("op_1757_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1757_cast_fp16 = squeeze(axes = var_1757_axes_0, x = var_1756_cast_fp16)[name = tensor<string, []>("op_1757_cast_fp16")];
            tensor<string, []> var_1763_pad_type_0 = const()[name = tensor<string, []>("op_1763_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1763_strides_0 = const()[name = tensor<string, []>("op_1763_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1763_pad_0 = const()[name = tensor<string, []>("op_1763_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1763_dilations_0 = const()[name = tensor<string, []>("op_1763_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1763_groups_0 = const()[name = tensor<string, []>("op_1763_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15976704)))];
            tensor<fp16, [1, 64, 1, 197]> var_1763_cast_fp16 = conv(dilations = var_1763_dilations_0, groups = var_1763_groups_0, pad = var_1763_pad_0, pad_type = var_1763_pad_type_0, strides = var_1763_strides_0, weight = model_blocks_1_attn_q_projs_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1763_cast_fp16")];
            tensor<int32, [1]> var_1764_axes_0 = const()[name = tensor<string, []>("op_1764_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1764_cast_fp16 = squeeze(axes = var_1764_axes_0, x = var_1763_cast_fp16)[name = tensor<string, []>("op_1764_cast_fp16")];
            tensor<int32, [3]> q_131_perm_0 = const()[name = tensor<string, []>("q_131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1771_pad_type_0 = const()[name = tensor<string, []>("op_1771_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1771_strides_0 = const()[name = tensor<string, []>("op_1771_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1771_pad_0 = const()[name = tensor<string, []>("op_1771_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1771_dilations_0 = const()[name = tensor<string, []>("op_1771_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1771_groups_0 = const()[name = tensor<string, []>("op_1771_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16075072)))];
            tensor<fp16, [1, 64, 1, 197]> var_1771_cast_fp16 = conv(dilations = var_1771_dilations_0, groups = var_1771_groups_0, pad = var_1771_pad_0, pad_type = var_1771_pad_type_0, strides = var_1771_strides_0, weight = model_blocks_1_attn_k_projs_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1771_cast_fp16")];
            tensor<int32, [1]> var_1772_axes_0 = const()[name = tensor<string, []>("op_1772_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1772_cast_fp16 = squeeze(axes = var_1772_axes_0, x = var_1771_cast_fp16)[name = tensor<string, []>("op_1772_cast_fp16")];
            tensor<int32, [3]> k_79_perm_0 = const()[name = tensor<string, []>("k_79_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1779_pad_type_0 = const()[name = tensor<string, []>("op_1779_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1779_strides_0 = const()[name = tensor<string, []>("op_1779_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1779_pad_0 = const()[name = tensor<string, []>("op_1779_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1779_dilations_0 = const()[name = tensor<string, []>("op_1779_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1779_groups_0 = const()[name = tensor<string, []>("op_1779_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16173440)))];
            tensor<fp16, [1, 64, 1, 197]> var_1779_cast_fp16 = conv(dilations = var_1779_dilations_0, groups = var_1779_groups_0, pad = var_1779_pad_0, pad_type = var_1779_pad_type_0, strides = var_1779_strides_0, weight = model_blocks_1_attn_v_projs_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1779_cast_fp16")];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = tensor<string, []>("op_1780_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1780_cast_fp16 = squeeze(axes = var_1780_axes_0, x = var_1779_cast_fp16)[name = tensor<string, []>("op_1780_cast_fp16")];
            tensor<int32, [3]> v_79_perm_0 = const()[name = tensor<string, []>("v_79_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16271808)))];
            tensor<fp16, [1, 197, 64]> q_131_cast_fp16 = transpose(perm = q_131_perm_0, x = var_1764_cast_fp16)[name = tensor<string, []>("transpose_468")];
            tensor<fp16, [1, 197, 64]> q_133_cast_fp16 = add(x = q_131_cast_fp16, y = model_blocks_1_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_133_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16272000)))];
            tensor<fp16, [1, 197, 64]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = var_1780_cast_fp16)[name = tensor<string, []>("transpose_466")];
            tensor<fp16, [1, 197, 64]> v_81_cast_fp16 = add(x = v_79_cast_fp16, y = model_blocks_1_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_81_cast_fp16")];
            tensor<int32, [1]> q_135_axes_0 = const()[name = tensor<string, []>("q_135_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_135_cast_fp16 = expand_dims(axes = q_135_axes_0, x = q_133_cast_fp16)[name = tensor<string, []>("q_135_cast_fp16")];
            tensor<int32, [1]> k_81_axes_0 = const()[name = tensor<string, []>("k_81_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_79_cast_fp16 = transpose(perm = k_79_perm_0, x = var_1772_cast_fp16)[name = tensor<string, []>("transpose_467")];
            tensor<fp16, [1, 1, 197, 64]> k_81_cast_fp16 = expand_dims(axes = k_81_axes_0, x = k_79_cast_fp16)[name = tensor<string, []>("k_81_cast_fp16")];
            tensor<int32, [1]> v_83_axes_0 = const()[name = tensor<string, []>("v_83_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_83_cast_fp16 = expand_dims(axes = v_83_axes_0, x = v_81_cast_fp16)[name = tensor<string, []>("v_83_cast_fp16")];
            tensor<int32, [4]> var_1789_begin_0 = const()[name = tensor<string, []>("op_1789_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1789_end_0 = const()[name = tensor<string, []>("op_1789_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1789_end_mask_0 = const()[name = tensor<string, []>("op_1789_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1789_cast_fp16 = slice_by_index(begin = var_1789_begin_0, end = var_1789_end_0, end_mask = var_1789_end_mask_0, x = q_135_cast_fp16)[name = tensor<string, []>("op_1789_cast_fp16")];
            tensor<int32, [4]> var_1793_begin_0 = const()[name = tensor<string, []>("op_1793_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1793_end_0 = const()[name = tensor<string, []>("op_1793_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1793_end_mask_0 = const()[name = tensor<string, []>("op_1793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1793_cast_fp16 = slice_by_index(begin = var_1793_begin_0, end = var_1793_end_0, end_mask = var_1793_end_mask_0, x = k_81_cast_fp16)[name = tensor<string, []>("op_1793_cast_fp16")];
            tensor<int32, [4]> var_1797_begin_0 = const()[name = tensor<string, []>("op_1797_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1797_end_0 = const()[name = tensor<string, []>("op_1797_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1797_end_mask_0 = const()[name = tensor<string, []>("op_1797_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1797_cast_fp16 = slice_by_index(begin = var_1797_begin_0, end = var_1797_end_0, end_mask = var_1797_end_mask_0, x = q_135_cast_fp16)[name = tensor<string, []>("op_1797_cast_fp16")];
            tensor<int32, [4]> var_1801_begin_0 = const()[name = tensor<string, []>("op_1801_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1801_end_0 = const()[name = tensor<string, []>("op_1801_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1801_end_mask_0 = const()[name = tensor<string, []>("op_1801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1801_cast_fp16 = slice_by_index(begin = var_1801_begin_0, end = var_1801_end_0, end_mask = var_1801_end_mask_0, x = k_81_cast_fp16)[name = tensor<string, []>("op_1801_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1803_cast_fp16 = mul(x = var_1797_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1803_cast_fp16")];
            tensor<int32, [5]> var_1811 = const()[name = tensor<string, []>("op_1811"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_187_cast_fp16 = reshape(shape = var_1811, x = var_1797_cast_fp16)[name = tensor<string, []>("x_187_cast_fp16")];
            tensor<int32, [2]> var_1813_split_sizes_0 = const()[name = tensor<string, []>("op_1813_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1813_axis_0 = const()[name = tensor<string, []>("op_1813_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1813_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1813_cast_fp16_1 = split(axis = var_1813_axis_0, split_sizes = var_1813_split_sizes_0, x = x_187_cast_fp16)[name = tensor<string, []>("op_1813_cast_fp16")];
            tensor<int32, [1]> squeeze_52_axes_0 = const()[name = tensor<string, []>("squeeze_52_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_52_cast_fp16 = squeeze(axes = squeeze_52_axes_0, x = var_1813_cast_fp16_0)[name = tensor<string, []>("squeeze_52_cast_fp16")];
            tensor<int32, [1]> squeeze_53_axes_0 = const()[name = tensor<string, []>("squeeze_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_53_cast_fp16 = squeeze(axes = squeeze_53_axes_0, x = var_1813_cast_fp16_1)[name = tensor<string, []>("squeeze_53_cast_fp16")];
            tensor<fp16, []> const_219_promoted_to_fp16 = const()[name = tensor<string, []>("const_219_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1816_cast_fp16 = mul(x = squeeze_53_cast_fp16, y = const_219_promoted_to_fp16)[name = tensor<string, []>("op_1816_cast_fp16")];
            tensor<int32, []> x_189_axis_0 = const()[name = tensor<string, []>("x_189_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_189_cast_fp16 = stack(axis = x_189_axis_0, values = (var_1816_cast_fp16, squeeze_52_cast_fp16))[name = tensor<string, []>("x_189_cast_fp16")];
            tensor<int32, [4]> var_1822 = const()[name = tensor<string, []>("op_1822"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1823_cast_fp16 = reshape(shape = var_1822, x = x_189_cast_fp16)[name = tensor<string, []>("op_1823_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1824_cast_fp16 = mul(x = var_1823_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1824_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_27_cast_fp16 = add(x = var_1803_cast_fp16, y = var_1824_cast_fp16)[name = tensor<string, []>("q_patches_27_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1826_cast_fp16 = mul(x = var_1801_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1826_cast_fp16")];
            tensor<int32, [5]> var_1834 = const()[name = tensor<string, []>("op_1834"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_193_cast_fp16 = reshape(shape = var_1834, x = var_1801_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
            tensor<int32, [2]> var_1836_split_sizes_0 = const()[name = tensor<string, []>("op_1836_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1836_axis_0 = const()[name = tensor<string, []>("op_1836_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1836_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1836_cast_fp16_1 = split(axis = var_1836_axis_0, split_sizes = var_1836_split_sizes_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1836_cast_fp16")];
            tensor<int32, [1]> squeeze_54_axes_0 = const()[name = tensor<string, []>("squeeze_54_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_54_cast_fp16 = squeeze(axes = squeeze_54_axes_0, x = var_1836_cast_fp16_0)[name = tensor<string, []>("squeeze_54_cast_fp16")];
            tensor<int32, [1]> squeeze_55_axes_0 = const()[name = tensor<string, []>("squeeze_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_55_cast_fp16 = squeeze(axes = squeeze_55_axes_0, x = var_1836_cast_fp16_1)[name = tensor<string, []>("squeeze_55_cast_fp16")];
            tensor<fp16, []> const_227_promoted_to_fp16 = const()[name = tensor<string, []>("const_227_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1839_cast_fp16 = mul(x = squeeze_55_cast_fp16, y = const_227_promoted_to_fp16)[name = tensor<string, []>("op_1839_cast_fp16")];
            tensor<int32, []> x_195_axis_0 = const()[name = tensor<string, []>("x_195_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_195_cast_fp16 = stack(axis = x_195_axis_0, values = (var_1839_cast_fp16, squeeze_54_cast_fp16))[name = tensor<string, []>("x_195_cast_fp16")];
            tensor<int32, [4]> var_1845 = const()[name = tensor<string, []>("op_1845"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1846_cast_fp16 = reshape(shape = var_1845, x = x_195_cast_fp16)[name = tensor<string, []>("op_1846_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1847_cast_fp16 = mul(x = var_1846_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1847_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_27_cast_fp16 = add(x = var_1826_cast_fp16, y = var_1847_cast_fp16)[name = tensor<string, []>("k_patches_27_cast_fp16")];
            tensor<bool, []> var_1850_interleave_0 = const()[name = tensor<string, []>("op_1850_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1850_cast_fp16 = concat(axis = var_20, interleave = var_1850_interleave_0, values = (var_1789_cast_fp16, q_patches_27_cast_fp16))[name = tensor<string, []>("op_1850_cast_fp16")];
            tensor<bool, []> var_1853_interleave_0 = const()[name = tensor<string, []>("op_1853_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1853_cast_fp16 = concat(axis = var_20, interleave = var_1853_interleave_0, values = (var_1793_cast_fp16, k_patches_27_cast_fp16))[name = tensor<string, []>("op_1853_cast_fp16")];
            tensor<fp16, []> var_1855_to_fp16 = const()[name = tensor<string, []>("op_1855_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_139_cast_fp16 = mul(x = var_1850_cast_fp16, y = var_1855_to_fp16)[name = tensor<string, []>("q_139_cast_fp16")];
            tensor<bool, []> attn_53_transpose_x_1 = const()[name = tensor<string, []>("attn_53_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_53_transpose_y_1 = const()[name = tensor<string, []>("attn_53_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_1, transpose_y = attn_53_transpose_y_1, x = q_139_cast_fp16, y = var_1853_cast_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_55_cast_fp16 = softmax(axis = var_21, x = attn_53_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
            tensor<bool, []> var_1860_transpose_x_0 = const()[name = tensor<string, []>("op_1860_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1860_transpose_y_0 = const()[name = tensor<string, []>("op_1860_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1860_cast_fp16 = matmul(transpose_x = var_1860_transpose_x_0, transpose_y = var_1860_transpose_y_0, x = attn_55_cast_fp16, y = v_83_cast_fp16)[name = tensor<string, []>("op_1860_cast_fp16")];
            tensor<int32, [1]> var_1861_axes_0 = const()[name = tensor<string, []>("op_1861_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1861_cast_fp16 = squeeze(axes = var_1861_axes_0, x = var_1860_cast_fp16)[name = tensor<string, []>("op_1861_cast_fp16")];
            tensor<string, []> var_1867_pad_type_0 = const()[name = tensor<string, []>("op_1867_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1867_strides_0 = const()[name = tensor<string, []>("op_1867_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1867_pad_0 = const()[name = tensor<string, []>("op_1867_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1867_dilations_0 = const()[name = tensor<string, []>("op_1867_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1867_groups_0 = const()[name = tensor<string, []>("op_1867_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16272192)))];
            tensor<fp16, [1, 64, 1, 197]> var_1867_cast_fp16 = conv(dilations = var_1867_dilations_0, groups = var_1867_groups_0, pad = var_1867_pad_0, pad_type = var_1867_pad_type_0, strides = var_1867_strides_0, weight = model_blocks_1_attn_q_projs_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1867_cast_fp16")];
            tensor<int32, [1]> var_1868_axes_0 = const()[name = tensor<string, []>("op_1868_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1868_cast_fp16 = squeeze(axes = var_1868_axes_0, x = var_1867_cast_fp16)[name = tensor<string, []>("op_1868_cast_fp16")];
            tensor<int32, [3]> q_141_perm_0 = const()[name = tensor<string, []>("q_141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1875_pad_type_0 = const()[name = tensor<string, []>("op_1875_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1875_strides_0 = const()[name = tensor<string, []>("op_1875_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1875_pad_0 = const()[name = tensor<string, []>("op_1875_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1875_dilations_0 = const()[name = tensor<string, []>("op_1875_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1875_groups_0 = const()[name = tensor<string, []>("op_1875_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16370560)))];
            tensor<fp16, [1, 64, 1, 197]> var_1875_cast_fp16 = conv(dilations = var_1875_dilations_0, groups = var_1875_groups_0, pad = var_1875_pad_0, pad_type = var_1875_pad_type_0, strides = var_1875_strides_0, weight = model_blocks_1_attn_k_projs_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1875_cast_fp16")];
            tensor<int32, [1]> var_1876_axes_0 = const()[name = tensor<string, []>("op_1876_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1876_cast_fp16 = squeeze(axes = var_1876_axes_0, x = var_1875_cast_fp16)[name = tensor<string, []>("op_1876_cast_fp16")];
            tensor<int32, [3]> k_85_perm_0 = const()[name = tensor<string, []>("k_85_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1883_pad_type_0 = const()[name = tensor<string, []>("op_1883_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1883_strides_0 = const()[name = tensor<string, []>("op_1883_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1883_pad_0 = const()[name = tensor<string, []>("op_1883_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1883_dilations_0 = const()[name = tensor<string, []>("op_1883_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1883_groups_0 = const()[name = tensor<string, []>("op_1883_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16468928)))];
            tensor<fp16, [1, 64, 1, 197]> var_1883_cast_fp16 = conv(dilations = var_1883_dilations_0, groups = var_1883_groups_0, pad = var_1883_pad_0, pad_type = var_1883_pad_type_0, strides = var_1883_strides_0, weight = model_blocks_1_attn_v_projs_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1883_cast_fp16")];
            tensor<int32, [1]> var_1884_axes_0 = const()[name = tensor<string, []>("op_1884_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1884_cast_fp16 = squeeze(axes = var_1884_axes_0, x = var_1883_cast_fp16)[name = tensor<string, []>("op_1884_cast_fp16")];
            tensor<int32, [3]> v_85_perm_0 = const()[name = tensor<string, []>("v_85_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16567296)))];
            tensor<fp16, [1, 197, 64]> q_141_cast_fp16 = transpose(perm = q_141_perm_0, x = var_1868_cast_fp16)[name = tensor<string, []>("transpose_465")];
            tensor<fp16, [1, 197, 64]> q_143_cast_fp16 = add(x = q_141_cast_fp16, y = model_blocks_1_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_143_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16567488)))];
            tensor<fp16, [1, 197, 64]> v_85_cast_fp16 = transpose(perm = v_85_perm_0, x = var_1884_cast_fp16)[name = tensor<string, []>("transpose_463")];
            tensor<fp16, [1, 197, 64]> v_87_cast_fp16 = add(x = v_85_cast_fp16, y = model_blocks_1_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_87_cast_fp16")];
            tensor<int32, [1]> q_145_axes_0 = const()[name = tensor<string, []>("q_145_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_145_cast_fp16 = expand_dims(axes = q_145_axes_0, x = q_143_cast_fp16)[name = tensor<string, []>("q_145_cast_fp16")];
            tensor<int32, [1]> k_87_axes_0 = const()[name = tensor<string, []>("k_87_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_85_cast_fp16 = transpose(perm = k_85_perm_0, x = var_1876_cast_fp16)[name = tensor<string, []>("transpose_464")];
            tensor<fp16, [1, 1, 197, 64]> k_87_cast_fp16 = expand_dims(axes = k_87_axes_0, x = k_85_cast_fp16)[name = tensor<string, []>("k_87_cast_fp16")];
            tensor<int32, [1]> v_89_axes_0 = const()[name = tensor<string, []>("v_89_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_89_cast_fp16 = expand_dims(axes = v_89_axes_0, x = v_87_cast_fp16)[name = tensor<string, []>("v_89_cast_fp16")];
            tensor<int32, [4]> var_1893_begin_0 = const()[name = tensor<string, []>("op_1893_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1893_end_0 = const()[name = tensor<string, []>("op_1893_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1893_end_mask_0 = const()[name = tensor<string, []>("op_1893_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1893_cast_fp16 = slice_by_index(begin = var_1893_begin_0, end = var_1893_end_0, end_mask = var_1893_end_mask_0, x = q_145_cast_fp16)[name = tensor<string, []>("op_1893_cast_fp16")];
            tensor<int32, [4]> var_1897_begin_0 = const()[name = tensor<string, []>("op_1897_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1897_end_0 = const()[name = tensor<string, []>("op_1897_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1897_end_mask_0 = const()[name = tensor<string, []>("op_1897_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1897_cast_fp16 = slice_by_index(begin = var_1897_begin_0, end = var_1897_end_0, end_mask = var_1897_end_mask_0, x = k_87_cast_fp16)[name = tensor<string, []>("op_1897_cast_fp16")];
            tensor<int32, [4]> var_1901_begin_0 = const()[name = tensor<string, []>("op_1901_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1901_end_0 = const()[name = tensor<string, []>("op_1901_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1901_end_mask_0 = const()[name = tensor<string, []>("op_1901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1901_cast_fp16 = slice_by_index(begin = var_1901_begin_0, end = var_1901_end_0, end_mask = var_1901_end_mask_0, x = q_145_cast_fp16)[name = tensor<string, []>("op_1901_cast_fp16")];
            tensor<int32, [4]> var_1905_begin_0 = const()[name = tensor<string, []>("op_1905_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1905_end_0 = const()[name = tensor<string, []>("op_1905_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_1905_end_mask_0 = const()[name = tensor<string, []>("op_1905_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_1905_cast_fp16 = slice_by_index(begin = var_1905_begin_0, end = var_1905_end_0, end_mask = var_1905_end_mask_0, x = k_87_cast_fp16)[name = tensor<string, []>("op_1905_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1907_cast_fp16 = mul(x = var_1901_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1907_cast_fp16")];
            tensor<int32, [5]> var_1915 = const()[name = tensor<string, []>("op_1915"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_199_cast_fp16 = reshape(shape = var_1915, x = var_1901_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
            tensor<int32, [2]> var_1917_split_sizes_0 = const()[name = tensor<string, []>("op_1917_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1917_axis_0 = const()[name = tensor<string, []>("op_1917_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1917_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1917_cast_fp16_1 = split(axis = var_1917_axis_0, split_sizes = var_1917_split_sizes_0, x = x_199_cast_fp16)[name = tensor<string, []>("op_1917_cast_fp16")];
            tensor<int32, [1]> squeeze_56_axes_0 = const()[name = tensor<string, []>("squeeze_56_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_56_cast_fp16 = squeeze(axes = squeeze_56_axes_0, x = var_1917_cast_fp16_0)[name = tensor<string, []>("squeeze_56_cast_fp16")];
            tensor<int32, [1]> squeeze_57_axes_0 = const()[name = tensor<string, []>("squeeze_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_57_cast_fp16 = squeeze(axes = squeeze_57_axes_0, x = var_1917_cast_fp16_1)[name = tensor<string, []>("squeeze_57_cast_fp16")];
            tensor<fp16, []> const_235_promoted_to_fp16 = const()[name = tensor<string, []>("const_235_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1920_cast_fp16 = mul(x = squeeze_57_cast_fp16, y = const_235_promoted_to_fp16)[name = tensor<string, []>("op_1920_cast_fp16")];
            tensor<int32, []> x_201_axis_0 = const()[name = tensor<string, []>("x_201_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_201_cast_fp16 = stack(axis = x_201_axis_0, values = (var_1920_cast_fp16, squeeze_56_cast_fp16))[name = tensor<string, []>("x_201_cast_fp16")];
            tensor<int32, [4]> var_1926 = const()[name = tensor<string, []>("op_1926"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1927_cast_fp16 = reshape(shape = var_1926, x = x_201_cast_fp16)[name = tensor<string, []>("op_1927_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1928_cast_fp16 = mul(x = var_1927_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1928_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_29_cast_fp16 = add(x = var_1907_cast_fp16, y = var_1928_cast_fp16)[name = tensor<string, []>("q_patches_29_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1930_cast_fp16 = mul(x = var_1905_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_1930_cast_fp16")];
            tensor<int32, [5]> var_1938 = const()[name = tensor<string, []>("op_1938"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_205_cast_fp16 = reshape(shape = var_1938, x = var_1905_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
            tensor<int32, [2]> var_1940_split_sizes_0 = const()[name = tensor<string, []>("op_1940_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1940_axis_0 = const()[name = tensor<string, []>("op_1940_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_1940_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_1940_cast_fp16_1 = split(axis = var_1940_axis_0, split_sizes = var_1940_split_sizes_0, x = x_205_cast_fp16)[name = tensor<string, []>("op_1940_cast_fp16")];
            tensor<int32, [1]> squeeze_58_axes_0 = const()[name = tensor<string, []>("squeeze_58_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_58_cast_fp16 = squeeze(axes = squeeze_58_axes_0, x = var_1940_cast_fp16_0)[name = tensor<string, []>("squeeze_58_cast_fp16")];
            tensor<int32, [1]> squeeze_59_axes_0 = const()[name = tensor<string, []>("squeeze_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_59_cast_fp16 = squeeze(axes = squeeze_59_axes_0, x = var_1940_cast_fp16_1)[name = tensor<string, []>("squeeze_59_cast_fp16")];
            tensor<fp16, []> const_243_promoted_to_fp16 = const()[name = tensor<string, []>("const_243_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_1943_cast_fp16 = mul(x = squeeze_59_cast_fp16, y = const_243_promoted_to_fp16)[name = tensor<string, []>("op_1943_cast_fp16")];
            tensor<int32, []> x_207_axis_0 = const()[name = tensor<string, []>("x_207_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_207_cast_fp16 = stack(axis = x_207_axis_0, values = (var_1943_cast_fp16, squeeze_58_cast_fp16))[name = tensor<string, []>("x_207_cast_fp16")];
            tensor<int32, [4]> var_1949 = const()[name = tensor<string, []>("op_1949"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_1950_cast_fp16 = reshape(shape = var_1949, x = x_207_cast_fp16)[name = tensor<string, []>("op_1950_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_1951_cast_fp16 = mul(x = var_1950_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_1951_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_29_cast_fp16 = add(x = var_1930_cast_fp16, y = var_1951_cast_fp16)[name = tensor<string, []>("k_patches_29_cast_fp16")];
            tensor<bool, []> var_1954_interleave_0 = const()[name = tensor<string, []>("op_1954_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1954_cast_fp16 = concat(axis = var_20, interleave = var_1954_interleave_0, values = (var_1893_cast_fp16, q_patches_29_cast_fp16))[name = tensor<string, []>("op_1954_cast_fp16")];
            tensor<bool, []> var_1957_interleave_0 = const()[name = tensor<string, []>("op_1957_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1957_cast_fp16 = concat(axis = var_20, interleave = var_1957_interleave_0, values = (var_1897_cast_fp16, k_patches_29_cast_fp16))[name = tensor<string, []>("op_1957_cast_fp16")];
            tensor<fp16, []> var_1959_to_fp16 = const()[name = tensor<string, []>("op_1959_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_149_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1959_to_fp16)[name = tensor<string, []>("q_149_cast_fp16")];
            tensor<bool, []> attn_57_transpose_x_1 = const()[name = tensor<string, []>("attn_57_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_57_transpose_y_1 = const()[name = tensor<string, []>("attn_57_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_1, transpose_y = attn_57_transpose_y_1, x = q_149_cast_fp16, y = var_1957_cast_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_59_cast_fp16 = softmax(axis = var_21, x = attn_57_cast_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
            tensor<bool, []> var_1964_transpose_x_0 = const()[name = tensor<string, []>("op_1964_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1964_transpose_y_0 = const()[name = tensor<string, []>("op_1964_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_1964_cast_fp16 = matmul(transpose_x = var_1964_transpose_x_0, transpose_y = var_1964_transpose_y_0, x = attn_59_cast_fp16, y = v_89_cast_fp16)[name = tensor<string, []>("op_1964_cast_fp16")];
            tensor<int32, [1]> var_1965_axes_0 = const()[name = tensor<string, []>("op_1965_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_1965_cast_fp16 = squeeze(axes = var_1965_axes_0, x = var_1964_cast_fp16)[name = tensor<string, []>("op_1965_cast_fp16")];
            tensor<string, []> var_1971_pad_type_0 = const()[name = tensor<string, []>("op_1971_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1971_strides_0 = const()[name = tensor<string, []>("op_1971_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1971_pad_0 = const()[name = tensor<string, []>("op_1971_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1971_dilations_0 = const()[name = tensor<string, []>("op_1971_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1971_groups_0 = const()[name = tensor<string, []>("op_1971_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16567680)))];
            tensor<fp16, [1, 64, 1, 197]> var_1971_cast_fp16 = conv(dilations = var_1971_dilations_0, groups = var_1971_groups_0, pad = var_1971_pad_0, pad_type = var_1971_pad_type_0, strides = var_1971_strides_0, weight = model_blocks_1_attn_q_projs_3_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1971_cast_fp16")];
            tensor<int32, [1]> var_1972_axes_0 = const()[name = tensor<string, []>("op_1972_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1972_cast_fp16 = squeeze(axes = var_1972_axes_0, x = var_1971_cast_fp16)[name = tensor<string, []>("op_1972_cast_fp16")];
            tensor<int32, [3]> q_151_perm_0 = const()[name = tensor<string, []>("q_151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1979_pad_type_0 = const()[name = tensor<string, []>("op_1979_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1979_strides_0 = const()[name = tensor<string, []>("op_1979_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1979_pad_0 = const()[name = tensor<string, []>("op_1979_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1979_dilations_0 = const()[name = tensor<string, []>("op_1979_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1979_groups_0 = const()[name = tensor<string, []>("op_1979_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16666048)))];
            tensor<fp16, [1, 64, 1, 197]> var_1979_cast_fp16 = conv(dilations = var_1979_dilations_0, groups = var_1979_groups_0, pad = var_1979_pad_0, pad_type = var_1979_pad_type_0, strides = var_1979_strides_0, weight = model_blocks_1_attn_k_projs_3_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1979_cast_fp16")];
            tensor<int32, [1]> var_1980_axes_0 = const()[name = tensor<string, []>("op_1980_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1980_cast_fp16 = squeeze(axes = var_1980_axes_0, x = var_1979_cast_fp16)[name = tensor<string, []>("op_1980_cast_fp16")];
            tensor<int32, [3]> k_91_perm_0 = const()[name = tensor<string, []>("k_91_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_1987_pad_type_0 = const()[name = tensor<string, []>("op_1987_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1987_strides_0 = const()[name = tensor<string, []>("op_1987_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1987_pad_0 = const()[name = tensor<string, []>("op_1987_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1987_dilations_0 = const()[name = tensor<string, []>("op_1987_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1987_groups_0 = const()[name = tensor<string, []>("op_1987_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16764416)))];
            tensor<fp16, [1, 64, 1, 197]> var_1987_cast_fp16 = conv(dilations = var_1987_dilations_0, groups = var_1987_groups_0, pad = var_1987_pad_0, pad_type = var_1987_pad_type_0, strides = var_1987_strides_0, weight = model_blocks_1_attn_v_projs_3_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_1987_cast_fp16")];
            tensor<int32, [1]> var_1988_axes_0 = const()[name = tensor<string, []>("op_1988_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_1988_cast_fp16 = squeeze(axes = var_1988_axes_0, x = var_1987_cast_fp16)[name = tensor<string, []>("op_1988_cast_fp16")];
            tensor<int32, [3]> v_91_perm_0 = const()[name = tensor<string, []>("v_91_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16862784)))];
            tensor<fp16, [1, 197, 64]> q_151_cast_fp16 = transpose(perm = q_151_perm_0, x = var_1972_cast_fp16)[name = tensor<string, []>("transpose_462")];
            tensor<fp16, [1, 197, 64]> q_153_cast_fp16 = add(x = q_151_cast_fp16, y = model_blocks_1_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_153_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16862976)))];
            tensor<fp16, [1, 197, 64]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = var_1988_cast_fp16)[name = tensor<string, []>("transpose_460")];
            tensor<fp16, [1, 197, 64]> v_93_cast_fp16 = add(x = v_91_cast_fp16, y = model_blocks_1_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_93_cast_fp16")];
            tensor<int32, [1]> q_155_axes_0 = const()[name = tensor<string, []>("q_155_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_155_cast_fp16 = expand_dims(axes = q_155_axes_0, x = q_153_cast_fp16)[name = tensor<string, []>("q_155_cast_fp16")];
            tensor<int32, [1]> k_93_axes_0 = const()[name = tensor<string, []>("k_93_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_91_cast_fp16 = transpose(perm = k_91_perm_0, x = var_1980_cast_fp16)[name = tensor<string, []>("transpose_461")];
            tensor<fp16, [1, 1, 197, 64]> k_93_cast_fp16 = expand_dims(axes = k_93_axes_0, x = k_91_cast_fp16)[name = tensor<string, []>("k_93_cast_fp16")];
            tensor<int32, [1]> v_95_axes_0 = const()[name = tensor<string, []>("v_95_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_95_cast_fp16 = expand_dims(axes = v_95_axes_0, x = v_93_cast_fp16)[name = tensor<string, []>("v_95_cast_fp16")];
            tensor<int32, [4]> var_1997_begin_0 = const()[name = tensor<string, []>("op_1997_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1997_end_0 = const()[name = tensor<string, []>("op_1997_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_1997_end_mask_0 = const()[name = tensor<string, []>("op_1997_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_1997_cast_fp16 = slice_by_index(begin = var_1997_begin_0, end = var_1997_end_0, end_mask = var_1997_end_mask_0, x = q_155_cast_fp16)[name = tensor<string, []>("op_1997_cast_fp16")];
            tensor<int32, [4]> var_2001_begin_0 = const()[name = tensor<string, []>("op_2001_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2001_end_0 = const()[name = tensor<string, []>("op_2001_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2001_end_mask_0 = const()[name = tensor<string, []>("op_2001_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2001_cast_fp16 = slice_by_index(begin = var_2001_begin_0, end = var_2001_end_0, end_mask = var_2001_end_mask_0, x = k_93_cast_fp16)[name = tensor<string, []>("op_2001_cast_fp16")];
            tensor<int32, [4]> var_2005_begin_0 = const()[name = tensor<string, []>("op_2005_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2005_end_0 = const()[name = tensor<string, []>("op_2005_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2005_end_mask_0 = const()[name = tensor<string, []>("op_2005_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2005_cast_fp16 = slice_by_index(begin = var_2005_begin_0, end = var_2005_end_0, end_mask = var_2005_end_mask_0, x = q_155_cast_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
            tensor<int32, [4]> var_2009_begin_0 = const()[name = tensor<string, []>("op_2009_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2009_end_0 = const()[name = tensor<string, []>("op_2009_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2009_end_mask_0 = const()[name = tensor<string, []>("op_2009_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2009_cast_fp16 = slice_by_index(begin = var_2009_begin_0, end = var_2009_end_0, end_mask = var_2009_end_mask_0, x = k_93_cast_fp16)[name = tensor<string, []>("op_2009_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2011_cast_fp16 = mul(x = var_2005_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2011_cast_fp16")];
            tensor<int32, [5]> var_2019 = const()[name = tensor<string, []>("op_2019"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_211_cast_fp16 = reshape(shape = var_2019, x = var_2005_cast_fp16)[name = tensor<string, []>("x_211_cast_fp16")];
            tensor<int32, [2]> var_2021_split_sizes_0 = const()[name = tensor<string, []>("op_2021_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2021_axis_0 = const()[name = tensor<string, []>("op_2021_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2021_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2021_cast_fp16_1 = split(axis = var_2021_axis_0, split_sizes = var_2021_split_sizes_0, x = x_211_cast_fp16)[name = tensor<string, []>("op_2021_cast_fp16")];
            tensor<int32, [1]> squeeze_60_axes_0 = const()[name = tensor<string, []>("squeeze_60_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_60_cast_fp16 = squeeze(axes = squeeze_60_axes_0, x = var_2021_cast_fp16_0)[name = tensor<string, []>("squeeze_60_cast_fp16")];
            tensor<int32, [1]> squeeze_61_axes_0 = const()[name = tensor<string, []>("squeeze_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_61_cast_fp16 = squeeze(axes = squeeze_61_axes_0, x = var_2021_cast_fp16_1)[name = tensor<string, []>("squeeze_61_cast_fp16")];
            tensor<fp16, []> const_251_promoted_to_fp16 = const()[name = tensor<string, []>("const_251_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2024_cast_fp16 = mul(x = squeeze_61_cast_fp16, y = const_251_promoted_to_fp16)[name = tensor<string, []>("op_2024_cast_fp16")];
            tensor<int32, []> x_213_axis_0 = const()[name = tensor<string, []>("x_213_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_213_cast_fp16 = stack(axis = x_213_axis_0, values = (var_2024_cast_fp16, squeeze_60_cast_fp16))[name = tensor<string, []>("x_213_cast_fp16")];
            tensor<int32, [4]> var_2030 = const()[name = tensor<string, []>("op_2030"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2031_cast_fp16 = reshape(shape = var_2030, x = x_213_cast_fp16)[name = tensor<string, []>("op_2031_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2032_cast_fp16 = mul(x = var_2031_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2032_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_31_cast_fp16 = add(x = var_2011_cast_fp16, y = var_2032_cast_fp16)[name = tensor<string, []>("q_patches_31_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2034_cast_fp16 = mul(x = var_2009_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2034_cast_fp16")];
            tensor<int32, [5]> var_2042 = const()[name = tensor<string, []>("op_2042"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_217_cast_fp16 = reshape(shape = var_2042, x = var_2009_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
            tensor<int32, [2]> var_2044_split_sizes_0 = const()[name = tensor<string, []>("op_2044_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2044_axis_0 = const()[name = tensor<string, []>("op_2044_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2044_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2044_cast_fp16_1 = split(axis = var_2044_axis_0, split_sizes = var_2044_split_sizes_0, x = x_217_cast_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
            tensor<int32, [1]> squeeze_62_axes_0 = const()[name = tensor<string, []>("squeeze_62_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_62_cast_fp16 = squeeze(axes = squeeze_62_axes_0, x = var_2044_cast_fp16_0)[name = tensor<string, []>("squeeze_62_cast_fp16")];
            tensor<int32, [1]> squeeze_63_axes_0 = const()[name = tensor<string, []>("squeeze_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_63_cast_fp16 = squeeze(axes = squeeze_63_axes_0, x = var_2044_cast_fp16_1)[name = tensor<string, []>("squeeze_63_cast_fp16")];
            tensor<fp16, []> const_259_promoted_to_fp16 = const()[name = tensor<string, []>("const_259_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2047_cast_fp16 = mul(x = squeeze_63_cast_fp16, y = const_259_promoted_to_fp16)[name = tensor<string, []>("op_2047_cast_fp16")];
            tensor<int32, []> x_219_axis_0 = const()[name = tensor<string, []>("x_219_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_219_cast_fp16 = stack(axis = x_219_axis_0, values = (var_2047_cast_fp16, squeeze_62_cast_fp16))[name = tensor<string, []>("x_219_cast_fp16")];
            tensor<int32, [4]> var_2053 = const()[name = tensor<string, []>("op_2053"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2054_cast_fp16 = reshape(shape = var_2053, x = x_219_cast_fp16)[name = tensor<string, []>("op_2054_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2055_cast_fp16 = mul(x = var_2054_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2055_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2055_cast_fp16)[name = tensor<string, []>("k_patches_31_cast_fp16")];
            tensor<bool, []> var_2058_interleave_0 = const()[name = tensor<string, []>("op_2058_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2058_cast_fp16 = concat(axis = var_20, interleave = var_2058_interleave_0, values = (var_1997_cast_fp16, q_patches_31_cast_fp16))[name = tensor<string, []>("op_2058_cast_fp16")];
            tensor<bool, []> var_2061_interleave_0 = const()[name = tensor<string, []>("op_2061_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2061_cast_fp16 = concat(axis = var_20, interleave = var_2061_interleave_0, values = (var_2001_cast_fp16, k_patches_31_cast_fp16))[name = tensor<string, []>("op_2061_cast_fp16")];
            tensor<fp16, []> var_2063_to_fp16 = const()[name = tensor<string, []>("op_2063_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_159_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_2063_to_fp16)[name = tensor<string, []>("q_159_cast_fp16")];
            tensor<bool, []> attn_61_transpose_x_1 = const()[name = tensor<string, []>("attn_61_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_61_transpose_y_1 = const()[name = tensor<string, []>("attn_61_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_1, transpose_y = attn_61_transpose_y_1, x = q_159_cast_fp16, y = var_2061_cast_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_63_cast_fp16 = softmax(axis = var_21, x = attn_61_cast_fp16)[name = tensor<string, []>("attn_63_cast_fp16")];
            tensor<bool, []> var_2068_transpose_x_0 = const()[name = tensor<string, []>("op_2068_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2068_transpose_y_0 = const()[name = tensor<string, []>("op_2068_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2068_cast_fp16 = matmul(transpose_x = var_2068_transpose_x_0, transpose_y = var_2068_transpose_y_0, x = attn_63_cast_fp16, y = v_95_cast_fp16)[name = tensor<string, []>("op_2068_cast_fp16")];
            tensor<int32, [1]> var_2069_axes_0 = const()[name = tensor<string, []>("op_2069_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2069_cast_fp16 = squeeze(axes = var_2069_axes_0, x = var_2068_cast_fp16)[name = tensor<string, []>("op_2069_cast_fp16")];
            tensor<string, []> var_2075_pad_type_0 = const()[name = tensor<string, []>("op_2075_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2075_strides_0 = const()[name = tensor<string, []>("op_2075_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2075_pad_0 = const()[name = tensor<string, []>("op_2075_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2075_dilations_0 = const()[name = tensor<string, []>("op_2075_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2075_groups_0 = const()[name = tensor<string, []>("op_2075_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16863168)))];
            tensor<fp16, [1, 64, 1, 197]> var_2075_cast_fp16 = conv(dilations = var_2075_dilations_0, groups = var_2075_groups_0, pad = var_2075_pad_0, pad_type = var_2075_pad_type_0, strides = var_2075_strides_0, weight = model_blocks_1_attn_q_projs_4_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2075_cast_fp16")];
            tensor<int32, [1]> var_2076_axes_0 = const()[name = tensor<string, []>("op_2076_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2076_cast_fp16 = squeeze(axes = var_2076_axes_0, x = var_2075_cast_fp16)[name = tensor<string, []>("op_2076_cast_fp16")];
            tensor<int32, [3]> q_161_perm_0 = const()[name = tensor<string, []>("q_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2083_pad_type_0 = const()[name = tensor<string, []>("op_2083_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2083_strides_0 = const()[name = tensor<string, []>("op_2083_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2083_pad_0 = const()[name = tensor<string, []>("op_2083_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2083_dilations_0 = const()[name = tensor<string, []>("op_2083_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2083_groups_0 = const()[name = tensor<string, []>("op_2083_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16961536)))];
            tensor<fp16, [1, 64, 1, 197]> var_2083_cast_fp16 = conv(dilations = var_2083_dilations_0, groups = var_2083_groups_0, pad = var_2083_pad_0, pad_type = var_2083_pad_type_0, strides = var_2083_strides_0, weight = model_blocks_1_attn_k_projs_4_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2083_cast_fp16")];
            tensor<int32, [1]> var_2084_axes_0 = const()[name = tensor<string, []>("op_2084_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2084_cast_fp16 = squeeze(axes = var_2084_axes_0, x = var_2083_cast_fp16)[name = tensor<string, []>("op_2084_cast_fp16")];
            tensor<int32, [3]> k_97_perm_0 = const()[name = tensor<string, []>("k_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2091_pad_type_0 = const()[name = tensor<string, []>("op_2091_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2091_strides_0 = const()[name = tensor<string, []>("op_2091_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2091_pad_0 = const()[name = tensor<string, []>("op_2091_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2091_dilations_0 = const()[name = tensor<string, []>("op_2091_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2091_groups_0 = const()[name = tensor<string, []>("op_2091_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17059904)))];
            tensor<fp16, [1, 64, 1, 197]> var_2091_cast_fp16 = conv(dilations = var_2091_dilations_0, groups = var_2091_groups_0, pad = var_2091_pad_0, pad_type = var_2091_pad_type_0, strides = var_2091_strides_0, weight = model_blocks_1_attn_v_projs_4_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2091_cast_fp16")];
            tensor<int32, [1]> var_2092_axes_0 = const()[name = tensor<string, []>("op_2092_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2092_cast_fp16 = squeeze(axes = var_2092_axes_0, x = var_2091_cast_fp16)[name = tensor<string, []>("op_2092_cast_fp16")];
            tensor<int32, [3]> v_97_perm_0 = const()[name = tensor<string, []>("v_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17158272)))];
            tensor<fp16, [1, 197, 64]> q_161_cast_fp16 = transpose(perm = q_161_perm_0, x = var_2076_cast_fp16)[name = tensor<string, []>("transpose_459")];
            tensor<fp16, [1, 197, 64]> q_163_cast_fp16 = add(x = q_161_cast_fp16, y = model_blocks_1_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_163_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17158464)))];
            tensor<fp16, [1, 197, 64]> v_97_cast_fp16 = transpose(perm = v_97_perm_0, x = var_2092_cast_fp16)[name = tensor<string, []>("transpose_457")];
            tensor<fp16, [1, 197, 64]> v_99_cast_fp16 = add(x = v_97_cast_fp16, y = model_blocks_1_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_99_cast_fp16")];
            tensor<int32, [1]> q_165_axes_0 = const()[name = tensor<string, []>("q_165_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_165_cast_fp16 = expand_dims(axes = q_165_axes_0, x = q_163_cast_fp16)[name = tensor<string, []>("q_165_cast_fp16")];
            tensor<int32, [1]> k_99_axes_0 = const()[name = tensor<string, []>("k_99_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_97_cast_fp16 = transpose(perm = k_97_perm_0, x = var_2084_cast_fp16)[name = tensor<string, []>("transpose_458")];
            tensor<fp16, [1, 1, 197, 64]> k_99_cast_fp16 = expand_dims(axes = k_99_axes_0, x = k_97_cast_fp16)[name = tensor<string, []>("k_99_cast_fp16")];
            tensor<int32, [1]> v_101_axes_0 = const()[name = tensor<string, []>("v_101_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_101_cast_fp16 = expand_dims(axes = v_101_axes_0, x = v_99_cast_fp16)[name = tensor<string, []>("v_101_cast_fp16")];
            tensor<int32, [4]> var_2101_begin_0 = const()[name = tensor<string, []>("op_2101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2101_end_0 = const()[name = tensor<string, []>("op_2101_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2101_end_mask_0 = const()[name = tensor<string, []>("op_2101_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2101_cast_fp16 = slice_by_index(begin = var_2101_begin_0, end = var_2101_end_0, end_mask = var_2101_end_mask_0, x = q_165_cast_fp16)[name = tensor<string, []>("op_2101_cast_fp16")];
            tensor<int32, [4]> var_2105_begin_0 = const()[name = tensor<string, []>("op_2105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2105_end_0 = const()[name = tensor<string, []>("op_2105_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2105_end_mask_0 = const()[name = tensor<string, []>("op_2105_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2105_cast_fp16 = slice_by_index(begin = var_2105_begin_0, end = var_2105_end_0, end_mask = var_2105_end_mask_0, x = k_99_cast_fp16)[name = tensor<string, []>("op_2105_cast_fp16")];
            tensor<int32, [4]> var_2109_begin_0 = const()[name = tensor<string, []>("op_2109_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2109_end_0 = const()[name = tensor<string, []>("op_2109_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2109_end_mask_0 = const()[name = tensor<string, []>("op_2109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2109_cast_fp16 = slice_by_index(begin = var_2109_begin_0, end = var_2109_end_0, end_mask = var_2109_end_mask_0, x = q_165_cast_fp16)[name = tensor<string, []>("op_2109_cast_fp16")];
            tensor<int32, [4]> var_2113_begin_0 = const()[name = tensor<string, []>("op_2113_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2113_end_0 = const()[name = tensor<string, []>("op_2113_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2113_end_mask_0 = const()[name = tensor<string, []>("op_2113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2113_cast_fp16 = slice_by_index(begin = var_2113_begin_0, end = var_2113_end_0, end_mask = var_2113_end_mask_0, x = k_99_cast_fp16)[name = tensor<string, []>("op_2113_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2115_cast_fp16 = mul(x = var_2109_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2115_cast_fp16")];
            tensor<int32, [5]> var_2123 = const()[name = tensor<string, []>("op_2123"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_223_cast_fp16 = reshape(shape = var_2123, x = var_2109_cast_fp16)[name = tensor<string, []>("x_223_cast_fp16")];
            tensor<int32, [2]> var_2125_split_sizes_0 = const()[name = tensor<string, []>("op_2125_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2125_axis_0 = const()[name = tensor<string, []>("op_2125_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2125_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2125_cast_fp16_1 = split(axis = var_2125_axis_0, split_sizes = var_2125_split_sizes_0, x = x_223_cast_fp16)[name = tensor<string, []>("op_2125_cast_fp16")];
            tensor<int32, [1]> squeeze_64_axes_0 = const()[name = tensor<string, []>("squeeze_64_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_64_cast_fp16 = squeeze(axes = squeeze_64_axes_0, x = var_2125_cast_fp16_0)[name = tensor<string, []>("squeeze_64_cast_fp16")];
            tensor<int32, [1]> squeeze_65_axes_0 = const()[name = tensor<string, []>("squeeze_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_65_cast_fp16 = squeeze(axes = squeeze_65_axes_0, x = var_2125_cast_fp16_1)[name = tensor<string, []>("squeeze_65_cast_fp16")];
            tensor<fp16, []> const_267_promoted_to_fp16 = const()[name = tensor<string, []>("const_267_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2128_cast_fp16 = mul(x = squeeze_65_cast_fp16, y = const_267_promoted_to_fp16)[name = tensor<string, []>("op_2128_cast_fp16")];
            tensor<int32, []> x_225_axis_0 = const()[name = tensor<string, []>("x_225_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_225_cast_fp16 = stack(axis = x_225_axis_0, values = (var_2128_cast_fp16, squeeze_64_cast_fp16))[name = tensor<string, []>("x_225_cast_fp16")];
            tensor<int32, [4]> var_2134 = const()[name = tensor<string, []>("op_2134"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2135_cast_fp16 = reshape(shape = var_2134, x = x_225_cast_fp16)[name = tensor<string, []>("op_2135_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2136_cast_fp16 = mul(x = var_2135_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2136_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_33_cast_fp16 = add(x = var_2115_cast_fp16, y = var_2136_cast_fp16)[name = tensor<string, []>("q_patches_33_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2138_cast_fp16 = mul(x = var_2113_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2138_cast_fp16")];
            tensor<int32, [5]> var_2146 = const()[name = tensor<string, []>("op_2146"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_229_cast_fp16 = reshape(shape = var_2146, x = var_2113_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
            tensor<int32, [2]> var_2148_split_sizes_0 = const()[name = tensor<string, []>("op_2148_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2148_axis_0 = const()[name = tensor<string, []>("op_2148_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2148_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2148_cast_fp16_1 = split(axis = var_2148_axis_0, split_sizes = var_2148_split_sizes_0, x = x_229_cast_fp16)[name = tensor<string, []>("op_2148_cast_fp16")];
            tensor<int32, [1]> squeeze_66_axes_0 = const()[name = tensor<string, []>("squeeze_66_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_66_cast_fp16 = squeeze(axes = squeeze_66_axes_0, x = var_2148_cast_fp16_0)[name = tensor<string, []>("squeeze_66_cast_fp16")];
            tensor<int32, [1]> squeeze_67_axes_0 = const()[name = tensor<string, []>("squeeze_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_67_cast_fp16 = squeeze(axes = squeeze_67_axes_0, x = var_2148_cast_fp16_1)[name = tensor<string, []>("squeeze_67_cast_fp16")];
            tensor<fp16, []> const_275_promoted_to_fp16 = const()[name = tensor<string, []>("const_275_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2151_cast_fp16 = mul(x = squeeze_67_cast_fp16, y = const_275_promoted_to_fp16)[name = tensor<string, []>("op_2151_cast_fp16")];
            tensor<int32, []> x_231_axis_0 = const()[name = tensor<string, []>("x_231_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_231_cast_fp16 = stack(axis = x_231_axis_0, values = (var_2151_cast_fp16, squeeze_66_cast_fp16))[name = tensor<string, []>("x_231_cast_fp16")];
            tensor<int32, [4]> var_2157 = const()[name = tensor<string, []>("op_2157"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2158_cast_fp16 = reshape(shape = var_2157, x = x_231_cast_fp16)[name = tensor<string, []>("op_2158_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2159_cast_fp16 = mul(x = var_2158_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2159_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_33_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2159_cast_fp16)[name = tensor<string, []>("k_patches_33_cast_fp16")];
            tensor<bool, []> var_2162_interleave_0 = const()[name = tensor<string, []>("op_2162_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2162_cast_fp16 = concat(axis = var_20, interleave = var_2162_interleave_0, values = (var_2101_cast_fp16, q_patches_33_cast_fp16))[name = tensor<string, []>("op_2162_cast_fp16")];
            tensor<bool, []> var_2165_interleave_0 = const()[name = tensor<string, []>("op_2165_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2165_cast_fp16 = concat(axis = var_20, interleave = var_2165_interleave_0, values = (var_2105_cast_fp16, k_patches_33_cast_fp16))[name = tensor<string, []>("op_2165_cast_fp16")];
            tensor<fp16, []> var_2167_to_fp16 = const()[name = tensor<string, []>("op_2167_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_169_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2167_to_fp16)[name = tensor<string, []>("q_169_cast_fp16")];
            tensor<bool, []> attn_65_transpose_x_1 = const()[name = tensor<string, []>("attn_65_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_65_transpose_y_1 = const()[name = tensor<string, []>("attn_65_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_1, transpose_y = attn_65_transpose_y_1, x = q_169_cast_fp16, y = var_2165_cast_fp16)[name = tensor<string, []>("attn_65_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_67_cast_fp16 = softmax(axis = var_21, x = attn_65_cast_fp16)[name = tensor<string, []>("attn_67_cast_fp16")];
            tensor<bool, []> var_2172_transpose_x_0 = const()[name = tensor<string, []>("op_2172_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2172_transpose_y_0 = const()[name = tensor<string, []>("op_2172_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2172_cast_fp16 = matmul(transpose_x = var_2172_transpose_x_0, transpose_y = var_2172_transpose_y_0, x = attn_67_cast_fp16, y = v_101_cast_fp16)[name = tensor<string, []>("op_2172_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = tensor<string, []>("op_2173_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2173_cast_fp16 = squeeze(axes = var_2173_axes_0, x = var_2172_cast_fp16)[name = tensor<string, []>("op_2173_cast_fp16")];
            tensor<string, []> var_2179_pad_type_0 = const()[name = tensor<string, []>("op_2179_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2179_strides_0 = const()[name = tensor<string, []>("op_2179_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2179_pad_0 = const()[name = tensor<string, []>("op_2179_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2179_dilations_0 = const()[name = tensor<string, []>("op_2179_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2179_groups_0 = const()[name = tensor<string, []>("op_2179_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17158656)))];
            tensor<fp16, [1, 64, 1, 197]> var_2179_cast_fp16 = conv(dilations = var_2179_dilations_0, groups = var_2179_groups_0, pad = var_2179_pad_0, pad_type = var_2179_pad_type_0, strides = var_2179_strides_0, weight = model_blocks_1_attn_q_projs_5_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2179_cast_fp16")];
            tensor<int32, [1]> var_2180_axes_0 = const()[name = tensor<string, []>("op_2180_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2180_cast_fp16 = squeeze(axes = var_2180_axes_0, x = var_2179_cast_fp16)[name = tensor<string, []>("op_2180_cast_fp16")];
            tensor<int32, [3]> q_171_perm_0 = const()[name = tensor<string, []>("q_171_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2187_pad_type_0 = const()[name = tensor<string, []>("op_2187_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2187_strides_0 = const()[name = tensor<string, []>("op_2187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2187_pad_0 = const()[name = tensor<string, []>("op_2187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2187_dilations_0 = const()[name = tensor<string, []>("op_2187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2187_groups_0 = const()[name = tensor<string, []>("op_2187_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17257024)))];
            tensor<fp16, [1, 64, 1, 197]> var_2187_cast_fp16 = conv(dilations = var_2187_dilations_0, groups = var_2187_groups_0, pad = var_2187_pad_0, pad_type = var_2187_pad_type_0, strides = var_2187_strides_0, weight = model_blocks_1_attn_k_projs_5_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2188_axes_0 = const()[name = tensor<string, []>("op_2188_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2188_cast_fp16 = squeeze(axes = var_2188_axes_0, x = var_2187_cast_fp16)[name = tensor<string, []>("op_2188_cast_fp16")];
            tensor<int32, [3]> k_103_perm_0 = const()[name = tensor<string, []>("k_103_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2195_pad_type_0 = const()[name = tensor<string, []>("op_2195_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2195_strides_0 = const()[name = tensor<string, []>("op_2195_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2195_pad_0 = const()[name = tensor<string, []>("op_2195_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2195_dilations_0 = const()[name = tensor<string, []>("op_2195_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2195_groups_0 = const()[name = tensor<string, []>("op_2195_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17355392)))];
            tensor<fp16, [1, 64, 1, 197]> var_2195_cast_fp16 = conv(dilations = var_2195_dilations_0, groups = var_2195_groups_0, pad = var_2195_pad_0, pad_type = var_2195_pad_type_0, strides = var_2195_strides_0, weight = model_blocks_1_attn_v_projs_5_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2195_cast_fp16")];
            tensor<int32, [1]> var_2196_axes_0 = const()[name = tensor<string, []>("op_2196_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2196_cast_fp16 = squeeze(axes = var_2196_axes_0, x = var_2195_cast_fp16)[name = tensor<string, []>("op_2196_cast_fp16")];
            tensor<int32, [3]> v_103_perm_0 = const()[name = tensor<string, []>("v_103_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17453760)))];
            tensor<fp16, [1, 197, 64]> q_171_cast_fp16 = transpose(perm = q_171_perm_0, x = var_2180_cast_fp16)[name = tensor<string, []>("transpose_456")];
            tensor<fp16, [1, 197, 64]> q_173_cast_fp16 = add(x = q_171_cast_fp16, y = model_blocks_1_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_173_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17453952)))];
            tensor<fp16, [1, 197, 64]> v_103_cast_fp16 = transpose(perm = v_103_perm_0, x = var_2196_cast_fp16)[name = tensor<string, []>("transpose_454")];
            tensor<fp16, [1, 197, 64]> v_105_cast_fp16 = add(x = v_103_cast_fp16, y = model_blocks_1_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_105_cast_fp16")];
            tensor<int32, [1]> q_175_axes_0 = const()[name = tensor<string, []>("q_175_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_175_cast_fp16 = expand_dims(axes = q_175_axes_0, x = q_173_cast_fp16)[name = tensor<string, []>("q_175_cast_fp16")];
            tensor<int32, [1]> k_105_axes_0 = const()[name = tensor<string, []>("k_105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_103_cast_fp16 = transpose(perm = k_103_perm_0, x = var_2188_cast_fp16)[name = tensor<string, []>("transpose_455")];
            tensor<fp16, [1, 1, 197, 64]> k_105_cast_fp16 = expand_dims(axes = k_105_axes_0, x = k_103_cast_fp16)[name = tensor<string, []>("k_105_cast_fp16")];
            tensor<int32, [1]> v_107_axes_0 = const()[name = tensor<string, []>("v_107_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_107_cast_fp16 = expand_dims(axes = v_107_axes_0, x = v_105_cast_fp16)[name = tensor<string, []>("v_107_cast_fp16")];
            tensor<int32, [4]> var_2205_begin_0 = const()[name = tensor<string, []>("op_2205_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2205_end_0 = const()[name = tensor<string, []>("op_2205_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2205_end_mask_0 = const()[name = tensor<string, []>("op_2205_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2205_cast_fp16 = slice_by_index(begin = var_2205_begin_0, end = var_2205_end_0, end_mask = var_2205_end_mask_0, x = q_175_cast_fp16)[name = tensor<string, []>("op_2205_cast_fp16")];
            tensor<int32, [4]> var_2209_begin_0 = const()[name = tensor<string, []>("op_2209_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2209_end_0 = const()[name = tensor<string, []>("op_2209_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2209_end_mask_0 = const()[name = tensor<string, []>("op_2209_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2209_cast_fp16 = slice_by_index(begin = var_2209_begin_0, end = var_2209_end_0, end_mask = var_2209_end_mask_0, x = k_105_cast_fp16)[name = tensor<string, []>("op_2209_cast_fp16")];
            tensor<int32, [4]> var_2213_begin_0 = const()[name = tensor<string, []>("op_2213_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2213_end_0 = const()[name = tensor<string, []>("op_2213_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2213_end_mask_0 = const()[name = tensor<string, []>("op_2213_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2213_cast_fp16 = slice_by_index(begin = var_2213_begin_0, end = var_2213_end_0, end_mask = var_2213_end_mask_0, x = q_175_cast_fp16)[name = tensor<string, []>("op_2213_cast_fp16")];
            tensor<int32, [4]> var_2217_begin_0 = const()[name = tensor<string, []>("op_2217_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2217_end_0 = const()[name = tensor<string, []>("op_2217_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2217_end_mask_0 = const()[name = tensor<string, []>("op_2217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2217_cast_fp16 = slice_by_index(begin = var_2217_begin_0, end = var_2217_end_0, end_mask = var_2217_end_mask_0, x = k_105_cast_fp16)[name = tensor<string, []>("op_2217_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2219_cast_fp16 = mul(x = var_2213_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2219_cast_fp16")];
            tensor<int32, [5]> var_2227 = const()[name = tensor<string, []>("op_2227"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_235_cast_fp16 = reshape(shape = var_2227, x = var_2213_cast_fp16)[name = tensor<string, []>("x_235_cast_fp16")];
            tensor<int32, [2]> var_2229_split_sizes_0 = const()[name = tensor<string, []>("op_2229_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2229_axis_0 = const()[name = tensor<string, []>("op_2229_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2229_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2229_cast_fp16_1 = split(axis = var_2229_axis_0, split_sizes = var_2229_split_sizes_0, x = x_235_cast_fp16)[name = tensor<string, []>("op_2229_cast_fp16")];
            tensor<int32, [1]> squeeze_68_axes_0 = const()[name = tensor<string, []>("squeeze_68_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_68_cast_fp16 = squeeze(axes = squeeze_68_axes_0, x = var_2229_cast_fp16_0)[name = tensor<string, []>("squeeze_68_cast_fp16")];
            tensor<int32, [1]> squeeze_69_axes_0 = const()[name = tensor<string, []>("squeeze_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_69_cast_fp16 = squeeze(axes = squeeze_69_axes_0, x = var_2229_cast_fp16_1)[name = tensor<string, []>("squeeze_69_cast_fp16")];
            tensor<fp16, []> const_283_promoted_to_fp16 = const()[name = tensor<string, []>("const_283_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2232_cast_fp16 = mul(x = squeeze_69_cast_fp16, y = const_283_promoted_to_fp16)[name = tensor<string, []>("op_2232_cast_fp16")];
            tensor<int32, []> x_237_axis_0 = const()[name = tensor<string, []>("x_237_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_237_cast_fp16 = stack(axis = x_237_axis_0, values = (var_2232_cast_fp16, squeeze_68_cast_fp16))[name = tensor<string, []>("x_237_cast_fp16")];
            tensor<int32, [4]> var_2238 = const()[name = tensor<string, []>("op_2238"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2239_cast_fp16 = reshape(shape = var_2238, x = x_237_cast_fp16)[name = tensor<string, []>("op_2239_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2240_cast_fp16 = mul(x = var_2239_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2240_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_35_cast_fp16 = add(x = var_2219_cast_fp16, y = var_2240_cast_fp16)[name = tensor<string, []>("q_patches_35_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2242_cast_fp16 = mul(x = var_2217_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2242_cast_fp16")];
            tensor<int32, [5]> var_2250 = const()[name = tensor<string, []>("op_2250"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_241_cast_fp16 = reshape(shape = var_2250, x = var_2217_cast_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
            tensor<int32, [2]> var_2252_split_sizes_0 = const()[name = tensor<string, []>("op_2252_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2252_axis_0 = const()[name = tensor<string, []>("op_2252_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2252_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2252_cast_fp16_1 = split(axis = var_2252_axis_0, split_sizes = var_2252_split_sizes_0, x = x_241_cast_fp16)[name = tensor<string, []>("op_2252_cast_fp16")];
            tensor<int32, [1]> squeeze_70_axes_0 = const()[name = tensor<string, []>("squeeze_70_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_70_cast_fp16 = squeeze(axes = squeeze_70_axes_0, x = var_2252_cast_fp16_0)[name = tensor<string, []>("squeeze_70_cast_fp16")];
            tensor<int32, [1]> squeeze_71_axes_0 = const()[name = tensor<string, []>("squeeze_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_71_cast_fp16 = squeeze(axes = squeeze_71_axes_0, x = var_2252_cast_fp16_1)[name = tensor<string, []>("squeeze_71_cast_fp16")];
            tensor<fp16, []> const_291_promoted_to_fp16 = const()[name = tensor<string, []>("const_291_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2255_cast_fp16 = mul(x = squeeze_71_cast_fp16, y = const_291_promoted_to_fp16)[name = tensor<string, []>("op_2255_cast_fp16")];
            tensor<int32, []> x_243_axis_0 = const()[name = tensor<string, []>("x_243_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_243_cast_fp16 = stack(axis = x_243_axis_0, values = (var_2255_cast_fp16, squeeze_70_cast_fp16))[name = tensor<string, []>("x_243_cast_fp16")];
            tensor<int32, [4]> var_2261 = const()[name = tensor<string, []>("op_2261"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2262_cast_fp16 = reshape(shape = var_2261, x = x_243_cast_fp16)[name = tensor<string, []>("op_2262_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2263_cast_fp16 = mul(x = var_2262_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2263_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_35_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2263_cast_fp16)[name = tensor<string, []>("k_patches_35_cast_fp16")];
            tensor<bool, []> var_2266_interleave_0 = const()[name = tensor<string, []>("op_2266_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2266_cast_fp16 = concat(axis = var_20, interleave = var_2266_interleave_0, values = (var_2205_cast_fp16, q_patches_35_cast_fp16))[name = tensor<string, []>("op_2266_cast_fp16")];
            tensor<bool, []> var_2269_interleave_0 = const()[name = tensor<string, []>("op_2269_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2269_cast_fp16 = concat(axis = var_20, interleave = var_2269_interleave_0, values = (var_2209_cast_fp16, k_patches_35_cast_fp16))[name = tensor<string, []>("op_2269_cast_fp16")];
            tensor<fp16, []> var_2271_to_fp16 = const()[name = tensor<string, []>("op_2271_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_179_cast_fp16 = mul(x = var_2266_cast_fp16, y = var_2271_to_fp16)[name = tensor<string, []>("q_179_cast_fp16")];
            tensor<bool, []> attn_69_transpose_x_1 = const()[name = tensor<string, []>("attn_69_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_69_transpose_y_1 = const()[name = tensor<string, []>("attn_69_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_1, transpose_y = attn_69_transpose_y_1, x = q_179_cast_fp16, y = var_2269_cast_fp16)[name = tensor<string, []>("attn_69_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_71_cast_fp16 = softmax(axis = var_21, x = attn_69_cast_fp16)[name = tensor<string, []>("attn_71_cast_fp16")];
            tensor<bool, []> var_2276_transpose_x_0 = const()[name = tensor<string, []>("op_2276_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2276_transpose_y_0 = const()[name = tensor<string, []>("op_2276_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2276_cast_fp16 = matmul(transpose_x = var_2276_transpose_x_0, transpose_y = var_2276_transpose_y_0, x = attn_71_cast_fp16, y = v_107_cast_fp16)[name = tensor<string, []>("op_2276_cast_fp16")];
            tensor<int32, [1]> var_2277_axes_0 = const()[name = tensor<string, []>("op_2277_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2277_cast_fp16 = squeeze(axes = var_2277_axes_0, x = var_2276_cast_fp16)[name = tensor<string, []>("op_2277_cast_fp16")];
            tensor<string, []> var_2283_pad_type_0 = const()[name = tensor<string, []>("op_2283_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2283_strides_0 = const()[name = tensor<string, []>("op_2283_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2283_pad_0 = const()[name = tensor<string, []>("op_2283_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2283_dilations_0 = const()[name = tensor<string, []>("op_2283_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2283_groups_0 = const()[name = tensor<string, []>("op_2283_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17454144)))];
            tensor<fp16, [1, 64, 1, 197]> var_2283_cast_fp16 = conv(dilations = var_2283_dilations_0, groups = var_2283_groups_0, pad = var_2283_pad_0, pad_type = var_2283_pad_type_0, strides = var_2283_strides_0, weight = model_blocks_1_attn_q_projs_6_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2283_cast_fp16")];
            tensor<int32, [1]> var_2284_axes_0 = const()[name = tensor<string, []>("op_2284_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2284_cast_fp16 = squeeze(axes = var_2284_axes_0, x = var_2283_cast_fp16)[name = tensor<string, []>("op_2284_cast_fp16")];
            tensor<int32, [3]> q_181_perm_0 = const()[name = tensor<string, []>("q_181_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2291_pad_type_0 = const()[name = tensor<string, []>("op_2291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2291_strides_0 = const()[name = tensor<string, []>("op_2291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2291_pad_0 = const()[name = tensor<string, []>("op_2291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2291_dilations_0 = const()[name = tensor<string, []>("op_2291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2291_groups_0 = const()[name = tensor<string, []>("op_2291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17552512)))];
            tensor<fp16, [1, 64, 1, 197]> var_2291_cast_fp16 = conv(dilations = var_2291_dilations_0, groups = var_2291_groups_0, pad = var_2291_pad_0, pad_type = var_2291_pad_type_0, strides = var_2291_strides_0, weight = model_blocks_1_attn_k_projs_6_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2291_cast_fp16")];
            tensor<int32, [1]> var_2292_axes_0 = const()[name = tensor<string, []>("op_2292_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2292_cast_fp16 = squeeze(axes = var_2292_axes_0, x = var_2291_cast_fp16)[name = tensor<string, []>("op_2292_cast_fp16")];
            tensor<int32, [3]> k_109_perm_0 = const()[name = tensor<string, []>("k_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2299_pad_type_0 = const()[name = tensor<string, []>("op_2299_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2299_strides_0 = const()[name = tensor<string, []>("op_2299_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2299_pad_0 = const()[name = tensor<string, []>("op_2299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2299_dilations_0 = const()[name = tensor<string, []>("op_2299_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2299_groups_0 = const()[name = tensor<string, []>("op_2299_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17650880)))];
            tensor<fp16, [1, 64, 1, 197]> var_2299_cast_fp16 = conv(dilations = var_2299_dilations_0, groups = var_2299_groups_0, pad = var_2299_pad_0, pad_type = var_2299_pad_type_0, strides = var_2299_strides_0, weight = model_blocks_1_attn_v_projs_6_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2299_cast_fp16")];
            tensor<int32, [1]> var_2300_axes_0 = const()[name = tensor<string, []>("op_2300_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2300_cast_fp16 = squeeze(axes = var_2300_axes_0, x = var_2299_cast_fp16)[name = tensor<string, []>("op_2300_cast_fp16")];
            tensor<int32, [3]> v_109_perm_0 = const()[name = tensor<string, []>("v_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17749248)))];
            tensor<fp16, [1, 197, 64]> q_181_cast_fp16 = transpose(perm = q_181_perm_0, x = var_2284_cast_fp16)[name = tensor<string, []>("transpose_453")];
            tensor<fp16, [1, 197, 64]> q_183_cast_fp16 = add(x = q_181_cast_fp16, y = model_blocks_1_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_183_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17749440)))];
            tensor<fp16, [1, 197, 64]> v_109_cast_fp16 = transpose(perm = v_109_perm_0, x = var_2300_cast_fp16)[name = tensor<string, []>("transpose_451")];
            tensor<fp16, [1, 197, 64]> v_111_cast_fp16 = add(x = v_109_cast_fp16, y = model_blocks_1_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_111_cast_fp16")];
            tensor<int32, [1]> q_185_axes_0 = const()[name = tensor<string, []>("q_185_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_185_cast_fp16 = expand_dims(axes = q_185_axes_0, x = q_183_cast_fp16)[name = tensor<string, []>("q_185_cast_fp16")];
            tensor<int32, [1]> k_111_axes_0 = const()[name = tensor<string, []>("k_111_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_109_cast_fp16 = transpose(perm = k_109_perm_0, x = var_2292_cast_fp16)[name = tensor<string, []>("transpose_452")];
            tensor<fp16, [1, 1, 197, 64]> k_111_cast_fp16 = expand_dims(axes = k_111_axes_0, x = k_109_cast_fp16)[name = tensor<string, []>("k_111_cast_fp16")];
            tensor<int32, [1]> v_113_axes_0 = const()[name = tensor<string, []>("v_113_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_113_cast_fp16 = expand_dims(axes = v_113_axes_0, x = v_111_cast_fp16)[name = tensor<string, []>("v_113_cast_fp16")];
            tensor<int32, [4]> var_2309_begin_0 = const()[name = tensor<string, []>("op_2309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2309_end_0 = const()[name = tensor<string, []>("op_2309_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2309_end_mask_0 = const()[name = tensor<string, []>("op_2309_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2309_cast_fp16 = slice_by_index(begin = var_2309_begin_0, end = var_2309_end_0, end_mask = var_2309_end_mask_0, x = q_185_cast_fp16)[name = tensor<string, []>("op_2309_cast_fp16")];
            tensor<int32, [4]> var_2313_begin_0 = const()[name = tensor<string, []>("op_2313_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2313_end_0 = const()[name = tensor<string, []>("op_2313_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2313_end_mask_0 = const()[name = tensor<string, []>("op_2313_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2313_cast_fp16 = slice_by_index(begin = var_2313_begin_0, end = var_2313_end_0, end_mask = var_2313_end_mask_0, x = k_111_cast_fp16)[name = tensor<string, []>("op_2313_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = tensor<string, []>("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = tensor<string, []>("op_2317_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = tensor<string, []>("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_185_cast_fp16)[name = tensor<string, []>("op_2317_cast_fp16")];
            tensor<int32, [4]> var_2321_begin_0 = const()[name = tensor<string, []>("op_2321_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2321_end_0 = const()[name = tensor<string, []>("op_2321_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2321_end_mask_0 = const()[name = tensor<string, []>("op_2321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2321_cast_fp16 = slice_by_index(begin = var_2321_begin_0, end = var_2321_end_0, end_mask = var_2321_end_mask_0, x = k_111_cast_fp16)[name = tensor<string, []>("op_2321_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2323_cast_fp16 = mul(x = var_2317_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2323_cast_fp16")];
            tensor<int32, [5]> var_2331 = const()[name = tensor<string, []>("op_2331"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_247_cast_fp16 = reshape(shape = var_2331, x = var_2317_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
            tensor<int32, [2]> var_2333_split_sizes_0 = const()[name = tensor<string, []>("op_2333_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2333_axis_0 = const()[name = tensor<string, []>("op_2333_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2333_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2333_cast_fp16_1 = split(axis = var_2333_axis_0, split_sizes = var_2333_split_sizes_0, x = x_247_cast_fp16)[name = tensor<string, []>("op_2333_cast_fp16")];
            tensor<int32, [1]> squeeze_72_axes_0 = const()[name = tensor<string, []>("squeeze_72_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_72_cast_fp16 = squeeze(axes = squeeze_72_axes_0, x = var_2333_cast_fp16_0)[name = tensor<string, []>("squeeze_72_cast_fp16")];
            tensor<int32, [1]> squeeze_73_axes_0 = const()[name = tensor<string, []>("squeeze_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_73_cast_fp16 = squeeze(axes = squeeze_73_axes_0, x = var_2333_cast_fp16_1)[name = tensor<string, []>("squeeze_73_cast_fp16")];
            tensor<fp16, []> const_299_promoted_to_fp16 = const()[name = tensor<string, []>("const_299_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2336_cast_fp16 = mul(x = squeeze_73_cast_fp16, y = const_299_promoted_to_fp16)[name = tensor<string, []>("op_2336_cast_fp16")];
            tensor<int32, []> x_249_axis_0 = const()[name = tensor<string, []>("x_249_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_249_cast_fp16 = stack(axis = x_249_axis_0, values = (var_2336_cast_fp16, squeeze_72_cast_fp16))[name = tensor<string, []>("x_249_cast_fp16")];
            tensor<int32, [4]> var_2342 = const()[name = tensor<string, []>("op_2342"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2343_cast_fp16 = reshape(shape = var_2342, x = x_249_cast_fp16)[name = tensor<string, []>("op_2343_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2344_cast_fp16 = mul(x = var_2343_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2344_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_37_cast_fp16 = add(x = var_2323_cast_fp16, y = var_2344_cast_fp16)[name = tensor<string, []>("q_patches_37_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2346_cast_fp16 = mul(x = var_2321_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2346_cast_fp16")];
            tensor<int32, [5]> var_2354 = const()[name = tensor<string, []>("op_2354"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_253_cast_fp16 = reshape(shape = var_2354, x = var_2321_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
            tensor<int32, [2]> var_2356_split_sizes_0 = const()[name = tensor<string, []>("op_2356_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2356_axis_0 = const()[name = tensor<string, []>("op_2356_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2356_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2356_cast_fp16_1 = split(axis = var_2356_axis_0, split_sizes = var_2356_split_sizes_0, x = x_253_cast_fp16)[name = tensor<string, []>("op_2356_cast_fp16")];
            tensor<int32, [1]> squeeze_74_axes_0 = const()[name = tensor<string, []>("squeeze_74_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_74_cast_fp16 = squeeze(axes = squeeze_74_axes_0, x = var_2356_cast_fp16_0)[name = tensor<string, []>("squeeze_74_cast_fp16")];
            tensor<int32, [1]> squeeze_75_axes_0 = const()[name = tensor<string, []>("squeeze_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_75_cast_fp16 = squeeze(axes = squeeze_75_axes_0, x = var_2356_cast_fp16_1)[name = tensor<string, []>("squeeze_75_cast_fp16")];
            tensor<fp16, []> const_307_promoted_to_fp16 = const()[name = tensor<string, []>("const_307_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2359_cast_fp16 = mul(x = squeeze_75_cast_fp16, y = const_307_promoted_to_fp16)[name = tensor<string, []>("op_2359_cast_fp16")];
            tensor<int32, []> x_255_axis_0 = const()[name = tensor<string, []>("x_255_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_255_cast_fp16 = stack(axis = x_255_axis_0, values = (var_2359_cast_fp16, squeeze_74_cast_fp16))[name = tensor<string, []>("x_255_cast_fp16")];
            tensor<int32, [4]> var_2365 = const()[name = tensor<string, []>("op_2365"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2366_cast_fp16 = reshape(shape = var_2365, x = x_255_cast_fp16)[name = tensor<string, []>("op_2366_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2367_cast_fp16 = mul(x = var_2366_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2367_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_37_cast_fp16 = add(x = var_2346_cast_fp16, y = var_2367_cast_fp16)[name = tensor<string, []>("k_patches_37_cast_fp16")];
            tensor<bool, []> var_2370_interleave_0 = const()[name = tensor<string, []>("op_2370_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2370_cast_fp16 = concat(axis = var_20, interleave = var_2370_interleave_0, values = (var_2309_cast_fp16, q_patches_37_cast_fp16))[name = tensor<string, []>("op_2370_cast_fp16")];
            tensor<bool, []> var_2373_interleave_0 = const()[name = tensor<string, []>("op_2373_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2373_cast_fp16 = concat(axis = var_20, interleave = var_2373_interleave_0, values = (var_2313_cast_fp16, k_patches_37_cast_fp16))[name = tensor<string, []>("op_2373_cast_fp16")];
            tensor<fp16, []> var_2375_to_fp16 = const()[name = tensor<string, []>("op_2375_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_189_cast_fp16 = mul(x = var_2370_cast_fp16, y = var_2375_to_fp16)[name = tensor<string, []>("q_189_cast_fp16")];
            tensor<bool, []> attn_73_transpose_x_1 = const()[name = tensor<string, []>("attn_73_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_73_transpose_y_1 = const()[name = tensor<string, []>("attn_73_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_1, transpose_y = attn_73_transpose_y_1, x = q_189_cast_fp16, y = var_2373_cast_fp16)[name = tensor<string, []>("attn_73_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_75_cast_fp16 = softmax(axis = var_21, x = attn_73_cast_fp16)[name = tensor<string, []>("attn_75_cast_fp16")];
            tensor<bool, []> var_2380_transpose_x_0 = const()[name = tensor<string, []>("op_2380_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2380_transpose_y_0 = const()[name = tensor<string, []>("op_2380_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2380_cast_fp16 = matmul(transpose_x = var_2380_transpose_x_0, transpose_y = var_2380_transpose_y_0, x = attn_75_cast_fp16, y = v_113_cast_fp16)[name = tensor<string, []>("op_2380_cast_fp16")];
            tensor<int32, [1]> var_2381_axes_0 = const()[name = tensor<string, []>("op_2381_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2381_cast_fp16 = squeeze(axes = var_2381_axes_0, x = var_2380_cast_fp16)[name = tensor<string, []>("op_2381_cast_fp16")];
            tensor<string, []> var_2387_pad_type_0 = const()[name = tensor<string, []>("op_2387_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2387_strides_0 = const()[name = tensor<string, []>("op_2387_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2387_pad_0 = const()[name = tensor<string, []>("op_2387_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2387_dilations_0 = const()[name = tensor<string, []>("op_2387_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2387_groups_0 = const()[name = tensor<string, []>("op_2387_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17749632)))];
            tensor<fp16, [1, 64, 1, 197]> var_2387_cast_fp16 = conv(dilations = var_2387_dilations_0, groups = var_2387_groups_0, pad = var_2387_pad_0, pad_type = var_2387_pad_type_0, strides = var_2387_strides_0, weight = model_blocks_1_attn_q_projs_7_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2387_cast_fp16")];
            tensor<int32, [1]> var_2388_axes_0 = const()[name = tensor<string, []>("op_2388_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2388_cast_fp16 = squeeze(axes = var_2388_axes_0, x = var_2387_cast_fp16)[name = tensor<string, []>("op_2388_cast_fp16")];
            tensor<int32, [3]> q_191_perm_0 = const()[name = tensor<string, []>("q_191_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2395_pad_type_0 = const()[name = tensor<string, []>("op_2395_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2395_strides_0 = const()[name = tensor<string, []>("op_2395_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2395_pad_0 = const()[name = tensor<string, []>("op_2395_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2395_dilations_0 = const()[name = tensor<string, []>("op_2395_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2395_groups_0 = const()[name = tensor<string, []>("op_2395_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17848000)))];
            tensor<fp16, [1, 64, 1, 197]> var_2395_cast_fp16 = conv(dilations = var_2395_dilations_0, groups = var_2395_groups_0, pad = var_2395_pad_0, pad_type = var_2395_pad_type_0, strides = var_2395_strides_0, weight = model_blocks_1_attn_k_projs_7_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2395_cast_fp16")];
            tensor<int32, [1]> var_2396_axes_0 = const()[name = tensor<string, []>("op_2396_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2396_cast_fp16 = squeeze(axes = var_2396_axes_0, x = var_2395_cast_fp16)[name = tensor<string, []>("op_2396_cast_fp16")];
            tensor<int32, [3]> k_115_perm_0 = const()[name = tensor<string, []>("k_115_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2403_pad_type_0 = const()[name = tensor<string, []>("op_2403_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2403_strides_0 = const()[name = tensor<string, []>("op_2403_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2403_pad_0 = const()[name = tensor<string, []>("op_2403_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2403_dilations_0 = const()[name = tensor<string, []>("op_2403_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2403_groups_0 = const()[name = tensor<string, []>("op_2403_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17946368)))];
            tensor<fp16, [1, 64, 1, 197]> var_2403_cast_fp16 = conv(dilations = var_2403_dilations_0, groups = var_2403_groups_0, pad = var_2403_pad_0, pad_type = var_2403_pad_type_0, strides = var_2403_strides_0, weight = model_blocks_1_attn_v_projs_7_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2403_cast_fp16")];
            tensor<int32, [1]> var_2404_axes_0 = const()[name = tensor<string, []>("op_2404_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2404_cast_fp16 = squeeze(axes = var_2404_axes_0, x = var_2403_cast_fp16)[name = tensor<string, []>("op_2404_cast_fp16")];
            tensor<int32, [3]> v_115_perm_0 = const()[name = tensor<string, []>("v_115_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18044736)))];
            tensor<fp16, [1, 197, 64]> q_191_cast_fp16 = transpose(perm = q_191_perm_0, x = var_2388_cast_fp16)[name = tensor<string, []>("transpose_450")];
            tensor<fp16, [1, 197, 64]> q_193_cast_fp16 = add(x = q_191_cast_fp16, y = model_blocks_1_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_193_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18044928)))];
            tensor<fp16, [1, 197, 64]> v_115_cast_fp16 = transpose(perm = v_115_perm_0, x = var_2404_cast_fp16)[name = tensor<string, []>("transpose_448")];
            tensor<fp16, [1, 197, 64]> v_117_cast_fp16 = add(x = v_115_cast_fp16, y = model_blocks_1_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_117_cast_fp16")];
            tensor<int32, [1]> q_195_axes_0 = const()[name = tensor<string, []>("q_195_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_195_cast_fp16 = expand_dims(axes = q_195_axes_0, x = q_193_cast_fp16)[name = tensor<string, []>("q_195_cast_fp16")];
            tensor<int32, [1]> k_117_axes_0 = const()[name = tensor<string, []>("k_117_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_115_cast_fp16 = transpose(perm = k_115_perm_0, x = var_2396_cast_fp16)[name = tensor<string, []>("transpose_449")];
            tensor<fp16, [1, 1, 197, 64]> k_117_cast_fp16 = expand_dims(axes = k_117_axes_0, x = k_115_cast_fp16)[name = tensor<string, []>("k_117_cast_fp16")];
            tensor<int32, [1]> v_119_axes_0 = const()[name = tensor<string, []>("v_119_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_119_cast_fp16 = expand_dims(axes = v_119_axes_0, x = v_117_cast_fp16)[name = tensor<string, []>("v_119_cast_fp16")];
            tensor<int32, [4]> var_2413_begin_0 = const()[name = tensor<string, []>("op_2413_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2413_end_0 = const()[name = tensor<string, []>("op_2413_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2413_end_mask_0 = const()[name = tensor<string, []>("op_2413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2413_cast_fp16 = slice_by_index(begin = var_2413_begin_0, end = var_2413_end_0, end_mask = var_2413_end_mask_0, x = q_195_cast_fp16)[name = tensor<string, []>("op_2413_cast_fp16")];
            tensor<int32, [4]> var_2417_begin_0 = const()[name = tensor<string, []>("op_2417_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2417_end_0 = const()[name = tensor<string, []>("op_2417_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2417_end_mask_0 = const()[name = tensor<string, []>("op_2417_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2417_cast_fp16 = slice_by_index(begin = var_2417_begin_0, end = var_2417_end_0, end_mask = var_2417_end_mask_0, x = k_117_cast_fp16)[name = tensor<string, []>("op_2417_cast_fp16")];
            tensor<int32, [4]> var_2421_begin_0 = const()[name = tensor<string, []>("op_2421_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2421_end_0 = const()[name = tensor<string, []>("op_2421_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2421_end_mask_0 = const()[name = tensor<string, []>("op_2421_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2421_cast_fp16 = slice_by_index(begin = var_2421_begin_0, end = var_2421_end_0, end_mask = var_2421_end_mask_0, x = q_195_cast_fp16)[name = tensor<string, []>("op_2421_cast_fp16")];
            tensor<int32, [4]> var_2425_begin_0 = const()[name = tensor<string, []>("op_2425_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2425_end_0 = const()[name = tensor<string, []>("op_2425_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2425_end_mask_0 = const()[name = tensor<string, []>("op_2425_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, x = k_117_cast_fp16)[name = tensor<string, []>("op_2425_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2427_cast_fp16 = mul(x = var_2421_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2427_cast_fp16")];
            tensor<int32, [5]> var_2435 = const()[name = tensor<string, []>("op_2435"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_259_cast_fp16 = reshape(shape = var_2435, x = var_2421_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
            tensor<int32, [2]> var_2437_split_sizes_0 = const()[name = tensor<string, []>("op_2437_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2437_axis_0 = const()[name = tensor<string, []>("op_2437_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2437_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2437_cast_fp16_1 = split(axis = var_2437_axis_0, split_sizes = var_2437_split_sizes_0, x = x_259_cast_fp16)[name = tensor<string, []>("op_2437_cast_fp16")];
            tensor<int32, [1]> squeeze_76_axes_0 = const()[name = tensor<string, []>("squeeze_76_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_76_cast_fp16 = squeeze(axes = squeeze_76_axes_0, x = var_2437_cast_fp16_0)[name = tensor<string, []>("squeeze_76_cast_fp16")];
            tensor<int32, [1]> squeeze_77_axes_0 = const()[name = tensor<string, []>("squeeze_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_77_cast_fp16 = squeeze(axes = squeeze_77_axes_0, x = var_2437_cast_fp16_1)[name = tensor<string, []>("squeeze_77_cast_fp16")];
            tensor<fp16, []> const_315_promoted_to_fp16 = const()[name = tensor<string, []>("const_315_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2440_cast_fp16 = mul(x = squeeze_77_cast_fp16, y = const_315_promoted_to_fp16)[name = tensor<string, []>("op_2440_cast_fp16")];
            tensor<int32, []> x_261_axis_0 = const()[name = tensor<string, []>("x_261_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_261_cast_fp16 = stack(axis = x_261_axis_0, values = (var_2440_cast_fp16, squeeze_76_cast_fp16))[name = tensor<string, []>("x_261_cast_fp16")];
            tensor<int32, [4]> var_2446 = const()[name = tensor<string, []>("op_2446"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2447_cast_fp16 = reshape(shape = var_2446, x = x_261_cast_fp16)[name = tensor<string, []>("op_2447_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2448_cast_fp16 = mul(x = var_2447_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2448_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_39_cast_fp16 = add(x = var_2427_cast_fp16, y = var_2448_cast_fp16)[name = tensor<string, []>("q_patches_39_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2450_cast_fp16 = mul(x = var_2425_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2450_cast_fp16")];
            tensor<int32, [5]> var_2458 = const()[name = tensor<string, []>("op_2458"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_265_cast_fp16 = reshape(shape = var_2458, x = var_2425_cast_fp16)[name = tensor<string, []>("x_265_cast_fp16")];
            tensor<int32, [2]> var_2460_split_sizes_0 = const()[name = tensor<string, []>("op_2460_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2460_axis_0 = const()[name = tensor<string, []>("op_2460_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2460_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2460_cast_fp16_1 = split(axis = var_2460_axis_0, split_sizes = var_2460_split_sizes_0, x = x_265_cast_fp16)[name = tensor<string, []>("op_2460_cast_fp16")];
            tensor<int32, [1]> squeeze_78_axes_0 = const()[name = tensor<string, []>("squeeze_78_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_78_cast_fp16 = squeeze(axes = squeeze_78_axes_0, x = var_2460_cast_fp16_0)[name = tensor<string, []>("squeeze_78_cast_fp16")];
            tensor<int32, [1]> squeeze_79_axes_0 = const()[name = tensor<string, []>("squeeze_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_79_cast_fp16 = squeeze(axes = squeeze_79_axes_0, x = var_2460_cast_fp16_1)[name = tensor<string, []>("squeeze_79_cast_fp16")];
            tensor<fp16, []> const_323_promoted_to_fp16 = const()[name = tensor<string, []>("const_323_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2463_cast_fp16 = mul(x = squeeze_79_cast_fp16, y = const_323_promoted_to_fp16)[name = tensor<string, []>("op_2463_cast_fp16")];
            tensor<int32, []> x_267_axis_0 = const()[name = tensor<string, []>("x_267_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_267_cast_fp16 = stack(axis = x_267_axis_0, values = (var_2463_cast_fp16, squeeze_78_cast_fp16))[name = tensor<string, []>("x_267_cast_fp16")];
            tensor<int32, [4]> var_2469 = const()[name = tensor<string, []>("op_2469"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2470_cast_fp16 = reshape(shape = var_2469, x = x_267_cast_fp16)[name = tensor<string, []>("op_2470_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2471_cast_fp16 = mul(x = var_2470_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2471_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_39_cast_fp16 = add(x = var_2450_cast_fp16, y = var_2471_cast_fp16)[name = tensor<string, []>("k_patches_39_cast_fp16")];
            tensor<bool, []> var_2474_interleave_0 = const()[name = tensor<string, []>("op_2474_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2474_cast_fp16 = concat(axis = var_20, interleave = var_2474_interleave_0, values = (var_2413_cast_fp16, q_patches_39_cast_fp16))[name = tensor<string, []>("op_2474_cast_fp16")];
            tensor<bool, []> var_2477_interleave_0 = const()[name = tensor<string, []>("op_2477_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2477_cast_fp16 = concat(axis = var_20, interleave = var_2477_interleave_0, values = (var_2417_cast_fp16, k_patches_39_cast_fp16))[name = tensor<string, []>("op_2477_cast_fp16")];
            tensor<fp16, []> var_2479_to_fp16 = const()[name = tensor<string, []>("op_2479_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_199_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2479_to_fp16)[name = tensor<string, []>("q_199_cast_fp16")];
            tensor<bool, []> attn_77_transpose_x_1 = const()[name = tensor<string, []>("attn_77_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_77_transpose_y_1 = const()[name = tensor<string, []>("attn_77_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_1, transpose_y = attn_77_transpose_y_1, x = q_199_cast_fp16, y = var_2477_cast_fp16)[name = tensor<string, []>("attn_77_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_79_cast_fp16 = softmax(axis = var_21, x = attn_77_cast_fp16)[name = tensor<string, []>("attn_79_cast_fp16")];
            tensor<bool, []> var_2484_transpose_x_0 = const()[name = tensor<string, []>("op_2484_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2484_transpose_y_0 = const()[name = tensor<string, []>("op_2484_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2484_cast_fp16 = matmul(transpose_x = var_2484_transpose_x_0, transpose_y = var_2484_transpose_y_0, x = attn_79_cast_fp16, y = v_119_cast_fp16)[name = tensor<string, []>("op_2484_cast_fp16")];
            tensor<int32, [1]> var_2485_axes_0 = const()[name = tensor<string, []>("op_2485_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2485_cast_fp16 = squeeze(axes = var_2485_axes_0, x = var_2484_cast_fp16)[name = tensor<string, []>("op_2485_cast_fp16")];
            tensor<string, []> var_2491_pad_type_0 = const()[name = tensor<string, []>("op_2491_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2491_strides_0 = const()[name = tensor<string, []>("op_2491_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2491_pad_0 = const()[name = tensor<string, []>("op_2491_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2491_dilations_0 = const()[name = tensor<string, []>("op_2491_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2491_groups_0 = const()[name = tensor<string, []>("op_2491_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18045120)))];
            tensor<fp16, [1, 64, 1, 197]> var_2491_cast_fp16 = conv(dilations = var_2491_dilations_0, groups = var_2491_groups_0, pad = var_2491_pad_0, pad_type = var_2491_pad_type_0, strides = var_2491_strides_0, weight = model_blocks_1_attn_q_projs_8_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2492_axes_0 = const()[name = tensor<string, []>("op_2492_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2492_cast_fp16 = squeeze(axes = var_2492_axes_0, x = var_2491_cast_fp16)[name = tensor<string, []>("op_2492_cast_fp16")];
            tensor<int32, [3]> q_201_perm_0 = const()[name = tensor<string, []>("q_201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2499_pad_type_0 = const()[name = tensor<string, []>("op_2499_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2499_strides_0 = const()[name = tensor<string, []>("op_2499_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2499_pad_0 = const()[name = tensor<string, []>("op_2499_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2499_dilations_0 = const()[name = tensor<string, []>("op_2499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2499_groups_0 = const()[name = tensor<string, []>("op_2499_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18143488)))];
            tensor<fp16, [1, 64, 1, 197]> var_2499_cast_fp16 = conv(dilations = var_2499_dilations_0, groups = var_2499_groups_0, pad = var_2499_pad_0, pad_type = var_2499_pad_type_0, strides = var_2499_strides_0, weight = model_blocks_1_attn_k_projs_8_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2499_cast_fp16")];
            tensor<int32, [1]> var_2500_axes_0 = const()[name = tensor<string, []>("op_2500_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2500_cast_fp16 = squeeze(axes = var_2500_axes_0, x = var_2499_cast_fp16)[name = tensor<string, []>("op_2500_cast_fp16")];
            tensor<int32, [3]> k_121_perm_0 = const()[name = tensor<string, []>("k_121_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2507_pad_type_0 = const()[name = tensor<string, []>("op_2507_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2507_strides_0 = const()[name = tensor<string, []>("op_2507_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2507_pad_0 = const()[name = tensor<string, []>("op_2507_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2507_dilations_0 = const()[name = tensor<string, []>("op_2507_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2507_groups_0 = const()[name = tensor<string, []>("op_2507_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18241856)))];
            tensor<fp16, [1, 64, 1, 197]> var_2507_cast_fp16 = conv(dilations = var_2507_dilations_0, groups = var_2507_groups_0, pad = var_2507_pad_0, pad_type = var_2507_pad_type_0, strides = var_2507_strides_0, weight = model_blocks_1_attn_v_projs_8_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2507_cast_fp16")];
            tensor<int32, [1]> var_2508_axes_0 = const()[name = tensor<string, []>("op_2508_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2508_cast_fp16 = squeeze(axes = var_2508_axes_0, x = var_2507_cast_fp16)[name = tensor<string, []>("op_2508_cast_fp16")];
            tensor<int32, [3]> v_121_perm_0 = const()[name = tensor<string, []>("v_121_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18340224)))];
            tensor<fp16, [1, 197, 64]> q_201_cast_fp16 = transpose(perm = q_201_perm_0, x = var_2492_cast_fp16)[name = tensor<string, []>("transpose_447")];
            tensor<fp16, [1, 197, 64]> q_203_cast_fp16 = add(x = q_201_cast_fp16, y = model_blocks_1_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_203_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18340416)))];
            tensor<fp16, [1, 197, 64]> v_121_cast_fp16 = transpose(perm = v_121_perm_0, x = var_2508_cast_fp16)[name = tensor<string, []>("transpose_445")];
            tensor<fp16, [1, 197, 64]> v_123_cast_fp16 = add(x = v_121_cast_fp16, y = model_blocks_1_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_123_cast_fp16")];
            tensor<int32, [1]> q_205_axes_0 = const()[name = tensor<string, []>("q_205_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_205_cast_fp16 = expand_dims(axes = q_205_axes_0, x = q_203_cast_fp16)[name = tensor<string, []>("q_205_cast_fp16")];
            tensor<int32, [1]> k_123_axes_0 = const()[name = tensor<string, []>("k_123_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_121_cast_fp16 = transpose(perm = k_121_perm_0, x = var_2500_cast_fp16)[name = tensor<string, []>("transpose_446")];
            tensor<fp16, [1, 1, 197, 64]> k_123_cast_fp16 = expand_dims(axes = k_123_axes_0, x = k_121_cast_fp16)[name = tensor<string, []>("k_123_cast_fp16")];
            tensor<int32, [1]> v_125_axes_0 = const()[name = tensor<string, []>("v_125_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_125_cast_fp16 = expand_dims(axes = v_125_axes_0, x = v_123_cast_fp16)[name = tensor<string, []>("v_125_cast_fp16")];
            tensor<int32, [4]> var_2517_begin_0 = const()[name = tensor<string, []>("op_2517_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2517_end_0 = const()[name = tensor<string, []>("op_2517_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2517_end_mask_0 = const()[name = tensor<string, []>("op_2517_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2517_cast_fp16 = slice_by_index(begin = var_2517_begin_0, end = var_2517_end_0, end_mask = var_2517_end_mask_0, x = q_205_cast_fp16)[name = tensor<string, []>("op_2517_cast_fp16")];
            tensor<int32, [4]> var_2521_begin_0 = const()[name = tensor<string, []>("op_2521_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2521_end_0 = const()[name = tensor<string, []>("op_2521_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2521_end_mask_0 = const()[name = tensor<string, []>("op_2521_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2521_cast_fp16 = slice_by_index(begin = var_2521_begin_0, end = var_2521_end_0, end_mask = var_2521_end_mask_0, x = k_123_cast_fp16)[name = tensor<string, []>("op_2521_cast_fp16")];
            tensor<int32, [4]> var_2525_begin_0 = const()[name = tensor<string, []>("op_2525_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2525_end_0 = const()[name = tensor<string, []>("op_2525_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2525_end_mask_0 = const()[name = tensor<string, []>("op_2525_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2525_cast_fp16 = slice_by_index(begin = var_2525_begin_0, end = var_2525_end_0, end_mask = var_2525_end_mask_0, x = q_205_cast_fp16)[name = tensor<string, []>("op_2525_cast_fp16")];
            tensor<int32, [4]> var_2529_begin_0 = const()[name = tensor<string, []>("op_2529_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2529_end_0 = const()[name = tensor<string, []>("op_2529_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2529_end_mask_0 = const()[name = tensor<string, []>("op_2529_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2529_cast_fp16 = slice_by_index(begin = var_2529_begin_0, end = var_2529_end_0, end_mask = var_2529_end_mask_0, x = k_123_cast_fp16)[name = tensor<string, []>("op_2529_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2531_cast_fp16 = mul(x = var_2525_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2531_cast_fp16")];
            tensor<int32, [5]> var_2539 = const()[name = tensor<string, []>("op_2539"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_271_cast_fp16 = reshape(shape = var_2539, x = var_2525_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
            tensor<int32, [2]> var_2541_split_sizes_0 = const()[name = tensor<string, []>("op_2541_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2541_axis_0 = const()[name = tensor<string, []>("op_2541_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2541_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2541_cast_fp16_1 = split(axis = var_2541_axis_0, split_sizes = var_2541_split_sizes_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2541_cast_fp16")];
            tensor<int32, [1]> squeeze_80_axes_0 = const()[name = tensor<string, []>("squeeze_80_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_80_cast_fp16 = squeeze(axes = squeeze_80_axes_0, x = var_2541_cast_fp16_0)[name = tensor<string, []>("squeeze_80_cast_fp16")];
            tensor<int32, [1]> squeeze_81_axes_0 = const()[name = tensor<string, []>("squeeze_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_81_cast_fp16 = squeeze(axes = squeeze_81_axes_0, x = var_2541_cast_fp16_1)[name = tensor<string, []>("squeeze_81_cast_fp16")];
            tensor<fp16, []> const_331_promoted_to_fp16 = const()[name = tensor<string, []>("const_331_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2544_cast_fp16 = mul(x = squeeze_81_cast_fp16, y = const_331_promoted_to_fp16)[name = tensor<string, []>("op_2544_cast_fp16")];
            tensor<int32, []> x_273_axis_0 = const()[name = tensor<string, []>("x_273_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_273_cast_fp16 = stack(axis = x_273_axis_0, values = (var_2544_cast_fp16, squeeze_80_cast_fp16))[name = tensor<string, []>("x_273_cast_fp16")];
            tensor<int32, [4]> var_2550 = const()[name = tensor<string, []>("op_2550"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2551_cast_fp16 = reshape(shape = var_2550, x = x_273_cast_fp16)[name = tensor<string, []>("op_2551_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2552_cast_fp16 = mul(x = var_2551_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2552_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_41_cast_fp16 = add(x = var_2531_cast_fp16, y = var_2552_cast_fp16)[name = tensor<string, []>("q_patches_41_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2554_cast_fp16 = mul(x = var_2529_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2554_cast_fp16")];
            tensor<int32, [5]> var_2562 = const()[name = tensor<string, []>("op_2562"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_277_cast_fp16 = reshape(shape = var_2562, x = var_2529_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = tensor<string, []>("op_2564_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2564_axis_0 = const()[name = tensor<string, []>("op_2564_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2564_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = x_277_cast_fp16)[name = tensor<string, []>("op_2564_cast_fp16")];
            tensor<int32, [1]> squeeze_82_axes_0 = const()[name = tensor<string, []>("squeeze_82_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_82_cast_fp16 = squeeze(axes = squeeze_82_axes_0, x = var_2564_cast_fp16_0)[name = tensor<string, []>("squeeze_82_cast_fp16")];
            tensor<int32, [1]> squeeze_83_axes_0 = const()[name = tensor<string, []>("squeeze_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_83_cast_fp16 = squeeze(axes = squeeze_83_axes_0, x = var_2564_cast_fp16_1)[name = tensor<string, []>("squeeze_83_cast_fp16")];
            tensor<fp16, []> const_339_promoted_to_fp16 = const()[name = tensor<string, []>("const_339_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2567_cast_fp16 = mul(x = squeeze_83_cast_fp16, y = const_339_promoted_to_fp16)[name = tensor<string, []>("op_2567_cast_fp16")];
            tensor<int32, []> x_279_axis_0 = const()[name = tensor<string, []>("x_279_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_279_cast_fp16 = stack(axis = x_279_axis_0, values = (var_2567_cast_fp16, squeeze_82_cast_fp16))[name = tensor<string, []>("x_279_cast_fp16")];
            tensor<int32, [4]> var_2573 = const()[name = tensor<string, []>("op_2573"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2574_cast_fp16 = reshape(shape = var_2573, x = x_279_cast_fp16)[name = tensor<string, []>("op_2574_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2575_cast_fp16 = mul(x = var_2574_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2575_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_41_cast_fp16 = add(x = var_2554_cast_fp16, y = var_2575_cast_fp16)[name = tensor<string, []>("k_patches_41_cast_fp16")];
            tensor<bool, []> var_2578_interleave_0 = const()[name = tensor<string, []>("op_2578_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2578_cast_fp16 = concat(axis = var_20, interleave = var_2578_interleave_0, values = (var_2517_cast_fp16, q_patches_41_cast_fp16))[name = tensor<string, []>("op_2578_cast_fp16")];
            tensor<bool, []> var_2581_interleave_0 = const()[name = tensor<string, []>("op_2581_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2581_cast_fp16 = concat(axis = var_20, interleave = var_2581_interleave_0, values = (var_2521_cast_fp16, k_patches_41_cast_fp16))[name = tensor<string, []>("op_2581_cast_fp16")];
            tensor<fp16, []> var_2583_to_fp16 = const()[name = tensor<string, []>("op_2583_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_209_cast_fp16 = mul(x = var_2578_cast_fp16, y = var_2583_to_fp16)[name = tensor<string, []>("q_209_cast_fp16")];
            tensor<bool, []> attn_81_transpose_x_1 = const()[name = tensor<string, []>("attn_81_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_81_transpose_y_1 = const()[name = tensor<string, []>("attn_81_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_1, transpose_y = attn_81_transpose_y_1, x = q_209_cast_fp16, y = var_2581_cast_fp16)[name = tensor<string, []>("attn_81_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_83_cast_fp16 = softmax(axis = var_21, x = attn_81_cast_fp16)[name = tensor<string, []>("attn_83_cast_fp16")];
            tensor<bool, []> var_2588_transpose_x_0 = const()[name = tensor<string, []>("op_2588_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2588_transpose_y_0 = const()[name = tensor<string, []>("op_2588_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2588_cast_fp16 = matmul(transpose_x = var_2588_transpose_x_0, transpose_y = var_2588_transpose_y_0, x = attn_83_cast_fp16, y = v_125_cast_fp16)[name = tensor<string, []>("op_2588_cast_fp16")];
            tensor<int32, [1]> var_2589_axes_0 = const()[name = tensor<string, []>("op_2589_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2589_cast_fp16 = squeeze(axes = var_2589_axes_0, x = var_2588_cast_fp16)[name = tensor<string, []>("op_2589_cast_fp16")];
            tensor<string, []> var_2595_pad_type_0 = const()[name = tensor<string, []>("op_2595_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2595_strides_0 = const()[name = tensor<string, []>("op_2595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2595_pad_0 = const()[name = tensor<string, []>("op_2595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2595_dilations_0 = const()[name = tensor<string, []>("op_2595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2595_groups_0 = const()[name = tensor<string, []>("op_2595_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18340608)))];
            tensor<fp16, [1, 64, 1, 197]> var_2595_cast_fp16 = conv(dilations = var_2595_dilations_0, groups = var_2595_groups_0, pad = var_2595_pad_0, pad_type = var_2595_pad_type_0, strides = var_2595_strides_0, weight = model_blocks_1_attn_q_projs_9_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2595_cast_fp16")];
            tensor<int32, [1]> var_2596_axes_0 = const()[name = tensor<string, []>("op_2596_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2596_cast_fp16 = squeeze(axes = var_2596_axes_0, x = var_2595_cast_fp16)[name = tensor<string, []>("op_2596_cast_fp16")];
            tensor<int32, [3]> q_211_perm_0 = const()[name = tensor<string, []>("q_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2603_pad_type_0 = const()[name = tensor<string, []>("op_2603_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2603_strides_0 = const()[name = tensor<string, []>("op_2603_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2603_pad_0 = const()[name = tensor<string, []>("op_2603_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2603_dilations_0 = const()[name = tensor<string, []>("op_2603_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2603_groups_0 = const()[name = tensor<string, []>("op_2603_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18438976)))];
            tensor<fp16, [1, 64, 1, 197]> var_2603_cast_fp16 = conv(dilations = var_2603_dilations_0, groups = var_2603_groups_0, pad = var_2603_pad_0, pad_type = var_2603_pad_type_0, strides = var_2603_strides_0, weight = model_blocks_1_attn_k_projs_9_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2603_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = tensor<string, []>("op_2604_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2604_cast_fp16 = squeeze(axes = var_2604_axes_0, x = var_2603_cast_fp16)[name = tensor<string, []>("op_2604_cast_fp16")];
            tensor<int32, [3]> k_127_perm_0 = const()[name = tensor<string, []>("k_127_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2611_pad_type_0 = const()[name = tensor<string, []>("op_2611_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2611_strides_0 = const()[name = tensor<string, []>("op_2611_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2611_pad_0 = const()[name = tensor<string, []>("op_2611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2611_dilations_0 = const()[name = tensor<string, []>("op_2611_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2611_groups_0 = const()[name = tensor<string, []>("op_2611_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18537344)))];
            tensor<fp16, [1, 64, 1, 197]> var_2611_cast_fp16 = conv(dilations = var_2611_dilations_0, groups = var_2611_groups_0, pad = var_2611_pad_0, pad_type = var_2611_pad_type_0, strides = var_2611_strides_0, weight = model_blocks_1_attn_v_projs_9_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2611_cast_fp16")];
            tensor<int32, [1]> var_2612_axes_0 = const()[name = tensor<string, []>("op_2612_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2612_cast_fp16 = squeeze(axes = var_2612_axes_0, x = var_2611_cast_fp16)[name = tensor<string, []>("op_2612_cast_fp16")];
            tensor<int32, [3]> v_127_perm_0 = const()[name = tensor<string, []>("v_127_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18635712)))];
            tensor<fp16, [1, 197, 64]> q_211_cast_fp16 = transpose(perm = q_211_perm_0, x = var_2596_cast_fp16)[name = tensor<string, []>("transpose_444")];
            tensor<fp16, [1, 197, 64]> q_213_cast_fp16 = add(x = q_211_cast_fp16, y = model_blocks_1_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_213_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18635904)))];
            tensor<fp16, [1, 197, 64]> v_127_cast_fp16 = transpose(perm = v_127_perm_0, x = var_2612_cast_fp16)[name = tensor<string, []>("transpose_442")];
            tensor<fp16, [1, 197, 64]> v_129_cast_fp16 = add(x = v_127_cast_fp16, y = model_blocks_1_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_129_cast_fp16")];
            tensor<int32, [1]> q_215_axes_0 = const()[name = tensor<string, []>("q_215_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_215_cast_fp16 = expand_dims(axes = q_215_axes_0, x = q_213_cast_fp16)[name = tensor<string, []>("q_215_cast_fp16")];
            tensor<int32, [1]> k_129_axes_0 = const()[name = tensor<string, []>("k_129_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_127_cast_fp16 = transpose(perm = k_127_perm_0, x = var_2604_cast_fp16)[name = tensor<string, []>("transpose_443")];
            tensor<fp16, [1, 1, 197, 64]> k_129_cast_fp16 = expand_dims(axes = k_129_axes_0, x = k_127_cast_fp16)[name = tensor<string, []>("k_129_cast_fp16")];
            tensor<int32, [1]> v_131_axes_0 = const()[name = tensor<string, []>("v_131_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_131_cast_fp16 = expand_dims(axes = v_131_axes_0, x = v_129_cast_fp16)[name = tensor<string, []>("v_131_cast_fp16")];
            tensor<int32, [4]> var_2621_begin_0 = const()[name = tensor<string, []>("op_2621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2621_end_0 = const()[name = tensor<string, []>("op_2621_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2621_end_mask_0 = const()[name = tensor<string, []>("op_2621_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2621_cast_fp16 = slice_by_index(begin = var_2621_begin_0, end = var_2621_end_0, end_mask = var_2621_end_mask_0, x = q_215_cast_fp16)[name = tensor<string, []>("op_2621_cast_fp16")];
            tensor<int32, [4]> var_2625_begin_0 = const()[name = tensor<string, []>("op_2625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2625_end_0 = const()[name = tensor<string, []>("op_2625_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2625_end_mask_0 = const()[name = tensor<string, []>("op_2625_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2625_cast_fp16 = slice_by_index(begin = var_2625_begin_0, end = var_2625_end_0, end_mask = var_2625_end_mask_0, x = k_129_cast_fp16)[name = tensor<string, []>("op_2625_cast_fp16")];
            tensor<int32, [4]> var_2629_begin_0 = const()[name = tensor<string, []>("op_2629_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2629_end_0 = const()[name = tensor<string, []>("op_2629_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2629_end_mask_0 = const()[name = tensor<string, []>("op_2629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2629_cast_fp16 = slice_by_index(begin = var_2629_begin_0, end = var_2629_end_0, end_mask = var_2629_end_mask_0, x = q_215_cast_fp16)[name = tensor<string, []>("op_2629_cast_fp16")];
            tensor<int32, [4]> var_2633_begin_0 = const()[name = tensor<string, []>("op_2633_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2633_end_0 = const()[name = tensor<string, []>("op_2633_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2633_end_mask_0 = const()[name = tensor<string, []>("op_2633_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2633_cast_fp16 = slice_by_index(begin = var_2633_begin_0, end = var_2633_end_0, end_mask = var_2633_end_mask_0, x = k_129_cast_fp16)[name = tensor<string, []>("op_2633_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2635_cast_fp16 = mul(x = var_2629_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2635_cast_fp16")];
            tensor<int32, [5]> var_2643 = const()[name = tensor<string, []>("op_2643"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_283_cast_fp16 = reshape(shape = var_2643, x = var_2629_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
            tensor<int32, [2]> var_2645_split_sizes_0 = const()[name = tensor<string, []>("op_2645_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2645_axis_0 = const()[name = tensor<string, []>("op_2645_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2645_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2645_cast_fp16_1 = split(axis = var_2645_axis_0, split_sizes = var_2645_split_sizes_0, x = x_283_cast_fp16)[name = tensor<string, []>("op_2645_cast_fp16")];
            tensor<int32, [1]> squeeze_84_axes_0 = const()[name = tensor<string, []>("squeeze_84_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_84_cast_fp16 = squeeze(axes = squeeze_84_axes_0, x = var_2645_cast_fp16_0)[name = tensor<string, []>("squeeze_84_cast_fp16")];
            tensor<int32, [1]> squeeze_85_axes_0 = const()[name = tensor<string, []>("squeeze_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_85_cast_fp16 = squeeze(axes = squeeze_85_axes_0, x = var_2645_cast_fp16_1)[name = tensor<string, []>("squeeze_85_cast_fp16")];
            tensor<fp16, []> const_347_promoted_to_fp16 = const()[name = tensor<string, []>("const_347_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2648_cast_fp16 = mul(x = squeeze_85_cast_fp16, y = const_347_promoted_to_fp16)[name = tensor<string, []>("op_2648_cast_fp16")];
            tensor<int32, []> x_285_axis_0 = const()[name = tensor<string, []>("x_285_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_285_cast_fp16 = stack(axis = x_285_axis_0, values = (var_2648_cast_fp16, squeeze_84_cast_fp16))[name = tensor<string, []>("x_285_cast_fp16")];
            tensor<int32, [4]> var_2654 = const()[name = tensor<string, []>("op_2654"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2655_cast_fp16 = reshape(shape = var_2654, x = x_285_cast_fp16)[name = tensor<string, []>("op_2655_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2656_cast_fp16 = mul(x = var_2655_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2656_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_43_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2656_cast_fp16)[name = tensor<string, []>("q_patches_43_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2658_cast_fp16 = mul(x = var_2633_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2658_cast_fp16")];
            tensor<int32, [5]> var_2666 = const()[name = tensor<string, []>("op_2666"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_289_cast_fp16 = reshape(shape = var_2666, x = var_2633_cast_fp16)[name = tensor<string, []>("x_289_cast_fp16")];
            tensor<int32, [2]> var_2668_split_sizes_0 = const()[name = tensor<string, []>("op_2668_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2668_axis_0 = const()[name = tensor<string, []>("op_2668_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2668_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2668_cast_fp16_1 = split(axis = var_2668_axis_0, split_sizes = var_2668_split_sizes_0, x = x_289_cast_fp16)[name = tensor<string, []>("op_2668_cast_fp16")];
            tensor<int32, [1]> squeeze_86_axes_0 = const()[name = tensor<string, []>("squeeze_86_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_86_cast_fp16 = squeeze(axes = squeeze_86_axes_0, x = var_2668_cast_fp16_0)[name = tensor<string, []>("squeeze_86_cast_fp16")];
            tensor<int32, [1]> squeeze_87_axes_0 = const()[name = tensor<string, []>("squeeze_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_87_cast_fp16 = squeeze(axes = squeeze_87_axes_0, x = var_2668_cast_fp16_1)[name = tensor<string, []>("squeeze_87_cast_fp16")];
            tensor<fp16, []> const_355_promoted_to_fp16 = const()[name = tensor<string, []>("const_355_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2671_cast_fp16 = mul(x = squeeze_87_cast_fp16, y = const_355_promoted_to_fp16)[name = tensor<string, []>("op_2671_cast_fp16")];
            tensor<int32, []> x_291_axis_0 = const()[name = tensor<string, []>("x_291_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_291_cast_fp16 = stack(axis = x_291_axis_0, values = (var_2671_cast_fp16, squeeze_86_cast_fp16))[name = tensor<string, []>("x_291_cast_fp16")];
            tensor<int32, [4]> var_2677 = const()[name = tensor<string, []>("op_2677"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2678_cast_fp16 = reshape(shape = var_2677, x = x_291_cast_fp16)[name = tensor<string, []>("op_2678_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2679_cast_fp16 = mul(x = var_2678_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2679_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_43_cast_fp16 = add(x = var_2658_cast_fp16, y = var_2679_cast_fp16)[name = tensor<string, []>("k_patches_43_cast_fp16")];
            tensor<bool, []> var_2682_interleave_0 = const()[name = tensor<string, []>("op_2682_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2682_cast_fp16 = concat(axis = var_20, interleave = var_2682_interleave_0, values = (var_2621_cast_fp16, q_patches_43_cast_fp16))[name = tensor<string, []>("op_2682_cast_fp16")];
            tensor<bool, []> var_2685_interleave_0 = const()[name = tensor<string, []>("op_2685_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2685_cast_fp16 = concat(axis = var_20, interleave = var_2685_interleave_0, values = (var_2625_cast_fp16, k_patches_43_cast_fp16))[name = tensor<string, []>("op_2685_cast_fp16")];
            tensor<fp16, []> var_2687_to_fp16 = const()[name = tensor<string, []>("op_2687_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_219_cast_fp16 = mul(x = var_2682_cast_fp16, y = var_2687_to_fp16)[name = tensor<string, []>("q_219_cast_fp16")];
            tensor<bool, []> attn_85_transpose_x_1 = const()[name = tensor<string, []>("attn_85_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_85_transpose_y_1 = const()[name = tensor<string, []>("attn_85_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_1, transpose_y = attn_85_transpose_y_1, x = q_219_cast_fp16, y = var_2685_cast_fp16)[name = tensor<string, []>("attn_85_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_87_cast_fp16 = softmax(axis = var_21, x = attn_85_cast_fp16)[name = tensor<string, []>("attn_87_cast_fp16")];
            tensor<bool, []> var_2692_transpose_x_0 = const()[name = tensor<string, []>("op_2692_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2692_transpose_y_0 = const()[name = tensor<string, []>("op_2692_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2692_cast_fp16 = matmul(transpose_x = var_2692_transpose_x_0, transpose_y = var_2692_transpose_y_0, x = attn_87_cast_fp16, y = v_131_cast_fp16)[name = tensor<string, []>("op_2692_cast_fp16")];
            tensor<int32, [1]> var_2693_axes_0 = const()[name = tensor<string, []>("op_2693_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2693_cast_fp16 = squeeze(axes = var_2693_axes_0, x = var_2692_cast_fp16)[name = tensor<string, []>("op_2693_cast_fp16")];
            tensor<string, []> var_2699_pad_type_0 = const()[name = tensor<string, []>("op_2699_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2699_strides_0 = const()[name = tensor<string, []>("op_2699_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2699_pad_0 = const()[name = tensor<string, []>("op_2699_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2699_dilations_0 = const()[name = tensor<string, []>("op_2699_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2699_groups_0 = const()[name = tensor<string, []>("op_2699_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18636096)))];
            tensor<fp16, [1, 64, 1, 197]> var_2699_cast_fp16 = conv(dilations = var_2699_dilations_0, groups = var_2699_groups_0, pad = var_2699_pad_0, pad_type = var_2699_pad_type_0, strides = var_2699_strides_0, weight = model_blocks_1_attn_q_projs_10_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2699_cast_fp16")];
            tensor<int32, [1]> var_2700_axes_0 = const()[name = tensor<string, []>("op_2700_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2700_cast_fp16 = squeeze(axes = var_2700_axes_0, x = var_2699_cast_fp16)[name = tensor<string, []>("op_2700_cast_fp16")];
            tensor<int32, [3]> q_221_perm_0 = const()[name = tensor<string, []>("q_221_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2707_pad_type_0 = const()[name = tensor<string, []>("op_2707_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2707_strides_0 = const()[name = tensor<string, []>("op_2707_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2707_pad_0 = const()[name = tensor<string, []>("op_2707_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2707_dilations_0 = const()[name = tensor<string, []>("op_2707_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2707_groups_0 = const()[name = tensor<string, []>("op_2707_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18734464)))];
            tensor<fp16, [1, 64, 1, 197]> var_2707_cast_fp16 = conv(dilations = var_2707_dilations_0, groups = var_2707_groups_0, pad = var_2707_pad_0, pad_type = var_2707_pad_type_0, strides = var_2707_strides_0, weight = model_blocks_1_attn_k_projs_10_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2707_cast_fp16")];
            tensor<int32, [1]> var_2708_axes_0 = const()[name = tensor<string, []>("op_2708_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2708_cast_fp16 = squeeze(axes = var_2708_axes_0, x = var_2707_cast_fp16)[name = tensor<string, []>("op_2708_cast_fp16")];
            tensor<int32, [3]> k_133_perm_0 = const()[name = tensor<string, []>("k_133_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2715_pad_type_0 = const()[name = tensor<string, []>("op_2715_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2715_strides_0 = const()[name = tensor<string, []>("op_2715_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2715_pad_0 = const()[name = tensor<string, []>("op_2715_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2715_dilations_0 = const()[name = tensor<string, []>("op_2715_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2715_groups_0 = const()[name = tensor<string, []>("op_2715_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18832832)))];
            tensor<fp16, [1, 64, 1, 197]> var_2715_cast_fp16 = conv(dilations = var_2715_dilations_0, groups = var_2715_groups_0, pad = var_2715_pad_0, pad_type = var_2715_pad_type_0, strides = var_2715_strides_0, weight = model_blocks_1_attn_v_projs_10_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2715_cast_fp16")];
            tensor<int32, [1]> var_2716_axes_0 = const()[name = tensor<string, []>("op_2716_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2716_cast_fp16 = squeeze(axes = var_2716_axes_0, x = var_2715_cast_fp16)[name = tensor<string, []>("op_2716_cast_fp16")];
            tensor<int32, [3]> v_133_perm_0 = const()[name = tensor<string, []>("v_133_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18931200)))];
            tensor<fp16, [1, 197, 64]> q_221_cast_fp16 = transpose(perm = q_221_perm_0, x = var_2700_cast_fp16)[name = tensor<string, []>("transpose_441")];
            tensor<fp16, [1, 197, 64]> q_223_cast_fp16 = add(x = q_221_cast_fp16, y = model_blocks_1_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_223_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18931392)))];
            tensor<fp16, [1, 197, 64]> v_133_cast_fp16 = transpose(perm = v_133_perm_0, x = var_2716_cast_fp16)[name = tensor<string, []>("transpose_439")];
            tensor<fp16, [1, 197, 64]> v_135_cast_fp16 = add(x = v_133_cast_fp16, y = model_blocks_1_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_135_cast_fp16")];
            tensor<int32, [1]> q_225_axes_0 = const()[name = tensor<string, []>("q_225_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_225_cast_fp16 = expand_dims(axes = q_225_axes_0, x = q_223_cast_fp16)[name = tensor<string, []>("q_225_cast_fp16")];
            tensor<int32, [1]> k_135_axes_0 = const()[name = tensor<string, []>("k_135_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_133_cast_fp16 = transpose(perm = k_133_perm_0, x = var_2708_cast_fp16)[name = tensor<string, []>("transpose_440")];
            tensor<fp16, [1, 1, 197, 64]> k_135_cast_fp16 = expand_dims(axes = k_135_axes_0, x = k_133_cast_fp16)[name = tensor<string, []>("k_135_cast_fp16")];
            tensor<int32, [1]> v_137_axes_0 = const()[name = tensor<string, []>("v_137_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_137_cast_fp16 = expand_dims(axes = v_137_axes_0, x = v_135_cast_fp16)[name = tensor<string, []>("v_137_cast_fp16")];
            tensor<int32, [4]> var_2725_begin_0 = const()[name = tensor<string, []>("op_2725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2725_end_0 = const()[name = tensor<string, []>("op_2725_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2725_end_mask_0 = const()[name = tensor<string, []>("op_2725_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2725_cast_fp16 = slice_by_index(begin = var_2725_begin_0, end = var_2725_end_0, end_mask = var_2725_end_mask_0, x = q_225_cast_fp16)[name = tensor<string, []>("op_2725_cast_fp16")];
            tensor<int32, [4]> var_2729_begin_0 = const()[name = tensor<string, []>("op_2729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2729_end_0 = const()[name = tensor<string, []>("op_2729_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2729_end_mask_0 = const()[name = tensor<string, []>("op_2729_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2729_cast_fp16 = slice_by_index(begin = var_2729_begin_0, end = var_2729_end_0, end_mask = var_2729_end_mask_0, x = k_135_cast_fp16)[name = tensor<string, []>("op_2729_cast_fp16")];
            tensor<int32, [4]> var_2733_begin_0 = const()[name = tensor<string, []>("op_2733_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2733_end_0 = const()[name = tensor<string, []>("op_2733_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2733_end_mask_0 = const()[name = tensor<string, []>("op_2733_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2733_cast_fp16 = slice_by_index(begin = var_2733_begin_0, end = var_2733_end_0, end_mask = var_2733_end_mask_0, x = q_225_cast_fp16)[name = tensor<string, []>("op_2733_cast_fp16")];
            tensor<int32, [4]> var_2737_begin_0 = const()[name = tensor<string, []>("op_2737_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2737_end_0 = const()[name = tensor<string, []>("op_2737_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2737_end_mask_0 = const()[name = tensor<string, []>("op_2737_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2737_cast_fp16 = slice_by_index(begin = var_2737_begin_0, end = var_2737_end_0, end_mask = var_2737_end_mask_0, x = k_135_cast_fp16)[name = tensor<string, []>("op_2737_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2739_cast_fp16 = mul(x = var_2733_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2739_cast_fp16")];
            tensor<int32, [5]> var_2747 = const()[name = tensor<string, []>("op_2747"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_295_cast_fp16 = reshape(shape = var_2747, x = var_2733_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
            tensor<int32, [2]> var_2749_split_sizes_0 = const()[name = tensor<string, []>("op_2749_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2749_axis_0 = const()[name = tensor<string, []>("op_2749_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2749_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2749_cast_fp16_1 = split(axis = var_2749_axis_0, split_sizes = var_2749_split_sizes_0, x = x_295_cast_fp16)[name = tensor<string, []>("op_2749_cast_fp16")];
            tensor<int32, [1]> squeeze_88_axes_0 = const()[name = tensor<string, []>("squeeze_88_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_88_cast_fp16 = squeeze(axes = squeeze_88_axes_0, x = var_2749_cast_fp16_0)[name = tensor<string, []>("squeeze_88_cast_fp16")];
            tensor<int32, [1]> squeeze_89_axes_0 = const()[name = tensor<string, []>("squeeze_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_89_cast_fp16 = squeeze(axes = squeeze_89_axes_0, x = var_2749_cast_fp16_1)[name = tensor<string, []>("squeeze_89_cast_fp16")];
            tensor<fp16, []> const_363_promoted_to_fp16 = const()[name = tensor<string, []>("const_363_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2752_cast_fp16 = mul(x = squeeze_89_cast_fp16, y = const_363_promoted_to_fp16)[name = tensor<string, []>("op_2752_cast_fp16")];
            tensor<int32, []> x_297_axis_0 = const()[name = tensor<string, []>("x_297_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_297_cast_fp16 = stack(axis = x_297_axis_0, values = (var_2752_cast_fp16, squeeze_88_cast_fp16))[name = tensor<string, []>("x_297_cast_fp16")];
            tensor<int32, [4]> var_2758 = const()[name = tensor<string, []>("op_2758"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2759_cast_fp16 = reshape(shape = var_2758, x = x_297_cast_fp16)[name = tensor<string, []>("op_2759_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2760_cast_fp16 = mul(x = var_2759_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2760_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_45_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2760_cast_fp16)[name = tensor<string, []>("q_patches_45_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2762_cast_fp16 = mul(x = var_2737_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2762_cast_fp16")];
            tensor<int32, [5]> var_2770 = const()[name = tensor<string, []>("op_2770"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_301_cast_fp16 = reshape(shape = var_2770, x = var_2737_cast_fp16)[name = tensor<string, []>("x_301_cast_fp16")];
            tensor<int32, [2]> var_2772_split_sizes_0 = const()[name = tensor<string, []>("op_2772_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2772_axis_0 = const()[name = tensor<string, []>("op_2772_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2772_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2772_cast_fp16_1 = split(axis = var_2772_axis_0, split_sizes = var_2772_split_sizes_0, x = x_301_cast_fp16)[name = tensor<string, []>("op_2772_cast_fp16")];
            tensor<int32, [1]> squeeze_90_axes_0 = const()[name = tensor<string, []>("squeeze_90_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_90_cast_fp16 = squeeze(axes = squeeze_90_axes_0, x = var_2772_cast_fp16_0)[name = tensor<string, []>("squeeze_90_cast_fp16")];
            tensor<int32, [1]> squeeze_91_axes_0 = const()[name = tensor<string, []>("squeeze_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_91_cast_fp16 = squeeze(axes = squeeze_91_axes_0, x = var_2772_cast_fp16_1)[name = tensor<string, []>("squeeze_91_cast_fp16")];
            tensor<fp16, []> const_371_promoted_to_fp16 = const()[name = tensor<string, []>("const_371_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2775_cast_fp16 = mul(x = squeeze_91_cast_fp16, y = const_371_promoted_to_fp16)[name = tensor<string, []>("op_2775_cast_fp16")];
            tensor<int32, []> x_303_axis_0 = const()[name = tensor<string, []>("x_303_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_303_cast_fp16 = stack(axis = x_303_axis_0, values = (var_2775_cast_fp16, squeeze_90_cast_fp16))[name = tensor<string, []>("x_303_cast_fp16")];
            tensor<int32, [4]> var_2781 = const()[name = tensor<string, []>("op_2781"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2782_cast_fp16 = reshape(shape = var_2781, x = x_303_cast_fp16)[name = tensor<string, []>("op_2782_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2783_cast_fp16 = mul(x = var_2782_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2783_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_45_cast_fp16 = add(x = var_2762_cast_fp16, y = var_2783_cast_fp16)[name = tensor<string, []>("k_patches_45_cast_fp16")];
            tensor<bool, []> var_2786_interleave_0 = const()[name = tensor<string, []>("op_2786_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2786_cast_fp16 = concat(axis = var_20, interleave = var_2786_interleave_0, values = (var_2725_cast_fp16, q_patches_45_cast_fp16))[name = tensor<string, []>("op_2786_cast_fp16")];
            tensor<bool, []> var_2789_interleave_0 = const()[name = tensor<string, []>("op_2789_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2789_cast_fp16 = concat(axis = var_20, interleave = var_2789_interleave_0, values = (var_2729_cast_fp16, k_patches_45_cast_fp16))[name = tensor<string, []>("op_2789_cast_fp16")];
            tensor<fp16, []> var_2791_to_fp16 = const()[name = tensor<string, []>("op_2791_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_229_cast_fp16 = mul(x = var_2786_cast_fp16, y = var_2791_to_fp16)[name = tensor<string, []>("q_229_cast_fp16")];
            tensor<bool, []> attn_89_transpose_x_1 = const()[name = tensor<string, []>("attn_89_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_89_transpose_y_1 = const()[name = tensor<string, []>("attn_89_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_1, transpose_y = attn_89_transpose_y_1, x = q_229_cast_fp16, y = var_2789_cast_fp16)[name = tensor<string, []>("attn_89_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_91_cast_fp16 = softmax(axis = var_21, x = attn_89_cast_fp16)[name = tensor<string, []>("attn_91_cast_fp16")];
            tensor<bool, []> var_2796_transpose_x_0 = const()[name = tensor<string, []>("op_2796_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2796_transpose_y_0 = const()[name = tensor<string, []>("op_2796_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2796_cast_fp16 = matmul(transpose_x = var_2796_transpose_x_0, transpose_y = var_2796_transpose_y_0, x = attn_91_cast_fp16, y = v_137_cast_fp16)[name = tensor<string, []>("op_2796_cast_fp16")];
            tensor<int32, [1]> var_2797_axes_0 = const()[name = tensor<string, []>("op_2797_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_2797_cast_fp16 = squeeze(axes = var_2797_axes_0, x = var_2796_cast_fp16)[name = tensor<string, []>("op_2797_cast_fp16")];
            tensor<string, []> var_2803_pad_type_0 = const()[name = tensor<string, []>("op_2803_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2803_strides_0 = const()[name = tensor<string, []>("op_2803_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2803_pad_0 = const()[name = tensor<string, []>("op_2803_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2803_dilations_0 = const()[name = tensor<string, []>("op_2803_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2803_groups_0 = const()[name = tensor<string, []>("op_2803_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18931584)))];
            tensor<fp16, [1, 64, 1, 197]> var_2803_cast_fp16 = conv(dilations = var_2803_dilations_0, groups = var_2803_groups_0, pad = var_2803_pad_0, pad_type = var_2803_pad_type_0, strides = var_2803_strides_0, weight = model_blocks_1_attn_q_projs_11_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2803_cast_fp16")];
            tensor<int32, [1]> var_2804_axes_0 = const()[name = tensor<string, []>("op_2804_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2804_cast_fp16 = squeeze(axes = var_2804_axes_0, x = var_2803_cast_fp16)[name = tensor<string, []>("op_2804_cast_fp16")];
            tensor<int32, [3]> q_231_perm_0 = const()[name = tensor<string, []>("q_231_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2811_pad_type_0 = const()[name = tensor<string, []>("op_2811_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2811_strides_0 = const()[name = tensor<string, []>("op_2811_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2811_pad_0 = const()[name = tensor<string, []>("op_2811_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2811_dilations_0 = const()[name = tensor<string, []>("op_2811_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2811_groups_0 = const()[name = tensor<string, []>("op_2811_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19029952)))];
            tensor<fp16, [1, 64, 1, 197]> var_2811_cast_fp16 = conv(dilations = var_2811_dilations_0, groups = var_2811_groups_0, pad = var_2811_pad_0, pad_type = var_2811_pad_type_0, strides = var_2811_strides_0, weight = model_blocks_1_attn_k_projs_11_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2811_cast_fp16")];
            tensor<int32, [1]> var_2812_axes_0 = const()[name = tensor<string, []>("op_2812_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2812_cast_fp16 = squeeze(axes = var_2812_axes_0, x = var_2811_cast_fp16)[name = tensor<string, []>("op_2812_cast_fp16")];
            tensor<int32, [3]> k_139_perm_0 = const()[name = tensor<string, []>("k_139_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2819_pad_type_0 = const()[name = tensor<string, []>("op_2819_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2819_strides_0 = const()[name = tensor<string, []>("op_2819_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2819_pad_0 = const()[name = tensor<string, []>("op_2819_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2819_dilations_0 = const()[name = tensor<string, []>("op_2819_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2819_groups_0 = const()[name = tensor<string, []>("op_2819_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_1_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19128320)))];
            tensor<fp16, [1, 64, 1, 197]> var_2819_cast_fp16 = conv(dilations = var_2819_dilations_0, groups = var_2819_groups_0, pad = var_2819_pad_0, pad_type = var_2819_pad_type_0, strides = var_2819_strides_0, weight = model_blocks_1_attn_v_projs_11_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_2819_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = tensor<string, []>("op_2820_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_2820_cast_fp16 = squeeze(axes = var_2820_axes_0, x = var_2819_cast_fp16)[name = tensor<string, []>("op_2820_cast_fp16")];
            tensor<int32, [3]> v_139_perm_0 = const()[name = tensor<string, []>("v_139_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_1_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19226688)))];
            tensor<fp16, [1, 197, 64]> q_231_cast_fp16 = transpose(perm = q_231_perm_0, x = var_2804_cast_fp16)[name = tensor<string, []>("transpose_438")];
            tensor<fp16, [1, 197, 64]> q_233_cast_fp16 = add(x = q_231_cast_fp16, y = model_blocks_1_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_233_cast_fp16")];
            tensor<fp16, [64]> model_blocks_1_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19226880)))];
            tensor<fp16, [1, 197, 64]> v_139_cast_fp16 = transpose(perm = v_139_perm_0, x = var_2820_cast_fp16)[name = tensor<string, []>("transpose_436")];
            tensor<fp16, [1, 197, 64]> v_141_cast_fp16 = add(x = v_139_cast_fp16, y = model_blocks_1_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_141_cast_fp16")];
            tensor<int32, [1]> q_235_axes_0 = const()[name = tensor<string, []>("q_235_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_235_cast_fp16 = expand_dims(axes = q_235_axes_0, x = q_233_cast_fp16)[name = tensor<string, []>("q_235_cast_fp16")];
            tensor<int32, [1]> k_141_axes_0 = const()[name = tensor<string, []>("k_141_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_139_cast_fp16 = transpose(perm = k_139_perm_0, x = var_2812_cast_fp16)[name = tensor<string, []>("transpose_437")];
            tensor<fp16, [1, 1, 197, 64]> k_141_cast_fp16 = expand_dims(axes = k_141_axes_0, x = k_139_cast_fp16)[name = tensor<string, []>("k_141_cast_fp16")];
            tensor<int32, [1]> v_143_axes_0 = const()[name = tensor<string, []>("v_143_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_143_cast_fp16 = expand_dims(axes = v_143_axes_0, x = v_141_cast_fp16)[name = tensor<string, []>("v_143_cast_fp16")];
            tensor<int32, [4]> var_2829_begin_0 = const()[name = tensor<string, []>("op_2829_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2829_end_0 = const()[name = tensor<string, []>("op_2829_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2829_end_mask_0 = const()[name = tensor<string, []>("op_2829_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2829_cast_fp16 = slice_by_index(begin = var_2829_begin_0, end = var_2829_end_0, end_mask = var_2829_end_mask_0, x = q_235_cast_fp16)[name = tensor<string, []>("op_2829_cast_fp16")];
            tensor<int32, [4]> var_2833_begin_0 = const()[name = tensor<string, []>("op_2833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2833_end_0 = const()[name = tensor<string, []>("op_2833_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_2833_end_mask_0 = const()[name = tensor<string, []>("op_2833_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_2833_cast_fp16 = slice_by_index(begin = var_2833_begin_0, end = var_2833_end_0, end_mask = var_2833_end_mask_0, x = k_141_cast_fp16)[name = tensor<string, []>("op_2833_cast_fp16")];
            tensor<int32, [4]> var_2837_begin_0 = const()[name = tensor<string, []>("op_2837_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2837_end_0 = const()[name = tensor<string, []>("op_2837_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2837_end_mask_0 = const()[name = tensor<string, []>("op_2837_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2837_cast_fp16 = slice_by_index(begin = var_2837_begin_0, end = var_2837_end_0, end_mask = var_2837_end_mask_0, x = q_235_cast_fp16)[name = tensor<string, []>("op_2837_cast_fp16")];
            tensor<int32, [4]> var_2841_begin_0 = const()[name = tensor<string, []>("op_2841_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2841_end_0 = const()[name = tensor<string, []>("op_2841_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_2841_end_mask_0 = const()[name = tensor<string, []>("op_2841_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_2841_cast_fp16 = slice_by_index(begin = var_2841_begin_0, end = var_2841_end_0, end_mask = var_2841_end_mask_0, x = k_141_cast_fp16)[name = tensor<string, []>("op_2841_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2843_cast_fp16 = mul(x = var_2837_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2843_cast_fp16")];
            tensor<int32, [5]> var_2851 = const()[name = tensor<string, []>("op_2851"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_307_cast_fp16 = reshape(shape = var_2851, x = var_2837_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
            tensor<int32, [2]> var_2853_split_sizes_0 = const()[name = tensor<string, []>("op_2853_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2853_axis_0 = const()[name = tensor<string, []>("op_2853_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2853_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2853_cast_fp16_1 = split(axis = var_2853_axis_0, split_sizes = var_2853_split_sizes_0, x = x_307_cast_fp16)[name = tensor<string, []>("op_2853_cast_fp16")];
            tensor<int32, [1]> squeeze_92_axes_0 = const()[name = tensor<string, []>("squeeze_92_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_92_cast_fp16 = squeeze(axes = squeeze_92_axes_0, x = var_2853_cast_fp16_0)[name = tensor<string, []>("squeeze_92_cast_fp16")];
            tensor<int32, [1]> squeeze_93_axes_0 = const()[name = tensor<string, []>("squeeze_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_93_cast_fp16 = squeeze(axes = squeeze_93_axes_0, x = var_2853_cast_fp16_1)[name = tensor<string, []>("squeeze_93_cast_fp16")];
            tensor<fp16, []> const_379_promoted_to_fp16 = const()[name = tensor<string, []>("const_379_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2856_cast_fp16 = mul(x = squeeze_93_cast_fp16, y = const_379_promoted_to_fp16)[name = tensor<string, []>("op_2856_cast_fp16")];
            tensor<int32, []> x_309_axis_0 = const()[name = tensor<string, []>("x_309_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_309_cast_fp16 = stack(axis = x_309_axis_0, values = (var_2856_cast_fp16, squeeze_92_cast_fp16))[name = tensor<string, []>("x_309_cast_fp16")];
            tensor<int32, [4]> var_2862 = const()[name = tensor<string, []>("op_2862"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2863_cast_fp16 = reshape(shape = var_2862, x = x_309_cast_fp16)[name = tensor<string, []>("op_2863_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2864_cast_fp16 = mul(x = var_2863_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2864_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_47_cast_fp16 = add(x = var_2843_cast_fp16, y = var_2864_cast_fp16)[name = tensor<string, []>("q_patches_47_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2866_cast_fp16 = mul(x = var_2841_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_2866_cast_fp16")];
            tensor<int32, [5]> var_2874 = const()[name = tensor<string, []>("op_2874"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_313_cast_fp16 = reshape(shape = var_2874, x = var_2841_cast_fp16)[name = tensor<string, []>("x_313_cast_fp16")];
            tensor<int32, [2]> var_2876_split_sizes_0 = const()[name = tensor<string, []>("op_2876_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2876_axis_0 = const()[name = tensor<string, []>("op_2876_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_2876_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_2876_cast_fp16_1 = split(axis = var_2876_axis_0, split_sizes = var_2876_split_sizes_0, x = x_313_cast_fp16)[name = tensor<string, []>("op_2876_cast_fp16")];
            tensor<int32, [1]> squeeze_94_axes_0 = const()[name = tensor<string, []>("squeeze_94_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_94_cast_fp16 = squeeze(axes = squeeze_94_axes_0, x = var_2876_cast_fp16_0)[name = tensor<string, []>("squeeze_94_cast_fp16")];
            tensor<int32, [1]> squeeze_95_axes_0 = const()[name = tensor<string, []>("squeeze_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_95_cast_fp16 = squeeze(axes = squeeze_95_axes_0, x = var_2876_cast_fp16_1)[name = tensor<string, []>("squeeze_95_cast_fp16")];
            tensor<fp16, []> const_387_promoted_to_fp16 = const()[name = tensor<string, []>("const_387_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_2879_cast_fp16 = mul(x = squeeze_95_cast_fp16, y = const_387_promoted_to_fp16)[name = tensor<string, []>("op_2879_cast_fp16")];
            tensor<int32, []> x_315_axis_0 = const()[name = tensor<string, []>("x_315_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_315_cast_fp16 = stack(axis = x_315_axis_0, values = (var_2879_cast_fp16, squeeze_94_cast_fp16))[name = tensor<string, []>("x_315_cast_fp16")];
            tensor<int32, [4]> var_2885 = const()[name = tensor<string, []>("op_2885"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_2886_cast_fp16 = reshape(shape = var_2885, x = x_315_cast_fp16)[name = tensor<string, []>("op_2886_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_2887_cast_fp16 = mul(x = var_2886_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_2887_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_47_cast_fp16 = add(x = var_2866_cast_fp16, y = var_2887_cast_fp16)[name = tensor<string, []>("k_patches_47_cast_fp16")];
            tensor<bool, []> var_2890_interleave_0 = const()[name = tensor<string, []>("op_2890_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2890_cast_fp16 = concat(axis = var_20, interleave = var_2890_interleave_0, values = (var_2829_cast_fp16, q_patches_47_cast_fp16))[name = tensor<string, []>("op_2890_cast_fp16")];
            tensor<bool, []> var_2893_interleave_0 = const()[name = tensor<string, []>("op_2893_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2893_cast_fp16 = concat(axis = var_20, interleave = var_2893_interleave_0, values = (var_2833_cast_fp16, k_patches_47_cast_fp16))[name = tensor<string, []>("op_2893_cast_fp16")];
            tensor<fp16, []> var_2895_to_fp16 = const()[name = tensor<string, []>("op_2895_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_239_cast_fp16 = mul(x = var_2890_cast_fp16, y = var_2895_to_fp16)[name = tensor<string, []>("q_239_cast_fp16")];
            tensor<bool, []> attn_93_transpose_x_1 = const()[name = tensor<string, []>("attn_93_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_93_transpose_y_1 = const()[name = tensor<string, []>("attn_93_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_1, transpose_y = attn_93_transpose_y_1, x = q_239_cast_fp16, y = var_2893_cast_fp16)[name = tensor<string, []>("attn_93_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_95_cast_fp16 = softmax(axis = var_21, x = attn_93_cast_fp16)[name = tensor<string, []>("attn_95_cast_fp16")];
            tensor<bool, []> var_2900_transpose_x_0 = const()[name = tensor<string, []>("op_2900_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2900_transpose_y_0 = const()[name = tensor<string, []>("op_2900_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_2900_cast_fp16 = matmul(transpose_x = var_2900_transpose_x_0, transpose_y = var_2900_transpose_y_0, x = attn_95_cast_fp16, y = v_143_cast_fp16)[name = tensor<string, []>("op_2900_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_3_cast_fp16 = squeeze(axes = out_3_axes_0, x = var_2900_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<bool, []> input_23_interleave_0 = const()[name = tensor<string, []>("input_23_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_23_cast_fp16 = concat(axis = var_21, interleave = input_23_interleave_0, values = (var_1757_cast_fp16, var_1861_cast_fp16, var_1965_cast_fp16, var_2069_cast_fp16, var_2173_cast_fp16, var_2277_cast_fp16, var_2381_cast_fp16, var_2485_cast_fp16, var_2589_cast_fp16, var_2693_cast_fp16, var_2797_cast_fp16, out_3_cast_fp16))[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [1]> x_317_axes_0 = const()[name = tensor<string, []>("x_317_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_1_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19227072)))];
            tensor<fp16, [768]> model_blocks_1_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19228672)))];
            tensor<fp16, [1, 197, 768]> x_317_cast_fp16 = layer_norm(axes = x_317_axes_0, beta = model_blocks_1_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_1_attn_inner_attn_ln_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("x_317_cast_fp16")];
            tensor<int32, [4]> var_2908 = const()[name = tensor<string, []>("op_2908"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_319_cast_fp16 = reshape(shape = var_2908, x = x_317_cast_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
            tensor<int32, [4]> input_25_perm_0 = const()[name = tensor<string, []>("input_25_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_321_pad_type_0 = const()[name = tensor<string, []>("x_321_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_321_strides_0 = const()[name = tensor<string, []>("x_321_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_321_dilations_0 = const()[name = tensor<string, []>("x_321_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_321_groups_0 = const()[name = tensor<string, []>("x_321_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_1_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19230272)))];
            tensor<fp16, [768]> model_blocks_1_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20409984)))];
            tensor<fp16, [1, 768, 1, 197]> input_25_cast_fp16 = transpose(perm = input_25_perm_0, x = x_319_cast_fp16)[name = tensor<string, []>("transpose_435")];
            tensor<fp16, [1, 768, 1, 197]> x_321_cast_fp16 = conv(bias = model_blocks_1_attn_proj_bias_to_fp16, dilations = x_321_dilations_0, groups = x_321_groups_0, pad = x_321_pad_0, pad_type = x_321_pad_type_0, strides = x_321_strides_0, weight = model_blocks_1_attn_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
            tensor<int32, [4]> x_323_perm_0 = const()[name = tensor<string, []>("x_323_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_2919 = const()[name = tensor<string, []>("op_2919"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_323_cast_fp16 = transpose(perm = x_323_perm_0, x = x_321_cast_fp16)[name = tensor<string, []>("transpose_434")];
            tensor<fp16, [1, 197, 768]> var_2920_cast_fp16 = reshape(shape = var_2919, x = x_323_cast_fp16)[name = tensor<string, []>("op_2920_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_27_cast_fp16 = add(x = input_19_cast_fp16, y = var_2920_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<int32, [1]> x_325_axes_0 = const()[name = tensor<string, []>("x_325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20411584)))];
            tensor<fp16, [768]> model_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20413184)))];
            tensor<fp16, [1, 197, 768]> x_325_cast_fp16 = layer_norm(axes = x_325_axes_0, beta = model_blocks_1_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_1_norm2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
            tensor<int32, [4]> var_2932 = const()[name = tensor<string, []>("op_2932"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_327_cast_fp16 = reshape(shape = var_2932, x = x_325_cast_fp16)[name = tensor<string, []>("x_327_cast_fp16")];
            tensor<int32, [4]> input_29_perm_0 = const()[name = tensor<string, []>("input_29_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_31_pad_type_0 = const()[name = tensor<string, []>("input_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_31_strides_0 = const()[name = tensor<string, []>("input_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_31_pad_0 = const()[name = tensor<string, []>("input_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_31_dilations_0 = const()[name = tensor<string, []>("input_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_31_groups_0 = const()[name = tensor<string, []>("input_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_1_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20414784)))];
            tensor<fp16, [2048]> model_blocks_1_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23560576)))];
            tensor<fp16, [1, 768, 1, 197]> input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = x_327_cast_fp16)[name = tensor<string, []>("transpose_433")];
            tensor<fp16, [1, 2048, 1, 197]> input_31_cast_fp16 = conv(bias = model_blocks_1_mlp_w1_bias_to_fp16, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = model_blocks_1_mlp_w1_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<string, []> x2_99_pad_type_0 = const()[name = tensor<string, []>("x2_99_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_99_strides_0 = const()[name = tensor<string, []>("x2_99_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_99_pad_0 = const()[name = tensor<string, []>("x2_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_99_dilations_0 = const()[name = tensor<string, []>("x2_99_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_99_groups_0 = const()[name = tensor<string, []>("x2_99_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_1_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23564736)))];
            tensor<fp16, [2048]> model_blocks_1_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26710528)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_99_cast_fp16 = conv(bias = model_blocks_1_mlp_w2_bias_to_fp16, dilations = x2_99_dilations_0, groups = x2_99_groups_0, pad = x2_99_pad_0, pad_type = x2_99_pad_type_0, strides = x2_99_strides_0, weight = model_blocks_1_mlp_w2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("x2_99_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_2949_cast_fp16 = silu(x = input_31_cast_fp16)[name = tensor<string, []>("op_2949_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_9_cast_fp16 = mul(x = var_2949_cast_fp16, y = x2_99_cast_fp16)[name = tensor<string, []>("hidden_9_cast_fp16")];
            tensor<int32, [4]> hidden_11_perm_0 = const()[name = tensor<string, []>("hidden_11_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_2952 = const()[name = tensor<string, []>("op_2952"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_11_cast_fp16 = transpose(perm = hidden_11_perm_0, x = hidden_9_cast_fp16)[name = tensor<string, []>("transpose_432")];
            tensor<fp16, [1, 197, 2048]> input_33_cast_fp16 = reshape(shape = var_2952, x = hidden_11_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<int32, [1]> hidden_13_axes_0 = const()[name = tensor<string, []>("hidden_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_1_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26714688)))];
            tensor<fp16, [2048]> model_blocks_1_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26718848)))];
            tensor<fp16, [1, 197, 2048]> hidden_13_cast_fp16 = layer_norm(axes = hidden_13_axes_0, beta = model_blocks_1_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_1_mlp_ffn_ln_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("hidden_13_cast_fp16")];
            tensor<int32, [4]> var_2958 = const()[name = tensor<string, []>("op_2958"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_15_cast_fp16 = reshape(shape = var_2958, x = hidden_13_cast_fp16)[name = tensor<string, []>("hidden_15_cast_fp16")];
            tensor<int32, [4]> input_35_perm_0 = const()[name = tensor<string, []>("input_35_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_329_pad_type_0 = const()[name = tensor<string, []>("x_329_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_329_strides_0 = const()[name = tensor<string, []>("x_329_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_329_pad_0 = const()[name = tensor<string, []>("x_329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_329_dilations_0 = const()[name = tensor<string, []>("x_329_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_329_groups_0 = const()[name = tensor<string, []>("x_329_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_1_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26723008)))];
            tensor<fp16, [768]> model_blocks_1_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_1_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29868800)))];
            tensor<fp16, [1, 2048, 1, 197]> input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = hidden_15_cast_fp16)[name = tensor<string, []>("transpose_431")];
            tensor<fp16, [1, 768, 1, 197]> x_329_cast_fp16 = conv(bias = model_blocks_1_mlp_w3_bias_to_fp16, dilations = x_329_dilations_0, groups = x_329_groups_0, pad = x_329_pad_0, pad_type = x_329_pad_type_0, strides = x_329_strides_0, weight = model_blocks_1_mlp_w3_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = tensor<string, []>("x_331_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_2969 = const()[name = tensor<string, []>("op_2969"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_430")];
            tensor<fp16, [1, 197, 768]> var_2970_cast_fp16 = reshape(shape = var_2969, x = x_331_cast_fp16)[name = tensor<string, []>("op_2970_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_37_cast_fp16 = add(x = input_27_cast_fp16, y = var_2970_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<int32, [1]> x_333_axes_0 = const()[name = tensor<string, []>("x_333_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29870400)))];
            tensor<fp16, [768]> model_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29872000)))];
            tensor<fp16, [1, 197, 768]> x_333_cast_fp16 = layer_norm(axes = x_333_axes_0, beta = model_blocks_2_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_2_norm1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
            tensor<int32, [4]> var_3104 = const()[name = tensor<string, []>("op_3104"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_335_cast_fp16 = reshape(shape = var_3104, x = x_333_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
            tensor<int32, [4]> input_39_perm_0 = const()[name = tensor<string, []>("input_39_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_3112_pad_type_0 = const()[name = tensor<string, []>("op_3112_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3112_strides_0 = const()[name = tensor<string, []>("op_3112_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3112_pad_0 = const()[name = tensor<string, []>("op_3112_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3112_dilations_0 = const()[name = tensor<string, []>("op_3112_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3112_groups_0 = const()[name = tensor<string, []>("op_3112_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29873600)))];
            tensor<fp16, [1, 768, 1, 197]> input_39_cast_fp16 = transpose(perm = input_39_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_429")];
            tensor<fp16, [1, 64, 1, 197]> var_3112_cast_fp16 = conv(dilations = var_3112_dilations_0, groups = var_3112_groups_0, pad = var_3112_pad_0, pad_type = var_3112_pad_type_0, strides = var_3112_strides_0, weight = model_blocks_2_attn_q_projs_0_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3112_cast_fp16")];
            tensor<int32, [1]> var_3113_axes_0 = const()[name = tensor<string, []>("op_3113_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3113_cast_fp16 = squeeze(axes = var_3113_axes_0, x = var_3112_cast_fp16)[name = tensor<string, []>("op_3113_cast_fp16")];
            tensor<int32, [3]> q_241_perm_0 = const()[name = tensor<string, []>("q_241_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3120_pad_type_0 = const()[name = tensor<string, []>("op_3120_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3120_strides_0 = const()[name = tensor<string, []>("op_3120_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3120_pad_0 = const()[name = tensor<string, []>("op_3120_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3120_dilations_0 = const()[name = tensor<string, []>("op_3120_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3120_groups_0 = const()[name = tensor<string, []>("op_3120_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29971968)))];
            tensor<fp16, [1, 64, 1, 197]> var_3120_cast_fp16 = conv(dilations = var_3120_dilations_0, groups = var_3120_groups_0, pad = var_3120_pad_0, pad_type = var_3120_pad_type_0, strides = var_3120_strides_0, weight = model_blocks_2_attn_k_projs_0_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3121_axes_0 = const()[name = tensor<string, []>("op_3121_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3121_cast_fp16 = squeeze(axes = var_3121_axes_0, x = var_3120_cast_fp16)[name = tensor<string, []>("op_3121_cast_fp16")];
            tensor<int32, [3]> k_145_perm_0 = const()[name = tensor<string, []>("k_145_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3128_pad_type_0 = const()[name = tensor<string, []>("op_3128_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3128_strides_0 = const()[name = tensor<string, []>("op_3128_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3128_pad_0 = const()[name = tensor<string, []>("op_3128_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3128_dilations_0 = const()[name = tensor<string, []>("op_3128_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3128_groups_0 = const()[name = tensor<string, []>("op_3128_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30070336)))];
            tensor<fp16, [1, 64, 1, 197]> var_3128_cast_fp16 = conv(dilations = var_3128_dilations_0, groups = var_3128_groups_0, pad = var_3128_pad_0, pad_type = var_3128_pad_type_0, strides = var_3128_strides_0, weight = model_blocks_2_attn_v_projs_0_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3128_cast_fp16")];
            tensor<int32, [1]> var_3129_axes_0 = const()[name = tensor<string, []>("op_3129_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3129_cast_fp16 = squeeze(axes = var_3129_axes_0, x = var_3128_cast_fp16)[name = tensor<string, []>("op_3129_cast_fp16")];
            tensor<int32, [3]> v_145_perm_0 = const()[name = tensor<string, []>("v_145_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30168704)))];
            tensor<fp16, [1, 197, 64]> q_241_cast_fp16 = transpose(perm = q_241_perm_0, x = var_3113_cast_fp16)[name = tensor<string, []>("transpose_428")];
            tensor<fp16, [1, 197, 64]> q_243_cast_fp16 = add(x = q_241_cast_fp16, y = model_blocks_2_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_243_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30168896)))];
            tensor<fp16, [1, 197, 64]> v_145_cast_fp16 = transpose(perm = v_145_perm_0, x = var_3129_cast_fp16)[name = tensor<string, []>("transpose_426")];
            tensor<fp16, [1, 197, 64]> v_147_cast_fp16 = add(x = v_145_cast_fp16, y = model_blocks_2_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_147_cast_fp16")];
            tensor<int32, [1]> q_245_axes_0 = const()[name = tensor<string, []>("q_245_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_245_cast_fp16 = expand_dims(axes = q_245_axes_0, x = q_243_cast_fp16)[name = tensor<string, []>("q_245_cast_fp16")];
            tensor<int32, [1]> k_147_axes_0 = const()[name = tensor<string, []>("k_147_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_145_cast_fp16 = transpose(perm = k_145_perm_0, x = var_3121_cast_fp16)[name = tensor<string, []>("transpose_427")];
            tensor<fp16, [1, 1, 197, 64]> k_147_cast_fp16 = expand_dims(axes = k_147_axes_0, x = k_145_cast_fp16)[name = tensor<string, []>("k_147_cast_fp16")];
            tensor<int32, [1]> v_149_axes_0 = const()[name = tensor<string, []>("v_149_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_149_cast_fp16 = expand_dims(axes = v_149_axes_0, x = v_147_cast_fp16)[name = tensor<string, []>("v_149_cast_fp16")];
            tensor<int32, [4]> var_3138_begin_0 = const()[name = tensor<string, []>("op_3138_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3138_end_0 = const()[name = tensor<string, []>("op_3138_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3138_end_mask_0 = const()[name = tensor<string, []>("op_3138_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3138_cast_fp16 = slice_by_index(begin = var_3138_begin_0, end = var_3138_end_0, end_mask = var_3138_end_mask_0, x = q_245_cast_fp16)[name = tensor<string, []>("op_3138_cast_fp16")];
            tensor<int32, [4]> var_3142_begin_0 = const()[name = tensor<string, []>("op_3142_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3142_end_0 = const()[name = tensor<string, []>("op_3142_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3142_end_mask_0 = const()[name = tensor<string, []>("op_3142_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3142_cast_fp16 = slice_by_index(begin = var_3142_begin_0, end = var_3142_end_0, end_mask = var_3142_end_mask_0, x = k_147_cast_fp16)[name = tensor<string, []>("op_3142_cast_fp16")];
            tensor<int32, [4]> var_3146_begin_0 = const()[name = tensor<string, []>("op_3146_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3146_end_0 = const()[name = tensor<string, []>("op_3146_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3146_end_mask_0 = const()[name = tensor<string, []>("op_3146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3146_cast_fp16 = slice_by_index(begin = var_3146_begin_0, end = var_3146_end_0, end_mask = var_3146_end_mask_0, x = q_245_cast_fp16)[name = tensor<string, []>("op_3146_cast_fp16")];
            tensor<int32, [4]> var_3150_begin_0 = const()[name = tensor<string, []>("op_3150_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3150_end_0 = const()[name = tensor<string, []>("op_3150_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3150_end_mask_0 = const()[name = tensor<string, []>("op_3150_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3150_cast_fp16 = slice_by_index(begin = var_3150_begin_0, end = var_3150_end_0, end_mask = var_3150_end_mask_0, x = k_147_cast_fp16)[name = tensor<string, []>("op_3150_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3152_cast_fp16 = mul(x = var_3146_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3152_cast_fp16")];
            tensor<int32, [5]> var_3160 = const()[name = tensor<string, []>("op_3160"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_339_cast_fp16 = reshape(shape = var_3160, x = var_3146_cast_fp16)[name = tensor<string, []>("x_339_cast_fp16")];
            tensor<int32, [2]> var_3162_split_sizes_0 = const()[name = tensor<string, []>("op_3162_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3162_axis_0 = const()[name = tensor<string, []>("op_3162_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3162_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3162_cast_fp16_1 = split(axis = var_3162_axis_0, split_sizes = var_3162_split_sizes_0, x = x_339_cast_fp16)[name = tensor<string, []>("op_3162_cast_fp16")];
            tensor<int32, [1]> squeeze_96_axes_0 = const()[name = tensor<string, []>("squeeze_96_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_96_cast_fp16 = squeeze(axes = squeeze_96_axes_0, x = var_3162_cast_fp16_0)[name = tensor<string, []>("squeeze_96_cast_fp16")];
            tensor<int32, [1]> squeeze_97_axes_0 = const()[name = tensor<string, []>("squeeze_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_97_cast_fp16 = squeeze(axes = squeeze_97_axes_0, x = var_3162_cast_fp16_1)[name = tensor<string, []>("squeeze_97_cast_fp16")];
            tensor<fp16, []> const_399_promoted_to_fp16 = const()[name = tensor<string, []>("const_399_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3165_cast_fp16 = mul(x = squeeze_97_cast_fp16, y = const_399_promoted_to_fp16)[name = tensor<string, []>("op_3165_cast_fp16")];
            tensor<int32, []> x_341_axis_0 = const()[name = tensor<string, []>("x_341_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_341_cast_fp16 = stack(axis = x_341_axis_0, values = (var_3165_cast_fp16, squeeze_96_cast_fp16))[name = tensor<string, []>("x_341_cast_fp16")];
            tensor<int32, [4]> var_3171 = const()[name = tensor<string, []>("op_3171"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3172_cast_fp16 = reshape(shape = var_3171, x = x_341_cast_fp16)[name = tensor<string, []>("op_3172_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3173_cast_fp16 = mul(x = var_3172_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3173_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_49_cast_fp16 = add(x = var_3152_cast_fp16, y = var_3173_cast_fp16)[name = tensor<string, []>("q_patches_49_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3175_cast_fp16 = mul(x = var_3150_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3175_cast_fp16")];
            tensor<int32, [5]> var_3183 = const()[name = tensor<string, []>("op_3183"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_345_cast_fp16 = reshape(shape = var_3183, x = var_3150_cast_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
            tensor<int32, [2]> var_3185_split_sizes_0 = const()[name = tensor<string, []>("op_3185_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3185_axis_0 = const()[name = tensor<string, []>("op_3185_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3185_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3185_cast_fp16_1 = split(axis = var_3185_axis_0, split_sizes = var_3185_split_sizes_0, x = x_345_cast_fp16)[name = tensor<string, []>("op_3185_cast_fp16")];
            tensor<int32, [1]> squeeze_98_axes_0 = const()[name = tensor<string, []>("squeeze_98_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_98_cast_fp16 = squeeze(axes = squeeze_98_axes_0, x = var_3185_cast_fp16_0)[name = tensor<string, []>("squeeze_98_cast_fp16")];
            tensor<int32, [1]> squeeze_99_axes_0 = const()[name = tensor<string, []>("squeeze_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_99_cast_fp16 = squeeze(axes = squeeze_99_axes_0, x = var_3185_cast_fp16_1)[name = tensor<string, []>("squeeze_99_cast_fp16")];
            tensor<fp16, []> const_407_promoted_to_fp16 = const()[name = tensor<string, []>("const_407_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3188_cast_fp16 = mul(x = squeeze_99_cast_fp16, y = const_407_promoted_to_fp16)[name = tensor<string, []>("op_3188_cast_fp16")];
            tensor<int32, []> x_347_axis_0 = const()[name = tensor<string, []>("x_347_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_347_cast_fp16 = stack(axis = x_347_axis_0, values = (var_3188_cast_fp16, squeeze_98_cast_fp16))[name = tensor<string, []>("x_347_cast_fp16")];
            tensor<int32, [4]> var_3194 = const()[name = tensor<string, []>("op_3194"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3195_cast_fp16 = reshape(shape = var_3194, x = x_347_cast_fp16)[name = tensor<string, []>("op_3195_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3196_cast_fp16 = mul(x = var_3195_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3196_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_49_cast_fp16 = add(x = var_3175_cast_fp16, y = var_3196_cast_fp16)[name = tensor<string, []>("k_patches_49_cast_fp16")];
            tensor<bool, []> var_3199_interleave_0 = const()[name = tensor<string, []>("op_3199_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3199_cast_fp16 = concat(axis = var_20, interleave = var_3199_interleave_0, values = (var_3138_cast_fp16, q_patches_49_cast_fp16))[name = tensor<string, []>("op_3199_cast_fp16")];
            tensor<bool, []> var_3202_interleave_0 = const()[name = tensor<string, []>("op_3202_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3202_cast_fp16 = concat(axis = var_20, interleave = var_3202_interleave_0, values = (var_3142_cast_fp16, k_patches_49_cast_fp16))[name = tensor<string, []>("op_3202_cast_fp16")];
            tensor<fp16, []> var_3204_to_fp16 = const()[name = tensor<string, []>("op_3204_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_249_cast_fp16 = mul(x = var_3199_cast_fp16, y = var_3204_to_fp16)[name = tensor<string, []>("q_249_cast_fp16")];
            tensor<bool, []> attn_97_transpose_x_1 = const()[name = tensor<string, []>("attn_97_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_97_transpose_y_1 = const()[name = tensor<string, []>("attn_97_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_1, transpose_y = attn_97_transpose_y_1, x = q_249_cast_fp16, y = var_3202_cast_fp16)[name = tensor<string, []>("attn_97_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_99_cast_fp16 = softmax(axis = var_21, x = attn_97_cast_fp16)[name = tensor<string, []>("attn_99_cast_fp16")];
            tensor<bool, []> var_3209_transpose_x_0 = const()[name = tensor<string, []>("op_3209_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3209_transpose_y_0 = const()[name = tensor<string, []>("op_3209_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3209_cast_fp16 = matmul(transpose_x = var_3209_transpose_x_0, transpose_y = var_3209_transpose_y_0, x = attn_99_cast_fp16, y = v_149_cast_fp16)[name = tensor<string, []>("op_3209_cast_fp16")];
            tensor<int32, [1]> var_3210_axes_0 = const()[name = tensor<string, []>("op_3210_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3210_cast_fp16 = squeeze(axes = var_3210_axes_0, x = var_3209_cast_fp16)[name = tensor<string, []>("op_3210_cast_fp16")];
            tensor<string, []> var_3216_pad_type_0 = const()[name = tensor<string, []>("op_3216_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3216_strides_0 = const()[name = tensor<string, []>("op_3216_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3216_pad_0 = const()[name = tensor<string, []>("op_3216_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3216_dilations_0 = const()[name = tensor<string, []>("op_3216_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3216_groups_0 = const()[name = tensor<string, []>("op_3216_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30169088)))];
            tensor<fp16, [1, 64, 1, 197]> var_3216_cast_fp16 = conv(dilations = var_3216_dilations_0, groups = var_3216_groups_0, pad = var_3216_pad_0, pad_type = var_3216_pad_type_0, strides = var_3216_strides_0, weight = model_blocks_2_attn_q_projs_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3216_cast_fp16")];
            tensor<int32, [1]> var_3217_axes_0 = const()[name = tensor<string, []>("op_3217_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3217_cast_fp16 = squeeze(axes = var_3217_axes_0, x = var_3216_cast_fp16)[name = tensor<string, []>("op_3217_cast_fp16")];
            tensor<int32, [3]> q_251_perm_0 = const()[name = tensor<string, []>("q_251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3224_pad_type_0 = const()[name = tensor<string, []>("op_3224_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3224_strides_0 = const()[name = tensor<string, []>("op_3224_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3224_pad_0 = const()[name = tensor<string, []>("op_3224_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3224_dilations_0 = const()[name = tensor<string, []>("op_3224_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3224_groups_0 = const()[name = tensor<string, []>("op_3224_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30267456)))];
            tensor<fp16, [1, 64, 1, 197]> var_3224_cast_fp16 = conv(dilations = var_3224_dilations_0, groups = var_3224_groups_0, pad = var_3224_pad_0, pad_type = var_3224_pad_type_0, strides = var_3224_strides_0, weight = model_blocks_2_attn_k_projs_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3224_cast_fp16")];
            tensor<int32, [1]> var_3225_axes_0 = const()[name = tensor<string, []>("op_3225_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3225_cast_fp16 = squeeze(axes = var_3225_axes_0, x = var_3224_cast_fp16)[name = tensor<string, []>("op_3225_cast_fp16")];
            tensor<int32, [3]> k_151_perm_0 = const()[name = tensor<string, []>("k_151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3232_pad_type_0 = const()[name = tensor<string, []>("op_3232_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3232_strides_0 = const()[name = tensor<string, []>("op_3232_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3232_pad_0 = const()[name = tensor<string, []>("op_3232_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3232_dilations_0 = const()[name = tensor<string, []>("op_3232_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3232_groups_0 = const()[name = tensor<string, []>("op_3232_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30365824)))];
            tensor<fp16, [1, 64, 1, 197]> var_3232_cast_fp16 = conv(dilations = var_3232_dilations_0, groups = var_3232_groups_0, pad = var_3232_pad_0, pad_type = var_3232_pad_type_0, strides = var_3232_strides_0, weight = model_blocks_2_attn_v_projs_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3232_cast_fp16")];
            tensor<int32, [1]> var_3233_axes_0 = const()[name = tensor<string, []>("op_3233_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3233_cast_fp16 = squeeze(axes = var_3233_axes_0, x = var_3232_cast_fp16)[name = tensor<string, []>("op_3233_cast_fp16")];
            tensor<int32, [3]> v_151_perm_0 = const()[name = tensor<string, []>("v_151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30464192)))];
            tensor<fp16, [1, 197, 64]> q_251_cast_fp16 = transpose(perm = q_251_perm_0, x = var_3217_cast_fp16)[name = tensor<string, []>("transpose_425")];
            tensor<fp16, [1, 197, 64]> q_253_cast_fp16 = add(x = q_251_cast_fp16, y = model_blocks_2_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_253_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30464384)))];
            tensor<fp16, [1, 197, 64]> v_151_cast_fp16 = transpose(perm = v_151_perm_0, x = var_3233_cast_fp16)[name = tensor<string, []>("transpose_423")];
            tensor<fp16, [1, 197, 64]> v_153_cast_fp16 = add(x = v_151_cast_fp16, y = model_blocks_2_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_153_cast_fp16")];
            tensor<int32, [1]> q_255_axes_0 = const()[name = tensor<string, []>("q_255_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_255_cast_fp16 = expand_dims(axes = q_255_axes_0, x = q_253_cast_fp16)[name = tensor<string, []>("q_255_cast_fp16")];
            tensor<int32, [1]> k_153_axes_0 = const()[name = tensor<string, []>("k_153_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_151_cast_fp16 = transpose(perm = k_151_perm_0, x = var_3225_cast_fp16)[name = tensor<string, []>("transpose_424")];
            tensor<fp16, [1, 1, 197, 64]> k_153_cast_fp16 = expand_dims(axes = k_153_axes_0, x = k_151_cast_fp16)[name = tensor<string, []>("k_153_cast_fp16")];
            tensor<int32, [1]> v_155_axes_0 = const()[name = tensor<string, []>("v_155_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_155_cast_fp16 = expand_dims(axes = v_155_axes_0, x = v_153_cast_fp16)[name = tensor<string, []>("v_155_cast_fp16")];
            tensor<int32, [4]> var_3242_begin_0 = const()[name = tensor<string, []>("op_3242_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3242_end_0 = const()[name = tensor<string, []>("op_3242_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3242_end_mask_0 = const()[name = tensor<string, []>("op_3242_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3242_cast_fp16 = slice_by_index(begin = var_3242_begin_0, end = var_3242_end_0, end_mask = var_3242_end_mask_0, x = q_255_cast_fp16)[name = tensor<string, []>("op_3242_cast_fp16")];
            tensor<int32, [4]> var_3246_begin_0 = const()[name = tensor<string, []>("op_3246_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3246_end_0 = const()[name = tensor<string, []>("op_3246_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3246_end_mask_0 = const()[name = tensor<string, []>("op_3246_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3246_cast_fp16 = slice_by_index(begin = var_3246_begin_0, end = var_3246_end_0, end_mask = var_3246_end_mask_0, x = k_153_cast_fp16)[name = tensor<string, []>("op_3246_cast_fp16")];
            tensor<int32, [4]> var_3250_begin_0 = const()[name = tensor<string, []>("op_3250_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3250_end_0 = const()[name = tensor<string, []>("op_3250_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3250_end_mask_0 = const()[name = tensor<string, []>("op_3250_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3250_cast_fp16 = slice_by_index(begin = var_3250_begin_0, end = var_3250_end_0, end_mask = var_3250_end_mask_0, x = q_255_cast_fp16)[name = tensor<string, []>("op_3250_cast_fp16")];
            tensor<int32, [4]> var_3254_begin_0 = const()[name = tensor<string, []>("op_3254_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3254_end_0 = const()[name = tensor<string, []>("op_3254_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3254_end_mask_0 = const()[name = tensor<string, []>("op_3254_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3254_cast_fp16 = slice_by_index(begin = var_3254_begin_0, end = var_3254_end_0, end_mask = var_3254_end_mask_0, x = k_153_cast_fp16)[name = tensor<string, []>("op_3254_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3256_cast_fp16 = mul(x = var_3250_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3256_cast_fp16")];
            tensor<int32, [5]> var_3264 = const()[name = tensor<string, []>("op_3264"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_351_cast_fp16 = reshape(shape = var_3264, x = var_3250_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
            tensor<int32, [2]> var_3266_split_sizes_0 = const()[name = tensor<string, []>("op_3266_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3266_axis_0 = const()[name = tensor<string, []>("op_3266_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3266_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3266_cast_fp16_1 = split(axis = var_3266_axis_0, split_sizes = var_3266_split_sizes_0, x = x_351_cast_fp16)[name = tensor<string, []>("op_3266_cast_fp16")];
            tensor<int32, [1]> squeeze_100_axes_0 = const()[name = tensor<string, []>("squeeze_100_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_100_cast_fp16 = squeeze(axes = squeeze_100_axes_0, x = var_3266_cast_fp16_0)[name = tensor<string, []>("squeeze_100_cast_fp16")];
            tensor<int32, [1]> squeeze_101_axes_0 = const()[name = tensor<string, []>("squeeze_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_101_cast_fp16 = squeeze(axes = squeeze_101_axes_0, x = var_3266_cast_fp16_1)[name = tensor<string, []>("squeeze_101_cast_fp16")];
            tensor<fp16, []> const_415_promoted_to_fp16 = const()[name = tensor<string, []>("const_415_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3269_cast_fp16 = mul(x = squeeze_101_cast_fp16, y = const_415_promoted_to_fp16)[name = tensor<string, []>("op_3269_cast_fp16")];
            tensor<int32, []> x_353_axis_0 = const()[name = tensor<string, []>("x_353_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_353_cast_fp16 = stack(axis = x_353_axis_0, values = (var_3269_cast_fp16, squeeze_100_cast_fp16))[name = tensor<string, []>("x_353_cast_fp16")];
            tensor<int32, [4]> var_3275 = const()[name = tensor<string, []>("op_3275"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3276_cast_fp16 = reshape(shape = var_3275, x = x_353_cast_fp16)[name = tensor<string, []>("op_3276_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3277_cast_fp16 = mul(x = var_3276_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3277_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_51_cast_fp16 = add(x = var_3256_cast_fp16, y = var_3277_cast_fp16)[name = tensor<string, []>("q_patches_51_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3279_cast_fp16 = mul(x = var_3254_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3279_cast_fp16")];
            tensor<int32, [5]> var_3287 = const()[name = tensor<string, []>("op_3287"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_357_cast_fp16 = reshape(shape = var_3287, x = var_3254_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
            tensor<int32, [2]> var_3289_split_sizes_0 = const()[name = tensor<string, []>("op_3289_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3289_axis_0 = const()[name = tensor<string, []>("op_3289_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3289_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3289_cast_fp16_1 = split(axis = var_3289_axis_0, split_sizes = var_3289_split_sizes_0, x = x_357_cast_fp16)[name = tensor<string, []>("op_3289_cast_fp16")];
            tensor<int32, [1]> squeeze_102_axes_0 = const()[name = tensor<string, []>("squeeze_102_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_102_cast_fp16 = squeeze(axes = squeeze_102_axes_0, x = var_3289_cast_fp16_0)[name = tensor<string, []>("squeeze_102_cast_fp16")];
            tensor<int32, [1]> squeeze_103_axes_0 = const()[name = tensor<string, []>("squeeze_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_103_cast_fp16 = squeeze(axes = squeeze_103_axes_0, x = var_3289_cast_fp16_1)[name = tensor<string, []>("squeeze_103_cast_fp16")];
            tensor<fp16, []> const_423_promoted_to_fp16 = const()[name = tensor<string, []>("const_423_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3292_cast_fp16 = mul(x = squeeze_103_cast_fp16, y = const_423_promoted_to_fp16)[name = tensor<string, []>("op_3292_cast_fp16")];
            tensor<int32, []> x_359_axis_0 = const()[name = tensor<string, []>("x_359_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_359_cast_fp16 = stack(axis = x_359_axis_0, values = (var_3292_cast_fp16, squeeze_102_cast_fp16))[name = tensor<string, []>("x_359_cast_fp16")];
            tensor<int32, [4]> var_3298 = const()[name = tensor<string, []>("op_3298"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3299_cast_fp16 = reshape(shape = var_3298, x = x_359_cast_fp16)[name = tensor<string, []>("op_3299_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3300_cast_fp16 = mul(x = var_3299_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3300_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_51_cast_fp16 = add(x = var_3279_cast_fp16, y = var_3300_cast_fp16)[name = tensor<string, []>("k_patches_51_cast_fp16")];
            tensor<bool, []> var_3303_interleave_0 = const()[name = tensor<string, []>("op_3303_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3303_cast_fp16 = concat(axis = var_20, interleave = var_3303_interleave_0, values = (var_3242_cast_fp16, q_patches_51_cast_fp16))[name = tensor<string, []>("op_3303_cast_fp16")];
            tensor<bool, []> var_3306_interleave_0 = const()[name = tensor<string, []>("op_3306_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3306_cast_fp16 = concat(axis = var_20, interleave = var_3306_interleave_0, values = (var_3246_cast_fp16, k_patches_51_cast_fp16))[name = tensor<string, []>("op_3306_cast_fp16")];
            tensor<fp16, []> var_3308_to_fp16 = const()[name = tensor<string, []>("op_3308_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_259_cast_fp16 = mul(x = var_3303_cast_fp16, y = var_3308_to_fp16)[name = tensor<string, []>("q_259_cast_fp16")];
            tensor<bool, []> attn_101_transpose_x_1 = const()[name = tensor<string, []>("attn_101_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_101_transpose_y_1 = const()[name = tensor<string, []>("attn_101_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_1, transpose_y = attn_101_transpose_y_1, x = q_259_cast_fp16, y = var_3306_cast_fp16)[name = tensor<string, []>("attn_101_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_103_cast_fp16 = softmax(axis = var_21, x = attn_101_cast_fp16)[name = tensor<string, []>("attn_103_cast_fp16")];
            tensor<bool, []> var_3313_transpose_x_0 = const()[name = tensor<string, []>("op_3313_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3313_transpose_y_0 = const()[name = tensor<string, []>("op_3313_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3313_cast_fp16 = matmul(transpose_x = var_3313_transpose_x_0, transpose_y = var_3313_transpose_y_0, x = attn_103_cast_fp16, y = v_155_cast_fp16)[name = tensor<string, []>("op_3313_cast_fp16")];
            tensor<int32, [1]> var_3314_axes_0 = const()[name = tensor<string, []>("op_3314_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3314_cast_fp16 = squeeze(axes = var_3314_axes_0, x = var_3313_cast_fp16)[name = tensor<string, []>("op_3314_cast_fp16")];
            tensor<string, []> var_3320_pad_type_0 = const()[name = tensor<string, []>("op_3320_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3320_strides_0 = const()[name = tensor<string, []>("op_3320_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3320_pad_0 = const()[name = tensor<string, []>("op_3320_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3320_dilations_0 = const()[name = tensor<string, []>("op_3320_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3320_groups_0 = const()[name = tensor<string, []>("op_3320_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30464576)))];
            tensor<fp16, [1, 64, 1, 197]> var_3320_cast_fp16 = conv(dilations = var_3320_dilations_0, groups = var_3320_groups_0, pad = var_3320_pad_0, pad_type = var_3320_pad_type_0, strides = var_3320_strides_0, weight = model_blocks_2_attn_q_projs_2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3320_cast_fp16")];
            tensor<int32, [1]> var_3321_axes_0 = const()[name = tensor<string, []>("op_3321_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3321_cast_fp16 = squeeze(axes = var_3321_axes_0, x = var_3320_cast_fp16)[name = tensor<string, []>("op_3321_cast_fp16")];
            tensor<int32, [3]> q_261_perm_0 = const()[name = tensor<string, []>("q_261_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3328_pad_type_0 = const()[name = tensor<string, []>("op_3328_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3328_strides_0 = const()[name = tensor<string, []>("op_3328_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3328_pad_0 = const()[name = tensor<string, []>("op_3328_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3328_dilations_0 = const()[name = tensor<string, []>("op_3328_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3328_groups_0 = const()[name = tensor<string, []>("op_3328_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30562944)))];
            tensor<fp16, [1, 64, 1, 197]> var_3328_cast_fp16 = conv(dilations = var_3328_dilations_0, groups = var_3328_groups_0, pad = var_3328_pad_0, pad_type = var_3328_pad_type_0, strides = var_3328_strides_0, weight = model_blocks_2_attn_k_projs_2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3328_cast_fp16")];
            tensor<int32, [1]> var_3329_axes_0 = const()[name = tensor<string, []>("op_3329_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3329_cast_fp16 = squeeze(axes = var_3329_axes_0, x = var_3328_cast_fp16)[name = tensor<string, []>("op_3329_cast_fp16")];
            tensor<int32, [3]> k_157_perm_0 = const()[name = tensor<string, []>("k_157_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3336_pad_type_0 = const()[name = tensor<string, []>("op_3336_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3336_strides_0 = const()[name = tensor<string, []>("op_3336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3336_pad_0 = const()[name = tensor<string, []>("op_3336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3336_dilations_0 = const()[name = tensor<string, []>("op_3336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3336_groups_0 = const()[name = tensor<string, []>("op_3336_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30661312)))];
            tensor<fp16, [1, 64, 1, 197]> var_3336_cast_fp16 = conv(dilations = var_3336_dilations_0, groups = var_3336_groups_0, pad = var_3336_pad_0, pad_type = var_3336_pad_type_0, strides = var_3336_strides_0, weight = model_blocks_2_attn_v_projs_2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3336_cast_fp16")];
            tensor<int32, [1]> var_3337_axes_0 = const()[name = tensor<string, []>("op_3337_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3337_cast_fp16 = squeeze(axes = var_3337_axes_0, x = var_3336_cast_fp16)[name = tensor<string, []>("op_3337_cast_fp16")];
            tensor<int32, [3]> v_157_perm_0 = const()[name = tensor<string, []>("v_157_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30759680)))];
            tensor<fp16, [1, 197, 64]> q_261_cast_fp16 = transpose(perm = q_261_perm_0, x = var_3321_cast_fp16)[name = tensor<string, []>("transpose_422")];
            tensor<fp16, [1, 197, 64]> q_263_cast_fp16 = add(x = q_261_cast_fp16, y = model_blocks_2_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_263_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30759872)))];
            tensor<fp16, [1, 197, 64]> v_157_cast_fp16 = transpose(perm = v_157_perm_0, x = var_3337_cast_fp16)[name = tensor<string, []>("transpose_420")];
            tensor<fp16, [1, 197, 64]> v_159_cast_fp16 = add(x = v_157_cast_fp16, y = model_blocks_2_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_159_cast_fp16")];
            tensor<int32, [1]> q_265_axes_0 = const()[name = tensor<string, []>("q_265_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_265_cast_fp16 = expand_dims(axes = q_265_axes_0, x = q_263_cast_fp16)[name = tensor<string, []>("q_265_cast_fp16")];
            tensor<int32, [1]> k_159_axes_0 = const()[name = tensor<string, []>("k_159_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_157_cast_fp16 = transpose(perm = k_157_perm_0, x = var_3329_cast_fp16)[name = tensor<string, []>("transpose_421")];
            tensor<fp16, [1, 1, 197, 64]> k_159_cast_fp16 = expand_dims(axes = k_159_axes_0, x = k_157_cast_fp16)[name = tensor<string, []>("k_159_cast_fp16")];
            tensor<int32, [1]> v_161_axes_0 = const()[name = tensor<string, []>("v_161_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_161_cast_fp16 = expand_dims(axes = v_161_axes_0, x = v_159_cast_fp16)[name = tensor<string, []>("v_161_cast_fp16")];
            tensor<int32, [4]> var_3346_begin_0 = const()[name = tensor<string, []>("op_3346_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3346_end_0 = const()[name = tensor<string, []>("op_3346_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3346_end_mask_0 = const()[name = tensor<string, []>("op_3346_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3346_cast_fp16 = slice_by_index(begin = var_3346_begin_0, end = var_3346_end_0, end_mask = var_3346_end_mask_0, x = q_265_cast_fp16)[name = tensor<string, []>("op_3346_cast_fp16")];
            tensor<int32, [4]> var_3350_begin_0 = const()[name = tensor<string, []>("op_3350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3350_end_0 = const()[name = tensor<string, []>("op_3350_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3350_end_mask_0 = const()[name = tensor<string, []>("op_3350_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3350_cast_fp16 = slice_by_index(begin = var_3350_begin_0, end = var_3350_end_0, end_mask = var_3350_end_mask_0, x = k_159_cast_fp16)[name = tensor<string, []>("op_3350_cast_fp16")];
            tensor<int32, [4]> var_3354_begin_0 = const()[name = tensor<string, []>("op_3354_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3354_end_0 = const()[name = tensor<string, []>("op_3354_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3354_end_mask_0 = const()[name = tensor<string, []>("op_3354_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3354_cast_fp16 = slice_by_index(begin = var_3354_begin_0, end = var_3354_end_0, end_mask = var_3354_end_mask_0, x = q_265_cast_fp16)[name = tensor<string, []>("op_3354_cast_fp16")];
            tensor<int32, [4]> var_3358_begin_0 = const()[name = tensor<string, []>("op_3358_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3358_end_0 = const()[name = tensor<string, []>("op_3358_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3358_end_mask_0 = const()[name = tensor<string, []>("op_3358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3358_cast_fp16 = slice_by_index(begin = var_3358_begin_0, end = var_3358_end_0, end_mask = var_3358_end_mask_0, x = k_159_cast_fp16)[name = tensor<string, []>("op_3358_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3360_cast_fp16 = mul(x = var_3354_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3360_cast_fp16")];
            tensor<int32, [5]> var_3368 = const()[name = tensor<string, []>("op_3368"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_363_cast_fp16 = reshape(shape = var_3368, x = var_3354_cast_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
            tensor<int32, [2]> var_3370_split_sizes_0 = const()[name = tensor<string, []>("op_3370_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3370_axis_0 = const()[name = tensor<string, []>("op_3370_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3370_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3370_cast_fp16_1 = split(axis = var_3370_axis_0, split_sizes = var_3370_split_sizes_0, x = x_363_cast_fp16)[name = tensor<string, []>("op_3370_cast_fp16")];
            tensor<int32, [1]> squeeze_104_axes_0 = const()[name = tensor<string, []>("squeeze_104_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_104_cast_fp16 = squeeze(axes = squeeze_104_axes_0, x = var_3370_cast_fp16_0)[name = tensor<string, []>("squeeze_104_cast_fp16")];
            tensor<int32, [1]> squeeze_105_axes_0 = const()[name = tensor<string, []>("squeeze_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_105_cast_fp16 = squeeze(axes = squeeze_105_axes_0, x = var_3370_cast_fp16_1)[name = tensor<string, []>("squeeze_105_cast_fp16")];
            tensor<fp16, []> const_431_promoted_to_fp16 = const()[name = tensor<string, []>("const_431_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3373_cast_fp16 = mul(x = squeeze_105_cast_fp16, y = const_431_promoted_to_fp16)[name = tensor<string, []>("op_3373_cast_fp16")];
            tensor<int32, []> x_365_axis_0 = const()[name = tensor<string, []>("x_365_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_365_cast_fp16 = stack(axis = x_365_axis_0, values = (var_3373_cast_fp16, squeeze_104_cast_fp16))[name = tensor<string, []>("x_365_cast_fp16")];
            tensor<int32, [4]> var_3379 = const()[name = tensor<string, []>("op_3379"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3380_cast_fp16 = reshape(shape = var_3379, x = x_365_cast_fp16)[name = tensor<string, []>("op_3380_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3381_cast_fp16 = mul(x = var_3380_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3381_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_53_cast_fp16 = add(x = var_3360_cast_fp16, y = var_3381_cast_fp16)[name = tensor<string, []>("q_patches_53_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3383_cast_fp16 = mul(x = var_3358_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3383_cast_fp16")];
            tensor<int32, [5]> var_3391 = const()[name = tensor<string, []>("op_3391"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_369_cast_fp16 = reshape(shape = var_3391, x = var_3358_cast_fp16)[name = tensor<string, []>("x_369_cast_fp16")];
            tensor<int32, [2]> var_3393_split_sizes_0 = const()[name = tensor<string, []>("op_3393_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3393_axis_0 = const()[name = tensor<string, []>("op_3393_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3393_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3393_cast_fp16_1 = split(axis = var_3393_axis_0, split_sizes = var_3393_split_sizes_0, x = x_369_cast_fp16)[name = tensor<string, []>("op_3393_cast_fp16")];
            tensor<int32, [1]> squeeze_106_axes_0 = const()[name = tensor<string, []>("squeeze_106_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_106_cast_fp16 = squeeze(axes = squeeze_106_axes_0, x = var_3393_cast_fp16_0)[name = tensor<string, []>("squeeze_106_cast_fp16")];
            tensor<int32, [1]> squeeze_107_axes_0 = const()[name = tensor<string, []>("squeeze_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_107_cast_fp16 = squeeze(axes = squeeze_107_axes_0, x = var_3393_cast_fp16_1)[name = tensor<string, []>("squeeze_107_cast_fp16")];
            tensor<fp16, []> const_439_promoted_to_fp16 = const()[name = tensor<string, []>("const_439_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3396_cast_fp16 = mul(x = squeeze_107_cast_fp16, y = const_439_promoted_to_fp16)[name = tensor<string, []>("op_3396_cast_fp16")];
            tensor<int32, []> x_371_axis_0 = const()[name = tensor<string, []>("x_371_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_371_cast_fp16 = stack(axis = x_371_axis_0, values = (var_3396_cast_fp16, squeeze_106_cast_fp16))[name = tensor<string, []>("x_371_cast_fp16")];
            tensor<int32, [4]> var_3402 = const()[name = tensor<string, []>("op_3402"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3403_cast_fp16 = reshape(shape = var_3402, x = x_371_cast_fp16)[name = tensor<string, []>("op_3403_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3404_cast_fp16 = mul(x = var_3403_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3404_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_53_cast_fp16 = add(x = var_3383_cast_fp16, y = var_3404_cast_fp16)[name = tensor<string, []>("k_patches_53_cast_fp16")];
            tensor<bool, []> var_3407_interleave_0 = const()[name = tensor<string, []>("op_3407_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3407_cast_fp16 = concat(axis = var_20, interleave = var_3407_interleave_0, values = (var_3346_cast_fp16, q_patches_53_cast_fp16))[name = tensor<string, []>("op_3407_cast_fp16")];
            tensor<bool, []> var_3410_interleave_0 = const()[name = tensor<string, []>("op_3410_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3410_cast_fp16 = concat(axis = var_20, interleave = var_3410_interleave_0, values = (var_3350_cast_fp16, k_patches_53_cast_fp16))[name = tensor<string, []>("op_3410_cast_fp16")];
            tensor<fp16, []> var_3412_to_fp16 = const()[name = tensor<string, []>("op_3412_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_269_cast_fp16 = mul(x = var_3407_cast_fp16, y = var_3412_to_fp16)[name = tensor<string, []>("q_269_cast_fp16")];
            tensor<bool, []> attn_105_transpose_x_1 = const()[name = tensor<string, []>("attn_105_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_105_transpose_y_1 = const()[name = tensor<string, []>("attn_105_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_1, transpose_y = attn_105_transpose_y_1, x = q_269_cast_fp16, y = var_3410_cast_fp16)[name = tensor<string, []>("attn_105_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_107_cast_fp16 = softmax(axis = var_21, x = attn_105_cast_fp16)[name = tensor<string, []>("attn_107_cast_fp16")];
            tensor<bool, []> var_3417_transpose_x_0 = const()[name = tensor<string, []>("op_3417_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3417_transpose_y_0 = const()[name = tensor<string, []>("op_3417_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3417_cast_fp16 = matmul(transpose_x = var_3417_transpose_x_0, transpose_y = var_3417_transpose_y_0, x = attn_107_cast_fp16, y = v_161_cast_fp16)[name = tensor<string, []>("op_3417_cast_fp16")];
            tensor<int32, [1]> var_3418_axes_0 = const()[name = tensor<string, []>("op_3418_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3418_cast_fp16 = squeeze(axes = var_3418_axes_0, x = var_3417_cast_fp16)[name = tensor<string, []>("op_3418_cast_fp16")];
            tensor<string, []> var_3424_pad_type_0 = const()[name = tensor<string, []>("op_3424_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3424_strides_0 = const()[name = tensor<string, []>("op_3424_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3424_pad_0 = const()[name = tensor<string, []>("op_3424_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3424_dilations_0 = const()[name = tensor<string, []>("op_3424_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3424_groups_0 = const()[name = tensor<string, []>("op_3424_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30760064)))];
            tensor<fp16, [1, 64, 1, 197]> var_3424_cast_fp16 = conv(dilations = var_3424_dilations_0, groups = var_3424_groups_0, pad = var_3424_pad_0, pad_type = var_3424_pad_type_0, strides = var_3424_strides_0, weight = model_blocks_2_attn_q_projs_3_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3424_cast_fp16")];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = tensor<string, []>("op_3425_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3425_cast_fp16 = squeeze(axes = var_3425_axes_0, x = var_3424_cast_fp16)[name = tensor<string, []>("op_3425_cast_fp16")];
            tensor<int32, [3]> q_271_perm_0 = const()[name = tensor<string, []>("q_271_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3432_pad_type_0 = const()[name = tensor<string, []>("op_3432_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3432_strides_0 = const()[name = tensor<string, []>("op_3432_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3432_pad_0 = const()[name = tensor<string, []>("op_3432_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3432_dilations_0 = const()[name = tensor<string, []>("op_3432_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3432_groups_0 = const()[name = tensor<string, []>("op_3432_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30858432)))];
            tensor<fp16, [1, 64, 1, 197]> var_3432_cast_fp16 = conv(dilations = var_3432_dilations_0, groups = var_3432_groups_0, pad = var_3432_pad_0, pad_type = var_3432_pad_type_0, strides = var_3432_strides_0, weight = model_blocks_2_attn_k_projs_3_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3432_cast_fp16")];
            tensor<int32, [1]> var_3433_axes_0 = const()[name = tensor<string, []>("op_3433_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3433_cast_fp16 = squeeze(axes = var_3433_axes_0, x = var_3432_cast_fp16)[name = tensor<string, []>("op_3433_cast_fp16")];
            tensor<int32, [3]> k_163_perm_0 = const()[name = tensor<string, []>("k_163_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3440_pad_type_0 = const()[name = tensor<string, []>("op_3440_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3440_strides_0 = const()[name = tensor<string, []>("op_3440_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3440_pad_0 = const()[name = tensor<string, []>("op_3440_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3440_dilations_0 = const()[name = tensor<string, []>("op_3440_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3440_groups_0 = const()[name = tensor<string, []>("op_3440_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30956800)))];
            tensor<fp16, [1, 64, 1, 197]> var_3440_cast_fp16 = conv(dilations = var_3440_dilations_0, groups = var_3440_groups_0, pad = var_3440_pad_0, pad_type = var_3440_pad_type_0, strides = var_3440_strides_0, weight = model_blocks_2_attn_v_projs_3_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3440_cast_fp16")];
            tensor<int32, [1]> var_3441_axes_0 = const()[name = tensor<string, []>("op_3441_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3441_cast_fp16 = squeeze(axes = var_3441_axes_0, x = var_3440_cast_fp16)[name = tensor<string, []>("op_3441_cast_fp16")];
            tensor<int32, [3]> v_163_perm_0 = const()[name = tensor<string, []>("v_163_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31055168)))];
            tensor<fp16, [1, 197, 64]> q_271_cast_fp16 = transpose(perm = q_271_perm_0, x = var_3425_cast_fp16)[name = tensor<string, []>("transpose_419")];
            tensor<fp16, [1, 197, 64]> q_273_cast_fp16 = add(x = q_271_cast_fp16, y = model_blocks_2_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_273_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31055360)))];
            tensor<fp16, [1, 197, 64]> v_163_cast_fp16 = transpose(perm = v_163_perm_0, x = var_3441_cast_fp16)[name = tensor<string, []>("transpose_417")];
            tensor<fp16, [1, 197, 64]> v_165_cast_fp16 = add(x = v_163_cast_fp16, y = model_blocks_2_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_165_cast_fp16")];
            tensor<int32, [1]> q_275_axes_0 = const()[name = tensor<string, []>("q_275_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_275_cast_fp16 = expand_dims(axes = q_275_axes_0, x = q_273_cast_fp16)[name = tensor<string, []>("q_275_cast_fp16")];
            tensor<int32, [1]> k_165_axes_0 = const()[name = tensor<string, []>("k_165_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_163_cast_fp16 = transpose(perm = k_163_perm_0, x = var_3433_cast_fp16)[name = tensor<string, []>("transpose_418")];
            tensor<fp16, [1, 1, 197, 64]> k_165_cast_fp16 = expand_dims(axes = k_165_axes_0, x = k_163_cast_fp16)[name = tensor<string, []>("k_165_cast_fp16")];
            tensor<int32, [1]> v_167_axes_0 = const()[name = tensor<string, []>("v_167_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_167_cast_fp16 = expand_dims(axes = v_167_axes_0, x = v_165_cast_fp16)[name = tensor<string, []>("v_167_cast_fp16")];
            tensor<int32, [4]> var_3450_begin_0 = const()[name = tensor<string, []>("op_3450_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3450_end_0 = const()[name = tensor<string, []>("op_3450_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3450_end_mask_0 = const()[name = tensor<string, []>("op_3450_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3450_cast_fp16 = slice_by_index(begin = var_3450_begin_0, end = var_3450_end_0, end_mask = var_3450_end_mask_0, x = q_275_cast_fp16)[name = tensor<string, []>("op_3450_cast_fp16")];
            tensor<int32, [4]> var_3454_begin_0 = const()[name = tensor<string, []>("op_3454_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3454_end_0 = const()[name = tensor<string, []>("op_3454_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3454_end_mask_0 = const()[name = tensor<string, []>("op_3454_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3454_cast_fp16 = slice_by_index(begin = var_3454_begin_0, end = var_3454_end_0, end_mask = var_3454_end_mask_0, x = k_165_cast_fp16)[name = tensor<string, []>("op_3454_cast_fp16")];
            tensor<int32, [4]> var_3458_begin_0 = const()[name = tensor<string, []>("op_3458_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3458_end_0 = const()[name = tensor<string, []>("op_3458_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3458_end_mask_0 = const()[name = tensor<string, []>("op_3458_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3458_cast_fp16 = slice_by_index(begin = var_3458_begin_0, end = var_3458_end_0, end_mask = var_3458_end_mask_0, x = q_275_cast_fp16)[name = tensor<string, []>("op_3458_cast_fp16")];
            tensor<int32, [4]> var_3462_begin_0 = const()[name = tensor<string, []>("op_3462_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3462_end_0 = const()[name = tensor<string, []>("op_3462_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3462_end_mask_0 = const()[name = tensor<string, []>("op_3462_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3462_cast_fp16 = slice_by_index(begin = var_3462_begin_0, end = var_3462_end_0, end_mask = var_3462_end_mask_0, x = k_165_cast_fp16)[name = tensor<string, []>("op_3462_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3464_cast_fp16 = mul(x = var_3458_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3464_cast_fp16")];
            tensor<int32, [5]> var_3472 = const()[name = tensor<string, []>("op_3472"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_375_cast_fp16 = reshape(shape = var_3472, x = var_3458_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
            tensor<int32, [2]> var_3474_split_sizes_0 = const()[name = tensor<string, []>("op_3474_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3474_axis_0 = const()[name = tensor<string, []>("op_3474_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3474_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3474_cast_fp16_1 = split(axis = var_3474_axis_0, split_sizes = var_3474_split_sizes_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3474_cast_fp16")];
            tensor<int32, [1]> squeeze_108_axes_0 = const()[name = tensor<string, []>("squeeze_108_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_108_cast_fp16 = squeeze(axes = squeeze_108_axes_0, x = var_3474_cast_fp16_0)[name = tensor<string, []>("squeeze_108_cast_fp16")];
            tensor<int32, [1]> squeeze_109_axes_0 = const()[name = tensor<string, []>("squeeze_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_109_cast_fp16 = squeeze(axes = squeeze_109_axes_0, x = var_3474_cast_fp16_1)[name = tensor<string, []>("squeeze_109_cast_fp16")];
            tensor<fp16, []> const_447_promoted_to_fp16 = const()[name = tensor<string, []>("const_447_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3477_cast_fp16 = mul(x = squeeze_109_cast_fp16, y = const_447_promoted_to_fp16)[name = tensor<string, []>("op_3477_cast_fp16")];
            tensor<int32, []> x_377_axis_0 = const()[name = tensor<string, []>("x_377_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_377_cast_fp16 = stack(axis = x_377_axis_0, values = (var_3477_cast_fp16, squeeze_108_cast_fp16))[name = tensor<string, []>("x_377_cast_fp16")];
            tensor<int32, [4]> var_3483 = const()[name = tensor<string, []>("op_3483"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3484_cast_fp16 = reshape(shape = var_3483, x = x_377_cast_fp16)[name = tensor<string, []>("op_3484_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3485_cast_fp16 = mul(x = var_3484_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3485_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_55_cast_fp16 = add(x = var_3464_cast_fp16, y = var_3485_cast_fp16)[name = tensor<string, []>("q_patches_55_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3487_cast_fp16 = mul(x = var_3462_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3487_cast_fp16")];
            tensor<int32, [5]> var_3495 = const()[name = tensor<string, []>("op_3495"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_381_cast_fp16 = reshape(shape = var_3495, x = var_3462_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
            tensor<int32, [2]> var_3497_split_sizes_0 = const()[name = tensor<string, []>("op_3497_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3497_axis_0 = const()[name = tensor<string, []>("op_3497_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3497_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3497_cast_fp16_1 = split(axis = var_3497_axis_0, split_sizes = var_3497_split_sizes_0, x = x_381_cast_fp16)[name = tensor<string, []>("op_3497_cast_fp16")];
            tensor<int32, [1]> squeeze_110_axes_0 = const()[name = tensor<string, []>("squeeze_110_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_110_cast_fp16 = squeeze(axes = squeeze_110_axes_0, x = var_3497_cast_fp16_0)[name = tensor<string, []>("squeeze_110_cast_fp16")];
            tensor<int32, [1]> squeeze_111_axes_0 = const()[name = tensor<string, []>("squeeze_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_111_cast_fp16 = squeeze(axes = squeeze_111_axes_0, x = var_3497_cast_fp16_1)[name = tensor<string, []>("squeeze_111_cast_fp16")];
            tensor<fp16, []> const_455_promoted_to_fp16 = const()[name = tensor<string, []>("const_455_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3500_cast_fp16 = mul(x = squeeze_111_cast_fp16, y = const_455_promoted_to_fp16)[name = tensor<string, []>("op_3500_cast_fp16")];
            tensor<int32, []> x_383_axis_0 = const()[name = tensor<string, []>("x_383_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_383_cast_fp16 = stack(axis = x_383_axis_0, values = (var_3500_cast_fp16, squeeze_110_cast_fp16))[name = tensor<string, []>("x_383_cast_fp16")];
            tensor<int32, [4]> var_3506 = const()[name = tensor<string, []>("op_3506"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3507_cast_fp16 = reshape(shape = var_3506, x = x_383_cast_fp16)[name = tensor<string, []>("op_3507_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3508_cast_fp16 = mul(x = var_3507_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3508_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_55_cast_fp16 = add(x = var_3487_cast_fp16, y = var_3508_cast_fp16)[name = tensor<string, []>("k_patches_55_cast_fp16")];
            tensor<bool, []> var_3511_interleave_0 = const()[name = tensor<string, []>("op_3511_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3511_cast_fp16 = concat(axis = var_20, interleave = var_3511_interleave_0, values = (var_3450_cast_fp16, q_patches_55_cast_fp16))[name = tensor<string, []>("op_3511_cast_fp16")];
            tensor<bool, []> var_3514_interleave_0 = const()[name = tensor<string, []>("op_3514_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3514_cast_fp16 = concat(axis = var_20, interleave = var_3514_interleave_0, values = (var_3454_cast_fp16, k_patches_55_cast_fp16))[name = tensor<string, []>("op_3514_cast_fp16")];
            tensor<fp16, []> var_3516_to_fp16 = const()[name = tensor<string, []>("op_3516_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_279_cast_fp16 = mul(x = var_3511_cast_fp16, y = var_3516_to_fp16)[name = tensor<string, []>("q_279_cast_fp16")];
            tensor<bool, []> attn_109_transpose_x_1 = const()[name = tensor<string, []>("attn_109_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_109_transpose_y_1 = const()[name = tensor<string, []>("attn_109_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_1, transpose_y = attn_109_transpose_y_1, x = q_279_cast_fp16, y = var_3514_cast_fp16)[name = tensor<string, []>("attn_109_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_111_cast_fp16 = softmax(axis = var_21, x = attn_109_cast_fp16)[name = tensor<string, []>("attn_111_cast_fp16")];
            tensor<bool, []> var_3521_transpose_x_0 = const()[name = tensor<string, []>("op_3521_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3521_transpose_y_0 = const()[name = tensor<string, []>("op_3521_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3521_cast_fp16 = matmul(transpose_x = var_3521_transpose_x_0, transpose_y = var_3521_transpose_y_0, x = attn_111_cast_fp16, y = v_167_cast_fp16)[name = tensor<string, []>("op_3521_cast_fp16")];
            tensor<int32, [1]> var_3522_axes_0 = const()[name = tensor<string, []>("op_3522_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3522_cast_fp16 = squeeze(axes = var_3522_axes_0, x = var_3521_cast_fp16)[name = tensor<string, []>("op_3522_cast_fp16")];
            tensor<string, []> var_3528_pad_type_0 = const()[name = tensor<string, []>("op_3528_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3528_strides_0 = const()[name = tensor<string, []>("op_3528_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3528_pad_0 = const()[name = tensor<string, []>("op_3528_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3528_dilations_0 = const()[name = tensor<string, []>("op_3528_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3528_groups_0 = const()[name = tensor<string, []>("op_3528_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31055552)))];
            tensor<fp16, [1, 64, 1, 197]> var_3528_cast_fp16 = conv(dilations = var_3528_dilations_0, groups = var_3528_groups_0, pad = var_3528_pad_0, pad_type = var_3528_pad_type_0, strides = var_3528_strides_0, weight = model_blocks_2_attn_q_projs_4_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3528_cast_fp16")];
            tensor<int32, [1]> var_3529_axes_0 = const()[name = tensor<string, []>("op_3529_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3529_cast_fp16 = squeeze(axes = var_3529_axes_0, x = var_3528_cast_fp16)[name = tensor<string, []>("op_3529_cast_fp16")];
            tensor<int32, [3]> q_281_perm_0 = const()[name = tensor<string, []>("q_281_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3536_pad_type_0 = const()[name = tensor<string, []>("op_3536_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3536_strides_0 = const()[name = tensor<string, []>("op_3536_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3536_pad_0 = const()[name = tensor<string, []>("op_3536_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3536_dilations_0 = const()[name = tensor<string, []>("op_3536_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3536_groups_0 = const()[name = tensor<string, []>("op_3536_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31153920)))];
            tensor<fp16, [1, 64, 1, 197]> var_3536_cast_fp16 = conv(dilations = var_3536_dilations_0, groups = var_3536_groups_0, pad = var_3536_pad_0, pad_type = var_3536_pad_type_0, strides = var_3536_strides_0, weight = model_blocks_2_attn_k_projs_4_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3536_cast_fp16")];
            tensor<int32, [1]> var_3537_axes_0 = const()[name = tensor<string, []>("op_3537_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3537_cast_fp16 = squeeze(axes = var_3537_axes_0, x = var_3536_cast_fp16)[name = tensor<string, []>("op_3537_cast_fp16")];
            tensor<int32, [3]> k_169_perm_0 = const()[name = tensor<string, []>("k_169_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3544_pad_type_0 = const()[name = tensor<string, []>("op_3544_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3544_strides_0 = const()[name = tensor<string, []>("op_3544_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3544_pad_0 = const()[name = tensor<string, []>("op_3544_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3544_dilations_0 = const()[name = tensor<string, []>("op_3544_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3544_groups_0 = const()[name = tensor<string, []>("op_3544_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31252288)))];
            tensor<fp16, [1, 64, 1, 197]> var_3544_cast_fp16 = conv(dilations = var_3544_dilations_0, groups = var_3544_groups_0, pad = var_3544_pad_0, pad_type = var_3544_pad_type_0, strides = var_3544_strides_0, weight = model_blocks_2_attn_v_projs_4_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3544_cast_fp16")];
            tensor<int32, [1]> var_3545_axes_0 = const()[name = tensor<string, []>("op_3545_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3545_cast_fp16 = squeeze(axes = var_3545_axes_0, x = var_3544_cast_fp16)[name = tensor<string, []>("op_3545_cast_fp16")];
            tensor<int32, [3]> v_169_perm_0 = const()[name = tensor<string, []>("v_169_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31350656)))];
            tensor<fp16, [1, 197, 64]> q_281_cast_fp16 = transpose(perm = q_281_perm_0, x = var_3529_cast_fp16)[name = tensor<string, []>("transpose_416")];
            tensor<fp16, [1, 197, 64]> q_283_cast_fp16 = add(x = q_281_cast_fp16, y = model_blocks_2_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_283_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31350848)))];
            tensor<fp16, [1, 197, 64]> v_169_cast_fp16 = transpose(perm = v_169_perm_0, x = var_3545_cast_fp16)[name = tensor<string, []>("transpose_414")];
            tensor<fp16, [1, 197, 64]> v_171_cast_fp16 = add(x = v_169_cast_fp16, y = model_blocks_2_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_171_cast_fp16")];
            tensor<int32, [1]> q_285_axes_0 = const()[name = tensor<string, []>("q_285_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_285_cast_fp16 = expand_dims(axes = q_285_axes_0, x = q_283_cast_fp16)[name = tensor<string, []>("q_285_cast_fp16")];
            tensor<int32, [1]> k_171_axes_0 = const()[name = tensor<string, []>("k_171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_169_cast_fp16 = transpose(perm = k_169_perm_0, x = var_3537_cast_fp16)[name = tensor<string, []>("transpose_415")];
            tensor<fp16, [1, 1, 197, 64]> k_171_cast_fp16 = expand_dims(axes = k_171_axes_0, x = k_169_cast_fp16)[name = tensor<string, []>("k_171_cast_fp16")];
            tensor<int32, [1]> v_173_axes_0 = const()[name = tensor<string, []>("v_173_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_173_cast_fp16 = expand_dims(axes = v_173_axes_0, x = v_171_cast_fp16)[name = tensor<string, []>("v_173_cast_fp16")];
            tensor<int32, [4]> var_3554_begin_0 = const()[name = tensor<string, []>("op_3554_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3554_end_0 = const()[name = tensor<string, []>("op_3554_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3554_end_mask_0 = const()[name = tensor<string, []>("op_3554_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3554_cast_fp16 = slice_by_index(begin = var_3554_begin_0, end = var_3554_end_0, end_mask = var_3554_end_mask_0, x = q_285_cast_fp16)[name = tensor<string, []>("op_3554_cast_fp16")];
            tensor<int32, [4]> var_3558_begin_0 = const()[name = tensor<string, []>("op_3558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3558_end_0 = const()[name = tensor<string, []>("op_3558_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3558_end_mask_0 = const()[name = tensor<string, []>("op_3558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3558_cast_fp16 = slice_by_index(begin = var_3558_begin_0, end = var_3558_end_0, end_mask = var_3558_end_mask_0, x = k_171_cast_fp16)[name = tensor<string, []>("op_3558_cast_fp16")];
            tensor<int32, [4]> var_3562_begin_0 = const()[name = tensor<string, []>("op_3562_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3562_end_0 = const()[name = tensor<string, []>("op_3562_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3562_end_mask_0 = const()[name = tensor<string, []>("op_3562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3562_cast_fp16 = slice_by_index(begin = var_3562_begin_0, end = var_3562_end_0, end_mask = var_3562_end_mask_0, x = q_285_cast_fp16)[name = tensor<string, []>("op_3562_cast_fp16")];
            tensor<int32, [4]> var_3566_begin_0 = const()[name = tensor<string, []>("op_3566_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3566_end_0 = const()[name = tensor<string, []>("op_3566_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3566_end_mask_0 = const()[name = tensor<string, []>("op_3566_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3566_cast_fp16 = slice_by_index(begin = var_3566_begin_0, end = var_3566_end_0, end_mask = var_3566_end_mask_0, x = k_171_cast_fp16)[name = tensor<string, []>("op_3566_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3568_cast_fp16 = mul(x = var_3562_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3568_cast_fp16")];
            tensor<int32, [5]> var_3576 = const()[name = tensor<string, []>("op_3576"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_387_cast_fp16 = reshape(shape = var_3576, x = var_3562_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
            tensor<int32, [2]> var_3578_split_sizes_0 = const()[name = tensor<string, []>("op_3578_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3578_axis_0 = const()[name = tensor<string, []>("op_3578_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3578_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3578_cast_fp16_1 = split(axis = var_3578_axis_0, split_sizes = var_3578_split_sizes_0, x = x_387_cast_fp16)[name = tensor<string, []>("op_3578_cast_fp16")];
            tensor<int32, [1]> squeeze_112_axes_0 = const()[name = tensor<string, []>("squeeze_112_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_112_cast_fp16 = squeeze(axes = squeeze_112_axes_0, x = var_3578_cast_fp16_0)[name = tensor<string, []>("squeeze_112_cast_fp16")];
            tensor<int32, [1]> squeeze_113_axes_0 = const()[name = tensor<string, []>("squeeze_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_113_cast_fp16 = squeeze(axes = squeeze_113_axes_0, x = var_3578_cast_fp16_1)[name = tensor<string, []>("squeeze_113_cast_fp16")];
            tensor<fp16, []> const_463_promoted_to_fp16 = const()[name = tensor<string, []>("const_463_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3581_cast_fp16 = mul(x = squeeze_113_cast_fp16, y = const_463_promoted_to_fp16)[name = tensor<string, []>("op_3581_cast_fp16")];
            tensor<int32, []> x_389_axis_0 = const()[name = tensor<string, []>("x_389_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_389_cast_fp16 = stack(axis = x_389_axis_0, values = (var_3581_cast_fp16, squeeze_112_cast_fp16))[name = tensor<string, []>("x_389_cast_fp16")];
            tensor<int32, [4]> var_3587 = const()[name = tensor<string, []>("op_3587"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3588_cast_fp16 = reshape(shape = var_3587, x = x_389_cast_fp16)[name = tensor<string, []>("op_3588_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3589_cast_fp16 = mul(x = var_3588_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3589_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_57_cast_fp16 = add(x = var_3568_cast_fp16, y = var_3589_cast_fp16)[name = tensor<string, []>("q_patches_57_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3591_cast_fp16 = mul(x = var_3566_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3591_cast_fp16")];
            tensor<int32, [5]> var_3599 = const()[name = tensor<string, []>("op_3599"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_393_cast_fp16 = reshape(shape = var_3599, x = var_3566_cast_fp16)[name = tensor<string, []>("x_393_cast_fp16")];
            tensor<int32, [2]> var_3601_split_sizes_0 = const()[name = tensor<string, []>("op_3601_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3601_axis_0 = const()[name = tensor<string, []>("op_3601_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3601_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3601_cast_fp16_1 = split(axis = var_3601_axis_0, split_sizes = var_3601_split_sizes_0, x = x_393_cast_fp16)[name = tensor<string, []>("op_3601_cast_fp16")];
            tensor<int32, [1]> squeeze_114_axes_0 = const()[name = tensor<string, []>("squeeze_114_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_114_cast_fp16 = squeeze(axes = squeeze_114_axes_0, x = var_3601_cast_fp16_0)[name = tensor<string, []>("squeeze_114_cast_fp16")];
            tensor<int32, [1]> squeeze_115_axes_0 = const()[name = tensor<string, []>("squeeze_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_115_cast_fp16 = squeeze(axes = squeeze_115_axes_0, x = var_3601_cast_fp16_1)[name = tensor<string, []>("squeeze_115_cast_fp16")];
            tensor<fp16, []> const_471_promoted_to_fp16 = const()[name = tensor<string, []>("const_471_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3604_cast_fp16 = mul(x = squeeze_115_cast_fp16, y = const_471_promoted_to_fp16)[name = tensor<string, []>("op_3604_cast_fp16")];
            tensor<int32, []> x_395_axis_0 = const()[name = tensor<string, []>("x_395_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_395_cast_fp16 = stack(axis = x_395_axis_0, values = (var_3604_cast_fp16, squeeze_114_cast_fp16))[name = tensor<string, []>("x_395_cast_fp16")];
            tensor<int32, [4]> var_3610 = const()[name = tensor<string, []>("op_3610"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3611_cast_fp16 = reshape(shape = var_3610, x = x_395_cast_fp16)[name = tensor<string, []>("op_3611_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3612_cast_fp16 = mul(x = var_3611_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3612_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_57_cast_fp16 = add(x = var_3591_cast_fp16, y = var_3612_cast_fp16)[name = tensor<string, []>("k_patches_57_cast_fp16")];
            tensor<bool, []> var_3615_interleave_0 = const()[name = tensor<string, []>("op_3615_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3615_cast_fp16 = concat(axis = var_20, interleave = var_3615_interleave_0, values = (var_3554_cast_fp16, q_patches_57_cast_fp16))[name = tensor<string, []>("op_3615_cast_fp16")];
            tensor<bool, []> var_3618_interleave_0 = const()[name = tensor<string, []>("op_3618_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3618_cast_fp16 = concat(axis = var_20, interleave = var_3618_interleave_0, values = (var_3558_cast_fp16, k_patches_57_cast_fp16))[name = tensor<string, []>("op_3618_cast_fp16")];
            tensor<fp16, []> var_3620_to_fp16 = const()[name = tensor<string, []>("op_3620_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_289_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3620_to_fp16)[name = tensor<string, []>("q_289_cast_fp16")];
            tensor<bool, []> attn_113_transpose_x_1 = const()[name = tensor<string, []>("attn_113_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_113_transpose_y_1 = const()[name = tensor<string, []>("attn_113_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_1, transpose_y = attn_113_transpose_y_1, x = q_289_cast_fp16, y = var_3618_cast_fp16)[name = tensor<string, []>("attn_113_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_115_cast_fp16 = softmax(axis = var_21, x = attn_113_cast_fp16)[name = tensor<string, []>("attn_115_cast_fp16")];
            tensor<bool, []> var_3625_transpose_x_0 = const()[name = tensor<string, []>("op_3625_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3625_transpose_y_0 = const()[name = tensor<string, []>("op_3625_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3625_cast_fp16 = matmul(transpose_x = var_3625_transpose_x_0, transpose_y = var_3625_transpose_y_0, x = attn_115_cast_fp16, y = v_173_cast_fp16)[name = tensor<string, []>("op_3625_cast_fp16")];
            tensor<int32, [1]> var_3626_axes_0 = const()[name = tensor<string, []>("op_3626_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3626_cast_fp16 = squeeze(axes = var_3626_axes_0, x = var_3625_cast_fp16)[name = tensor<string, []>("op_3626_cast_fp16")];
            tensor<string, []> var_3632_pad_type_0 = const()[name = tensor<string, []>("op_3632_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3632_strides_0 = const()[name = tensor<string, []>("op_3632_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3632_pad_0 = const()[name = tensor<string, []>("op_3632_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3632_dilations_0 = const()[name = tensor<string, []>("op_3632_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3632_groups_0 = const()[name = tensor<string, []>("op_3632_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31351040)))];
            tensor<fp16, [1, 64, 1, 197]> var_3632_cast_fp16 = conv(dilations = var_3632_dilations_0, groups = var_3632_groups_0, pad = var_3632_pad_0, pad_type = var_3632_pad_type_0, strides = var_3632_strides_0, weight = model_blocks_2_attn_q_projs_5_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3632_cast_fp16")];
            tensor<int32, [1]> var_3633_axes_0 = const()[name = tensor<string, []>("op_3633_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3633_cast_fp16 = squeeze(axes = var_3633_axes_0, x = var_3632_cast_fp16)[name = tensor<string, []>("op_3633_cast_fp16")];
            tensor<int32, [3]> q_291_perm_0 = const()[name = tensor<string, []>("q_291_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3640_pad_type_0 = const()[name = tensor<string, []>("op_3640_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3640_strides_0 = const()[name = tensor<string, []>("op_3640_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3640_pad_0 = const()[name = tensor<string, []>("op_3640_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3640_dilations_0 = const()[name = tensor<string, []>("op_3640_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3640_groups_0 = const()[name = tensor<string, []>("op_3640_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31449408)))];
            tensor<fp16, [1, 64, 1, 197]> var_3640_cast_fp16 = conv(dilations = var_3640_dilations_0, groups = var_3640_groups_0, pad = var_3640_pad_0, pad_type = var_3640_pad_type_0, strides = var_3640_strides_0, weight = model_blocks_2_attn_k_projs_5_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3640_cast_fp16")];
            tensor<int32, [1]> var_3641_axes_0 = const()[name = tensor<string, []>("op_3641_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3641_cast_fp16 = squeeze(axes = var_3641_axes_0, x = var_3640_cast_fp16)[name = tensor<string, []>("op_3641_cast_fp16")];
            tensor<int32, [3]> k_175_perm_0 = const()[name = tensor<string, []>("k_175_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3648_pad_type_0 = const()[name = tensor<string, []>("op_3648_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3648_strides_0 = const()[name = tensor<string, []>("op_3648_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3648_pad_0 = const()[name = tensor<string, []>("op_3648_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3648_dilations_0 = const()[name = tensor<string, []>("op_3648_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3648_groups_0 = const()[name = tensor<string, []>("op_3648_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31547776)))];
            tensor<fp16, [1, 64, 1, 197]> var_3648_cast_fp16 = conv(dilations = var_3648_dilations_0, groups = var_3648_groups_0, pad = var_3648_pad_0, pad_type = var_3648_pad_type_0, strides = var_3648_strides_0, weight = model_blocks_2_attn_v_projs_5_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3648_cast_fp16")];
            tensor<int32, [1]> var_3649_axes_0 = const()[name = tensor<string, []>("op_3649_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3649_cast_fp16 = squeeze(axes = var_3649_axes_0, x = var_3648_cast_fp16)[name = tensor<string, []>("op_3649_cast_fp16")];
            tensor<int32, [3]> v_175_perm_0 = const()[name = tensor<string, []>("v_175_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31646144)))];
            tensor<fp16, [1, 197, 64]> q_291_cast_fp16 = transpose(perm = q_291_perm_0, x = var_3633_cast_fp16)[name = tensor<string, []>("transpose_413")];
            tensor<fp16, [1, 197, 64]> q_293_cast_fp16 = add(x = q_291_cast_fp16, y = model_blocks_2_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_293_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31646336)))];
            tensor<fp16, [1, 197, 64]> v_175_cast_fp16 = transpose(perm = v_175_perm_0, x = var_3649_cast_fp16)[name = tensor<string, []>("transpose_411")];
            tensor<fp16, [1, 197, 64]> v_177_cast_fp16 = add(x = v_175_cast_fp16, y = model_blocks_2_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_177_cast_fp16")];
            tensor<int32, [1]> q_295_axes_0 = const()[name = tensor<string, []>("q_295_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_295_cast_fp16 = expand_dims(axes = q_295_axes_0, x = q_293_cast_fp16)[name = tensor<string, []>("q_295_cast_fp16")];
            tensor<int32, [1]> k_177_axes_0 = const()[name = tensor<string, []>("k_177_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_175_cast_fp16 = transpose(perm = k_175_perm_0, x = var_3641_cast_fp16)[name = tensor<string, []>("transpose_412")];
            tensor<fp16, [1, 1, 197, 64]> k_177_cast_fp16 = expand_dims(axes = k_177_axes_0, x = k_175_cast_fp16)[name = tensor<string, []>("k_177_cast_fp16")];
            tensor<int32, [1]> v_179_axes_0 = const()[name = tensor<string, []>("v_179_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_179_cast_fp16 = expand_dims(axes = v_179_axes_0, x = v_177_cast_fp16)[name = tensor<string, []>("v_179_cast_fp16")];
            tensor<int32, [4]> var_3658_begin_0 = const()[name = tensor<string, []>("op_3658_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3658_end_0 = const()[name = tensor<string, []>("op_3658_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3658_end_mask_0 = const()[name = tensor<string, []>("op_3658_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3658_cast_fp16 = slice_by_index(begin = var_3658_begin_0, end = var_3658_end_0, end_mask = var_3658_end_mask_0, x = q_295_cast_fp16)[name = tensor<string, []>("op_3658_cast_fp16")];
            tensor<int32, [4]> var_3662_begin_0 = const()[name = tensor<string, []>("op_3662_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3662_end_0 = const()[name = tensor<string, []>("op_3662_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3662_end_mask_0 = const()[name = tensor<string, []>("op_3662_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3662_cast_fp16 = slice_by_index(begin = var_3662_begin_0, end = var_3662_end_0, end_mask = var_3662_end_mask_0, x = k_177_cast_fp16)[name = tensor<string, []>("op_3662_cast_fp16")];
            tensor<int32, [4]> var_3666_begin_0 = const()[name = tensor<string, []>("op_3666_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3666_end_0 = const()[name = tensor<string, []>("op_3666_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3666_end_mask_0 = const()[name = tensor<string, []>("op_3666_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3666_cast_fp16 = slice_by_index(begin = var_3666_begin_0, end = var_3666_end_0, end_mask = var_3666_end_mask_0, x = q_295_cast_fp16)[name = tensor<string, []>("op_3666_cast_fp16")];
            tensor<int32, [4]> var_3670_begin_0 = const()[name = tensor<string, []>("op_3670_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3670_end_0 = const()[name = tensor<string, []>("op_3670_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3670_end_mask_0 = const()[name = tensor<string, []>("op_3670_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3670_cast_fp16 = slice_by_index(begin = var_3670_begin_0, end = var_3670_end_0, end_mask = var_3670_end_mask_0, x = k_177_cast_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3672_cast_fp16 = mul(x = var_3666_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3672_cast_fp16")];
            tensor<int32, [5]> var_3680 = const()[name = tensor<string, []>("op_3680"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_399_cast_fp16 = reshape(shape = var_3680, x = var_3666_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
            tensor<int32, [2]> var_3682_split_sizes_0 = const()[name = tensor<string, []>("op_3682_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3682_axis_0 = const()[name = tensor<string, []>("op_3682_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3682_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3682_cast_fp16_1 = split(axis = var_3682_axis_0, split_sizes = var_3682_split_sizes_0, x = x_399_cast_fp16)[name = tensor<string, []>("op_3682_cast_fp16")];
            tensor<int32, [1]> squeeze_116_axes_0 = const()[name = tensor<string, []>("squeeze_116_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_116_cast_fp16 = squeeze(axes = squeeze_116_axes_0, x = var_3682_cast_fp16_0)[name = tensor<string, []>("squeeze_116_cast_fp16")];
            tensor<int32, [1]> squeeze_117_axes_0 = const()[name = tensor<string, []>("squeeze_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_117_cast_fp16 = squeeze(axes = squeeze_117_axes_0, x = var_3682_cast_fp16_1)[name = tensor<string, []>("squeeze_117_cast_fp16")];
            tensor<fp16, []> const_479_promoted_to_fp16 = const()[name = tensor<string, []>("const_479_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3685_cast_fp16 = mul(x = squeeze_117_cast_fp16, y = const_479_promoted_to_fp16)[name = tensor<string, []>("op_3685_cast_fp16")];
            tensor<int32, []> x_401_axis_0 = const()[name = tensor<string, []>("x_401_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_401_cast_fp16 = stack(axis = x_401_axis_0, values = (var_3685_cast_fp16, squeeze_116_cast_fp16))[name = tensor<string, []>("x_401_cast_fp16")];
            tensor<int32, [4]> var_3691 = const()[name = tensor<string, []>("op_3691"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3692_cast_fp16 = reshape(shape = var_3691, x = x_401_cast_fp16)[name = tensor<string, []>("op_3692_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3693_cast_fp16 = mul(x = var_3692_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3693_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_59_cast_fp16 = add(x = var_3672_cast_fp16, y = var_3693_cast_fp16)[name = tensor<string, []>("q_patches_59_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3695_cast_fp16 = mul(x = var_3670_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3695_cast_fp16")];
            tensor<int32, [5]> var_3703 = const()[name = tensor<string, []>("op_3703"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_405_cast_fp16 = reshape(shape = var_3703, x = var_3670_cast_fp16)[name = tensor<string, []>("x_405_cast_fp16")];
            tensor<int32, [2]> var_3705_split_sizes_0 = const()[name = tensor<string, []>("op_3705_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3705_axis_0 = const()[name = tensor<string, []>("op_3705_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3705_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3705_cast_fp16_1 = split(axis = var_3705_axis_0, split_sizes = var_3705_split_sizes_0, x = x_405_cast_fp16)[name = tensor<string, []>("op_3705_cast_fp16")];
            tensor<int32, [1]> squeeze_118_axes_0 = const()[name = tensor<string, []>("squeeze_118_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_118_cast_fp16 = squeeze(axes = squeeze_118_axes_0, x = var_3705_cast_fp16_0)[name = tensor<string, []>("squeeze_118_cast_fp16")];
            tensor<int32, [1]> squeeze_119_axes_0 = const()[name = tensor<string, []>("squeeze_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_119_cast_fp16 = squeeze(axes = squeeze_119_axes_0, x = var_3705_cast_fp16_1)[name = tensor<string, []>("squeeze_119_cast_fp16")];
            tensor<fp16, []> const_487_promoted_to_fp16 = const()[name = tensor<string, []>("const_487_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3708_cast_fp16 = mul(x = squeeze_119_cast_fp16, y = const_487_promoted_to_fp16)[name = tensor<string, []>("op_3708_cast_fp16")];
            tensor<int32, []> x_407_axis_0 = const()[name = tensor<string, []>("x_407_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_407_cast_fp16 = stack(axis = x_407_axis_0, values = (var_3708_cast_fp16, squeeze_118_cast_fp16))[name = tensor<string, []>("x_407_cast_fp16")];
            tensor<int32, [4]> var_3714 = const()[name = tensor<string, []>("op_3714"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3715_cast_fp16 = reshape(shape = var_3714, x = x_407_cast_fp16)[name = tensor<string, []>("op_3715_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3716_cast_fp16 = mul(x = var_3715_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3716_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_59_cast_fp16 = add(x = var_3695_cast_fp16, y = var_3716_cast_fp16)[name = tensor<string, []>("k_patches_59_cast_fp16")];
            tensor<bool, []> var_3719_interleave_0 = const()[name = tensor<string, []>("op_3719_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3719_cast_fp16 = concat(axis = var_20, interleave = var_3719_interleave_0, values = (var_3658_cast_fp16, q_patches_59_cast_fp16))[name = tensor<string, []>("op_3719_cast_fp16")];
            tensor<bool, []> var_3722_interleave_0 = const()[name = tensor<string, []>("op_3722_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3722_cast_fp16 = concat(axis = var_20, interleave = var_3722_interleave_0, values = (var_3662_cast_fp16, k_patches_59_cast_fp16))[name = tensor<string, []>("op_3722_cast_fp16")];
            tensor<fp16, []> var_3724_to_fp16 = const()[name = tensor<string, []>("op_3724_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_299_cast_fp16 = mul(x = var_3719_cast_fp16, y = var_3724_to_fp16)[name = tensor<string, []>("q_299_cast_fp16")];
            tensor<bool, []> attn_117_transpose_x_1 = const()[name = tensor<string, []>("attn_117_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_117_transpose_y_1 = const()[name = tensor<string, []>("attn_117_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_1, transpose_y = attn_117_transpose_y_1, x = q_299_cast_fp16, y = var_3722_cast_fp16)[name = tensor<string, []>("attn_117_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_119_cast_fp16 = softmax(axis = var_21, x = attn_117_cast_fp16)[name = tensor<string, []>("attn_119_cast_fp16")];
            tensor<bool, []> var_3729_transpose_x_0 = const()[name = tensor<string, []>("op_3729_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3729_transpose_y_0 = const()[name = tensor<string, []>("op_3729_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3729_cast_fp16 = matmul(transpose_x = var_3729_transpose_x_0, transpose_y = var_3729_transpose_y_0, x = attn_119_cast_fp16, y = v_179_cast_fp16)[name = tensor<string, []>("op_3729_cast_fp16")];
            tensor<int32, [1]> var_3730_axes_0 = const()[name = tensor<string, []>("op_3730_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3730_cast_fp16 = squeeze(axes = var_3730_axes_0, x = var_3729_cast_fp16)[name = tensor<string, []>("op_3730_cast_fp16")];
            tensor<string, []> var_3736_pad_type_0 = const()[name = tensor<string, []>("op_3736_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3736_strides_0 = const()[name = tensor<string, []>("op_3736_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3736_pad_0 = const()[name = tensor<string, []>("op_3736_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3736_dilations_0 = const()[name = tensor<string, []>("op_3736_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3736_groups_0 = const()[name = tensor<string, []>("op_3736_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31646528)))];
            tensor<fp16, [1, 64, 1, 197]> var_3736_cast_fp16 = conv(dilations = var_3736_dilations_0, groups = var_3736_groups_0, pad = var_3736_pad_0, pad_type = var_3736_pad_type_0, strides = var_3736_strides_0, weight = model_blocks_2_attn_q_projs_6_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3736_cast_fp16")];
            tensor<int32, [1]> var_3737_axes_0 = const()[name = tensor<string, []>("op_3737_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3737_cast_fp16 = squeeze(axes = var_3737_axes_0, x = var_3736_cast_fp16)[name = tensor<string, []>("op_3737_cast_fp16")];
            tensor<int32, [3]> q_301_perm_0 = const()[name = tensor<string, []>("q_301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3744_pad_type_0 = const()[name = tensor<string, []>("op_3744_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3744_strides_0 = const()[name = tensor<string, []>("op_3744_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3744_pad_0 = const()[name = tensor<string, []>("op_3744_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3744_dilations_0 = const()[name = tensor<string, []>("op_3744_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3744_groups_0 = const()[name = tensor<string, []>("op_3744_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31744896)))];
            tensor<fp16, [1, 64, 1, 197]> var_3744_cast_fp16 = conv(dilations = var_3744_dilations_0, groups = var_3744_groups_0, pad = var_3744_pad_0, pad_type = var_3744_pad_type_0, strides = var_3744_strides_0, weight = model_blocks_2_attn_k_projs_6_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3744_cast_fp16")];
            tensor<int32, [1]> var_3745_axes_0 = const()[name = tensor<string, []>("op_3745_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3745_cast_fp16 = squeeze(axes = var_3745_axes_0, x = var_3744_cast_fp16)[name = tensor<string, []>("op_3745_cast_fp16")];
            tensor<int32, [3]> k_181_perm_0 = const()[name = tensor<string, []>("k_181_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3752_pad_type_0 = const()[name = tensor<string, []>("op_3752_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3752_strides_0 = const()[name = tensor<string, []>("op_3752_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3752_pad_0 = const()[name = tensor<string, []>("op_3752_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3752_dilations_0 = const()[name = tensor<string, []>("op_3752_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3752_groups_0 = const()[name = tensor<string, []>("op_3752_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31843264)))];
            tensor<fp16, [1, 64, 1, 197]> var_3752_cast_fp16 = conv(dilations = var_3752_dilations_0, groups = var_3752_groups_0, pad = var_3752_pad_0, pad_type = var_3752_pad_type_0, strides = var_3752_strides_0, weight = model_blocks_2_attn_v_projs_6_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3752_cast_fp16")];
            tensor<int32, [1]> var_3753_axes_0 = const()[name = tensor<string, []>("op_3753_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3753_cast_fp16 = squeeze(axes = var_3753_axes_0, x = var_3752_cast_fp16)[name = tensor<string, []>("op_3753_cast_fp16")];
            tensor<int32, [3]> v_181_perm_0 = const()[name = tensor<string, []>("v_181_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31941632)))];
            tensor<fp16, [1, 197, 64]> q_301_cast_fp16 = transpose(perm = q_301_perm_0, x = var_3737_cast_fp16)[name = tensor<string, []>("transpose_410")];
            tensor<fp16, [1, 197, 64]> q_303_cast_fp16 = add(x = q_301_cast_fp16, y = model_blocks_2_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_303_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31941824)))];
            tensor<fp16, [1, 197, 64]> v_181_cast_fp16 = transpose(perm = v_181_perm_0, x = var_3753_cast_fp16)[name = tensor<string, []>("transpose_408")];
            tensor<fp16, [1, 197, 64]> v_183_cast_fp16 = add(x = v_181_cast_fp16, y = model_blocks_2_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_183_cast_fp16")];
            tensor<int32, [1]> q_305_axes_0 = const()[name = tensor<string, []>("q_305_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_305_cast_fp16 = expand_dims(axes = q_305_axes_0, x = q_303_cast_fp16)[name = tensor<string, []>("q_305_cast_fp16")];
            tensor<int32, [1]> k_183_axes_0 = const()[name = tensor<string, []>("k_183_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_181_cast_fp16 = transpose(perm = k_181_perm_0, x = var_3745_cast_fp16)[name = tensor<string, []>("transpose_409")];
            tensor<fp16, [1, 1, 197, 64]> k_183_cast_fp16 = expand_dims(axes = k_183_axes_0, x = k_181_cast_fp16)[name = tensor<string, []>("k_183_cast_fp16")];
            tensor<int32, [1]> v_185_axes_0 = const()[name = tensor<string, []>("v_185_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_185_cast_fp16 = expand_dims(axes = v_185_axes_0, x = v_183_cast_fp16)[name = tensor<string, []>("v_185_cast_fp16")];
            tensor<int32, [4]> var_3762_begin_0 = const()[name = tensor<string, []>("op_3762_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3762_end_0 = const()[name = tensor<string, []>("op_3762_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3762_end_mask_0 = const()[name = tensor<string, []>("op_3762_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3762_cast_fp16 = slice_by_index(begin = var_3762_begin_0, end = var_3762_end_0, end_mask = var_3762_end_mask_0, x = q_305_cast_fp16)[name = tensor<string, []>("op_3762_cast_fp16")];
            tensor<int32, [4]> var_3766_begin_0 = const()[name = tensor<string, []>("op_3766_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3766_end_0 = const()[name = tensor<string, []>("op_3766_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3766_end_mask_0 = const()[name = tensor<string, []>("op_3766_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3766_cast_fp16 = slice_by_index(begin = var_3766_begin_0, end = var_3766_end_0, end_mask = var_3766_end_mask_0, x = k_183_cast_fp16)[name = tensor<string, []>("op_3766_cast_fp16")];
            tensor<int32, [4]> var_3770_begin_0 = const()[name = tensor<string, []>("op_3770_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3770_end_0 = const()[name = tensor<string, []>("op_3770_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3770_end_mask_0 = const()[name = tensor<string, []>("op_3770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3770_cast_fp16 = slice_by_index(begin = var_3770_begin_0, end = var_3770_end_0, end_mask = var_3770_end_mask_0, x = q_305_cast_fp16)[name = tensor<string, []>("op_3770_cast_fp16")];
            tensor<int32, [4]> var_3774_begin_0 = const()[name = tensor<string, []>("op_3774_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3774_end_0 = const()[name = tensor<string, []>("op_3774_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3774_end_mask_0 = const()[name = tensor<string, []>("op_3774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3774_cast_fp16 = slice_by_index(begin = var_3774_begin_0, end = var_3774_end_0, end_mask = var_3774_end_mask_0, x = k_183_cast_fp16)[name = tensor<string, []>("op_3774_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3776_cast_fp16 = mul(x = var_3770_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3776_cast_fp16")];
            tensor<int32, [5]> var_3784 = const()[name = tensor<string, []>("op_3784"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_411_cast_fp16 = reshape(shape = var_3784, x = var_3770_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
            tensor<int32, [2]> var_3786_split_sizes_0 = const()[name = tensor<string, []>("op_3786_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3786_axis_0 = const()[name = tensor<string, []>("op_3786_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3786_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3786_cast_fp16_1 = split(axis = var_3786_axis_0, split_sizes = var_3786_split_sizes_0, x = x_411_cast_fp16)[name = tensor<string, []>("op_3786_cast_fp16")];
            tensor<int32, [1]> squeeze_120_axes_0 = const()[name = tensor<string, []>("squeeze_120_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_120_cast_fp16 = squeeze(axes = squeeze_120_axes_0, x = var_3786_cast_fp16_0)[name = tensor<string, []>("squeeze_120_cast_fp16")];
            tensor<int32, [1]> squeeze_121_axes_0 = const()[name = tensor<string, []>("squeeze_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_121_cast_fp16 = squeeze(axes = squeeze_121_axes_0, x = var_3786_cast_fp16_1)[name = tensor<string, []>("squeeze_121_cast_fp16")];
            tensor<fp16, []> const_495_promoted_to_fp16 = const()[name = tensor<string, []>("const_495_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3789_cast_fp16 = mul(x = squeeze_121_cast_fp16, y = const_495_promoted_to_fp16)[name = tensor<string, []>("op_3789_cast_fp16")];
            tensor<int32, []> x_413_axis_0 = const()[name = tensor<string, []>("x_413_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_413_cast_fp16 = stack(axis = x_413_axis_0, values = (var_3789_cast_fp16, squeeze_120_cast_fp16))[name = tensor<string, []>("x_413_cast_fp16")];
            tensor<int32, [4]> var_3795 = const()[name = tensor<string, []>("op_3795"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3796_cast_fp16 = reshape(shape = var_3795, x = x_413_cast_fp16)[name = tensor<string, []>("op_3796_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3797_cast_fp16 = mul(x = var_3796_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3797_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_61_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3797_cast_fp16)[name = tensor<string, []>("q_patches_61_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3799_cast_fp16 = mul(x = var_3774_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3799_cast_fp16")];
            tensor<int32, [5]> var_3807 = const()[name = tensor<string, []>("op_3807"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_417_cast_fp16 = reshape(shape = var_3807, x = var_3774_cast_fp16)[name = tensor<string, []>("x_417_cast_fp16")];
            tensor<int32, [2]> var_3809_split_sizes_0 = const()[name = tensor<string, []>("op_3809_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3809_axis_0 = const()[name = tensor<string, []>("op_3809_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3809_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3809_cast_fp16_1 = split(axis = var_3809_axis_0, split_sizes = var_3809_split_sizes_0, x = x_417_cast_fp16)[name = tensor<string, []>("op_3809_cast_fp16")];
            tensor<int32, [1]> squeeze_122_axes_0 = const()[name = tensor<string, []>("squeeze_122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_122_cast_fp16 = squeeze(axes = squeeze_122_axes_0, x = var_3809_cast_fp16_0)[name = tensor<string, []>("squeeze_122_cast_fp16")];
            tensor<int32, [1]> squeeze_123_axes_0 = const()[name = tensor<string, []>("squeeze_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_123_cast_fp16 = squeeze(axes = squeeze_123_axes_0, x = var_3809_cast_fp16_1)[name = tensor<string, []>("squeeze_123_cast_fp16")];
            tensor<fp16, []> const_503_promoted_to_fp16 = const()[name = tensor<string, []>("const_503_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3812_cast_fp16 = mul(x = squeeze_123_cast_fp16, y = const_503_promoted_to_fp16)[name = tensor<string, []>("op_3812_cast_fp16")];
            tensor<int32, []> x_419_axis_0 = const()[name = tensor<string, []>("x_419_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_419_cast_fp16 = stack(axis = x_419_axis_0, values = (var_3812_cast_fp16, squeeze_122_cast_fp16))[name = tensor<string, []>("x_419_cast_fp16")];
            tensor<int32, [4]> var_3818 = const()[name = tensor<string, []>("op_3818"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3819_cast_fp16 = reshape(shape = var_3818, x = x_419_cast_fp16)[name = tensor<string, []>("op_3819_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3820_cast_fp16 = mul(x = var_3819_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3820_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_61_cast_fp16 = add(x = var_3799_cast_fp16, y = var_3820_cast_fp16)[name = tensor<string, []>("k_patches_61_cast_fp16")];
            tensor<bool, []> var_3823_interleave_0 = const()[name = tensor<string, []>("op_3823_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3823_cast_fp16 = concat(axis = var_20, interleave = var_3823_interleave_0, values = (var_3762_cast_fp16, q_patches_61_cast_fp16))[name = tensor<string, []>("op_3823_cast_fp16")];
            tensor<bool, []> var_3826_interleave_0 = const()[name = tensor<string, []>("op_3826_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3826_cast_fp16 = concat(axis = var_20, interleave = var_3826_interleave_0, values = (var_3766_cast_fp16, k_patches_61_cast_fp16))[name = tensor<string, []>("op_3826_cast_fp16")];
            tensor<fp16, []> var_3828_to_fp16 = const()[name = tensor<string, []>("op_3828_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_309_cast_fp16 = mul(x = var_3823_cast_fp16, y = var_3828_to_fp16)[name = tensor<string, []>("q_309_cast_fp16")];
            tensor<bool, []> attn_121_transpose_x_1 = const()[name = tensor<string, []>("attn_121_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_121_transpose_y_1 = const()[name = tensor<string, []>("attn_121_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_1, transpose_y = attn_121_transpose_y_1, x = q_309_cast_fp16, y = var_3826_cast_fp16)[name = tensor<string, []>("attn_121_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_123_cast_fp16 = softmax(axis = var_21, x = attn_121_cast_fp16)[name = tensor<string, []>("attn_123_cast_fp16")];
            tensor<bool, []> var_3833_transpose_x_0 = const()[name = tensor<string, []>("op_3833_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3833_transpose_y_0 = const()[name = tensor<string, []>("op_3833_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3833_cast_fp16 = matmul(transpose_x = var_3833_transpose_x_0, transpose_y = var_3833_transpose_y_0, x = attn_123_cast_fp16, y = v_185_cast_fp16)[name = tensor<string, []>("op_3833_cast_fp16")];
            tensor<int32, [1]> var_3834_axes_0 = const()[name = tensor<string, []>("op_3834_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3834_cast_fp16 = squeeze(axes = var_3834_axes_0, x = var_3833_cast_fp16)[name = tensor<string, []>("op_3834_cast_fp16")];
            tensor<string, []> var_3840_pad_type_0 = const()[name = tensor<string, []>("op_3840_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3840_strides_0 = const()[name = tensor<string, []>("op_3840_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3840_pad_0 = const()[name = tensor<string, []>("op_3840_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3840_dilations_0 = const()[name = tensor<string, []>("op_3840_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3840_groups_0 = const()[name = tensor<string, []>("op_3840_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31942016)))];
            tensor<fp16, [1, 64, 1, 197]> var_3840_cast_fp16 = conv(dilations = var_3840_dilations_0, groups = var_3840_groups_0, pad = var_3840_pad_0, pad_type = var_3840_pad_type_0, strides = var_3840_strides_0, weight = model_blocks_2_attn_q_projs_7_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3840_cast_fp16")];
            tensor<int32, [1]> var_3841_axes_0 = const()[name = tensor<string, []>("op_3841_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3841_cast_fp16 = squeeze(axes = var_3841_axes_0, x = var_3840_cast_fp16)[name = tensor<string, []>("op_3841_cast_fp16")];
            tensor<int32, [3]> q_311_perm_0 = const()[name = tensor<string, []>("q_311_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3848_pad_type_0 = const()[name = tensor<string, []>("op_3848_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3848_strides_0 = const()[name = tensor<string, []>("op_3848_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3848_pad_0 = const()[name = tensor<string, []>("op_3848_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3848_dilations_0 = const()[name = tensor<string, []>("op_3848_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3848_groups_0 = const()[name = tensor<string, []>("op_3848_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32040384)))];
            tensor<fp16, [1, 64, 1, 197]> var_3848_cast_fp16 = conv(dilations = var_3848_dilations_0, groups = var_3848_groups_0, pad = var_3848_pad_0, pad_type = var_3848_pad_type_0, strides = var_3848_strides_0, weight = model_blocks_2_attn_k_projs_7_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3848_cast_fp16")];
            tensor<int32, [1]> var_3849_axes_0 = const()[name = tensor<string, []>("op_3849_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3849_cast_fp16 = squeeze(axes = var_3849_axes_0, x = var_3848_cast_fp16)[name = tensor<string, []>("op_3849_cast_fp16")];
            tensor<int32, [3]> k_187_perm_0 = const()[name = tensor<string, []>("k_187_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3856_pad_type_0 = const()[name = tensor<string, []>("op_3856_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3856_strides_0 = const()[name = tensor<string, []>("op_3856_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3856_pad_0 = const()[name = tensor<string, []>("op_3856_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3856_dilations_0 = const()[name = tensor<string, []>("op_3856_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3856_groups_0 = const()[name = tensor<string, []>("op_3856_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32138752)))];
            tensor<fp16, [1, 64, 1, 197]> var_3856_cast_fp16 = conv(dilations = var_3856_dilations_0, groups = var_3856_groups_0, pad = var_3856_pad_0, pad_type = var_3856_pad_type_0, strides = var_3856_strides_0, weight = model_blocks_2_attn_v_projs_7_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3856_cast_fp16")];
            tensor<int32, [1]> var_3857_axes_0 = const()[name = tensor<string, []>("op_3857_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3857_cast_fp16 = squeeze(axes = var_3857_axes_0, x = var_3856_cast_fp16)[name = tensor<string, []>("op_3857_cast_fp16")];
            tensor<int32, [3]> v_187_perm_0 = const()[name = tensor<string, []>("v_187_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32237120)))];
            tensor<fp16, [1, 197, 64]> q_311_cast_fp16 = transpose(perm = q_311_perm_0, x = var_3841_cast_fp16)[name = tensor<string, []>("transpose_407")];
            tensor<fp16, [1, 197, 64]> q_313_cast_fp16 = add(x = q_311_cast_fp16, y = model_blocks_2_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_313_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32237312)))];
            tensor<fp16, [1, 197, 64]> v_187_cast_fp16 = transpose(perm = v_187_perm_0, x = var_3857_cast_fp16)[name = tensor<string, []>("transpose_405")];
            tensor<fp16, [1, 197, 64]> v_189_cast_fp16 = add(x = v_187_cast_fp16, y = model_blocks_2_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_189_cast_fp16")];
            tensor<int32, [1]> q_315_axes_0 = const()[name = tensor<string, []>("q_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_315_cast_fp16 = expand_dims(axes = q_315_axes_0, x = q_313_cast_fp16)[name = tensor<string, []>("q_315_cast_fp16")];
            tensor<int32, [1]> k_189_axes_0 = const()[name = tensor<string, []>("k_189_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_187_cast_fp16 = transpose(perm = k_187_perm_0, x = var_3849_cast_fp16)[name = tensor<string, []>("transpose_406")];
            tensor<fp16, [1, 1, 197, 64]> k_189_cast_fp16 = expand_dims(axes = k_189_axes_0, x = k_187_cast_fp16)[name = tensor<string, []>("k_189_cast_fp16")];
            tensor<int32, [1]> v_191_axes_0 = const()[name = tensor<string, []>("v_191_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_191_cast_fp16 = expand_dims(axes = v_191_axes_0, x = v_189_cast_fp16)[name = tensor<string, []>("v_191_cast_fp16")];
            tensor<int32, [4]> var_3866_begin_0 = const()[name = tensor<string, []>("op_3866_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3866_end_0 = const()[name = tensor<string, []>("op_3866_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3866_end_mask_0 = const()[name = tensor<string, []>("op_3866_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3866_cast_fp16 = slice_by_index(begin = var_3866_begin_0, end = var_3866_end_0, end_mask = var_3866_end_mask_0, x = q_315_cast_fp16)[name = tensor<string, []>("op_3866_cast_fp16")];
            tensor<int32, [4]> var_3870_begin_0 = const()[name = tensor<string, []>("op_3870_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3870_end_0 = const()[name = tensor<string, []>("op_3870_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3870_end_mask_0 = const()[name = tensor<string, []>("op_3870_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3870_cast_fp16 = slice_by_index(begin = var_3870_begin_0, end = var_3870_end_0, end_mask = var_3870_end_mask_0, x = k_189_cast_fp16)[name = tensor<string, []>("op_3870_cast_fp16")];
            tensor<int32, [4]> var_3874_begin_0 = const()[name = tensor<string, []>("op_3874_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3874_end_0 = const()[name = tensor<string, []>("op_3874_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3874_end_mask_0 = const()[name = tensor<string, []>("op_3874_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3874_cast_fp16 = slice_by_index(begin = var_3874_begin_0, end = var_3874_end_0, end_mask = var_3874_end_mask_0, x = q_315_cast_fp16)[name = tensor<string, []>("op_3874_cast_fp16")];
            tensor<int32, [4]> var_3878_begin_0 = const()[name = tensor<string, []>("op_3878_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3878_end_0 = const()[name = tensor<string, []>("op_3878_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3878_end_mask_0 = const()[name = tensor<string, []>("op_3878_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3878_cast_fp16 = slice_by_index(begin = var_3878_begin_0, end = var_3878_end_0, end_mask = var_3878_end_mask_0, x = k_189_cast_fp16)[name = tensor<string, []>("op_3878_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3880_cast_fp16 = mul(x = var_3874_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3880_cast_fp16")];
            tensor<int32, [5]> var_3888 = const()[name = tensor<string, []>("op_3888"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_423_cast_fp16 = reshape(shape = var_3888, x = var_3874_cast_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
            tensor<int32, [2]> var_3890_split_sizes_0 = const()[name = tensor<string, []>("op_3890_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3890_axis_0 = const()[name = tensor<string, []>("op_3890_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3890_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3890_cast_fp16_1 = split(axis = var_3890_axis_0, split_sizes = var_3890_split_sizes_0, x = x_423_cast_fp16)[name = tensor<string, []>("op_3890_cast_fp16")];
            tensor<int32, [1]> squeeze_124_axes_0 = const()[name = tensor<string, []>("squeeze_124_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_124_cast_fp16 = squeeze(axes = squeeze_124_axes_0, x = var_3890_cast_fp16_0)[name = tensor<string, []>("squeeze_124_cast_fp16")];
            tensor<int32, [1]> squeeze_125_axes_0 = const()[name = tensor<string, []>("squeeze_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_125_cast_fp16 = squeeze(axes = squeeze_125_axes_0, x = var_3890_cast_fp16_1)[name = tensor<string, []>("squeeze_125_cast_fp16")];
            tensor<fp16, []> const_511_promoted_to_fp16 = const()[name = tensor<string, []>("const_511_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3893_cast_fp16 = mul(x = squeeze_125_cast_fp16, y = const_511_promoted_to_fp16)[name = tensor<string, []>("op_3893_cast_fp16")];
            tensor<int32, []> x_425_axis_0 = const()[name = tensor<string, []>("x_425_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_425_cast_fp16 = stack(axis = x_425_axis_0, values = (var_3893_cast_fp16, squeeze_124_cast_fp16))[name = tensor<string, []>("x_425_cast_fp16")];
            tensor<int32, [4]> var_3899 = const()[name = tensor<string, []>("op_3899"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3900_cast_fp16 = reshape(shape = var_3899, x = x_425_cast_fp16)[name = tensor<string, []>("op_3900_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3901_cast_fp16 = mul(x = var_3900_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3901_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_63_cast_fp16 = add(x = var_3880_cast_fp16, y = var_3901_cast_fp16)[name = tensor<string, []>("q_patches_63_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3903_cast_fp16 = mul(x = var_3878_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3903_cast_fp16")];
            tensor<int32, [5]> var_3911 = const()[name = tensor<string, []>("op_3911"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_429_cast_fp16 = reshape(shape = var_3911, x = var_3878_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
            tensor<int32, [2]> var_3913_split_sizes_0 = const()[name = tensor<string, []>("op_3913_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3913_axis_0 = const()[name = tensor<string, []>("op_3913_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3913_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3913_cast_fp16_1 = split(axis = var_3913_axis_0, split_sizes = var_3913_split_sizes_0, x = x_429_cast_fp16)[name = tensor<string, []>("op_3913_cast_fp16")];
            tensor<int32, [1]> squeeze_126_axes_0 = const()[name = tensor<string, []>("squeeze_126_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_126_cast_fp16 = squeeze(axes = squeeze_126_axes_0, x = var_3913_cast_fp16_0)[name = tensor<string, []>("squeeze_126_cast_fp16")];
            tensor<int32, [1]> squeeze_127_axes_0 = const()[name = tensor<string, []>("squeeze_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_127_cast_fp16 = squeeze(axes = squeeze_127_axes_0, x = var_3913_cast_fp16_1)[name = tensor<string, []>("squeeze_127_cast_fp16")];
            tensor<fp16, []> const_519_promoted_to_fp16 = const()[name = tensor<string, []>("const_519_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3916_cast_fp16 = mul(x = squeeze_127_cast_fp16, y = const_519_promoted_to_fp16)[name = tensor<string, []>("op_3916_cast_fp16")];
            tensor<int32, []> x_431_axis_0 = const()[name = tensor<string, []>("x_431_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_431_cast_fp16 = stack(axis = x_431_axis_0, values = (var_3916_cast_fp16, squeeze_126_cast_fp16))[name = tensor<string, []>("x_431_cast_fp16")];
            tensor<int32, [4]> var_3922 = const()[name = tensor<string, []>("op_3922"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_3923_cast_fp16 = reshape(shape = var_3922, x = x_431_cast_fp16)[name = tensor<string, []>("op_3923_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3924_cast_fp16 = mul(x = var_3923_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_63_cast_fp16 = add(x = var_3903_cast_fp16, y = var_3924_cast_fp16)[name = tensor<string, []>("k_patches_63_cast_fp16")];
            tensor<bool, []> var_3927_interleave_0 = const()[name = tensor<string, []>("op_3927_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3927_cast_fp16 = concat(axis = var_20, interleave = var_3927_interleave_0, values = (var_3866_cast_fp16, q_patches_63_cast_fp16))[name = tensor<string, []>("op_3927_cast_fp16")];
            tensor<bool, []> var_3930_interleave_0 = const()[name = tensor<string, []>("op_3930_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3930_cast_fp16 = concat(axis = var_20, interleave = var_3930_interleave_0, values = (var_3870_cast_fp16, k_patches_63_cast_fp16))[name = tensor<string, []>("op_3930_cast_fp16")];
            tensor<fp16, []> var_3932_to_fp16 = const()[name = tensor<string, []>("op_3932_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_319_cast_fp16 = mul(x = var_3927_cast_fp16, y = var_3932_to_fp16)[name = tensor<string, []>("q_319_cast_fp16")];
            tensor<bool, []> attn_125_transpose_x_1 = const()[name = tensor<string, []>("attn_125_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_125_transpose_y_1 = const()[name = tensor<string, []>("attn_125_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_1, transpose_y = attn_125_transpose_y_1, x = q_319_cast_fp16, y = var_3930_cast_fp16)[name = tensor<string, []>("attn_125_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_127_cast_fp16 = softmax(axis = var_21, x = attn_125_cast_fp16)[name = tensor<string, []>("attn_127_cast_fp16")];
            tensor<bool, []> var_3937_transpose_x_0 = const()[name = tensor<string, []>("op_3937_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3937_transpose_y_0 = const()[name = tensor<string, []>("op_3937_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_3937_cast_fp16 = matmul(transpose_x = var_3937_transpose_x_0, transpose_y = var_3937_transpose_y_0, x = attn_127_cast_fp16, y = v_191_cast_fp16)[name = tensor<string, []>("op_3937_cast_fp16")];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = tensor<string, []>("op_3938_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_3938_cast_fp16 = squeeze(axes = var_3938_axes_0, x = var_3937_cast_fp16)[name = tensor<string, []>("op_3938_cast_fp16")];
            tensor<string, []> var_3944_pad_type_0 = const()[name = tensor<string, []>("op_3944_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3944_strides_0 = const()[name = tensor<string, []>("op_3944_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3944_pad_0 = const()[name = tensor<string, []>("op_3944_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3944_dilations_0 = const()[name = tensor<string, []>("op_3944_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3944_groups_0 = const()[name = tensor<string, []>("op_3944_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32237504)))];
            tensor<fp16, [1, 64, 1, 197]> var_3944_cast_fp16 = conv(dilations = var_3944_dilations_0, groups = var_3944_groups_0, pad = var_3944_pad_0, pad_type = var_3944_pad_type_0, strides = var_3944_strides_0, weight = model_blocks_2_attn_q_projs_8_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3944_cast_fp16")];
            tensor<int32, [1]> var_3945_axes_0 = const()[name = tensor<string, []>("op_3945_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3945_cast_fp16 = squeeze(axes = var_3945_axes_0, x = var_3944_cast_fp16)[name = tensor<string, []>("op_3945_cast_fp16")];
            tensor<int32, [3]> q_321_perm_0 = const()[name = tensor<string, []>("q_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3952_pad_type_0 = const()[name = tensor<string, []>("op_3952_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3952_strides_0 = const()[name = tensor<string, []>("op_3952_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3952_pad_0 = const()[name = tensor<string, []>("op_3952_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3952_dilations_0 = const()[name = tensor<string, []>("op_3952_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3952_groups_0 = const()[name = tensor<string, []>("op_3952_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32335872)))];
            tensor<fp16, [1, 64, 1, 197]> var_3952_cast_fp16 = conv(dilations = var_3952_dilations_0, groups = var_3952_groups_0, pad = var_3952_pad_0, pad_type = var_3952_pad_type_0, strides = var_3952_strides_0, weight = model_blocks_2_attn_k_projs_8_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3952_cast_fp16")];
            tensor<int32, [1]> var_3953_axes_0 = const()[name = tensor<string, []>("op_3953_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3953_cast_fp16 = squeeze(axes = var_3953_axes_0, x = var_3952_cast_fp16)[name = tensor<string, []>("op_3953_cast_fp16")];
            tensor<int32, [3]> k_193_perm_0 = const()[name = tensor<string, []>("k_193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_3960_pad_type_0 = const()[name = tensor<string, []>("op_3960_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3960_strides_0 = const()[name = tensor<string, []>("op_3960_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3960_pad_0 = const()[name = tensor<string, []>("op_3960_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3960_dilations_0 = const()[name = tensor<string, []>("op_3960_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3960_groups_0 = const()[name = tensor<string, []>("op_3960_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32434240)))];
            tensor<fp16, [1, 64, 1, 197]> var_3960_cast_fp16 = conv(dilations = var_3960_dilations_0, groups = var_3960_groups_0, pad = var_3960_pad_0, pad_type = var_3960_pad_type_0, strides = var_3960_strides_0, weight = model_blocks_2_attn_v_projs_8_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_3960_cast_fp16")];
            tensor<int32, [1]> var_3961_axes_0 = const()[name = tensor<string, []>("op_3961_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_3961_cast_fp16 = squeeze(axes = var_3961_axes_0, x = var_3960_cast_fp16)[name = tensor<string, []>("op_3961_cast_fp16")];
            tensor<int32, [3]> v_193_perm_0 = const()[name = tensor<string, []>("v_193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32532608)))];
            tensor<fp16, [1, 197, 64]> q_321_cast_fp16 = transpose(perm = q_321_perm_0, x = var_3945_cast_fp16)[name = tensor<string, []>("transpose_404")];
            tensor<fp16, [1, 197, 64]> q_323_cast_fp16 = add(x = q_321_cast_fp16, y = model_blocks_2_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_323_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32532800)))];
            tensor<fp16, [1, 197, 64]> v_193_cast_fp16 = transpose(perm = v_193_perm_0, x = var_3961_cast_fp16)[name = tensor<string, []>("transpose_402")];
            tensor<fp16, [1, 197, 64]> v_195_cast_fp16 = add(x = v_193_cast_fp16, y = model_blocks_2_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_195_cast_fp16")];
            tensor<int32, [1]> q_325_axes_0 = const()[name = tensor<string, []>("q_325_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_325_cast_fp16 = expand_dims(axes = q_325_axes_0, x = q_323_cast_fp16)[name = tensor<string, []>("q_325_cast_fp16")];
            tensor<int32, [1]> k_195_axes_0 = const()[name = tensor<string, []>("k_195_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_193_cast_fp16 = transpose(perm = k_193_perm_0, x = var_3953_cast_fp16)[name = tensor<string, []>("transpose_403")];
            tensor<fp16, [1, 1, 197, 64]> k_195_cast_fp16 = expand_dims(axes = k_195_axes_0, x = k_193_cast_fp16)[name = tensor<string, []>("k_195_cast_fp16")];
            tensor<int32, [1]> v_197_axes_0 = const()[name = tensor<string, []>("v_197_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_197_cast_fp16 = expand_dims(axes = v_197_axes_0, x = v_195_cast_fp16)[name = tensor<string, []>("v_197_cast_fp16")];
            tensor<int32, [4]> var_3970_begin_0 = const()[name = tensor<string, []>("op_3970_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3970_end_0 = const()[name = tensor<string, []>("op_3970_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3970_end_mask_0 = const()[name = tensor<string, []>("op_3970_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3970_cast_fp16 = slice_by_index(begin = var_3970_begin_0, end = var_3970_end_0, end_mask = var_3970_end_mask_0, x = q_325_cast_fp16)[name = tensor<string, []>("op_3970_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = tensor<string, []>("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = tensor<string, []>("op_3974_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = tensor<string, []>("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_3974_cast_fp16 = slice_by_index(begin = var_3974_begin_0, end = var_3974_end_0, end_mask = var_3974_end_mask_0, x = k_195_cast_fp16)[name = tensor<string, []>("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3978_begin_0 = const()[name = tensor<string, []>("op_3978_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3978_end_0 = const()[name = tensor<string, []>("op_3978_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3978_end_mask_0 = const()[name = tensor<string, []>("op_3978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3978_cast_fp16 = slice_by_index(begin = var_3978_begin_0, end = var_3978_end_0, end_mask = var_3978_end_mask_0, x = q_325_cast_fp16)[name = tensor<string, []>("op_3978_cast_fp16")];
            tensor<int32, [4]> var_3982_begin_0 = const()[name = tensor<string, []>("op_3982_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3982_end_0 = const()[name = tensor<string, []>("op_3982_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_3982_end_mask_0 = const()[name = tensor<string, []>("op_3982_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_3982_cast_fp16 = slice_by_index(begin = var_3982_begin_0, end = var_3982_end_0, end_mask = var_3982_end_mask_0, x = k_195_cast_fp16)[name = tensor<string, []>("op_3982_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_3984_cast_fp16 = mul(x = var_3978_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_3984_cast_fp16")];
            tensor<int32, [5]> var_3992 = const()[name = tensor<string, []>("op_3992"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_435_cast_fp16 = reshape(shape = var_3992, x = var_3978_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
            tensor<int32, [2]> var_3994_split_sizes_0 = const()[name = tensor<string, []>("op_3994_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3994_axis_0 = const()[name = tensor<string, []>("op_3994_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_3994_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_3994_cast_fp16_1 = split(axis = var_3994_axis_0, split_sizes = var_3994_split_sizes_0, x = x_435_cast_fp16)[name = tensor<string, []>("op_3994_cast_fp16")];
            tensor<int32, [1]> squeeze_128_axes_0 = const()[name = tensor<string, []>("squeeze_128_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_128_cast_fp16 = squeeze(axes = squeeze_128_axes_0, x = var_3994_cast_fp16_0)[name = tensor<string, []>("squeeze_128_cast_fp16")];
            tensor<int32, [1]> squeeze_129_axes_0 = const()[name = tensor<string, []>("squeeze_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_129_cast_fp16 = squeeze(axes = squeeze_129_axes_0, x = var_3994_cast_fp16_1)[name = tensor<string, []>("squeeze_129_cast_fp16")];
            tensor<fp16, []> const_527_promoted_to_fp16 = const()[name = tensor<string, []>("const_527_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_3997_cast_fp16 = mul(x = squeeze_129_cast_fp16, y = const_527_promoted_to_fp16)[name = tensor<string, []>("op_3997_cast_fp16")];
            tensor<int32, []> x_437_axis_0 = const()[name = tensor<string, []>("x_437_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_437_cast_fp16 = stack(axis = x_437_axis_0, values = (var_3997_cast_fp16, squeeze_128_cast_fp16))[name = tensor<string, []>("x_437_cast_fp16")];
            tensor<int32, [4]> var_4003 = const()[name = tensor<string, []>("op_4003"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4004_cast_fp16 = reshape(shape = var_4003, x = x_437_cast_fp16)[name = tensor<string, []>("op_4004_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4005_cast_fp16 = mul(x = var_4004_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4005_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_65_cast_fp16 = add(x = var_3984_cast_fp16, y = var_4005_cast_fp16)[name = tensor<string, []>("q_patches_65_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4007_cast_fp16 = mul(x = var_3982_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4007_cast_fp16")];
            tensor<int32, [5]> var_4015 = const()[name = tensor<string, []>("op_4015"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_441_cast_fp16 = reshape(shape = var_4015, x = var_3982_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
            tensor<int32, [2]> var_4017_split_sizes_0 = const()[name = tensor<string, []>("op_4017_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4017_axis_0 = const()[name = tensor<string, []>("op_4017_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4017_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4017_cast_fp16_1 = split(axis = var_4017_axis_0, split_sizes = var_4017_split_sizes_0, x = x_441_cast_fp16)[name = tensor<string, []>("op_4017_cast_fp16")];
            tensor<int32, [1]> squeeze_130_axes_0 = const()[name = tensor<string, []>("squeeze_130_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_130_cast_fp16 = squeeze(axes = squeeze_130_axes_0, x = var_4017_cast_fp16_0)[name = tensor<string, []>("squeeze_130_cast_fp16")];
            tensor<int32, [1]> squeeze_131_axes_0 = const()[name = tensor<string, []>("squeeze_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_131_cast_fp16 = squeeze(axes = squeeze_131_axes_0, x = var_4017_cast_fp16_1)[name = tensor<string, []>("squeeze_131_cast_fp16")];
            tensor<fp16, []> const_535_promoted_to_fp16 = const()[name = tensor<string, []>("const_535_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4020_cast_fp16 = mul(x = squeeze_131_cast_fp16, y = const_535_promoted_to_fp16)[name = tensor<string, []>("op_4020_cast_fp16")];
            tensor<int32, []> x_443_axis_0 = const()[name = tensor<string, []>("x_443_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_443_cast_fp16 = stack(axis = x_443_axis_0, values = (var_4020_cast_fp16, squeeze_130_cast_fp16))[name = tensor<string, []>("x_443_cast_fp16")];
            tensor<int32, [4]> var_4026 = const()[name = tensor<string, []>("op_4026"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4027_cast_fp16 = reshape(shape = var_4026, x = x_443_cast_fp16)[name = tensor<string, []>("op_4027_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4028_cast_fp16 = mul(x = var_4027_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4028_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_65_cast_fp16 = add(x = var_4007_cast_fp16, y = var_4028_cast_fp16)[name = tensor<string, []>("k_patches_65_cast_fp16")];
            tensor<bool, []> var_4031_interleave_0 = const()[name = tensor<string, []>("op_4031_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4031_cast_fp16 = concat(axis = var_20, interleave = var_4031_interleave_0, values = (var_3970_cast_fp16, q_patches_65_cast_fp16))[name = tensor<string, []>("op_4031_cast_fp16")];
            tensor<bool, []> var_4034_interleave_0 = const()[name = tensor<string, []>("op_4034_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4034_cast_fp16 = concat(axis = var_20, interleave = var_4034_interleave_0, values = (var_3974_cast_fp16, k_patches_65_cast_fp16))[name = tensor<string, []>("op_4034_cast_fp16")];
            tensor<fp16, []> var_4036_to_fp16 = const()[name = tensor<string, []>("op_4036_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_329_cast_fp16 = mul(x = var_4031_cast_fp16, y = var_4036_to_fp16)[name = tensor<string, []>("q_329_cast_fp16")];
            tensor<bool, []> attn_129_transpose_x_1 = const()[name = tensor<string, []>("attn_129_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_129_transpose_y_1 = const()[name = tensor<string, []>("attn_129_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_129_cast_fp16 = matmul(transpose_x = attn_129_transpose_x_1, transpose_y = attn_129_transpose_y_1, x = q_329_cast_fp16, y = var_4034_cast_fp16)[name = tensor<string, []>("attn_129_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_131_cast_fp16 = softmax(axis = var_21, x = attn_129_cast_fp16)[name = tensor<string, []>("attn_131_cast_fp16")];
            tensor<bool, []> var_4041_transpose_x_0 = const()[name = tensor<string, []>("op_4041_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4041_transpose_y_0 = const()[name = tensor<string, []>("op_4041_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4041_cast_fp16 = matmul(transpose_x = var_4041_transpose_x_0, transpose_y = var_4041_transpose_y_0, x = attn_131_cast_fp16, y = v_197_cast_fp16)[name = tensor<string, []>("op_4041_cast_fp16")];
            tensor<int32, [1]> var_4042_axes_0 = const()[name = tensor<string, []>("op_4042_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4042_cast_fp16 = squeeze(axes = var_4042_axes_0, x = var_4041_cast_fp16)[name = tensor<string, []>("op_4042_cast_fp16")];
            tensor<string, []> var_4048_pad_type_0 = const()[name = tensor<string, []>("op_4048_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4048_strides_0 = const()[name = tensor<string, []>("op_4048_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4048_pad_0 = const()[name = tensor<string, []>("op_4048_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4048_dilations_0 = const()[name = tensor<string, []>("op_4048_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4048_groups_0 = const()[name = tensor<string, []>("op_4048_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32532992)))];
            tensor<fp16, [1, 64, 1, 197]> var_4048_cast_fp16 = conv(dilations = var_4048_dilations_0, groups = var_4048_groups_0, pad = var_4048_pad_0, pad_type = var_4048_pad_type_0, strides = var_4048_strides_0, weight = model_blocks_2_attn_q_projs_9_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4048_cast_fp16")];
            tensor<int32, [1]> var_4049_axes_0 = const()[name = tensor<string, []>("op_4049_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4049_cast_fp16 = squeeze(axes = var_4049_axes_0, x = var_4048_cast_fp16)[name = tensor<string, []>("op_4049_cast_fp16")];
            tensor<int32, [3]> q_331_perm_0 = const()[name = tensor<string, []>("q_331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4056_pad_type_0 = const()[name = tensor<string, []>("op_4056_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4056_strides_0 = const()[name = tensor<string, []>("op_4056_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4056_pad_0 = const()[name = tensor<string, []>("op_4056_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4056_dilations_0 = const()[name = tensor<string, []>("op_4056_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4056_groups_0 = const()[name = tensor<string, []>("op_4056_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32631360)))];
            tensor<fp16, [1, 64, 1, 197]> var_4056_cast_fp16 = conv(dilations = var_4056_dilations_0, groups = var_4056_groups_0, pad = var_4056_pad_0, pad_type = var_4056_pad_type_0, strides = var_4056_strides_0, weight = model_blocks_2_attn_k_projs_9_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4056_cast_fp16")];
            tensor<int32, [1]> var_4057_axes_0 = const()[name = tensor<string, []>("op_4057_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4057_cast_fp16 = squeeze(axes = var_4057_axes_0, x = var_4056_cast_fp16)[name = tensor<string, []>("op_4057_cast_fp16")];
            tensor<int32, [3]> k_199_perm_0 = const()[name = tensor<string, []>("k_199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4064_pad_type_0 = const()[name = tensor<string, []>("op_4064_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4064_strides_0 = const()[name = tensor<string, []>("op_4064_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4064_pad_0 = const()[name = tensor<string, []>("op_4064_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4064_dilations_0 = const()[name = tensor<string, []>("op_4064_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4064_groups_0 = const()[name = tensor<string, []>("op_4064_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32729728)))];
            tensor<fp16, [1, 64, 1, 197]> var_4064_cast_fp16 = conv(dilations = var_4064_dilations_0, groups = var_4064_groups_0, pad = var_4064_pad_0, pad_type = var_4064_pad_type_0, strides = var_4064_strides_0, weight = model_blocks_2_attn_v_projs_9_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4064_cast_fp16")];
            tensor<int32, [1]> var_4065_axes_0 = const()[name = tensor<string, []>("op_4065_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4065_cast_fp16 = squeeze(axes = var_4065_axes_0, x = var_4064_cast_fp16)[name = tensor<string, []>("op_4065_cast_fp16")];
            tensor<int32, [3]> v_199_perm_0 = const()[name = tensor<string, []>("v_199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32828096)))];
            tensor<fp16, [1, 197, 64]> q_331_cast_fp16 = transpose(perm = q_331_perm_0, x = var_4049_cast_fp16)[name = tensor<string, []>("transpose_401")];
            tensor<fp16, [1, 197, 64]> q_333_cast_fp16 = add(x = q_331_cast_fp16, y = model_blocks_2_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_333_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32828288)))];
            tensor<fp16, [1, 197, 64]> v_199_cast_fp16 = transpose(perm = v_199_perm_0, x = var_4065_cast_fp16)[name = tensor<string, []>("transpose_399")];
            tensor<fp16, [1, 197, 64]> v_201_cast_fp16 = add(x = v_199_cast_fp16, y = model_blocks_2_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_201_cast_fp16")];
            tensor<int32, [1]> q_335_axes_0 = const()[name = tensor<string, []>("q_335_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_335_cast_fp16 = expand_dims(axes = q_335_axes_0, x = q_333_cast_fp16)[name = tensor<string, []>("q_335_cast_fp16")];
            tensor<int32, [1]> k_201_axes_0 = const()[name = tensor<string, []>("k_201_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_199_cast_fp16 = transpose(perm = k_199_perm_0, x = var_4057_cast_fp16)[name = tensor<string, []>("transpose_400")];
            tensor<fp16, [1, 1, 197, 64]> k_201_cast_fp16 = expand_dims(axes = k_201_axes_0, x = k_199_cast_fp16)[name = tensor<string, []>("k_201_cast_fp16")];
            tensor<int32, [1]> v_203_axes_0 = const()[name = tensor<string, []>("v_203_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_203_cast_fp16 = expand_dims(axes = v_203_axes_0, x = v_201_cast_fp16)[name = tensor<string, []>("v_203_cast_fp16")];
            tensor<int32, [4]> var_4074_begin_0 = const()[name = tensor<string, []>("op_4074_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4074_end_0 = const()[name = tensor<string, []>("op_4074_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4074_end_mask_0 = const()[name = tensor<string, []>("op_4074_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4074_cast_fp16 = slice_by_index(begin = var_4074_begin_0, end = var_4074_end_0, end_mask = var_4074_end_mask_0, x = q_335_cast_fp16)[name = tensor<string, []>("op_4074_cast_fp16")];
            tensor<int32, [4]> var_4078_begin_0 = const()[name = tensor<string, []>("op_4078_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4078_end_0 = const()[name = tensor<string, []>("op_4078_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4078_end_mask_0 = const()[name = tensor<string, []>("op_4078_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4078_cast_fp16 = slice_by_index(begin = var_4078_begin_0, end = var_4078_end_0, end_mask = var_4078_end_mask_0, x = k_201_cast_fp16)[name = tensor<string, []>("op_4078_cast_fp16")];
            tensor<int32, [4]> var_4082_begin_0 = const()[name = tensor<string, []>("op_4082_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4082_end_0 = const()[name = tensor<string, []>("op_4082_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4082_end_mask_0 = const()[name = tensor<string, []>("op_4082_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4082_cast_fp16 = slice_by_index(begin = var_4082_begin_0, end = var_4082_end_0, end_mask = var_4082_end_mask_0, x = q_335_cast_fp16)[name = tensor<string, []>("op_4082_cast_fp16")];
            tensor<int32, [4]> var_4086_begin_0 = const()[name = tensor<string, []>("op_4086_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4086_end_0 = const()[name = tensor<string, []>("op_4086_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4086_end_mask_0 = const()[name = tensor<string, []>("op_4086_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4086_cast_fp16 = slice_by_index(begin = var_4086_begin_0, end = var_4086_end_0, end_mask = var_4086_end_mask_0, x = k_201_cast_fp16)[name = tensor<string, []>("op_4086_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4088_cast_fp16 = mul(x = var_4082_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4088_cast_fp16")];
            tensor<int32, [5]> var_4096 = const()[name = tensor<string, []>("op_4096"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_447_cast_fp16 = reshape(shape = var_4096, x = var_4082_cast_fp16)[name = tensor<string, []>("x_447_cast_fp16")];
            tensor<int32, [2]> var_4098_split_sizes_0 = const()[name = tensor<string, []>("op_4098_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4098_axis_0 = const()[name = tensor<string, []>("op_4098_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4098_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4098_cast_fp16_1 = split(axis = var_4098_axis_0, split_sizes = var_4098_split_sizes_0, x = x_447_cast_fp16)[name = tensor<string, []>("op_4098_cast_fp16")];
            tensor<int32, [1]> squeeze_132_axes_0 = const()[name = tensor<string, []>("squeeze_132_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_132_cast_fp16 = squeeze(axes = squeeze_132_axes_0, x = var_4098_cast_fp16_0)[name = tensor<string, []>("squeeze_132_cast_fp16")];
            tensor<int32, [1]> squeeze_133_axes_0 = const()[name = tensor<string, []>("squeeze_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_133_cast_fp16 = squeeze(axes = squeeze_133_axes_0, x = var_4098_cast_fp16_1)[name = tensor<string, []>("squeeze_133_cast_fp16")];
            tensor<fp16, []> const_543_promoted_to_fp16 = const()[name = tensor<string, []>("const_543_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4101_cast_fp16 = mul(x = squeeze_133_cast_fp16, y = const_543_promoted_to_fp16)[name = tensor<string, []>("op_4101_cast_fp16")];
            tensor<int32, []> x_449_axis_0 = const()[name = tensor<string, []>("x_449_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_449_cast_fp16 = stack(axis = x_449_axis_0, values = (var_4101_cast_fp16, squeeze_132_cast_fp16))[name = tensor<string, []>("x_449_cast_fp16")];
            tensor<int32, [4]> var_4107 = const()[name = tensor<string, []>("op_4107"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4108_cast_fp16 = reshape(shape = var_4107, x = x_449_cast_fp16)[name = tensor<string, []>("op_4108_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4109_cast_fp16 = mul(x = var_4108_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4109_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_67_cast_fp16 = add(x = var_4088_cast_fp16, y = var_4109_cast_fp16)[name = tensor<string, []>("q_patches_67_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4111_cast_fp16 = mul(x = var_4086_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4111_cast_fp16")];
            tensor<int32, [5]> var_4119 = const()[name = tensor<string, []>("op_4119"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_453_cast_fp16 = reshape(shape = var_4119, x = var_4086_cast_fp16)[name = tensor<string, []>("x_453_cast_fp16")];
            tensor<int32, [2]> var_4121_split_sizes_0 = const()[name = tensor<string, []>("op_4121_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4121_axis_0 = const()[name = tensor<string, []>("op_4121_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4121_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4121_cast_fp16_1 = split(axis = var_4121_axis_0, split_sizes = var_4121_split_sizes_0, x = x_453_cast_fp16)[name = tensor<string, []>("op_4121_cast_fp16")];
            tensor<int32, [1]> squeeze_134_axes_0 = const()[name = tensor<string, []>("squeeze_134_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_134_cast_fp16 = squeeze(axes = squeeze_134_axes_0, x = var_4121_cast_fp16_0)[name = tensor<string, []>("squeeze_134_cast_fp16")];
            tensor<int32, [1]> squeeze_135_axes_0 = const()[name = tensor<string, []>("squeeze_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_135_cast_fp16 = squeeze(axes = squeeze_135_axes_0, x = var_4121_cast_fp16_1)[name = tensor<string, []>("squeeze_135_cast_fp16")];
            tensor<fp16, []> const_551_promoted_to_fp16 = const()[name = tensor<string, []>("const_551_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4124_cast_fp16 = mul(x = squeeze_135_cast_fp16, y = const_551_promoted_to_fp16)[name = tensor<string, []>("op_4124_cast_fp16")];
            tensor<int32, []> x_455_axis_0 = const()[name = tensor<string, []>("x_455_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_455_cast_fp16 = stack(axis = x_455_axis_0, values = (var_4124_cast_fp16, squeeze_134_cast_fp16))[name = tensor<string, []>("x_455_cast_fp16")];
            tensor<int32, [4]> var_4130 = const()[name = tensor<string, []>("op_4130"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4131_cast_fp16 = reshape(shape = var_4130, x = x_455_cast_fp16)[name = tensor<string, []>("op_4131_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4132_cast_fp16 = mul(x = var_4131_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4132_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_67_cast_fp16 = add(x = var_4111_cast_fp16, y = var_4132_cast_fp16)[name = tensor<string, []>("k_patches_67_cast_fp16")];
            tensor<bool, []> var_4135_interleave_0 = const()[name = tensor<string, []>("op_4135_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4135_cast_fp16 = concat(axis = var_20, interleave = var_4135_interleave_0, values = (var_4074_cast_fp16, q_patches_67_cast_fp16))[name = tensor<string, []>("op_4135_cast_fp16")];
            tensor<bool, []> var_4138_interleave_0 = const()[name = tensor<string, []>("op_4138_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4138_cast_fp16 = concat(axis = var_20, interleave = var_4138_interleave_0, values = (var_4078_cast_fp16, k_patches_67_cast_fp16))[name = tensor<string, []>("op_4138_cast_fp16")];
            tensor<fp16, []> var_4140_to_fp16 = const()[name = tensor<string, []>("op_4140_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_339_cast_fp16 = mul(x = var_4135_cast_fp16, y = var_4140_to_fp16)[name = tensor<string, []>("q_339_cast_fp16")];
            tensor<bool, []> attn_133_transpose_x_1 = const()[name = tensor<string, []>("attn_133_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_133_transpose_y_1 = const()[name = tensor<string, []>("attn_133_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_133_cast_fp16 = matmul(transpose_x = attn_133_transpose_x_1, transpose_y = attn_133_transpose_y_1, x = q_339_cast_fp16, y = var_4138_cast_fp16)[name = tensor<string, []>("attn_133_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_135_cast_fp16 = softmax(axis = var_21, x = attn_133_cast_fp16)[name = tensor<string, []>("attn_135_cast_fp16")];
            tensor<bool, []> var_4145_transpose_x_0 = const()[name = tensor<string, []>("op_4145_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4145_transpose_y_0 = const()[name = tensor<string, []>("op_4145_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4145_cast_fp16 = matmul(transpose_x = var_4145_transpose_x_0, transpose_y = var_4145_transpose_y_0, x = attn_135_cast_fp16, y = v_203_cast_fp16)[name = tensor<string, []>("op_4145_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = tensor<string, []>("op_4146_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4146_cast_fp16 = squeeze(axes = var_4146_axes_0, x = var_4145_cast_fp16)[name = tensor<string, []>("op_4146_cast_fp16")];
            tensor<string, []> var_4152_pad_type_0 = const()[name = tensor<string, []>("op_4152_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4152_strides_0 = const()[name = tensor<string, []>("op_4152_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4152_pad_0 = const()[name = tensor<string, []>("op_4152_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4152_dilations_0 = const()[name = tensor<string, []>("op_4152_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4152_groups_0 = const()[name = tensor<string, []>("op_4152_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32828480)))];
            tensor<fp16, [1, 64, 1, 197]> var_4152_cast_fp16 = conv(dilations = var_4152_dilations_0, groups = var_4152_groups_0, pad = var_4152_pad_0, pad_type = var_4152_pad_type_0, strides = var_4152_strides_0, weight = model_blocks_2_attn_q_projs_10_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4152_cast_fp16")];
            tensor<int32, [1]> var_4153_axes_0 = const()[name = tensor<string, []>("op_4153_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4153_cast_fp16 = squeeze(axes = var_4153_axes_0, x = var_4152_cast_fp16)[name = tensor<string, []>("op_4153_cast_fp16")];
            tensor<int32, [3]> q_341_perm_0 = const()[name = tensor<string, []>("q_341_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4160_pad_type_0 = const()[name = tensor<string, []>("op_4160_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4160_strides_0 = const()[name = tensor<string, []>("op_4160_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4160_pad_0 = const()[name = tensor<string, []>("op_4160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4160_dilations_0 = const()[name = tensor<string, []>("op_4160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4160_groups_0 = const()[name = tensor<string, []>("op_4160_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32926848)))];
            tensor<fp16, [1, 64, 1, 197]> var_4160_cast_fp16 = conv(dilations = var_4160_dilations_0, groups = var_4160_groups_0, pad = var_4160_pad_0, pad_type = var_4160_pad_type_0, strides = var_4160_strides_0, weight = model_blocks_2_attn_k_projs_10_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4160_cast_fp16")];
            tensor<int32, [1]> var_4161_axes_0 = const()[name = tensor<string, []>("op_4161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4161_cast_fp16 = squeeze(axes = var_4161_axes_0, x = var_4160_cast_fp16)[name = tensor<string, []>("op_4161_cast_fp16")];
            tensor<int32, [3]> k_205_perm_0 = const()[name = tensor<string, []>("k_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4168_pad_type_0 = const()[name = tensor<string, []>("op_4168_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4168_strides_0 = const()[name = tensor<string, []>("op_4168_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4168_pad_0 = const()[name = tensor<string, []>("op_4168_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4168_dilations_0 = const()[name = tensor<string, []>("op_4168_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4168_groups_0 = const()[name = tensor<string, []>("op_4168_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33025216)))];
            tensor<fp16, [1, 64, 1, 197]> var_4168_cast_fp16 = conv(dilations = var_4168_dilations_0, groups = var_4168_groups_0, pad = var_4168_pad_0, pad_type = var_4168_pad_type_0, strides = var_4168_strides_0, weight = model_blocks_2_attn_v_projs_10_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4168_cast_fp16")];
            tensor<int32, [1]> var_4169_axes_0 = const()[name = tensor<string, []>("op_4169_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4169_cast_fp16 = squeeze(axes = var_4169_axes_0, x = var_4168_cast_fp16)[name = tensor<string, []>("op_4169_cast_fp16")];
            tensor<int32, [3]> v_205_perm_0 = const()[name = tensor<string, []>("v_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33123584)))];
            tensor<fp16, [1, 197, 64]> q_341_cast_fp16 = transpose(perm = q_341_perm_0, x = var_4153_cast_fp16)[name = tensor<string, []>("transpose_398")];
            tensor<fp16, [1, 197, 64]> q_343_cast_fp16 = add(x = q_341_cast_fp16, y = model_blocks_2_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_343_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33123776)))];
            tensor<fp16, [1, 197, 64]> v_205_cast_fp16 = transpose(perm = v_205_perm_0, x = var_4169_cast_fp16)[name = tensor<string, []>("transpose_396")];
            tensor<fp16, [1, 197, 64]> v_207_cast_fp16 = add(x = v_205_cast_fp16, y = model_blocks_2_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_207_cast_fp16")];
            tensor<int32, [1]> q_345_axes_0 = const()[name = tensor<string, []>("q_345_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_345_cast_fp16 = expand_dims(axes = q_345_axes_0, x = q_343_cast_fp16)[name = tensor<string, []>("q_345_cast_fp16")];
            tensor<int32, [1]> k_207_axes_0 = const()[name = tensor<string, []>("k_207_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_205_cast_fp16 = transpose(perm = k_205_perm_0, x = var_4161_cast_fp16)[name = tensor<string, []>("transpose_397")];
            tensor<fp16, [1, 1, 197, 64]> k_207_cast_fp16 = expand_dims(axes = k_207_axes_0, x = k_205_cast_fp16)[name = tensor<string, []>("k_207_cast_fp16")];
            tensor<int32, [1]> v_209_axes_0 = const()[name = tensor<string, []>("v_209_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_209_cast_fp16 = expand_dims(axes = v_209_axes_0, x = v_207_cast_fp16)[name = tensor<string, []>("v_209_cast_fp16")];
            tensor<int32, [4]> var_4178_begin_0 = const()[name = tensor<string, []>("op_4178_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4178_end_0 = const()[name = tensor<string, []>("op_4178_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4178_end_mask_0 = const()[name = tensor<string, []>("op_4178_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, x = q_345_cast_fp16)[name = tensor<string, []>("op_4178_cast_fp16")];
            tensor<int32, [4]> var_4182_begin_0 = const()[name = tensor<string, []>("op_4182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4182_end_0 = const()[name = tensor<string, []>("op_4182_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4182_end_mask_0 = const()[name = tensor<string, []>("op_4182_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4182_cast_fp16 = slice_by_index(begin = var_4182_begin_0, end = var_4182_end_0, end_mask = var_4182_end_mask_0, x = k_207_cast_fp16)[name = tensor<string, []>("op_4182_cast_fp16")];
            tensor<int32, [4]> var_4186_begin_0 = const()[name = tensor<string, []>("op_4186_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4186_end_0 = const()[name = tensor<string, []>("op_4186_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4186_end_mask_0 = const()[name = tensor<string, []>("op_4186_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, x = q_345_cast_fp16)[name = tensor<string, []>("op_4186_cast_fp16")];
            tensor<int32, [4]> var_4190_begin_0 = const()[name = tensor<string, []>("op_4190_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4190_end_0 = const()[name = tensor<string, []>("op_4190_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4190_end_mask_0 = const()[name = tensor<string, []>("op_4190_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4190_cast_fp16 = slice_by_index(begin = var_4190_begin_0, end = var_4190_end_0, end_mask = var_4190_end_mask_0, x = k_207_cast_fp16)[name = tensor<string, []>("op_4190_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4192_cast_fp16 = mul(x = var_4186_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4192_cast_fp16")];
            tensor<int32, [5]> var_4200 = const()[name = tensor<string, []>("op_4200"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_459_cast_fp16 = reshape(shape = var_4200, x = var_4186_cast_fp16)[name = tensor<string, []>("x_459_cast_fp16")];
            tensor<int32, [2]> var_4202_split_sizes_0 = const()[name = tensor<string, []>("op_4202_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4202_axis_0 = const()[name = tensor<string, []>("op_4202_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4202_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4202_cast_fp16_1 = split(axis = var_4202_axis_0, split_sizes = var_4202_split_sizes_0, x = x_459_cast_fp16)[name = tensor<string, []>("op_4202_cast_fp16")];
            tensor<int32, [1]> squeeze_136_axes_0 = const()[name = tensor<string, []>("squeeze_136_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_136_cast_fp16 = squeeze(axes = squeeze_136_axes_0, x = var_4202_cast_fp16_0)[name = tensor<string, []>("squeeze_136_cast_fp16")];
            tensor<int32, [1]> squeeze_137_axes_0 = const()[name = tensor<string, []>("squeeze_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_137_cast_fp16 = squeeze(axes = squeeze_137_axes_0, x = var_4202_cast_fp16_1)[name = tensor<string, []>("squeeze_137_cast_fp16")];
            tensor<fp16, []> const_559_promoted_to_fp16 = const()[name = tensor<string, []>("const_559_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4205_cast_fp16 = mul(x = squeeze_137_cast_fp16, y = const_559_promoted_to_fp16)[name = tensor<string, []>("op_4205_cast_fp16")];
            tensor<int32, []> x_461_axis_0 = const()[name = tensor<string, []>("x_461_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_461_cast_fp16 = stack(axis = x_461_axis_0, values = (var_4205_cast_fp16, squeeze_136_cast_fp16))[name = tensor<string, []>("x_461_cast_fp16")];
            tensor<int32, [4]> var_4211 = const()[name = tensor<string, []>("op_4211"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4212_cast_fp16 = reshape(shape = var_4211, x = x_461_cast_fp16)[name = tensor<string, []>("op_4212_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4213_cast_fp16 = mul(x = var_4212_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4213_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_69_cast_fp16 = add(x = var_4192_cast_fp16, y = var_4213_cast_fp16)[name = tensor<string, []>("q_patches_69_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4215_cast_fp16 = mul(x = var_4190_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4215_cast_fp16")];
            tensor<int32, [5]> var_4223 = const()[name = tensor<string, []>("op_4223"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_465_cast_fp16 = reshape(shape = var_4223, x = var_4190_cast_fp16)[name = tensor<string, []>("x_465_cast_fp16")];
            tensor<int32, [2]> var_4225_split_sizes_0 = const()[name = tensor<string, []>("op_4225_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4225_axis_0 = const()[name = tensor<string, []>("op_4225_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = var_4225_split_sizes_0, x = x_465_cast_fp16)[name = tensor<string, []>("op_4225_cast_fp16")];
            tensor<int32, [1]> squeeze_138_axes_0 = const()[name = tensor<string, []>("squeeze_138_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_138_cast_fp16 = squeeze(axes = squeeze_138_axes_0, x = var_4225_cast_fp16_0)[name = tensor<string, []>("squeeze_138_cast_fp16")];
            tensor<int32, [1]> squeeze_139_axes_0 = const()[name = tensor<string, []>("squeeze_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_139_cast_fp16 = squeeze(axes = squeeze_139_axes_0, x = var_4225_cast_fp16_1)[name = tensor<string, []>("squeeze_139_cast_fp16")];
            tensor<fp16, []> const_567_promoted_to_fp16 = const()[name = tensor<string, []>("const_567_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4228_cast_fp16 = mul(x = squeeze_139_cast_fp16, y = const_567_promoted_to_fp16)[name = tensor<string, []>("op_4228_cast_fp16")];
            tensor<int32, []> x_467_axis_0 = const()[name = tensor<string, []>("x_467_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_467_cast_fp16 = stack(axis = x_467_axis_0, values = (var_4228_cast_fp16, squeeze_138_cast_fp16))[name = tensor<string, []>("x_467_cast_fp16")];
            tensor<int32, [4]> var_4234 = const()[name = tensor<string, []>("op_4234"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4235_cast_fp16 = reshape(shape = var_4234, x = x_467_cast_fp16)[name = tensor<string, []>("op_4235_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4236_cast_fp16 = mul(x = var_4235_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4236_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_69_cast_fp16 = add(x = var_4215_cast_fp16, y = var_4236_cast_fp16)[name = tensor<string, []>("k_patches_69_cast_fp16")];
            tensor<bool, []> var_4239_interleave_0 = const()[name = tensor<string, []>("op_4239_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4239_cast_fp16 = concat(axis = var_20, interleave = var_4239_interleave_0, values = (var_4178_cast_fp16, q_patches_69_cast_fp16))[name = tensor<string, []>("op_4239_cast_fp16")];
            tensor<bool, []> var_4242_interleave_0 = const()[name = tensor<string, []>("op_4242_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4242_cast_fp16 = concat(axis = var_20, interleave = var_4242_interleave_0, values = (var_4182_cast_fp16, k_patches_69_cast_fp16))[name = tensor<string, []>("op_4242_cast_fp16")];
            tensor<fp16, []> var_4244_to_fp16 = const()[name = tensor<string, []>("op_4244_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_349_cast_fp16 = mul(x = var_4239_cast_fp16, y = var_4244_to_fp16)[name = tensor<string, []>("q_349_cast_fp16")];
            tensor<bool, []> attn_137_transpose_x_1 = const()[name = tensor<string, []>("attn_137_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_137_transpose_y_1 = const()[name = tensor<string, []>("attn_137_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_137_cast_fp16 = matmul(transpose_x = attn_137_transpose_x_1, transpose_y = attn_137_transpose_y_1, x = q_349_cast_fp16, y = var_4242_cast_fp16)[name = tensor<string, []>("attn_137_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_139_cast_fp16 = softmax(axis = var_21, x = attn_137_cast_fp16)[name = tensor<string, []>("attn_139_cast_fp16")];
            tensor<bool, []> var_4249_transpose_x_0 = const()[name = tensor<string, []>("op_4249_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4249_transpose_y_0 = const()[name = tensor<string, []>("op_4249_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4249_cast_fp16 = matmul(transpose_x = var_4249_transpose_x_0, transpose_y = var_4249_transpose_y_0, x = attn_139_cast_fp16, y = v_209_cast_fp16)[name = tensor<string, []>("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4250_axes_0 = const()[name = tensor<string, []>("op_4250_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4250_cast_fp16 = squeeze(axes = var_4250_axes_0, x = var_4249_cast_fp16)[name = tensor<string, []>("op_4250_cast_fp16")];
            tensor<string, []> var_4256_pad_type_0 = const()[name = tensor<string, []>("op_4256_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4256_strides_0 = const()[name = tensor<string, []>("op_4256_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4256_pad_0 = const()[name = tensor<string, []>("op_4256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4256_dilations_0 = const()[name = tensor<string, []>("op_4256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4256_groups_0 = const()[name = tensor<string, []>("op_4256_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33123968)))];
            tensor<fp16, [1, 64, 1, 197]> var_4256_cast_fp16 = conv(dilations = var_4256_dilations_0, groups = var_4256_groups_0, pad = var_4256_pad_0, pad_type = var_4256_pad_type_0, strides = var_4256_strides_0, weight = model_blocks_2_attn_q_projs_11_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4256_cast_fp16")];
            tensor<int32, [1]> var_4257_axes_0 = const()[name = tensor<string, []>("op_4257_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4257_cast_fp16 = squeeze(axes = var_4257_axes_0, x = var_4256_cast_fp16)[name = tensor<string, []>("op_4257_cast_fp16")];
            tensor<int32, [3]> q_351_perm_0 = const()[name = tensor<string, []>("q_351_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4264_pad_type_0 = const()[name = tensor<string, []>("op_4264_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4264_strides_0 = const()[name = tensor<string, []>("op_4264_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4264_pad_0 = const()[name = tensor<string, []>("op_4264_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4264_dilations_0 = const()[name = tensor<string, []>("op_4264_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4264_groups_0 = const()[name = tensor<string, []>("op_4264_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33222336)))];
            tensor<fp16, [1, 64, 1, 197]> var_4264_cast_fp16 = conv(dilations = var_4264_dilations_0, groups = var_4264_groups_0, pad = var_4264_pad_0, pad_type = var_4264_pad_type_0, strides = var_4264_strides_0, weight = model_blocks_2_attn_k_projs_11_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4264_cast_fp16")];
            tensor<int32, [1]> var_4265_axes_0 = const()[name = tensor<string, []>("op_4265_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4265_cast_fp16 = squeeze(axes = var_4265_axes_0, x = var_4264_cast_fp16)[name = tensor<string, []>("op_4265_cast_fp16")];
            tensor<int32, [3]> k_211_perm_0 = const()[name = tensor<string, []>("k_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4272_pad_type_0 = const()[name = tensor<string, []>("op_4272_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4272_strides_0 = const()[name = tensor<string, []>("op_4272_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4272_pad_0 = const()[name = tensor<string, []>("op_4272_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4272_dilations_0 = const()[name = tensor<string, []>("op_4272_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4272_groups_0 = const()[name = tensor<string, []>("op_4272_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_2_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33320704)))];
            tensor<fp16, [1, 64, 1, 197]> var_4272_cast_fp16 = conv(dilations = var_4272_dilations_0, groups = var_4272_groups_0, pad = var_4272_pad_0, pad_type = var_4272_pad_type_0, strides = var_4272_strides_0, weight = model_blocks_2_attn_v_projs_11_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4273_axes_0 = const()[name = tensor<string, []>("op_4273_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4273_cast_fp16 = squeeze(axes = var_4273_axes_0, x = var_4272_cast_fp16)[name = tensor<string, []>("op_4273_cast_fp16")];
            tensor<int32, [3]> v_211_perm_0 = const()[name = tensor<string, []>("v_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_2_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33419072)))];
            tensor<fp16, [1, 197, 64]> q_351_cast_fp16 = transpose(perm = q_351_perm_0, x = var_4257_cast_fp16)[name = tensor<string, []>("transpose_395")];
            tensor<fp16, [1, 197, 64]> q_353_cast_fp16 = add(x = q_351_cast_fp16, y = model_blocks_2_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_353_cast_fp16")];
            tensor<fp16, [64]> model_blocks_2_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33419264)))];
            tensor<fp16, [1, 197, 64]> v_211_cast_fp16 = transpose(perm = v_211_perm_0, x = var_4273_cast_fp16)[name = tensor<string, []>("transpose_393")];
            tensor<fp16, [1, 197, 64]> v_213_cast_fp16 = add(x = v_211_cast_fp16, y = model_blocks_2_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_213_cast_fp16")];
            tensor<int32, [1]> q_355_axes_0 = const()[name = tensor<string, []>("q_355_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_355_cast_fp16 = expand_dims(axes = q_355_axes_0, x = q_353_cast_fp16)[name = tensor<string, []>("q_355_cast_fp16")];
            tensor<int32, [1]> k_213_axes_0 = const()[name = tensor<string, []>("k_213_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_211_cast_fp16 = transpose(perm = k_211_perm_0, x = var_4265_cast_fp16)[name = tensor<string, []>("transpose_394")];
            tensor<fp16, [1, 1, 197, 64]> k_213_cast_fp16 = expand_dims(axes = k_213_axes_0, x = k_211_cast_fp16)[name = tensor<string, []>("k_213_cast_fp16")];
            tensor<int32, [1]> v_215_axes_0 = const()[name = tensor<string, []>("v_215_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_215_cast_fp16 = expand_dims(axes = v_215_axes_0, x = v_213_cast_fp16)[name = tensor<string, []>("v_215_cast_fp16")];
            tensor<int32, [4]> var_4282_begin_0 = const()[name = tensor<string, []>("op_4282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4282_end_0 = const()[name = tensor<string, []>("op_4282_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4282_end_mask_0 = const()[name = tensor<string, []>("op_4282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4282_cast_fp16 = slice_by_index(begin = var_4282_begin_0, end = var_4282_end_0, end_mask = var_4282_end_mask_0, x = q_355_cast_fp16)[name = tensor<string, []>("op_4282_cast_fp16")];
            tensor<int32, [4]> var_4286_begin_0 = const()[name = tensor<string, []>("op_4286_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4286_end_0 = const()[name = tensor<string, []>("op_4286_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4286_end_mask_0 = const()[name = tensor<string, []>("op_4286_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4286_cast_fp16 = slice_by_index(begin = var_4286_begin_0, end = var_4286_end_0, end_mask = var_4286_end_mask_0, x = k_213_cast_fp16)[name = tensor<string, []>("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4290_begin_0 = const()[name = tensor<string, []>("op_4290_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4290_end_0 = const()[name = tensor<string, []>("op_4290_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4290_end_mask_0 = const()[name = tensor<string, []>("op_4290_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4290_cast_fp16 = slice_by_index(begin = var_4290_begin_0, end = var_4290_end_0, end_mask = var_4290_end_mask_0, x = q_355_cast_fp16)[name = tensor<string, []>("op_4290_cast_fp16")];
            tensor<int32, [4]> var_4294_begin_0 = const()[name = tensor<string, []>("op_4294_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4294_end_0 = const()[name = tensor<string, []>("op_4294_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4294_end_mask_0 = const()[name = tensor<string, []>("op_4294_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4294_cast_fp16 = slice_by_index(begin = var_4294_begin_0, end = var_4294_end_0, end_mask = var_4294_end_mask_0, x = k_213_cast_fp16)[name = tensor<string, []>("op_4294_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4296_cast_fp16 = mul(x = var_4290_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4296_cast_fp16")];
            tensor<int32, [5]> var_4304 = const()[name = tensor<string, []>("op_4304"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_471_cast_fp16 = reshape(shape = var_4304, x = var_4290_cast_fp16)[name = tensor<string, []>("x_471_cast_fp16")];
            tensor<int32, [2]> var_4306_split_sizes_0 = const()[name = tensor<string, []>("op_4306_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4306_axis_0 = const()[name = tensor<string, []>("op_4306_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4306_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4306_cast_fp16_1 = split(axis = var_4306_axis_0, split_sizes = var_4306_split_sizes_0, x = x_471_cast_fp16)[name = tensor<string, []>("op_4306_cast_fp16")];
            tensor<int32, [1]> squeeze_140_axes_0 = const()[name = tensor<string, []>("squeeze_140_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_140_cast_fp16 = squeeze(axes = squeeze_140_axes_0, x = var_4306_cast_fp16_0)[name = tensor<string, []>("squeeze_140_cast_fp16")];
            tensor<int32, [1]> squeeze_141_axes_0 = const()[name = tensor<string, []>("squeeze_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_141_cast_fp16 = squeeze(axes = squeeze_141_axes_0, x = var_4306_cast_fp16_1)[name = tensor<string, []>("squeeze_141_cast_fp16")];
            tensor<fp16, []> const_575_promoted_to_fp16 = const()[name = tensor<string, []>("const_575_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4309_cast_fp16 = mul(x = squeeze_141_cast_fp16, y = const_575_promoted_to_fp16)[name = tensor<string, []>("op_4309_cast_fp16")];
            tensor<int32, []> x_473_axis_0 = const()[name = tensor<string, []>("x_473_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_473_cast_fp16 = stack(axis = x_473_axis_0, values = (var_4309_cast_fp16, squeeze_140_cast_fp16))[name = tensor<string, []>("x_473_cast_fp16")];
            tensor<int32, [4]> var_4315 = const()[name = tensor<string, []>("op_4315"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4316_cast_fp16 = reshape(shape = var_4315, x = x_473_cast_fp16)[name = tensor<string, []>("op_4316_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4317_cast_fp16 = mul(x = var_4316_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4317_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_71_cast_fp16 = add(x = var_4296_cast_fp16, y = var_4317_cast_fp16)[name = tensor<string, []>("q_patches_71_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4319_cast_fp16 = mul(x = var_4294_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4319_cast_fp16")];
            tensor<int32, [5]> var_4327 = const()[name = tensor<string, []>("op_4327"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_477_cast_fp16 = reshape(shape = var_4327, x = var_4294_cast_fp16)[name = tensor<string, []>("x_477_cast_fp16")];
            tensor<int32, [2]> var_4329_split_sizes_0 = const()[name = tensor<string, []>("op_4329_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4329_axis_0 = const()[name = tensor<string, []>("op_4329_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4329_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4329_cast_fp16_1 = split(axis = var_4329_axis_0, split_sizes = var_4329_split_sizes_0, x = x_477_cast_fp16)[name = tensor<string, []>("op_4329_cast_fp16")];
            tensor<int32, [1]> squeeze_142_axes_0 = const()[name = tensor<string, []>("squeeze_142_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_142_cast_fp16 = squeeze(axes = squeeze_142_axes_0, x = var_4329_cast_fp16_0)[name = tensor<string, []>("squeeze_142_cast_fp16")];
            tensor<int32, [1]> squeeze_143_axes_0 = const()[name = tensor<string, []>("squeeze_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_143_cast_fp16 = squeeze(axes = squeeze_143_axes_0, x = var_4329_cast_fp16_1)[name = tensor<string, []>("squeeze_143_cast_fp16")];
            tensor<fp16, []> const_583_promoted_to_fp16 = const()[name = tensor<string, []>("const_583_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4332_cast_fp16 = mul(x = squeeze_143_cast_fp16, y = const_583_promoted_to_fp16)[name = tensor<string, []>("op_4332_cast_fp16")];
            tensor<int32, []> x_479_axis_0 = const()[name = tensor<string, []>("x_479_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_479_cast_fp16 = stack(axis = x_479_axis_0, values = (var_4332_cast_fp16, squeeze_142_cast_fp16))[name = tensor<string, []>("x_479_cast_fp16")];
            tensor<int32, [4]> var_4338 = const()[name = tensor<string, []>("op_4338"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4339_cast_fp16 = reshape(shape = var_4338, x = x_479_cast_fp16)[name = tensor<string, []>("op_4339_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4340_cast_fp16 = mul(x = var_4339_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4340_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_71_cast_fp16 = add(x = var_4319_cast_fp16, y = var_4340_cast_fp16)[name = tensor<string, []>("k_patches_71_cast_fp16")];
            tensor<bool, []> var_4343_interleave_0 = const()[name = tensor<string, []>("op_4343_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4343_cast_fp16 = concat(axis = var_20, interleave = var_4343_interleave_0, values = (var_4282_cast_fp16, q_patches_71_cast_fp16))[name = tensor<string, []>("op_4343_cast_fp16")];
            tensor<bool, []> var_4346_interleave_0 = const()[name = tensor<string, []>("op_4346_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4346_cast_fp16 = concat(axis = var_20, interleave = var_4346_interleave_0, values = (var_4286_cast_fp16, k_patches_71_cast_fp16))[name = tensor<string, []>("op_4346_cast_fp16")];
            tensor<fp16, []> var_4348_to_fp16 = const()[name = tensor<string, []>("op_4348_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_359_cast_fp16 = mul(x = var_4343_cast_fp16, y = var_4348_to_fp16)[name = tensor<string, []>("q_359_cast_fp16")];
            tensor<bool, []> attn_141_transpose_x_1 = const()[name = tensor<string, []>("attn_141_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_141_transpose_y_1 = const()[name = tensor<string, []>("attn_141_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_141_cast_fp16 = matmul(transpose_x = attn_141_transpose_x_1, transpose_y = attn_141_transpose_y_1, x = q_359_cast_fp16, y = var_4346_cast_fp16)[name = tensor<string, []>("attn_141_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_143_cast_fp16 = softmax(axis = var_21, x = attn_141_cast_fp16)[name = tensor<string, []>("attn_143_cast_fp16")];
            tensor<bool, []> var_4353_transpose_x_0 = const()[name = tensor<string, []>("op_4353_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4353_transpose_y_0 = const()[name = tensor<string, []>("op_4353_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4353_cast_fp16 = matmul(transpose_x = var_4353_transpose_x_0, transpose_y = var_4353_transpose_y_0, x = attn_143_cast_fp16, y = v_215_cast_fp16)[name = tensor<string, []>("op_4353_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_5_cast_fp16 = squeeze(axes = out_5_axes_0, x = var_4353_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<bool, []> input_41_interleave_0 = const()[name = tensor<string, []>("input_41_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_41_cast_fp16 = concat(axis = var_21, interleave = input_41_interleave_0, values = (var_3210_cast_fp16, var_3314_cast_fp16, var_3418_cast_fp16, var_3522_cast_fp16, var_3626_cast_fp16, var_3730_cast_fp16, var_3834_cast_fp16, var_3938_cast_fp16, var_4042_cast_fp16, var_4146_cast_fp16, var_4250_cast_fp16, out_5_cast_fp16))[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<int32, [1]> x_481_axes_0 = const()[name = tensor<string, []>("x_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_2_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33419456)))];
            tensor<fp16, [768]> model_blocks_2_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33421056)))];
            tensor<fp16, [1, 197, 768]> x_481_cast_fp16 = layer_norm(axes = x_481_axes_0, beta = model_blocks_2_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_2_attn_inner_attn_ln_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("x_481_cast_fp16")];
            tensor<int32, [4]> var_4361 = const()[name = tensor<string, []>("op_4361"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_483_cast_fp16 = reshape(shape = var_4361, x = x_481_cast_fp16)[name = tensor<string, []>("x_483_cast_fp16")];
            tensor<int32, [4]> input_43_perm_0 = const()[name = tensor<string, []>("input_43_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_485_pad_type_0 = const()[name = tensor<string, []>("x_485_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_485_strides_0 = const()[name = tensor<string, []>("x_485_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_485_pad_0 = const()[name = tensor<string, []>("x_485_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_485_dilations_0 = const()[name = tensor<string, []>("x_485_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_485_groups_0 = const()[name = tensor<string, []>("x_485_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_2_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33422656)))];
            tensor<fp16, [768]> model_blocks_2_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34602368)))];
            tensor<fp16, [1, 768, 1, 197]> input_43_cast_fp16 = transpose(perm = input_43_perm_0, x = x_483_cast_fp16)[name = tensor<string, []>("transpose_392")];
            tensor<fp16, [1, 768, 1, 197]> x_485_cast_fp16 = conv(bias = model_blocks_2_attn_proj_bias_to_fp16, dilations = x_485_dilations_0, groups = x_485_groups_0, pad = x_485_pad_0, pad_type = x_485_pad_type_0, strides = x_485_strides_0, weight = model_blocks_2_attn_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("x_485_cast_fp16")];
            tensor<int32, [4]> x_487_perm_0 = const()[name = tensor<string, []>("x_487_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_4372 = const()[name = tensor<string, []>("op_4372"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_487_cast_fp16 = transpose(perm = x_487_perm_0, x = x_485_cast_fp16)[name = tensor<string, []>("transpose_391")];
            tensor<fp16, [1, 197, 768]> var_4373_cast_fp16 = reshape(shape = var_4372, x = x_487_cast_fp16)[name = tensor<string, []>("op_4373_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_45_cast_fp16 = add(x = input_37_cast_fp16, y = var_4373_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [1]> x_489_axes_0 = const()[name = tensor<string, []>("x_489_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34603968)))];
            tensor<fp16, [768]> model_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34605568)))];
            tensor<fp16, [1, 197, 768]> x_489_cast_fp16 = layer_norm(axes = x_489_axes_0, beta = model_blocks_2_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_2_norm2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("x_489_cast_fp16")];
            tensor<int32, [4]> var_4385 = const()[name = tensor<string, []>("op_4385"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_491_cast_fp16 = reshape(shape = var_4385, x = x_489_cast_fp16)[name = tensor<string, []>("x_491_cast_fp16")];
            tensor<int32, [4]> input_47_perm_0 = const()[name = tensor<string, []>("input_47_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_49_pad_type_0 = const()[name = tensor<string, []>("input_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_49_strides_0 = const()[name = tensor<string, []>("input_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_49_dilations_0 = const()[name = tensor<string, []>("input_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_49_groups_0 = const()[name = tensor<string, []>("input_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_2_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34607168)))];
            tensor<fp16, [2048]> model_blocks_2_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37752960)))];
            tensor<fp16, [1, 768, 1, 197]> input_47_cast_fp16 = transpose(perm = input_47_perm_0, x = x_491_cast_fp16)[name = tensor<string, []>("transpose_390")];
            tensor<fp16, [1, 2048, 1, 197]> input_49_cast_fp16 = conv(bias = model_blocks_2_mlp_w1_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = model_blocks_2_mlp_w1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<string, []> x2_149_pad_type_0 = const()[name = tensor<string, []>("x2_149_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_149_strides_0 = const()[name = tensor<string, []>("x2_149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_149_pad_0 = const()[name = tensor<string, []>("x2_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_149_dilations_0 = const()[name = tensor<string, []>("x2_149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_149_groups_0 = const()[name = tensor<string, []>("x2_149_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_2_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37757120)))];
            tensor<fp16, [2048]> model_blocks_2_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40902912)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_149_cast_fp16 = conv(bias = model_blocks_2_mlp_w2_bias_to_fp16, dilations = x2_149_dilations_0, groups = x2_149_groups_0, pad = x2_149_pad_0, pad_type = x2_149_pad_type_0, strides = x2_149_strides_0, weight = model_blocks_2_mlp_w2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x2_149_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_4402_cast_fp16 = silu(x = input_49_cast_fp16)[name = tensor<string, []>("op_4402_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_17_cast_fp16 = mul(x = var_4402_cast_fp16, y = x2_149_cast_fp16)[name = tensor<string, []>("hidden_17_cast_fp16")];
            tensor<int32, [4]> hidden_19_perm_0 = const()[name = tensor<string, []>("hidden_19_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_4405 = const()[name = tensor<string, []>("op_4405"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_19_cast_fp16 = transpose(perm = hidden_19_perm_0, x = hidden_17_cast_fp16)[name = tensor<string, []>("transpose_389")];
            tensor<fp16, [1, 197, 2048]> input_51_cast_fp16 = reshape(shape = var_4405, x = hidden_19_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<int32, [1]> hidden_21_axes_0 = const()[name = tensor<string, []>("hidden_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_2_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40907072)))];
            tensor<fp16, [2048]> model_blocks_2_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40911232)))];
            tensor<fp16, [1, 197, 2048]> hidden_21_cast_fp16 = layer_norm(axes = hidden_21_axes_0, beta = model_blocks_2_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_2_mlp_ffn_ln_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("hidden_21_cast_fp16")];
            tensor<int32, [4]> var_4411 = const()[name = tensor<string, []>("op_4411"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_23_cast_fp16 = reshape(shape = var_4411, x = hidden_21_cast_fp16)[name = tensor<string, []>("hidden_23_cast_fp16")];
            tensor<int32, [4]> input_53_perm_0 = const()[name = tensor<string, []>("input_53_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_493_pad_type_0 = const()[name = tensor<string, []>("x_493_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_493_strides_0 = const()[name = tensor<string, []>("x_493_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_493_pad_0 = const()[name = tensor<string, []>("x_493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_493_dilations_0 = const()[name = tensor<string, []>("x_493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_493_groups_0 = const()[name = tensor<string, []>("x_493_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_2_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40915392)))];
            tensor<fp16, [768]> model_blocks_2_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_2_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44061184)))];
            tensor<fp16, [1, 2048, 1, 197]> input_53_cast_fp16 = transpose(perm = input_53_perm_0, x = hidden_23_cast_fp16)[name = tensor<string, []>("transpose_388")];
            tensor<fp16, [1, 768, 1, 197]> x_493_cast_fp16 = conv(bias = model_blocks_2_mlp_w3_bias_to_fp16, dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = model_blocks_2_mlp_w3_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("x_493_cast_fp16")];
            tensor<int32, [4]> x_495_perm_0 = const()[name = tensor<string, []>("x_495_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_4422 = const()[name = tensor<string, []>("op_4422"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_495_cast_fp16 = transpose(perm = x_495_perm_0, x = x_493_cast_fp16)[name = tensor<string, []>("transpose_387")];
            tensor<fp16, [1, 197, 768]> var_4423_cast_fp16 = reshape(shape = var_4422, x = x_495_cast_fp16)[name = tensor<string, []>("op_4423_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_55_cast_fp16 = add(x = input_45_cast_fp16, y = var_4423_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [1]> x_497_axes_0 = const()[name = tensor<string, []>("x_497_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44062784)))];
            tensor<fp16, [768]> model_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44064384)))];
            tensor<fp16, [1, 197, 768]> x_497_cast_fp16 = layer_norm(axes = x_497_axes_0, beta = model_blocks_3_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_3_norm1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_497_cast_fp16")];
            tensor<int32, [4]> var_4557 = const()[name = tensor<string, []>("op_4557"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_499_cast_fp16 = reshape(shape = var_4557, x = x_497_cast_fp16)[name = tensor<string, []>("x_499_cast_fp16")];
            tensor<int32, [4]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_4565_pad_type_0 = const()[name = tensor<string, []>("op_4565_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4565_strides_0 = const()[name = tensor<string, []>("op_4565_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4565_pad_0 = const()[name = tensor<string, []>("op_4565_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4565_dilations_0 = const()[name = tensor<string, []>("op_4565_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4565_groups_0 = const()[name = tensor<string, []>("op_4565_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44065984)))];
            tensor<fp16, [1, 768, 1, 197]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_499_cast_fp16)[name = tensor<string, []>("transpose_386")];
            tensor<fp16, [1, 64, 1, 197]> var_4565_cast_fp16 = conv(dilations = var_4565_dilations_0, groups = var_4565_groups_0, pad = var_4565_pad_0, pad_type = var_4565_pad_type_0, strides = var_4565_strides_0, weight = model_blocks_3_attn_q_projs_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4565_cast_fp16")];
            tensor<int32, [1]> var_4566_axes_0 = const()[name = tensor<string, []>("op_4566_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4566_cast_fp16 = squeeze(axes = var_4566_axes_0, x = var_4565_cast_fp16)[name = tensor<string, []>("op_4566_cast_fp16")];
            tensor<int32, [3]> q_361_perm_0 = const()[name = tensor<string, []>("q_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4573_pad_type_0 = const()[name = tensor<string, []>("op_4573_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4573_strides_0 = const()[name = tensor<string, []>("op_4573_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4573_pad_0 = const()[name = tensor<string, []>("op_4573_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4573_dilations_0 = const()[name = tensor<string, []>("op_4573_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4573_groups_0 = const()[name = tensor<string, []>("op_4573_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44164352)))];
            tensor<fp16, [1, 64, 1, 197]> var_4573_cast_fp16 = conv(dilations = var_4573_dilations_0, groups = var_4573_groups_0, pad = var_4573_pad_0, pad_type = var_4573_pad_type_0, strides = var_4573_strides_0, weight = model_blocks_3_attn_k_projs_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4573_cast_fp16")];
            tensor<int32, [1]> var_4574_axes_0 = const()[name = tensor<string, []>("op_4574_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4574_cast_fp16 = squeeze(axes = var_4574_axes_0, x = var_4573_cast_fp16)[name = tensor<string, []>("op_4574_cast_fp16")];
            tensor<int32, [3]> k_217_perm_0 = const()[name = tensor<string, []>("k_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4581_pad_type_0 = const()[name = tensor<string, []>("op_4581_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4581_strides_0 = const()[name = tensor<string, []>("op_4581_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4581_pad_0 = const()[name = tensor<string, []>("op_4581_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4581_dilations_0 = const()[name = tensor<string, []>("op_4581_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4581_groups_0 = const()[name = tensor<string, []>("op_4581_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44262720)))];
            tensor<fp16, [1, 64, 1, 197]> var_4581_cast_fp16 = conv(dilations = var_4581_dilations_0, groups = var_4581_groups_0, pad = var_4581_pad_0, pad_type = var_4581_pad_type_0, strides = var_4581_strides_0, weight = model_blocks_3_attn_v_projs_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4581_cast_fp16")];
            tensor<int32, [1]> var_4582_axes_0 = const()[name = tensor<string, []>("op_4582_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4582_cast_fp16 = squeeze(axes = var_4582_axes_0, x = var_4581_cast_fp16)[name = tensor<string, []>("op_4582_cast_fp16")];
            tensor<int32, [3]> v_217_perm_0 = const()[name = tensor<string, []>("v_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44361088)))];
            tensor<fp16, [1, 197, 64]> q_361_cast_fp16 = transpose(perm = q_361_perm_0, x = var_4566_cast_fp16)[name = tensor<string, []>("transpose_385")];
            tensor<fp16, [1, 197, 64]> q_363_cast_fp16 = add(x = q_361_cast_fp16, y = model_blocks_3_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_363_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44361280)))];
            tensor<fp16, [1, 197, 64]> v_217_cast_fp16 = transpose(perm = v_217_perm_0, x = var_4582_cast_fp16)[name = tensor<string, []>("transpose_383")];
            tensor<fp16, [1, 197, 64]> v_219_cast_fp16 = add(x = v_217_cast_fp16, y = model_blocks_3_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_219_cast_fp16")];
            tensor<int32, [1]> q_365_axes_0 = const()[name = tensor<string, []>("q_365_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_365_cast_fp16 = expand_dims(axes = q_365_axes_0, x = q_363_cast_fp16)[name = tensor<string, []>("q_365_cast_fp16")];
            tensor<int32, [1]> k_219_axes_0 = const()[name = tensor<string, []>("k_219_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_217_cast_fp16 = transpose(perm = k_217_perm_0, x = var_4574_cast_fp16)[name = tensor<string, []>("transpose_384")];
            tensor<fp16, [1, 1, 197, 64]> k_219_cast_fp16 = expand_dims(axes = k_219_axes_0, x = k_217_cast_fp16)[name = tensor<string, []>("k_219_cast_fp16")];
            tensor<int32, [1]> v_221_axes_0 = const()[name = tensor<string, []>("v_221_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_221_cast_fp16 = expand_dims(axes = v_221_axes_0, x = v_219_cast_fp16)[name = tensor<string, []>("v_221_cast_fp16")];
            tensor<int32, [4]> var_4591_begin_0 = const()[name = tensor<string, []>("op_4591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4591_end_0 = const()[name = tensor<string, []>("op_4591_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4591_end_mask_0 = const()[name = tensor<string, []>("op_4591_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4591_cast_fp16 = slice_by_index(begin = var_4591_begin_0, end = var_4591_end_0, end_mask = var_4591_end_mask_0, x = q_365_cast_fp16)[name = tensor<string, []>("op_4591_cast_fp16")];
            tensor<int32, [4]> var_4595_begin_0 = const()[name = tensor<string, []>("op_4595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4595_end_0 = const()[name = tensor<string, []>("op_4595_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4595_end_mask_0 = const()[name = tensor<string, []>("op_4595_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4595_cast_fp16 = slice_by_index(begin = var_4595_begin_0, end = var_4595_end_0, end_mask = var_4595_end_mask_0, x = k_219_cast_fp16)[name = tensor<string, []>("op_4595_cast_fp16")];
            tensor<int32, [4]> var_4599_begin_0 = const()[name = tensor<string, []>("op_4599_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4599_end_0 = const()[name = tensor<string, []>("op_4599_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4599_end_mask_0 = const()[name = tensor<string, []>("op_4599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4599_cast_fp16 = slice_by_index(begin = var_4599_begin_0, end = var_4599_end_0, end_mask = var_4599_end_mask_0, x = q_365_cast_fp16)[name = tensor<string, []>("op_4599_cast_fp16")];
            tensor<int32, [4]> var_4603_begin_0 = const()[name = tensor<string, []>("op_4603_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4603_end_0 = const()[name = tensor<string, []>("op_4603_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4603_end_mask_0 = const()[name = tensor<string, []>("op_4603_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4603_cast_fp16 = slice_by_index(begin = var_4603_begin_0, end = var_4603_end_0, end_mask = var_4603_end_mask_0, x = k_219_cast_fp16)[name = tensor<string, []>("op_4603_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4605_cast_fp16 = mul(x = var_4599_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4605_cast_fp16")];
            tensor<int32, [5]> var_4613 = const()[name = tensor<string, []>("op_4613"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_503_cast_fp16 = reshape(shape = var_4613, x = var_4599_cast_fp16)[name = tensor<string, []>("x_503_cast_fp16")];
            tensor<int32, [2]> var_4615_split_sizes_0 = const()[name = tensor<string, []>("op_4615_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4615_axis_0 = const()[name = tensor<string, []>("op_4615_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4615_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4615_cast_fp16_1 = split(axis = var_4615_axis_0, split_sizes = var_4615_split_sizes_0, x = x_503_cast_fp16)[name = tensor<string, []>("op_4615_cast_fp16")];
            tensor<int32, [1]> squeeze_144_axes_0 = const()[name = tensor<string, []>("squeeze_144_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_144_cast_fp16 = squeeze(axes = squeeze_144_axes_0, x = var_4615_cast_fp16_0)[name = tensor<string, []>("squeeze_144_cast_fp16")];
            tensor<int32, [1]> squeeze_145_axes_0 = const()[name = tensor<string, []>("squeeze_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_145_cast_fp16 = squeeze(axes = squeeze_145_axes_0, x = var_4615_cast_fp16_1)[name = tensor<string, []>("squeeze_145_cast_fp16")];
            tensor<fp16, []> const_595_promoted_to_fp16 = const()[name = tensor<string, []>("const_595_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4618_cast_fp16 = mul(x = squeeze_145_cast_fp16, y = const_595_promoted_to_fp16)[name = tensor<string, []>("op_4618_cast_fp16")];
            tensor<int32, []> x_505_axis_0 = const()[name = tensor<string, []>("x_505_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_505_cast_fp16 = stack(axis = x_505_axis_0, values = (var_4618_cast_fp16, squeeze_144_cast_fp16))[name = tensor<string, []>("x_505_cast_fp16")];
            tensor<int32, [4]> var_4624 = const()[name = tensor<string, []>("op_4624"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4625_cast_fp16 = reshape(shape = var_4624, x = x_505_cast_fp16)[name = tensor<string, []>("op_4625_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4626_cast_fp16 = mul(x = var_4625_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4626_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_73_cast_fp16 = add(x = var_4605_cast_fp16, y = var_4626_cast_fp16)[name = tensor<string, []>("q_patches_73_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4628_cast_fp16 = mul(x = var_4603_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4628_cast_fp16")];
            tensor<int32, [5]> var_4636 = const()[name = tensor<string, []>("op_4636"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_509_cast_fp16 = reshape(shape = var_4636, x = var_4603_cast_fp16)[name = tensor<string, []>("x_509_cast_fp16")];
            tensor<int32, [2]> var_4638_split_sizes_0 = const()[name = tensor<string, []>("op_4638_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4638_axis_0 = const()[name = tensor<string, []>("op_4638_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4638_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4638_cast_fp16_1 = split(axis = var_4638_axis_0, split_sizes = var_4638_split_sizes_0, x = x_509_cast_fp16)[name = tensor<string, []>("op_4638_cast_fp16")];
            tensor<int32, [1]> squeeze_146_axes_0 = const()[name = tensor<string, []>("squeeze_146_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_146_cast_fp16 = squeeze(axes = squeeze_146_axes_0, x = var_4638_cast_fp16_0)[name = tensor<string, []>("squeeze_146_cast_fp16")];
            tensor<int32, [1]> squeeze_147_axes_0 = const()[name = tensor<string, []>("squeeze_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_147_cast_fp16 = squeeze(axes = squeeze_147_axes_0, x = var_4638_cast_fp16_1)[name = tensor<string, []>("squeeze_147_cast_fp16")];
            tensor<fp16, []> const_603_promoted_to_fp16 = const()[name = tensor<string, []>("const_603_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4641_cast_fp16 = mul(x = squeeze_147_cast_fp16, y = const_603_promoted_to_fp16)[name = tensor<string, []>("op_4641_cast_fp16")];
            tensor<int32, []> x_511_axis_0 = const()[name = tensor<string, []>("x_511_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_511_cast_fp16 = stack(axis = x_511_axis_0, values = (var_4641_cast_fp16, squeeze_146_cast_fp16))[name = tensor<string, []>("x_511_cast_fp16")];
            tensor<int32, [4]> var_4647 = const()[name = tensor<string, []>("op_4647"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4648_cast_fp16 = reshape(shape = var_4647, x = x_511_cast_fp16)[name = tensor<string, []>("op_4648_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4649_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_73_cast_fp16 = add(x = var_4628_cast_fp16, y = var_4649_cast_fp16)[name = tensor<string, []>("k_patches_73_cast_fp16")];
            tensor<bool, []> var_4652_interleave_0 = const()[name = tensor<string, []>("op_4652_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4652_cast_fp16 = concat(axis = var_20, interleave = var_4652_interleave_0, values = (var_4591_cast_fp16, q_patches_73_cast_fp16))[name = tensor<string, []>("op_4652_cast_fp16")];
            tensor<bool, []> var_4655_interleave_0 = const()[name = tensor<string, []>("op_4655_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4655_cast_fp16 = concat(axis = var_20, interleave = var_4655_interleave_0, values = (var_4595_cast_fp16, k_patches_73_cast_fp16))[name = tensor<string, []>("op_4655_cast_fp16")];
            tensor<fp16, []> var_4657_to_fp16 = const()[name = tensor<string, []>("op_4657_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_369_cast_fp16 = mul(x = var_4652_cast_fp16, y = var_4657_to_fp16)[name = tensor<string, []>("q_369_cast_fp16")];
            tensor<bool, []> attn_145_transpose_x_1 = const()[name = tensor<string, []>("attn_145_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_145_transpose_y_1 = const()[name = tensor<string, []>("attn_145_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_145_cast_fp16 = matmul(transpose_x = attn_145_transpose_x_1, transpose_y = attn_145_transpose_y_1, x = q_369_cast_fp16, y = var_4655_cast_fp16)[name = tensor<string, []>("attn_145_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_147_cast_fp16 = softmax(axis = var_21, x = attn_145_cast_fp16)[name = tensor<string, []>("attn_147_cast_fp16")];
            tensor<bool, []> var_4662_transpose_x_0 = const()[name = tensor<string, []>("op_4662_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4662_transpose_y_0 = const()[name = tensor<string, []>("op_4662_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4662_cast_fp16 = matmul(transpose_x = var_4662_transpose_x_0, transpose_y = var_4662_transpose_y_0, x = attn_147_cast_fp16, y = v_221_cast_fp16)[name = tensor<string, []>("op_4662_cast_fp16")];
            tensor<int32, [1]> var_4663_axes_0 = const()[name = tensor<string, []>("op_4663_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4663_cast_fp16 = squeeze(axes = var_4663_axes_0, x = var_4662_cast_fp16)[name = tensor<string, []>("op_4663_cast_fp16")];
            tensor<string, []> var_4669_pad_type_0 = const()[name = tensor<string, []>("op_4669_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4669_strides_0 = const()[name = tensor<string, []>("op_4669_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4669_pad_0 = const()[name = tensor<string, []>("op_4669_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4669_dilations_0 = const()[name = tensor<string, []>("op_4669_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4669_groups_0 = const()[name = tensor<string, []>("op_4669_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44361472)))];
            tensor<fp16, [1, 64, 1, 197]> var_4669_cast_fp16 = conv(dilations = var_4669_dilations_0, groups = var_4669_groups_0, pad = var_4669_pad_0, pad_type = var_4669_pad_type_0, strides = var_4669_strides_0, weight = model_blocks_3_attn_q_projs_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4669_cast_fp16")];
            tensor<int32, [1]> var_4670_axes_0 = const()[name = tensor<string, []>("op_4670_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4670_cast_fp16 = squeeze(axes = var_4670_axes_0, x = var_4669_cast_fp16)[name = tensor<string, []>("op_4670_cast_fp16")];
            tensor<int32, [3]> q_371_perm_0 = const()[name = tensor<string, []>("q_371_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4677_pad_type_0 = const()[name = tensor<string, []>("op_4677_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4677_strides_0 = const()[name = tensor<string, []>("op_4677_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4677_pad_0 = const()[name = tensor<string, []>("op_4677_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4677_dilations_0 = const()[name = tensor<string, []>("op_4677_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4677_groups_0 = const()[name = tensor<string, []>("op_4677_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44459840)))];
            tensor<fp16, [1, 64, 1, 197]> var_4677_cast_fp16 = conv(dilations = var_4677_dilations_0, groups = var_4677_groups_0, pad = var_4677_pad_0, pad_type = var_4677_pad_type_0, strides = var_4677_strides_0, weight = model_blocks_3_attn_k_projs_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4677_cast_fp16")];
            tensor<int32, [1]> var_4678_axes_0 = const()[name = tensor<string, []>("op_4678_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4678_cast_fp16 = squeeze(axes = var_4678_axes_0, x = var_4677_cast_fp16)[name = tensor<string, []>("op_4678_cast_fp16")];
            tensor<int32, [3]> k_223_perm_0 = const()[name = tensor<string, []>("k_223_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4685_pad_type_0 = const()[name = tensor<string, []>("op_4685_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4685_strides_0 = const()[name = tensor<string, []>("op_4685_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4685_pad_0 = const()[name = tensor<string, []>("op_4685_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4685_dilations_0 = const()[name = tensor<string, []>("op_4685_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4685_groups_0 = const()[name = tensor<string, []>("op_4685_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44558208)))];
            tensor<fp16, [1, 64, 1, 197]> var_4685_cast_fp16 = conv(dilations = var_4685_dilations_0, groups = var_4685_groups_0, pad = var_4685_pad_0, pad_type = var_4685_pad_type_0, strides = var_4685_strides_0, weight = model_blocks_3_attn_v_projs_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4685_cast_fp16")];
            tensor<int32, [1]> var_4686_axes_0 = const()[name = tensor<string, []>("op_4686_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4686_cast_fp16 = squeeze(axes = var_4686_axes_0, x = var_4685_cast_fp16)[name = tensor<string, []>("op_4686_cast_fp16")];
            tensor<int32, [3]> v_223_perm_0 = const()[name = tensor<string, []>("v_223_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44656576)))];
            tensor<fp16, [1, 197, 64]> q_371_cast_fp16 = transpose(perm = q_371_perm_0, x = var_4670_cast_fp16)[name = tensor<string, []>("transpose_382")];
            tensor<fp16, [1, 197, 64]> q_373_cast_fp16 = add(x = q_371_cast_fp16, y = model_blocks_3_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_373_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44656768)))];
            tensor<fp16, [1, 197, 64]> v_223_cast_fp16 = transpose(perm = v_223_perm_0, x = var_4686_cast_fp16)[name = tensor<string, []>("transpose_380")];
            tensor<fp16, [1, 197, 64]> v_225_cast_fp16 = add(x = v_223_cast_fp16, y = model_blocks_3_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_225_cast_fp16")];
            tensor<int32, [1]> q_375_axes_0 = const()[name = tensor<string, []>("q_375_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_375_cast_fp16 = expand_dims(axes = q_375_axes_0, x = q_373_cast_fp16)[name = tensor<string, []>("q_375_cast_fp16")];
            tensor<int32, [1]> k_225_axes_0 = const()[name = tensor<string, []>("k_225_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_223_cast_fp16 = transpose(perm = k_223_perm_0, x = var_4678_cast_fp16)[name = tensor<string, []>("transpose_381")];
            tensor<fp16, [1, 1, 197, 64]> k_225_cast_fp16 = expand_dims(axes = k_225_axes_0, x = k_223_cast_fp16)[name = tensor<string, []>("k_225_cast_fp16")];
            tensor<int32, [1]> v_227_axes_0 = const()[name = tensor<string, []>("v_227_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_227_cast_fp16 = expand_dims(axes = v_227_axes_0, x = v_225_cast_fp16)[name = tensor<string, []>("v_227_cast_fp16")];
            tensor<int32, [4]> var_4695_begin_0 = const()[name = tensor<string, []>("op_4695_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4695_end_0 = const()[name = tensor<string, []>("op_4695_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4695_end_mask_0 = const()[name = tensor<string, []>("op_4695_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4695_cast_fp16 = slice_by_index(begin = var_4695_begin_0, end = var_4695_end_0, end_mask = var_4695_end_mask_0, x = q_375_cast_fp16)[name = tensor<string, []>("op_4695_cast_fp16")];
            tensor<int32, [4]> var_4699_begin_0 = const()[name = tensor<string, []>("op_4699_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4699_end_0 = const()[name = tensor<string, []>("op_4699_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4699_end_mask_0 = const()[name = tensor<string, []>("op_4699_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4699_cast_fp16 = slice_by_index(begin = var_4699_begin_0, end = var_4699_end_0, end_mask = var_4699_end_mask_0, x = k_225_cast_fp16)[name = tensor<string, []>("op_4699_cast_fp16")];
            tensor<int32, [4]> var_4703_begin_0 = const()[name = tensor<string, []>("op_4703_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4703_end_0 = const()[name = tensor<string, []>("op_4703_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4703_end_mask_0 = const()[name = tensor<string, []>("op_4703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4703_cast_fp16 = slice_by_index(begin = var_4703_begin_0, end = var_4703_end_0, end_mask = var_4703_end_mask_0, x = q_375_cast_fp16)[name = tensor<string, []>("op_4703_cast_fp16")];
            tensor<int32, [4]> var_4707_begin_0 = const()[name = tensor<string, []>("op_4707_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4707_end_0 = const()[name = tensor<string, []>("op_4707_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4707_end_mask_0 = const()[name = tensor<string, []>("op_4707_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4707_cast_fp16 = slice_by_index(begin = var_4707_begin_0, end = var_4707_end_0, end_mask = var_4707_end_mask_0, x = k_225_cast_fp16)[name = tensor<string, []>("op_4707_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4709_cast_fp16 = mul(x = var_4703_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4709_cast_fp16")];
            tensor<int32, [5]> var_4717 = const()[name = tensor<string, []>("op_4717"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_515_cast_fp16 = reshape(shape = var_4717, x = var_4703_cast_fp16)[name = tensor<string, []>("x_515_cast_fp16")];
            tensor<int32, [2]> var_4719_split_sizes_0 = const()[name = tensor<string, []>("op_4719_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4719_axis_0 = const()[name = tensor<string, []>("op_4719_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4719_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4719_cast_fp16_1 = split(axis = var_4719_axis_0, split_sizes = var_4719_split_sizes_0, x = x_515_cast_fp16)[name = tensor<string, []>("op_4719_cast_fp16")];
            tensor<int32, [1]> squeeze_148_axes_0 = const()[name = tensor<string, []>("squeeze_148_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_148_cast_fp16 = squeeze(axes = squeeze_148_axes_0, x = var_4719_cast_fp16_0)[name = tensor<string, []>("squeeze_148_cast_fp16")];
            tensor<int32, [1]> squeeze_149_axes_0 = const()[name = tensor<string, []>("squeeze_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_149_cast_fp16 = squeeze(axes = squeeze_149_axes_0, x = var_4719_cast_fp16_1)[name = tensor<string, []>("squeeze_149_cast_fp16")];
            tensor<fp16, []> const_611_promoted_to_fp16 = const()[name = tensor<string, []>("const_611_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4722_cast_fp16 = mul(x = squeeze_149_cast_fp16, y = const_611_promoted_to_fp16)[name = tensor<string, []>("op_4722_cast_fp16")];
            tensor<int32, []> x_517_axis_0 = const()[name = tensor<string, []>("x_517_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_517_cast_fp16 = stack(axis = x_517_axis_0, values = (var_4722_cast_fp16, squeeze_148_cast_fp16))[name = tensor<string, []>("x_517_cast_fp16")];
            tensor<int32, [4]> var_4728 = const()[name = tensor<string, []>("op_4728"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4729_cast_fp16 = reshape(shape = var_4728, x = x_517_cast_fp16)[name = tensor<string, []>("op_4729_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4730_cast_fp16 = mul(x = var_4729_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4730_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_75_cast_fp16 = add(x = var_4709_cast_fp16, y = var_4730_cast_fp16)[name = tensor<string, []>("q_patches_75_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4732_cast_fp16 = mul(x = var_4707_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4732_cast_fp16")];
            tensor<int32, [5]> var_4740 = const()[name = tensor<string, []>("op_4740"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_521_cast_fp16 = reshape(shape = var_4740, x = var_4707_cast_fp16)[name = tensor<string, []>("x_521_cast_fp16")];
            tensor<int32, [2]> var_4742_split_sizes_0 = const()[name = tensor<string, []>("op_4742_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4742_axis_0 = const()[name = tensor<string, []>("op_4742_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4742_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4742_cast_fp16_1 = split(axis = var_4742_axis_0, split_sizes = var_4742_split_sizes_0, x = x_521_cast_fp16)[name = tensor<string, []>("op_4742_cast_fp16")];
            tensor<int32, [1]> squeeze_150_axes_0 = const()[name = tensor<string, []>("squeeze_150_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_150_cast_fp16 = squeeze(axes = squeeze_150_axes_0, x = var_4742_cast_fp16_0)[name = tensor<string, []>("squeeze_150_cast_fp16")];
            tensor<int32, [1]> squeeze_151_axes_0 = const()[name = tensor<string, []>("squeeze_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_151_cast_fp16 = squeeze(axes = squeeze_151_axes_0, x = var_4742_cast_fp16_1)[name = tensor<string, []>("squeeze_151_cast_fp16")];
            tensor<fp16, []> const_619_promoted_to_fp16 = const()[name = tensor<string, []>("const_619_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4745_cast_fp16 = mul(x = squeeze_151_cast_fp16, y = const_619_promoted_to_fp16)[name = tensor<string, []>("op_4745_cast_fp16")];
            tensor<int32, []> x_523_axis_0 = const()[name = tensor<string, []>("x_523_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_523_cast_fp16 = stack(axis = x_523_axis_0, values = (var_4745_cast_fp16, squeeze_150_cast_fp16))[name = tensor<string, []>("x_523_cast_fp16")];
            tensor<int32, [4]> var_4751 = const()[name = tensor<string, []>("op_4751"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4752_cast_fp16 = reshape(shape = var_4751, x = x_523_cast_fp16)[name = tensor<string, []>("op_4752_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4753_cast_fp16 = mul(x = var_4752_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4753_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_75_cast_fp16 = add(x = var_4732_cast_fp16, y = var_4753_cast_fp16)[name = tensor<string, []>("k_patches_75_cast_fp16")];
            tensor<bool, []> var_4756_interleave_0 = const()[name = tensor<string, []>("op_4756_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4756_cast_fp16 = concat(axis = var_20, interleave = var_4756_interleave_0, values = (var_4695_cast_fp16, q_patches_75_cast_fp16))[name = tensor<string, []>("op_4756_cast_fp16")];
            tensor<bool, []> var_4759_interleave_0 = const()[name = tensor<string, []>("op_4759_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4759_cast_fp16 = concat(axis = var_20, interleave = var_4759_interleave_0, values = (var_4699_cast_fp16, k_patches_75_cast_fp16))[name = tensor<string, []>("op_4759_cast_fp16")];
            tensor<fp16, []> var_4761_to_fp16 = const()[name = tensor<string, []>("op_4761_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_379_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4761_to_fp16)[name = tensor<string, []>("q_379_cast_fp16")];
            tensor<bool, []> attn_149_transpose_x_1 = const()[name = tensor<string, []>("attn_149_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_149_transpose_y_1 = const()[name = tensor<string, []>("attn_149_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_149_cast_fp16 = matmul(transpose_x = attn_149_transpose_x_1, transpose_y = attn_149_transpose_y_1, x = q_379_cast_fp16, y = var_4759_cast_fp16)[name = tensor<string, []>("attn_149_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_151_cast_fp16 = softmax(axis = var_21, x = attn_149_cast_fp16)[name = tensor<string, []>("attn_151_cast_fp16")];
            tensor<bool, []> var_4766_transpose_x_0 = const()[name = tensor<string, []>("op_4766_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4766_transpose_y_0 = const()[name = tensor<string, []>("op_4766_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4766_cast_fp16 = matmul(transpose_x = var_4766_transpose_x_0, transpose_y = var_4766_transpose_y_0, x = attn_151_cast_fp16, y = v_227_cast_fp16)[name = tensor<string, []>("op_4766_cast_fp16")];
            tensor<int32, [1]> var_4767_axes_0 = const()[name = tensor<string, []>("op_4767_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4767_cast_fp16 = squeeze(axes = var_4767_axes_0, x = var_4766_cast_fp16)[name = tensor<string, []>("op_4767_cast_fp16")];
            tensor<string, []> var_4773_pad_type_0 = const()[name = tensor<string, []>("op_4773_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4773_strides_0 = const()[name = tensor<string, []>("op_4773_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4773_pad_0 = const()[name = tensor<string, []>("op_4773_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4773_dilations_0 = const()[name = tensor<string, []>("op_4773_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4773_groups_0 = const()[name = tensor<string, []>("op_4773_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44656960)))];
            tensor<fp16, [1, 64, 1, 197]> var_4773_cast_fp16 = conv(dilations = var_4773_dilations_0, groups = var_4773_groups_0, pad = var_4773_pad_0, pad_type = var_4773_pad_type_0, strides = var_4773_strides_0, weight = model_blocks_3_attn_q_projs_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4773_cast_fp16")];
            tensor<int32, [1]> var_4774_axes_0 = const()[name = tensor<string, []>("op_4774_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4774_cast_fp16 = squeeze(axes = var_4774_axes_0, x = var_4773_cast_fp16)[name = tensor<string, []>("op_4774_cast_fp16")];
            tensor<int32, [3]> q_381_perm_0 = const()[name = tensor<string, []>("q_381_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4781_pad_type_0 = const()[name = tensor<string, []>("op_4781_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4781_strides_0 = const()[name = tensor<string, []>("op_4781_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4781_pad_0 = const()[name = tensor<string, []>("op_4781_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4781_dilations_0 = const()[name = tensor<string, []>("op_4781_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4781_groups_0 = const()[name = tensor<string, []>("op_4781_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44755328)))];
            tensor<fp16, [1, 64, 1, 197]> var_4781_cast_fp16 = conv(dilations = var_4781_dilations_0, groups = var_4781_groups_0, pad = var_4781_pad_0, pad_type = var_4781_pad_type_0, strides = var_4781_strides_0, weight = model_blocks_3_attn_k_projs_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4781_cast_fp16")];
            tensor<int32, [1]> var_4782_axes_0 = const()[name = tensor<string, []>("op_4782_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4782_cast_fp16 = squeeze(axes = var_4782_axes_0, x = var_4781_cast_fp16)[name = tensor<string, []>("op_4782_cast_fp16")];
            tensor<int32, [3]> k_229_perm_0 = const()[name = tensor<string, []>("k_229_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4789_pad_type_0 = const()[name = tensor<string, []>("op_4789_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4789_strides_0 = const()[name = tensor<string, []>("op_4789_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4789_pad_0 = const()[name = tensor<string, []>("op_4789_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4789_dilations_0 = const()[name = tensor<string, []>("op_4789_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4789_groups_0 = const()[name = tensor<string, []>("op_4789_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44853696)))];
            tensor<fp16, [1, 64, 1, 197]> var_4789_cast_fp16 = conv(dilations = var_4789_dilations_0, groups = var_4789_groups_0, pad = var_4789_pad_0, pad_type = var_4789_pad_type_0, strides = var_4789_strides_0, weight = model_blocks_3_attn_v_projs_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4789_cast_fp16")];
            tensor<int32, [1]> var_4790_axes_0 = const()[name = tensor<string, []>("op_4790_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4790_cast_fp16 = squeeze(axes = var_4790_axes_0, x = var_4789_cast_fp16)[name = tensor<string, []>("op_4790_cast_fp16")];
            tensor<int32, [3]> v_229_perm_0 = const()[name = tensor<string, []>("v_229_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44952064)))];
            tensor<fp16, [1, 197, 64]> q_381_cast_fp16 = transpose(perm = q_381_perm_0, x = var_4774_cast_fp16)[name = tensor<string, []>("transpose_379")];
            tensor<fp16, [1, 197, 64]> q_383_cast_fp16 = add(x = q_381_cast_fp16, y = model_blocks_3_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_383_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44952256)))];
            tensor<fp16, [1, 197, 64]> v_229_cast_fp16 = transpose(perm = v_229_perm_0, x = var_4790_cast_fp16)[name = tensor<string, []>("transpose_377")];
            tensor<fp16, [1, 197, 64]> v_231_cast_fp16 = add(x = v_229_cast_fp16, y = model_blocks_3_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_231_cast_fp16")];
            tensor<int32, [1]> q_385_axes_0 = const()[name = tensor<string, []>("q_385_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_385_cast_fp16 = expand_dims(axes = q_385_axes_0, x = q_383_cast_fp16)[name = tensor<string, []>("q_385_cast_fp16")];
            tensor<int32, [1]> k_231_axes_0 = const()[name = tensor<string, []>("k_231_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_229_cast_fp16 = transpose(perm = k_229_perm_0, x = var_4782_cast_fp16)[name = tensor<string, []>("transpose_378")];
            tensor<fp16, [1, 1, 197, 64]> k_231_cast_fp16 = expand_dims(axes = k_231_axes_0, x = k_229_cast_fp16)[name = tensor<string, []>("k_231_cast_fp16")];
            tensor<int32, [1]> v_233_axes_0 = const()[name = tensor<string, []>("v_233_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_233_cast_fp16 = expand_dims(axes = v_233_axes_0, x = v_231_cast_fp16)[name = tensor<string, []>("v_233_cast_fp16")];
            tensor<int32, [4]> var_4799_begin_0 = const()[name = tensor<string, []>("op_4799_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4799_end_0 = const()[name = tensor<string, []>("op_4799_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4799_end_mask_0 = const()[name = tensor<string, []>("op_4799_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4799_cast_fp16 = slice_by_index(begin = var_4799_begin_0, end = var_4799_end_0, end_mask = var_4799_end_mask_0, x = q_385_cast_fp16)[name = tensor<string, []>("op_4799_cast_fp16")];
            tensor<int32, [4]> var_4803_begin_0 = const()[name = tensor<string, []>("op_4803_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4803_end_0 = const()[name = tensor<string, []>("op_4803_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4803_end_mask_0 = const()[name = tensor<string, []>("op_4803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4803_cast_fp16 = slice_by_index(begin = var_4803_begin_0, end = var_4803_end_0, end_mask = var_4803_end_mask_0, x = k_231_cast_fp16)[name = tensor<string, []>("op_4803_cast_fp16")];
            tensor<int32, [4]> var_4807_begin_0 = const()[name = tensor<string, []>("op_4807_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4807_end_0 = const()[name = tensor<string, []>("op_4807_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4807_end_mask_0 = const()[name = tensor<string, []>("op_4807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4807_cast_fp16 = slice_by_index(begin = var_4807_begin_0, end = var_4807_end_0, end_mask = var_4807_end_mask_0, x = q_385_cast_fp16)[name = tensor<string, []>("op_4807_cast_fp16")];
            tensor<int32, [4]> var_4811_begin_0 = const()[name = tensor<string, []>("op_4811_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4811_end_0 = const()[name = tensor<string, []>("op_4811_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4811_end_mask_0 = const()[name = tensor<string, []>("op_4811_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, x = k_231_cast_fp16)[name = tensor<string, []>("op_4811_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4813_cast_fp16 = mul(x = var_4807_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4813_cast_fp16")];
            tensor<int32, [5]> var_4821 = const()[name = tensor<string, []>("op_4821"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_527_cast_fp16 = reshape(shape = var_4821, x = var_4807_cast_fp16)[name = tensor<string, []>("x_527_cast_fp16")];
            tensor<int32, [2]> var_4823_split_sizes_0 = const()[name = tensor<string, []>("op_4823_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4823_axis_0 = const()[name = tensor<string, []>("op_4823_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4823_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4823_cast_fp16_1 = split(axis = var_4823_axis_0, split_sizes = var_4823_split_sizes_0, x = x_527_cast_fp16)[name = tensor<string, []>("op_4823_cast_fp16")];
            tensor<int32, [1]> squeeze_152_axes_0 = const()[name = tensor<string, []>("squeeze_152_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_152_cast_fp16 = squeeze(axes = squeeze_152_axes_0, x = var_4823_cast_fp16_0)[name = tensor<string, []>("squeeze_152_cast_fp16")];
            tensor<int32, [1]> squeeze_153_axes_0 = const()[name = tensor<string, []>("squeeze_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_153_cast_fp16 = squeeze(axes = squeeze_153_axes_0, x = var_4823_cast_fp16_1)[name = tensor<string, []>("squeeze_153_cast_fp16")];
            tensor<fp16, []> const_627_promoted_to_fp16 = const()[name = tensor<string, []>("const_627_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4826_cast_fp16 = mul(x = squeeze_153_cast_fp16, y = const_627_promoted_to_fp16)[name = tensor<string, []>("op_4826_cast_fp16")];
            tensor<int32, []> x_529_axis_0 = const()[name = tensor<string, []>("x_529_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_529_cast_fp16 = stack(axis = x_529_axis_0, values = (var_4826_cast_fp16, squeeze_152_cast_fp16))[name = tensor<string, []>("x_529_cast_fp16")];
            tensor<int32, [4]> var_4832 = const()[name = tensor<string, []>("op_4832"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4833_cast_fp16 = reshape(shape = var_4832, x = x_529_cast_fp16)[name = tensor<string, []>("op_4833_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4834_cast_fp16 = mul(x = var_4833_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4834_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_77_cast_fp16 = add(x = var_4813_cast_fp16, y = var_4834_cast_fp16)[name = tensor<string, []>("q_patches_77_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4836_cast_fp16 = mul(x = var_4811_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4836_cast_fp16")];
            tensor<int32, [5]> var_4844 = const()[name = tensor<string, []>("op_4844"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_533_cast_fp16 = reshape(shape = var_4844, x = var_4811_cast_fp16)[name = tensor<string, []>("x_533_cast_fp16")];
            tensor<int32, [2]> var_4846_split_sizes_0 = const()[name = tensor<string, []>("op_4846_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4846_axis_0 = const()[name = tensor<string, []>("op_4846_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4846_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4846_cast_fp16_1 = split(axis = var_4846_axis_0, split_sizes = var_4846_split_sizes_0, x = x_533_cast_fp16)[name = tensor<string, []>("op_4846_cast_fp16")];
            tensor<int32, [1]> squeeze_154_axes_0 = const()[name = tensor<string, []>("squeeze_154_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_154_cast_fp16 = squeeze(axes = squeeze_154_axes_0, x = var_4846_cast_fp16_0)[name = tensor<string, []>("squeeze_154_cast_fp16")];
            tensor<int32, [1]> squeeze_155_axes_0 = const()[name = tensor<string, []>("squeeze_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_155_cast_fp16 = squeeze(axes = squeeze_155_axes_0, x = var_4846_cast_fp16_1)[name = tensor<string, []>("squeeze_155_cast_fp16")];
            tensor<fp16, []> const_635_promoted_to_fp16 = const()[name = tensor<string, []>("const_635_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4849_cast_fp16 = mul(x = squeeze_155_cast_fp16, y = const_635_promoted_to_fp16)[name = tensor<string, []>("op_4849_cast_fp16")];
            tensor<int32, []> x_535_axis_0 = const()[name = tensor<string, []>("x_535_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_535_cast_fp16 = stack(axis = x_535_axis_0, values = (var_4849_cast_fp16, squeeze_154_cast_fp16))[name = tensor<string, []>("x_535_cast_fp16")];
            tensor<int32, [4]> var_4855 = const()[name = tensor<string, []>("op_4855"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4856_cast_fp16 = reshape(shape = var_4855, x = x_535_cast_fp16)[name = tensor<string, []>("op_4856_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4857_cast_fp16 = mul(x = var_4856_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4857_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_77_cast_fp16 = add(x = var_4836_cast_fp16, y = var_4857_cast_fp16)[name = tensor<string, []>("k_patches_77_cast_fp16")];
            tensor<bool, []> var_4860_interleave_0 = const()[name = tensor<string, []>("op_4860_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4860_cast_fp16 = concat(axis = var_20, interleave = var_4860_interleave_0, values = (var_4799_cast_fp16, q_patches_77_cast_fp16))[name = tensor<string, []>("op_4860_cast_fp16")];
            tensor<bool, []> var_4863_interleave_0 = const()[name = tensor<string, []>("op_4863_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4863_cast_fp16 = concat(axis = var_20, interleave = var_4863_interleave_0, values = (var_4803_cast_fp16, k_patches_77_cast_fp16))[name = tensor<string, []>("op_4863_cast_fp16")];
            tensor<fp16, []> var_4865_to_fp16 = const()[name = tensor<string, []>("op_4865_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_389_cast_fp16 = mul(x = var_4860_cast_fp16, y = var_4865_to_fp16)[name = tensor<string, []>("q_389_cast_fp16")];
            tensor<bool, []> attn_153_transpose_x_1 = const()[name = tensor<string, []>("attn_153_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_153_transpose_y_1 = const()[name = tensor<string, []>("attn_153_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_153_cast_fp16 = matmul(transpose_x = attn_153_transpose_x_1, transpose_y = attn_153_transpose_y_1, x = q_389_cast_fp16, y = var_4863_cast_fp16)[name = tensor<string, []>("attn_153_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_155_cast_fp16 = softmax(axis = var_21, x = attn_153_cast_fp16)[name = tensor<string, []>("attn_155_cast_fp16")];
            tensor<bool, []> var_4870_transpose_x_0 = const()[name = tensor<string, []>("op_4870_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4870_transpose_y_0 = const()[name = tensor<string, []>("op_4870_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4870_cast_fp16 = matmul(transpose_x = var_4870_transpose_x_0, transpose_y = var_4870_transpose_y_0, x = attn_155_cast_fp16, y = v_233_cast_fp16)[name = tensor<string, []>("op_4870_cast_fp16")];
            tensor<int32, [1]> var_4871_axes_0 = const()[name = tensor<string, []>("op_4871_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4871_cast_fp16 = squeeze(axes = var_4871_axes_0, x = var_4870_cast_fp16)[name = tensor<string, []>("op_4871_cast_fp16")];
            tensor<string, []> var_4877_pad_type_0 = const()[name = tensor<string, []>("op_4877_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4877_strides_0 = const()[name = tensor<string, []>("op_4877_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4877_pad_0 = const()[name = tensor<string, []>("op_4877_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4877_dilations_0 = const()[name = tensor<string, []>("op_4877_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4877_groups_0 = const()[name = tensor<string, []>("op_4877_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44952448)))];
            tensor<fp16, [1, 64, 1, 197]> var_4877_cast_fp16 = conv(dilations = var_4877_dilations_0, groups = var_4877_groups_0, pad = var_4877_pad_0, pad_type = var_4877_pad_type_0, strides = var_4877_strides_0, weight = model_blocks_3_attn_q_projs_3_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4877_cast_fp16")];
            tensor<int32, [1]> var_4878_axes_0 = const()[name = tensor<string, []>("op_4878_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4878_cast_fp16 = squeeze(axes = var_4878_axes_0, x = var_4877_cast_fp16)[name = tensor<string, []>("op_4878_cast_fp16")];
            tensor<int32, [3]> q_391_perm_0 = const()[name = tensor<string, []>("q_391_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4885_pad_type_0 = const()[name = tensor<string, []>("op_4885_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4885_strides_0 = const()[name = tensor<string, []>("op_4885_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4885_pad_0 = const()[name = tensor<string, []>("op_4885_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4885_dilations_0 = const()[name = tensor<string, []>("op_4885_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4885_groups_0 = const()[name = tensor<string, []>("op_4885_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45050816)))];
            tensor<fp16, [1, 64, 1, 197]> var_4885_cast_fp16 = conv(dilations = var_4885_dilations_0, groups = var_4885_groups_0, pad = var_4885_pad_0, pad_type = var_4885_pad_type_0, strides = var_4885_strides_0, weight = model_blocks_3_attn_k_projs_3_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4885_cast_fp16")];
            tensor<int32, [1]> var_4886_axes_0 = const()[name = tensor<string, []>("op_4886_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4886_cast_fp16 = squeeze(axes = var_4886_axes_0, x = var_4885_cast_fp16)[name = tensor<string, []>("op_4886_cast_fp16")];
            tensor<int32, [3]> k_235_perm_0 = const()[name = tensor<string, []>("k_235_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4893_pad_type_0 = const()[name = tensor<string, []>("op_4893_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4893_strides_0 = const()[name = tensor<string, []>("op_4893_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4893_pad_0 = const()[name = tensor<string, []>("op_4893_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4893_dilations_0 = const()[name = tensor<string, []>("op_4893_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4893_groups_0 = const()[name = tensor<string, []>("op_4893_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45149184)))];
            tensor<fp16, [1, 64, 1, 197]> var_4893_cast_fp16 = conv(dilations = var_4893_dilations_0, groups = var_4893_groups_0, pad = var_4893_pad_0, pad_type = var_4893_pad_type_0, strides = var_4893_strides_0, weight = model_blocks_3_attn_v_projs_3_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4893_cast_fp16")];
            tensor<int32, [1]> var_4894_axes_0 = const()[name = tensor<string, []>("op_4894_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4894_cast_fp16 = squeeze(axes = var_4894_axes_0, x = var_4893_cast_fp16)[name = tensor<string, []>("op_4894_cast_fp16")];
            tensor<int32, [3]> v_235_perm_0 = const()[name = tensor<string, []>("v_235_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45247552)))];
            tensor<fp16, [1, 197, 64]> q_391_cast_fp16 = transpose(perm = q_391_perm_0, x = var_4878_cast_fp16)[name = tensor<string, []>("transpose_376")];
            tensor<fp16, [1, 197, 64]> q_393_cast_fp16 = add(x = q_391_cast_fp16, y = model_blocks_3_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_393_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45247744)))];
            tensor<fp16, [1, 197, 64]> v_235_cast_fp16 = transpose(perm = v_235_perm_0, x = var_4894_cast_fp16)[name = tensor<string, []>("transpose_374")];
            tensor<fp16, [1, 197, 64]> v_237_cast_fp16 = add(x = v_235_cast_fp16, y = model_blocks_3_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_237_cast_fp16")];
            tensor<int32, [1]> q_395_axes_0 = const()[name = tensor<string, []>("q_395_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_395_cast_fp16 = expand_dims(axes = q_395_axes_0, x = q_393_cast_fp16)[name = tensor<string, []>("q_395_cast_fp16")];
            tensor<int32, [1]> k_237_axes_0 = const()[name = tensor<string, []>("k_237_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_235_cast_fp16 = transpose(perm = k_235_perm_0, x = var_4886_cast_fp16)[name = tensor<string, []>("transpose_375")];
            tensor<fp16, [1, 1, 197, 64]> k_237_cast_fp16 = expand_dims(axes = k_237_axes_0, x = k_235_cast_fp16)[name = tensor<string, []>("k_237_cast_fp16")];
            tensor<int32, [1]> v_239_axes_0 = const()[name = tensor<string, []>("v_239_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_239_cast_fp16 = expand_dims(axes = v_239_axes_0, x = v_237_cast_fp16)[name = tensor<string, []>("v_239_cast_fp16")];
            tensor<int32, [4]> var_4903_begin_0 = const()[name = tensor<string, []>("op_4903_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4903_end_0 = const()[name = tensor<string, []>("op_4903_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4903_end_mask_0 = const()[name = tensor<string, []>("op_4903_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4903_cast_fp16 = slice_by_index(begin = var_4903_begin_0, end = var_4903_end_0, end_mask = var_4903_end_mask_0, x = q_395_cast_fp16)[name = tensor<string, []>("op_4903_cast_fp16")];
            tensor<int32, [4]> var_4907_begin_0 = const()[name = tensor<string, []>("op_4907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4907_end_0 = const()[name = tensor<string, []>("op_4907_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_4907_end_mask_0 = const()[name = tensor<string, []>("op_4907_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_4907_cast_fp16 = slice_by_index(begin = var_4907_begin_0, end = var_4907_end_0, end_mask = var_4907_end_mask_0, x = k_237_cast_fp16)[name = tensor<string, []>("op_4907_cast_fp16")];
            tensor<int32, [4]> var_4911_begin_0 = const()[name = tensor<string, []>("op_4911_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4911_end_0 = const()[name = tensor<string, []>("op_4911_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4911_end_mask_0 = const()[name = tensor<string, []>("op_4911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4911_cast_fp16 = slice_by_index(begin = var_4911_begin_0, end = var_4911_end_0, end_mask = var_4911_end_mask_0, x = q_395_cast_fp16)[name = tensor<string, []>("op_4911_cast_fp16")];
            tensor<int32, [4]> var_4915_begin_0 = const()[name = tensor<string, []>("op_4915_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4915_end_0 = const()[name = tensor<string, []>("op_4915_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_4915_end_mask_0 = const()[name = tensor<string, []>("op_4915_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_4915_cast_fp16 = slice_by_index(begin = var_4915_begin_0, end = var_4915_end_0, end_mask = var_4915_end_mask_0, x = k_237_cast_fp16)[name = tensor<string, []>("op_4915_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4917_cast_fp16 = mul(x = var_4911_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4917_cast_fp16")];
            tensor<int32, [5]> var_4925 = const()[name = tensor<string, []>("op_4925"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_539_cast_fp16 = reshape(shape = var_4925, x = var_4911_cast_fp16)[name = tensor<string, []>("x_539_cast_fp16")];
            tensor<int32, [2]> var_4927_split_sizes_0 = const()[name = tensor<string, []>("op_4927_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4927_axis_0 = const()[name = tensor<string, []>("op_4927_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4927_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4927_cast_fp16_1 = split(axis = var_4927_axis_0, split_sizes = var_4927_split_sizes_0, x = x_539_cast_fp16)[name = tensor<string, []>("op_4927_cast_fp16")];
            tensor<int32, [1]> squeeze_156_axes_0 = const()[name = tensor<string, []>("squeeze_156_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_156_cast_fp16 = squeeze(axes = squeeze_156_axes_0, x = var_4927_cast_fp16_0)[name = tensor<string, []>("squeeze_156_cast_fp16")];
            tensor<int32, [1]> squeeze_157_axes_0 = const()[name = tensor<string, []>("squeeze_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_157_cast_fp16 = squeeze(axes = squeeze_157_axes_0, x = var_4927_cast_fp16_1)[name = tensor<string, []>("squeeze_157_cast_fp16")];
            tensor<fp16, []> const_643_promoted_to_fp16 = const()[name = tensor<string, []>("const_643_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4930_cast_fp16 = mul(x = squeeze_157_cast_fp16, y = const_643_promoted_to_fp16)[name = tensor<string, []>("op_4930_cast_fp16")];
            tensor<int32, []> x_541_axis_0 = const()[name = tensor<string, []>("x_541_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_541_cast_fp16 = stack(axis = x_541_axis_0, values = (var_4930_cast_fp16, squeeze_156_cast_fp16))[name = tensor<string, []>("x_541_cast_fp16")];
            tensor<int32, [4]> var_4936 = const()[name = tensor<string, []>("op_4936"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4937_cast_fp16 = reshape(shape = var_4936, x = x_541_cast_fp16)[name = tensor<string, []>("op_4937_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4938_cast_fp16 = mul(x = var_4937_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4938_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_79_cast_fp16 = add(x = var_4917_cast_fp16, y = var_4938_cast_fp16)[name = tensor<string, []>("q_patches_79_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4940_cast_fp16 = mul(x = var_4915_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_4940_cast_fp16")];
            tensor<int32, [5]> var_4948 = const()[name = tensor<string, []>("op_4948"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_545_cast_fp16 = reshape(shape = var_4948, x = var_4915_cast_fp16)[name = tensor<string, []>("x_545_cast_fp16")];
            tensor<int32, [2]> var_4950_split_sizes_0 = const()[name = tensor<string, []>("op_4950_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4950_axis_0 = const()[name = tensor<string, []>("op_4950_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_4950_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_4950_cast_fp16_1 = split(axis = var_4950_axis_0, split_sizes = var_4950_split_sizes_0, x = x_545_cast_fp16)[name = tensor<string, []>("op_4950_cast_fp16")];
            tensor<int32, [1]> squeeze_158_axes_0 = const()[name = tensor<string, []>("squeeze_158_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_158_cast_fp16 = squeeze(axes = squeeze_158_axes_0, x = var_4950_cast_fp16_0)[name = tensor<string, []>("squeeze_158_cast_fp16")];
            tensor<int32, [1]> squeeze_159_axes_0 = const()[name = tensor<string, []>("squeeze_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_159_cast_fp16 = squeeze(axes = squeeze_159_axes_0, x = var_4950_cast_fp16_1)[name = tensor<string, []>("squeeze_159_cast_fp16")];
            tensor<fp16, []> const_651_promoted_to_fp16 = const()[name = tensor<string, []>("const_651_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_4953_cast_fp16 = mul(x = squeeze_159_cast_fp16, y = const_651_promoted_to_fp16)[name = tensor<string, []>("op_4953_cast_fp16")];
            tensor<int32, []> x_547_axis_0 = const()[name = tensor<string, []>("x_547_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_547_cast_fp16 = stack(axis = x_547_axis_0, values = (var_4953_cast_fp16, squeeze_158_cast_fp16))[name = tensor<string, []>("x_547_cast_fp16")];
            tensor<int32, [4]> var_4959 = const()[name = tensor<string, []>("op_4959"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_4960_cast_fp16 = reshape(shape = var_4959, x = x_547_cast_fp16)[name = tensor<string, []>("op_4960_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_4961_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_79_cast_fp16 = add(x = var_4940_cast_fp16, y = var_4961_cast_fp16)[name = tensor<string, []>("k_patches_79_cast_fp16")];
            tensor<bool, []> var_4964_interleave_0 = const()[name = tensor<string, []>("op_4964_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4964_cast_fp16 = concat(axis = var_20, interleave = var_4964_interleave_0, values = (var_4903_cast_fp16, q_patches_79_cast_fp16))[name = tensor<string, []>("op_4964_cast_fp16")];
            tensor<bool, []> var_4967_interleave_0 = const()[name = tensor<string, []>("op_4967_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4967_cast_fp16 = concat(axis = var_20, interleave = var_4967_interleave_0, values = (var_4907_cast_fp16, k_patches_79_cast_fp16))[name = tensor<string, []>("op_4967_cast_fp16")];
            tensor<fp16, []> var_4969_to_fp16 = const()[name = tensor<string, []>("op_4969_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_399_cast_fp16 = mul(x = var_4964_cast_fp16, y = var_4969_to_fp16)[name = tensor<string, []>("q_399_cast_fp16")];
            tensor<bool, []> attn_157_transpose_x_1 = const()[name = tensor<string, []>("attn_157_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_157_transpose_y_1 = const()[name = tensor<string, []>("attn_157_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_157_cast_fp16 = matmul(transpose_x = attn_157_transpose_x_1, transpose_y = attn_157_transpose_y_1, x = q_399_cast_fp16, y = var_4967_cast_fp16)[name = tensor<string, []>("attn_157_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_159_cast_fp16 = softmax(axis = var_21, x = attn_157_cast_fp16)[name = tensor<string, []>("attn_159_cast_fp16")];
            tensor<bool, []> var_4974_transpose_x_0 = const()[name = tensor<string, []>("op_4974_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4974_transpose_y_0 = const()[name = tensor<string, []>("op_4974_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_4974_cast_fp16 = matmul(transpose_x = var_4974_transpose_x_0, transpose_y = var_4974_transpose_y_0, x = attn_159_cast_fp16, y = v_239_cast_fp16)[name = tensor<string, []>("op_4974_cast_fp16")];
            tensor<int32, [1]> var_4975_axes_0 = const()[name = tensor<string, []>("op_4975_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_4975_cast_fp16 = squeeze(axes = var_4975_axes_0, x = var_4974_cast_fp16)[name = tensor<string, []>("op_4975_cast_fp16")];
            tensor<string, []> var_4981_pad_type_0 = const()[name = tensor<string, []>("op_4981_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4981_strides_0 = const()[name = tensor<string, []>("op_4981_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4981_pad_0 = const()[name = tensor<string, []>("op_4981_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4981_dilations_0 = const()[name = tensor<string, []>("op_4981_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4981_groups_0 = const()[name = tensor<string, []>("op_4981_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45247936)))];
            tensor<fp16, [1, 64, 1, 197]> var_4981_cast_fp16 = conv(dilations = var_4981_dilations_0, groups = var_4981_groups_0, pad = var_4981_pad_0, pad_type = var_4981_pad_type_0, strides = var_4981_strides_0, weight = model_blocks_3_attn_q_projs_4_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4981_cast_fp16")];
            tensor<int32, [1]> var_4982_axes_0 = const()[name = tensor<string, []>("op_4982_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4982_cast_fp16 = squeeze(axes = var_4982_axes_0, x = var_4981_cast_fp16)[name = tensor<string, []>("op_4982_cast_fp16")];
            tensor<int32, [3]> q_401_perm_0 = const()[name = tensor<string, []>("q_401_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4989_pad_type_0 = const()[name = tensor<string, []>("op_4989_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4989_strides_0 = const()[name = tensor<string, []>("op_4989_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4989_pad_0 = const()[name = tensor<string, []>("op_4989_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4989_dilations_0 = const()[name = tensor<string, []>("op_4989_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4989_groups_0 = const()[name = tensor<string, []>("op_4989_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45346304)))];
            tensor<fp16, [1, 64, 1, 197]> var_4989_cast_fp16 = conv(dilations = var_4989_dilations_0, groups = var_4989_groups_0, pad = var_4989_pad_0, pad_type = var_4989_pad_type_0, strides = var_4989_strides_0, weight = model_blocks_3_attn_k_projs_4_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4989_cast_fp16")];
            tensor<int32, [1]> var_4990_axes_0 = const()[name = tensor<string, []>("op_4990_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4990_cast_fp16 = squeeze(axes = var_4990_axes_0, x = var_4989_cast_fp16)[name = tensor<string, []>("op_4990_cast_fp16")];
            tensor<int32, [3]> k_241_perm_0 = const()[name = tensor<string, []>("k_241_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_4997_pad_type_0 = const()[name = tensor<string, []>("op_4997_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4997_strides_0 = const()[name = tensor<string, []>("op_4997_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4997_pad_0 = const()[name = tensor<string, []>("op_4997_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4997_dilations_0 = const()[name = tensor<string, []>("op_4997_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4997_groups_0 = const()[name = tensor<string, []>("op_4997_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45444672)))];
            tensor<fp16, [1, 64, 1, 197]> var_4997_cast_fp16 = conv(dilations = var_4997_dilations_0, groups = var_4997_groups_0, pad = var_4997_pad_0, pad_type = var_4997_pad_type_0, strides = var_4997_strides_0, weight = model_blocks_3_attn_v_projs_4_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_4997_cast_fp16")];
            tensor<int32, [1]> var_4998_axes_0 = const()[name = tensor<string, []>("op_4998_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_4998_cast_fp16 = squeeze(axes = var_4998_axes_0, x = var_4997_cast_fp16)[name = tensor<string, []>("op_4998_cast_fp16")];
            tensor<int32, [3]> v_241_perm_0 = const()[name = tensor<string, []>("v_241_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45543040)))];
            tensor<fp16, [1, 197, 64]> q_401_cast_fp16 = transpose(perm = q_401_perm_0, x = var_4982_cast_fp16)[name = tensor<string, []>("transpose_373")];
            tensor<fp16, [1, 197, 64]> q_403_cast_fp16 = add(x = q_401_cast_fp16, y = model_blocks_3_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_403_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45543232)))];
            tensor<fp16, [1, 197, 64]> v_241_cast_fp16 = transpose(perm = v_241_perm_0, x = var_4998_cast_fp16)[name = tensor<string, []>("transpose_371")];
            tensor<fp16, [1, 197, 64]> v_243_cast_fp16 = add(x = v_241_cast_fp16, y = model_blocks_3_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_243_cast_fp16")];
            tensor<int32, [1]> q_405_axes_0 = const()[name = tensor<string, []>("q_405_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_405_cast_fp16 = expand_dims(axes = q_405_axes_0, x = q_403_cast_fp16)[name = tensor<string, []>("q_405_cast_fp16")];
            tensor<int32, [1]> k_243_axes_0 = const()[name = tensor<string, []>("k_243_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_241_cast_fp16 = transpose(perm = k_241_perm_0, x = var_4990_cast_fp16)[name = tensor<string, []>("transpose_372")];
            tensor<fp16, [1, 1, 197, 64]> k_243_cast_fp16 = expand_dims(axes = k_243_axes_0, x = k_241_cast_fp16)[name = tensor<string, []>("k_243_cast_fp16")];
            tensor<int32, [1]> v_245_axes_0 = const()[name = tensor<string, []>("v_245_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_245_cast_fp16 = expand_dims(axes = v_245_axes_0, x = v_243_cast_fp16)[name = tensor<string, []>("v_245_cast_fp16")];
            tensor<int32, [4]> var_5007_begin_0 = const()[name = tensor<string, []>("op_5007_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5007_end_0 = const()[name = tensor<string, []>("op_5007_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5007_end_mask_0 = const()[name = tensor<string, []>("op_5007_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5007_cast_fp16 = slice_by_index(begin = var_5007_begin_0, end = var_5007_end_0, end_mask = var_5007_end_mask_0, x = q_405_cast_fp16)[name = tensor<string, []>("op_5007_cast_fp16")];
            tensor<int32, [4]> var_5011_begin_0 = const()[name = tensor<string, []>("op_5011_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5011_end_0 = const()[name = tensor<string, []>("op_5011_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5011_end_mask_0 = const()[name = tensor<string, []>("op_5011_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5011_cast_fp16 = slice_by_index(begin = var_5011_begin_0, end = var_5011_end_0, end_mask = var_5011_end_mask_0, x = k_243_cast_fp16)[name = tensor<string, []>("op_5011_cast_fp16")];
            tensor<int32, [4]> var_5015_begin_0 = const()[name = tensor<string, []>("op_5015_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5015_end_0 = const()[name = tensor<string, []>("op_5015_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5015_end_mask_0 = const()[name = tensor<string, []>("op_5015_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5015_cast_fp16 = slice_by_index(begin = var_5015_begin_0, end = var_5015_end_0, end_mask = var_5015_end_mask_0, x = q_405_cast_fp16)[name = tensor<string, []>("op_5015_cast_fp16")];
            tensor<int32, [4]> var_5019_begin_0 = const()[name = tensor<string, []>("op_5019_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5019_end_0 = const()[name = tensor<string, []>("op_5019_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5019_end_mask_0 = const()[name = tensor<string, []>("op_5019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5019_cast_fp16 = slice_by_index(begin = var_5019_begin_0, end = var_5019_end_0, end_mask = var_5019_end_mask_0, x = k_243_cast_fp16)[name = tensor<string, []>("op_5019_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5021_cast_fp16 = mul(x = var_5015_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5021_cast_fp16")];
            tensor<int32, [5]> var_5029 = const()[name = tensor<string, []>("op_5029"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_551_cast_fp16 = reshape(shape = var_5029, x = var_5015_cast_fp16)[name = tensor<string, []>("x_551_cast_fp16")];
            tensor<int32, [2]> var_5031_split_sizes_0 = const()[name = tensor<string, []>("op_5031_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5031_axis_0 = const()[name = tensor<string, []>("op_5031_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5031_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5031_cast_fp16_1 = split(axis = var_5031_axis_0, split_sizes = var_5031_split_sizes_0, x = x_551_cast_fp16)[name = tensor<string, []>("op_5031_cast_fp16")];
            tensor<int32, [1]> squeeze_160_axes_0 = const()[name = tensor<string, []>("squeeze_160_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_160_cast_fp16 = squeeze(axes = squeeze_160_axes_0, x = var_5031_cast_fp16_0)[name = tensor<string, []>("squeeze_160_cast_fp16")];
            tensor<int32, [1]> squeeze_161_axes_0 = const()[name = tensor<string, []>("squeeze_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_161_cast_fp16 = squeeze(axes = squeeze_161_axes_0, x = var_5031_cast_fp16_1)[name = tensor<string, []>("squeeze_161_cast_fp16")];
            tensor<fp16, []> const_659_promoted_to_fp16 = const()[name = tensor<string, []>("const_659_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5034_cast_fp16 = mul(x = squeeze_161_cast_fp16, y = const_659_promoted_to_fp16)[name = tensor<string, []>("op_5034_cast_fp16")];
            tensor<int32, []> x_553_axis_0 = const()[name = tensor<string, []>("x_553_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_553_cast_fp16 = stack(axis = x_553_axis_0, values = (var_5034_cast_fp16, squeeze_160_cast_fp16))[name = tensor<string, []>("x_553_cast_fp16")];
            tensor<int32, [4]> var_5040 = const()[name = tensor<string, []>("op_5040"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5041_cast_fp16 = reshape(shape = var_5040, x = x_553_cast_fp16)[name = tensor<string, []>("op_5041_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5042_cast_fp16 = mul(x = var_5041_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5042_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_81_cast_fp16 = add(x = var_5021_cast_fp16, y = var_5042_cast_fp16)[name = tensor<string, []>("q_patches_81_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5044_cast_fp16 = mul(x = var_5019_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5044_cast_fp16")];
            tensor<int32, [5]> var_5052 = const()[name = tensor<string, []>("op_5052"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_557_cast_fp16 = reshape(shape = var_5052, x = var_5019_cast_fp16)[name = tensor<string, []>("x_557_cast_fp16")];
            tensor<int32, [2]> var_5054_split_sizes_0 = const()[name = tensor<string, []>("op_5054_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5054_axis_0 = const()[name = tensor<string, []>("op_5054_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5054_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5054_cast_fp16_1 = split(axis = var_5054_axis_0, split_sizes = var_5054_split_sizes_0, x = x_557_cast_fp16)[name = tensor<string, []>("op_5054_cast_fp16")];
            tensor<int32, [1]> squeeze_162_axes_0 = const()[name = tensor<string, []>("squeeze_162_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_162_cast_fp16 = squeeze(axes = squeeze_162_axes_0, x = var_5054_cast_fp16_0)[name = tensor<string, []>("squeeze_162_cast_fp16")];
            tensor<int32, [1]> squeeze_163_axes_0 = const()[name = tensor<string, []>("squeeze_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_163_cast_fp16 = squeeze(axes = squeeze_163_axes_0, x = var_5054_cast_fp16_1)[name = tensor<string, []>("squeeze_163_cast_fp16")];
            tensor<fp16, []> const_667_promoted_to_fp16 = const()[name = tensor<string, []>("const_667_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5057_cast_fp16 = mul(x = squeeze_163_cast_fp16, y = const_667_promoted_to_fp16)[name = tensor<string, []>("op_5057_cast_fp16")];
            tensor<int32, []> x_559_axis_0 = const()[name = tensor<string, []>("x_559_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_559_cast_fp16 = stack(axis = x_559_axis_0, values = (var_5057_cast_fp16, squeeze_162_cast_fp16))[name = tensor<string, []>("x_559_cast_fp16")];
            tensor<int32, [4]> var_5063 = const()[name = tensor<string, []>("op_5063"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5064_cast_fp16 = reshape(shape = var_5063, x = x_559_cast_fp16)[name = tensor<string, []>("op_5064_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5065_cast_fp16 = mul(x = var_5064_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5065_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_81_cast_fp16 = add(x = var_5044_cast_fp16, y = var_5065_cast_fp16)[name = tensor<string, []>("k_patches_81_cast_fp16")];
            tensor<bool, []> var_5068_interleave_0 = const()[name = tensor<string, []>("op_5068_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5068_cast_fp16 = concat(axis = var_20, interleave = var_5068_interleave_0, values = (var_5007_cast_fp16, q_patches_81_cast_fp16))[name = tensor<string, []>("op_5068_cast_fp16")];
            tensor<bool, []> var_5071_interleave_0 = const()[name = tensor<string, []>("op_5071_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5071_cast_fp16 = concat(axis = var_20, interleave = var_5071_interleave_0, values = (var_5011_cast_fp16, k_patches_81_cast_fp16))[name = tensor<string, []>("op_5071_cast_fp16")];
            tensor<fp16, []> var_5073_to_fp16 = const()[name = tensor<string, []>("op_5073_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_409_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5073_to_fp16)[name = tensor<string, []>("q_409_cast_fp16")];
            tensor<bool, []> attn_161_transpose_x_1 = const()[name = tensor<string, []>("attn_161_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_161_transpose_y_1 = const()[name = tensor<string, []>("attn_161_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_161_cast_fp16 = matmul(transpose_x = attn_161_transpose_x_1, transpose_y = attn_161_transpose_y_1, x = q_409_cast_fp16, y = var_5071_cast_fp16)[name = tensor<string, []>("attn_161_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_163_cast_fp16 = softmax(axis = var_21, x = attn_161_cast_fp16)[name = tensor<string, []>("attn_163_cast_fp16")];
            tensor<bool, []> var_5078_transpose_x_0 = const()[name = tensor<string, []>("op_5078_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5078_transpose_y_0 = const()[name = tensor<string, []>("op_5078_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5078_cast_fp16 = matmul(transpose_x = var_5078_transpose_x_0, transpose_y = var_5078_transpose_y_0, x = attn_163_cast_fp16, y = v_245_cast_fp16)[name = tensor<string, []>("op_5078_cast_fp16")];
            tensor<int32, [1]> var_5079_axes_0 = const()[name = tensor<string, []>("op_5079_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5079_cast_fp16 = squeeze(axes = var_5079_axes_0, x = var_5078_cast_fp16)[name = tensor<string, []>("op_5079_cast_fp16")];
            tensor<string, []> var_5085_pad_type_0 = const()[name = tensor<string, []>("op_5085_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5085_strides_0 = const()[name = tensor<string, []>("op_5085_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5085_pad_0 = const()[name = tensor<string, []>("op_5085_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5085_dilations_0 = const()[name = tensor<string, []>("op_5085_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5085_groups_0 = const()[name = tensor<string, []>("op_5085_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45543424)))];
            tensor<fp16, [1, 64, 1, 197]> var_5085_cast_fp16 = conv(dilations = var_5085_dilations_0, groups = var_5085_groups_0, pad = var_5085_pad_0, pad_type = var_5085_pad_type_0, strides = var_5085_strides_0, weight = model_blocks_3_attn_q_projs_5_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5085_cast_fp16")];
            tensor<int32, [1]> var_5086_axes_0 = const()[name = tensor<string, []>("op_5086_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5086_cast_fp16 = squeeze(axes = var_5086_axes_0, x = var_5085_cast_fp16)[name = tensor<string, []>("op_5086_cast_fp16")];
            tensor<int32, [3]> q_411_perm_0 = const()[name = tensor<string, []>("q_411_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5093_pad_type_0 = const()[name = tensor<string, []>("op_5093_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5093_strides_0 = const()[name = tensor<string, []>("op_5093_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5093_pad_0 = const()[name = tensor<string, []>("op_5093_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5093_dilations_0 = const()[name = tensor<string, []>("op_5093_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5093_groups_0 = const()[name = tensor<string, []>("op_5093_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45641792)))];
            tensor<fp16, [1, 64, 1, 197]> var_5093_cast_fp16 = conv(dilations = var_5093_dilations_0, groups = var_5093_groups_0, pad = var_5093_pad_0, pad_type = var_5093_pad_type_0, strides = var_5093_strides_0, weight = model_blocks_3_attn_k_projs_5_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5093_cast_fp16")];
            tensor<int32, [1]> var_5094_axes_0 = const()[name = tensor<string, []>("op_5094_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5094_cast_fp16 = squeeze(axes = var_5094_axes_0, x = var_5093_cast_fp16)[name = tensor<string, []>("op_5094_cast_fp16")];
            tensor<int32, [3]> k_247_perm_0 = const()[name = tensor<string, []>("k_247_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5101_pad_type_0 = const()[name = tensor<string, []>("op_5101_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5101_strides_0 = const()[name = tensor<string, []>("op_5101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5101_pad_0 = const()[name = tensor<string, []>("op_5101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5101_dilations_0 = const()[name = tensor<string, []>("op_5101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5101_groups_0 = const()[name = tensor<string, []>("op_5101_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45740160)))];
            tensor<fp16, [1, 64, 1, 197]> var_5101_cast_fp16 = conv(dilations = var_5101_dilations_0, groups = var_5101_groups_0, pad = var_5101_pad_0, pad_type = var_5101_pad_type_0, strides = var_5101_strides_0, weight = model_blocks_3_attn_v_projs_5_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5101_cast_fp16")];
            tensor<int32, [1]> var_5102_axes_0 = const()[name = tensor<string, []>("op_5102_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5102_cast_fp16 = squeeze(axes = var_5102_axes_0, x = var_5101_cast_fp16)[name = tensor<string, []>("op_5102_cast_fp16")];
            tensor<int32, [3]> v_247_perm_0 = const()[name = tensor<string, []>("v_247_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45838528)))];
            tensor<fp16, [1, 197, 64]> q_411_cast_fp16 = transpose(perm = q_411_perm_0, x = var_5086_cast_fp16)[name = tensor<string, []>("transpose_370")];
            tensor<fp16, [1, 197, 64]> q_413_cast_fp16 = add(x = q_411_cast_fp16, y = model_blocks_3_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_413_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45838720)))];
            tensor<fp16, [1, 197, 64]> v_247_cast_fp16 = transpose(perm = v_247_perm_0, x = var_5102_cast_fp16)[name = tensor<string, []>("transpose_368")];
            tensor<fp16, [1, 197, 64]> v_249_cast_fp16 = add(x = v_247_cast_fp16, y = model_blocks_3_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_249_cast_fp16")];
            tensor<int32, [1]> q_415_axes_0 = const()[name = tensor<string, []>("q_415_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_415_cast_fp16 = expand_dims(axes = q_415_axes_0, x = q_413_cast_fp16)[name = tensor<string, []>("q_415_cast_fp16")];
            tensor<int32, [1]> k_249_axes_0 = const()[name = tensor<string, []>("k_249_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_247_cast_fp16 = transpose(perm = k_247_perm_0, x = var_5094_cast_fp16)[name = tensor<string, []>("transpose_369")];
            tensor<fp16, [1, 1, 197, 64]> k_249_cast_fp16 = expand_dims(axes = k_249_axes_0, x = k_247_cast_fp16)[name = tensor<string, []>("k_249_cast_fp16")];
            tensor<int32, [1]> v_251_axes_0 = const()[name = tensor<string, []>("v_251_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_251_cast_fp16 = expand_dims(axes = v_251_axes_0, x = v_249_cast_fp16)[name = tensor<string, []>("v_251_cast_fp16")];
            tensor<int32, [4]> var_5111_begin_0 = const()[name = tensor<string, []>("op_5111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5111_end_0 = const()[name = tensor<string, []>("op_5111_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5111_end_mask_0 = const()[name = tensor<string, []>("op_5111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5111_cast_fp16 = slice_by_index(begin = var_5111_begin_0, end = var_5111_end_0, end_mask = var_5111_end_mask_0, x = q_415_cast_fp16)[name = tensor<string, []>("op_5111_cast_fp16")];
            tensor<int32, [4]> var_5115_begin_0 = const()[name = tensor<string, []>("op_5115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5115_end_0 = const()[name = tensor<string, []>("op_5115_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5115_end_mask_0 = const()[name = tensor<string, []>("op_5115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5115_cast_fp16 = slice_by_index(begin = var_5115_begin_0, end = var_5115_end_0, end_mask = var_5115_end_mask_0, x = k_249_cast_fp16)[name = tensor<string, []>("op_5115_cast_fp16")];
            tensor<int32, [4]> var_5119_begin_0 = const()[name = tensor<string, []>("op_5119_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5119_end_0 = const()[name = tensor<string, []>("op_5119_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5119_end_mask_0 = const()[name = tensor<string, []>("op_5119_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5119_cast_fp16 = slice_by_index(begin = var_5119_begin_0, end = var_5119_end_0, end_mask = var_5119_end_mask_0, x = q_415_cast_fp16)[name = tensor<string, []>("op_5119_cast_fp16")];
            tensor<int32, [4]> var_5123_begin_0 = const()[name = tensor<string, []>("op_5123_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5123_end_0 = const()[name = tensor<string, []>("op_5123_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5123_end_mask_0 = const()[name = tensor<string, []>("op_5123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5123_cast_fp16 = slice_by_index(begin = var_5123_begin_0, end = var_5123_end_0, end_mask = var_5123_end_mask_0, x = k_249_cast_fp16)[name = tensor<string, []>("op_5123_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5125_cast_fp16 = mul(x = var_5119_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5125_cast_fp16")];
            tensor<int32, [5]> var_5133 = const()[name = tensor<string, []>("op_5133"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_563_cast_fp16 = reshape(shape = var_5133, x = var_5119_cast_fp16)[name = tensor<string, []>("x_563_cast_fp16")];
            tensor<int32, [2]> var_5135_split_sizes_0 = const()[name = tensor<string, []>("op_5135_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5135_axis_0 = const()[name = tensor<string, []>("op_5135_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5135_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5135_cast_fp16_1 = split(axis = var_5135_axis_0, split_sizes = var_5135_split_sizes_0, x = x_563_cast_fp16)[name = tensor<string, []>("op_5135_cast_fp16")];
            tensor<int32, [1]> squeeze_164_axes_0 = const()[name = tensor<string, []>("squeeze_164_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_164_cast_fp16 = squeeze(axes = squeeze_164_axes_0, x = var_5135_cast_fp16_0)[name = tensor<string, []>("squeeze_164_cast_fp16")];
            tensor<int32, [1]> squeeze_165_axes_0 = const()[name = tensor<string, []>("squeeze_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_165_cast_fp16 = squeeze(axes = squeeze_165_axes_0, x = var_5135_cast_fp16_1)[name = tensor<string, []>("squeeze_165_cast_fp16")];
            tensor<fp16, []> const_675_promoted_to_fp16 = const()[name = tensor<string, []>("const_675_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5138_cast_fp16 = mul(x = squeeze_165_cast_fp16, y = const_675_promoted_to_fp16)[name = tensor<string, []>("op_5138_cast_fp16")];
            tensor<int32, []> x_565_axis_0 = const()[name = tensor<string, []>("x_565_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_565_cast_fp16 = stack(axis = x_565_axis_0, values = (var_5138_cast_fp16, squeeze_164_cast_fp16))[name = tensor<string, []>("x_565_cast_fp16")];
            tensor<int32, [4]> var_5144 = const()[name = tensor<string, []>("op_5144"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5145_cast_fp16 = reshape(shape = var_5144, x = x_565_cast_fp16)[name = tensor<string, []>("op_5145_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5146_cast_fp16 = mul(x = var_5145_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5146_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_83_cast_fp16 = add(x = var_5125_cast_fp16, y = var_5146_cast_fp16)[name = tensor<string, []>("q_patches_83_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5148_cast_fp16 = mul(x = var_5123_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5148_cast_fp16")];
            tensor<int32, [5]> var_5156 = const()[name = tensor<string, []>("op_5156"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_569_cast_fp16 = reshape(shape = var_5156, x = var_5123_cast_fp16)[name = tensor<string, []>("x_569_cast_fp16")];
            tensor<int32, [2]> var_5158_split_sizes_0 = const()[name = tensor<string, []>("op_5158_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5158_axis_0 = const()[name = tensor<string, []>("op_5158_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5158_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5158_cast_fp16_1 = split(axis = var_5158_axis_0, split_sizes = var_5158_split_sizes_0, x = x_569_cast_fp16)[name = tensor<string, []>("op_5158_cast_fp16")];
            tensor<int32, [1]> squeeze_166_axes_0 = const()[name = tensor<string, []>("squeeze_166_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_166_cast_fp16 = squeeze(axes = squeeze_166_axes_0, x = var_5158_cast_fp16_0)[name = tensor<string, []>("squeeze_166_cast_fp16")];
            tensor<int32, [1]> squeeze_167_axes_0 = const()[name = tensor<string, []>("squeeze_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_167_cast_fp16 = squeeze(axes = squeeze_167_axes_0, x = var_5158_cast_fp16_1)[name = tensor<string, []>("squeeze_167_cast_fp16")];
            tensor<fp16, []> const_683_promoted_to_fp16 = const()[name = tensor<string, []>("const_683_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5161_cast_fp16 = mul(x = squeeze_167_cast_fp16, y = const_683_promoted_to_fp16)[name = tensor<string, []>("op_5161_cast_fp16")];
            tensor<int32, []> x_571_axis_0 = const()[name = tensor<string, []>("x_571_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_571_cast_fp16 = stack(axis = x_571_axis_0, values = (var_5161_cast_fp16, squeeze_166_cast_fp16))[name = tensor<string, []>("x_571_cast_fp16")];
            tensor<int32, [4]> var_5167 = const()[name = tensor<string, []>("op_5167"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5168_cast_fp16 = reshape(shape = var_5167, x = x_571_cast_fp16)[name = tensor<string, []>("op_5168_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5169_cast_fp16 = mul(x = var_5168_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5169_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_83_cast_fp16 = add(x = var_5148_cast_fp16, y = var_5169_cast_fp16)[name = tensor<string, []>("k_patches_83_cast_fp16")];
            tensor<bool, []> var_5172_interleave_0 = const()[name = tensor<string, []>("op_5172_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5172_cast_fp16 = concat(axis = var_20, interleave = var_5172_interleave_0, values = (var_5111_cast_fp16, q_patches_83_cast_fp16))[name = tensor<string, []>("op_5172_cast_fp16")];
            tensor<bool, []> var_5175_interleave_0 = const()[name = tensor<string, []>("op_5175_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5175_cast_fp16 = concat(axis = var_20, interleave = var_5175_interleave_0, values = (var_5115_cast_fp16, k_patches_83_cast_fp16))[name = tensor<string, []>("op_5175_cast_fp16")];
            tensor<fp16, []> var_5177_to_fp16 = const()[name = tensor<string, []>("op_5177_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_419_cast_fp16 = mul(x = var_5172_cast_fp16, y = var_5177_to_fp16)[name = tensor<string, []>("q_419_cast_fp16")];
            tensor<bool, []> attn_165_transpose_x_1 = const()[name = tensor<string, []>("attn_165_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_165_transpose_y_1 = const()[name = tensor<string, []>("attn_165_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_165_cast_fp16 = matmul(transpose_x = attn_165_transpose_x_1, transpose_y = attn_165_transpose_y_1, x = q_419_cast_fp16, y = var_5175_cast_fp16)[name = tensor<string, []>("attn_165_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_167_cast_fp16 = softmax(axis = var_21, x = attn_165_cast_fp16)[name = tensor<string, []>("attn_167_cast_fp16")];
            tensor<bool, []> var_5182_transpose_x_0 = const()[name = tensor<string, []>("op_5182_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5182_transpose_y_0 = const()[name = tensor<string, []>("op_5182_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5182_cast_fp16 = matmul(transpose_x = var_5182_transpose_x_0, transpose_y = var_5182_transpose_y_0, x = attn_167_cast_fp16, y = v_251_cast_fp16)[name = tensor<string, []>("op_5182_cast_fp16")];
            tensor<int32, [1]> var_5183_axes_0 = const()[name = tensor<string, []>("op_5183_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5183_cast_fp16 = squeeze(axes = var_5183_axes_0, x = var_5182_cast_fp16)[name = tensor<string, []>("op_5183_cast_fp16")];
            tensor<string, []> var_5189_pad_type_0 = const()[name = tensor<string, []>("op_5189_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5189_strides_0 = const()[name = tensor<string, []>("op_5189_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5189_pad_0 = const()[name = tensor<string, []>("op_5189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5189_dilations_0 = const()[name = tensor<string, []>("op_5189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5189_groups_0 = const()[name = tensor<string, []>("op_5189_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45838912)))];
            tensor<fp16, [1, 64, 1, 197]> var_5189_cast_fp16 = conv(dilations = var_5189_dilations_0, groups = var_5189_groups_0, pad = var_5189_pad_0, pad_type = var_5189_pad_type_0, strides = var_5189_strides_0, weight = model_blocks_3_attn_q_projs_6_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5189_cast_fp16")];
            tensor<int32, [1]> var_5190_axes_0 = const()[name = tensor<string, []>("op_5190_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5190_cast_fp16 = squeeze(axes = var_5190_axes_0, x = var_5189_cast_fp16)[name = tensor<string, []>("op_5190_cast_fp16")];
            tensor<int32, [3]> q_421_perm_0 = const()[name = tensor<string, []>("q_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5197_pad_type_0 = const()[name = tensor<string, []>("op_5197_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5197_strides_0 = const()[name = tensor<string, []>("op_5197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5197_pad_0 = const()[name = tensor<string, []>("op_5197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5197_dilations_0 = const()[name = tensor<string, []>("op_5197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5197_groups_0 = const()[name = tensor<string, []>("op_5197_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45937280)))];
            tensor<fp16, [1, 64, 1, 197]> var_5197_cast_fp16 = conv(dilations = var_5197_dilations_0, groups = var_5197_groups_0, pad = var_5197_pad_0, pad_type = var_5197_pad_type_0, strides = var_5197_strides_0, weight = model_blocks_3_attn_k_projs_6_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5197_cast_fp16")];
            tensor<int32, [1]> var_5198_axes_0 = const()[name = tensor<string, []>("op_5198_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5198_cast_fp16 = squeeze(axes = var_5198_axes_0, x = var_5197_cast_fp16)[name = tensor<string, []>("op_5198_cast_fp16")];
            tensor<int32, [3]> k_253_perm_0 = const()[name = tensor<string, []>("k_253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5205_pad_type_0 = const()[name = tensor<string, []>("op_5205_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5205_strides_0 = const()[name = tensor<string, []>("op_5205_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5205_pad_0 = const()[name = tensor<string, []>("op_5205_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5205_dilations_0 = const()[name = tensor<string, []>("op_5205_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5205_groups_0 = const()[name = tensor<string, []>("op_5205_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46035648)))];
            tensor<fp16, [1, 64, 1, 197]> var_5205_cast_fp16 = conv(dilations = var_5205_dilations_0, groups = var_5205_groups_0, pad = var_5205_pad_0, pad_type = var_5205_pad_type_0, strides = var_5205_strides_0, weight = model_blocks_3_attn_v_projs_6_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5205_cast_fp16")];
            tensor<int32, [1]> var_5206_axes_0 = const()[name = tensor<string, []>("op_5206_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5206_cast_fp16 = squeeze(axes = var_5206_axes_0, x = var_5205_cast_fp16)[name = tensor<string, []>("op_5206_cast_fp16")];
            tensor<int32, [3]> v_253_perm_0 = const()[name = tensor<string, []>("v_253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46134016)))];
            tensor<fp16, [1, 197, 64]> q_421_cast_fp16 = transpose(perm = q_421_perm_0, x = var_5190_cast_fp16)[name = tensor<string, []>("transpose_367")];
            tensor<fp16, [1, 197, 64]> q_423_cast_fp16 = add(x = q_421_cast_fp16, y = model_blocks_3_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_423_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46134208)))];
            tensor<fp16, [1, 197, 64]> v_253_cast_fp16 = transpose(perm = v_253_perm_0, x = var_5206_cast_fp16)[name = tensor<string, []>("transpose_365")];
            tensor<fp16, [1, 197, 64]> v_255_cast_fp16 = add(x = v_253_cast_fp16, y = model_blocks_3_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_255_cast_fp16")];
            tensor<int32, [1]> q_425_axes_0 = const()[name = tensor<string, []>("q_425_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_425_cast_fp16 = expand_dims(axes = q_425_axes_0, x = q_423_cast_fp16)[name = tensor<string, []>("q_425_cast_fp16")];
            tensor<int32, [1]> k_255_axes_0 = const()[name = tensor<string, []>("k_255_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_253_cast_fp16 = transpose(perm = k_253_perm_0, x = var_5198_cast_fp16)[name = tensor<string, []>("transpose_366")];
            tensor<fp16, [1, 1, 197, 64]> k_255_cast_fp16 = expand_dims(axes = k_255_axes_0, x = k_253_cast_fp16)[name = tensor<string, []>("k_255_cast_fp16")];
            tensor<int32, [1]> v_257_axes_0 = const()[name = tensor<string, []>("v_257_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_257_cast_fp16 = expand_dims(axes = v_257_axes_0, x = v_255_cast_fp16)[name = tensor<string, []>("v_257_cast_fp16")];
            tensor<int32, [4]> var_5215_begin_0 = const()[name = tensor<string, []>("op_5215_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5215_end_0 = const()[name = tensor<string, []>("op_5215_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5215_end_mask_0 = const()[name = tensor<string, []>("op_5215_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, x = q_425_cast_fp16)[name = tensor<string, []>("op_5215_cast_fp16")];
            tensor<int32, [4]> var_5219_begin_0 = const()[name = tensor<string, []>("op_5219_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5219_end_0 = const()[name = tensor<string, []>("op_5219_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5219_end_mask_0 = const()[name = tensor<string, []>("op_5219_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5219_cast_fp16 = slice_by_index(begin = var_5219_begin_0, end = var_5219_end_0, end_mask = var_5219_end_mask_0, x = k_255_cast_fp16)[name = tensor<string, []>("op_5219_cast_fp16")];
            tensor<int32, [4]> var_5223_begin_0 = const()[name = tensor<string, []>("op_5223_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5223_end_0 = const()[name = tensor<string, []>("op_5223_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5223_end_mask_0 = const()[name = tensor<string, []>("op_5223_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5223_cast_fp16 = slice_by_index(begin = var_5223_begin_0, end = var_5223_end_0, end_mask = var_5223_end_mask_0, x = q_425_cast_fp16)[name = tensor<string, []>("op_5223_cast_fp16")];
            tensor<int32, [4]> var_5227_begin_0 = const()[name = tensor<string, []>("op_5227_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5227_end_0 = const()[name = tensor<string, []>("op_5227_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5227_end_mask_0 = const()[name = tensor<string, []>("op_5227_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5227_cast_fp16 = slice_by_index(begin = var_5227_begin_0, end = var_5227_end_0, end_mask = var_5227_end_mask_0, x = k_255_cast_fp16)[name = tensor<string, []>("op_5227_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5229_cast_fp16 = mul(x = var_5223_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5229_cast_fp16")];
            tensor<int32, [5]> var_5237 = const()[name = tensor<string, []>("op_5237"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_575_cast_fp16 = reshape(shape = var_5237, x = var_5223_cast_fp16)[name = tensor<string, []>("x_575_cast_fp16")];
            tensor<int32, [2]> var_5239_split_sizes_0 = const()[name = tensor<string, []>("op_5239_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5239_axis_0 = const()[name = tensor<string, []>("op_5239_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5239_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5239_cast_fp16_1 = split(axis = var_5239_axis_0, split_sizes = var_5239_split_sizes_0, x = x_575_cast_fp16)[name = tensor<string, []>("op_5239_cast_fp16")];
            tensor<int32, [1]> squeeze_168_axes_0 = const()[name = tensor<string, []>("squeeze_168_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_168_cast_fp16 = squeeze(axes = squeeze_168_axes_0, x = var_5239_cast_fp16_0)[name = tensor<string, []>("squeeze_168_cast_fp16")];
            tensor<int32, [1]> squeeze_169_axes_0 = const()[name = tensor<string, []>("squeeze_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_169_cast_fp16 = squeeze(axes = squeeze_169_axes_0, x = var_5239_cast_fp16_1)[name = tensor<string, []>("squeeze_169_cast_fp16")];
            tensor<fp16, []> const_691_promoted_to_fp16 = const()[name = tensor<string, []>("const_691_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5242_cast_fp16 = mul(x = squeeze_169_cast_fp16, y = const_691_promoted_to_fp16)[name = tensor<string, []>("op_5242_cast_fp16")];
            tensor<int32, []> x_577_axis_0 = const()[name = tensor<string, []>("x_577_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_577_cast_fp16 = stack(axis = x_577_axis_0, values = (var_5242_cast_fp16, squeeze_168_cast_fp16))[name = tensor<string, []>("x_577_cast_fp16")];
            tensor<int32, [4]> var_5248 = const()[name = tensor<string, []>("op_5248"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5249_cast_fp16 = reshape(shape = var_5248, x = x_577_cast_fp16)[name = tensor<string, []>("op_5249_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5250_cast_fp16 = mul(x = var_5249_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5250_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_85_cast_fp16 = add(x = var_5229_cast_fp16, y = var_5250_cast_fp16)[name = tensor<string, []>("q_patches_85_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5252_cast_fp16 = mul(x = var_5227_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5252_cast_fp16")];
            tensor<int32, [5]> var_5260 = const()[name = tensor<string, []>("op_5260"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_581_cast_fp16 = reshape(shape = var_5260, x = var_5227_cast_fp16)[name = tensor<string, []>("x_581_cast_fp16")];
            tensor<int32, [2]> var_5262_split_sizes_0 = const()[name = tensor<string, []>("op_5262_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5262_axis_0 = const()[name = tensor<string, []>("op_5262_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5262_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5262_cast_fp16_1 = split(axis = var_5262_axis_0, split_sizes = var_5262_split_sizes_0, x = x_581_cast_fp16)[name = tensor<string, []>("op_5262_cast_fp16")];
            tensor<int32, [1]> squeeze_170_axes_0 = const()[name = tensor<string, []>("squeeze_170_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_170_cast_fp16 = squeeze(axes = squeeze_170_axes_0, x = var_5262_cast_fp16_0)[name = tensor<string, []>("squeeze_170_cast_fp16")];
            tensor<int32, [1]> squeeze_171_axes_0 = const()[name = tensor<string, []>("squeeze_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_171_cast_fp16 = squeeze(axes = squeeze_171_axes_0, x = var_5262_cast_fp16_1)[name = tensor<string, []>("squeeze_171_cast_fp16")];
            tensor<fp16, []> const_699_promoted_to_fp16 = const()[name = tensor<string, []>("const_699_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5265_cast_fp16 = mul(x = squeeze_171_cast_fp16, y = const_699_promoted_to_fp16)[name = tensor<string, []>("op_5265_cast_fp16")];
            tensor<int32, []> x_583_axis_0 = const()[name = tensor<string, []>("x_583_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_583_cast_fp16 = stack(axis = x_583_axis_0, values = (var_5265_cast_fp16, squeeze_170_cast_fp16))[name = tensor<string, []>("x_583_cast_fp16")];
            tensor<int32, [4]> var_5271 = const()[name = tensor<string, []>("op_5271"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5272_cast_fp16 = reshape(shape = var_5271, x = x_583_cast_fp16)[name = tensor<string, []>("op_5272_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5273_cast_fp16 = mul(x = var_5272_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5273_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_85_cast_fp16 = add(x = var_5252_cast_fp16, y = var_5273_cast_fp16)[name = tensor<string, []>("k_patches_85_cast_fp16")];
            tensor<bool, []> var_5276_interleave_0 = const()[name = tensor<string, []>("op_5276_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5276_cast_fp16 = concat(axis = var_20, interleave = var_5276_interleave_0, values = (var_5215_cast_fp16, q_patches_85_cast_fp16))[name = tensor<string, []>("op_5276_cast_fp16")];
            tensor<bool, []> var_5279_interleave_0 = const()[name = tensor<string, []>("op_5279_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5279_cast_fp16 = concat(axis = var_20, interleave = var_5279_interleave_0, values = (var_5219_cast_fp16, k_patches_85_cast_fp16))[name = tensor<string, []>("op_5279_cast_fp16")];
            tensor<fp16, []> var_5281_to_fp16 = const()[name = tensor<string, []>("op_5281_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_429_cast_fp16 = mul(x = var_5276_cast_fp16, y = var_5281_to_fp16)[name = tensor<string, []>("q_429_cast_fp16")];
            tensor<bool, []> attn_169_transpose_x_1 = const()[name = tensor<string, []>("attn_169_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_169_transpose_y_1 = const()[name = tensor<string, []>("attn_169_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_169_cast_fp16 = matmul(transpose_x = attn_169_transpose_x_1, transpose_y = attn_169_transpose_y_1, x = q_429_cast_fp16, y = var_5279_cast_fp16)[name = tensor<string, []>("attn_169_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_171_cast_fp16 = softmax(axis = var_21, x = attn_169_cast_fp16)[name = tensor<string, []>("attn_171_cast_fp16")];
            tensor<bool, []> var_5286_transpose_x_0 = const()[name = tensor<string, []>("op_5286_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5286_transpose_y_0 = const()[name = tensor<string, []>("op_5286_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5286_cast_fp16 = matmul(transpose_x = var_5286_transpose_x_0, transpose_y = var_5286_transpose_y_0, x = attn_171_cast_fp16, y = v_257_cast_fp16)[name = tensor<string, []>("op_5286_cast_fp16")];
            tensor<int32, [1]> var_5287_axes_0 = const()[name = tensor<string, []>("op_5287_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5287_cast_fp16 = squeeze(axes = var_5287_axes_0, x = var_5286_cast_fp16)[name = tensor<string, []>("op_5287_cast_fp16")];
            tensor<string, []> var_5293_pad_type_0 = const()[name = tensor<string, []>("op_5293_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5293_strides_0 = const()[name = tensor<string, []>("op_5293_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5293_pad_0 = const()[name = tensor<string, []>("op_5293_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5293_dilations_0 = const()[name = tensor<string, []>("op_5293_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5293_groups_0 = const()[name = tensor<string, []>("op_5293_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46134400)))];
            tensor<fp16, [1, 64, 1, 197]> var_5293_cast_fp16 = conv(dilations = var_5293_dilations_0, groups = var_5293_groups_0, pad = var_5293_pad_0, pad_type = var_5293_pad_type_0, strides = var_5293_strides_0, weight = model_blocks_3_attn_q_projs_7_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5293_cast_fp16")];
            tensor<int32, [1]> var_5294_axes_0 = const()[name = tensor<string, []>("op_5294_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5294_cast_fp16 = squeeze(axes = var_5294_axes_0, x = var_5293_cast_fp16)[name = tensor<string, []>("op_5294_cast_fp16")];
            tensor<int32, [3]> q_431_perm_0 = const()[name = tensor<string, []>("q_431_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5301_pad_type_0 = const()[name = tensor<string, []>("op_5301_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5301_strides_0 = const()[name = tensor<string, []>("op_5301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5301_pad_0 = const()[name = tensor<string, []>("op_5301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5301_dilations_0 = const()[name = tensor<string, []>("op_5301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5301_groups_0 = const()[name = tensor<string, []>("op_5301_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46232768)))];
            tensor<fp16, [1, 64, 1, 197]> var_5301_cast_fp16 = conv(dilations = var_5301_dilations_0, groups = var_5301_groups_0, pad = var_5301_pad_0, pad_type = var_5301_pad_type_0, strides = var_5301_strides_0, weight = model_blocks_3_attn_k_projs_7_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5301_cast_fp16")];
            tensor<int32, [1]> var_5302_axes_0 = const()[name = tensor<string, []>("op_5302_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5302_cast_fp16 = squeeze(axes = var_5302_axes_0, x = var_5301_cast_fp16)[name = tensor<string, []>("op_5302_cast_fp16")];
            tensor<int32, [3]> k_259_perm_0 = const()[name = tensor<string, []>("k_259_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5309_pad_type_0 = const()[name = tensor<string, []>("op_5309_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5309_strides_0 = const()[name = tensor<string, []>("op_5309_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5309_pad_0 = const()[name = tensor<string, []>("op_5309_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5309_dilations_0 = const()[name = tensor<string, []>("op_5309_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5309_groups_0 = const()[name = tensor<string, []>("op_5309_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46331136)))];
            tensor<fp16, [1, 64, 1, 197]> var_5309_cast_fp16 = conv(dilations = var_5309_dilations_0, groups = var_5309_groups_0, pad = var_5309_pad_0, pad_type = var_5309_pad_type_0, strides = var_5309_strides_0, weight = model_blocks_3_attn_v_projs_7_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5309_cast_fp16")];
            tensor<int32, [1]> var_5310_axes_0 = const()[name = tensor<string, []>("op_5310_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5310_cast_fp16 = squeeze(axes = var_5310_axes_0, x = var_5309_cast_fp16)[name = tensor<string, []>("op_5310_cast_fp16")];
            tensor<int32, [3]> v_259_perm_0 = const()[name = tensor<string, []>("v_259_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46429504)))];
            tensor<fp16, [1, 197, 64]> q_431_cast_fp16 = transpose(perm = q_431_perm_0, x = var_5294_cast_fp16)[name = tensor<string, []>("transpose_364")];
            tensor<fp16, [1, 197, 64]> q_433_cast_fp16 = add(x = q_431_cast_fp16, y = model_blocks_3_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_433_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46429696)))];
            tensor<fp16, [1, 197, 64]> v_259_cast_fp16 = transpose(perm = v_259_perm_0, x = var_5310_cast_fp16)[name = tensor<string, []>("transpose_362")];
            tensor<fp16, [1, 197, 64]> v_261_cast_fp16 = add(x = v_259_cast_fp16, y = model_blocks_3_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_261_cast_fp16")];
            tensor<int32, [1]> q_435_axes_0 = const()[name = tensor<string, []>("q_435_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_435_cast_fp16 = expand_dims(axes = q_435_axes_0, x = q_433_cast_fp16)[name = tensor<string, []>("q_435_cast_fp16")];
            tensor<int32, [1]> k_261_axes_0 = const()[name = tensor<string, []>("k_261_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_259_cast_fp16 = transpose(perm = k_259_perm_0, x = var_5302_cast_fp16)[name = tensor<string, []>("transpose_363")];
            tensor<fp16, [1, 1, 197, 64]> k_261_cast_fp16 = expand_dims(axes = k_261_axes_0, x = k_259_cast_fp16)[name = tensor<string, []>("k_261_cast_fp16")];
            tensor<int32, [1]> v_263_axes_0 = const()[name = tensor<string, []>("v_263_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_263_cast_fp16 = expand_dims(axes = v_263_axes_0, x = v_261_cast_fp16)[name = tensor<string, []>("v_263_cast_fp16")];
            tensor<int32, [4]> var_5319_begin_0 = const()[name = tensor<string, []>("op_5319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5319_end_0 = const()[name = tensor<string, []>("op_5319_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5319_end_mask_0 = const()[name = tensor<string, []>("op_5319_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5319_cast_fp16 = slice_by_index(begin = var_5319_begin_0, end = var_5319_end_0, end_mask = var_5319_end_mask_0, x = q_435_cast_fp16)[name = tensor<string, []>("op_5319_cast_fp16")];
            tensor<int32, [4]> var_5323_begin_0 = const()[name = tensor<string, []>("op_5323_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5323_end_0 = const()[name = tensor<string, []>("op_5323_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5323_end_mask_0 = const()[name = tensor<string, []>("op_5323_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5323_cast_fp16 = slice_by_index(begin = var_5323_begin_0, end = var_5323_end_0, end_mask = var_5323_end_mask_0, x = k_261_cast_fp16)[name = tensor<string, []>("op_5323_cast_fp16")];
            tensor<int32, [4]> var_5327_begin_0 = const()[name = tensor<string, []>("op_5327_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = tensor<string, []>("op_5327_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = tensor<string, []>("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5327_cast_fp16 = slice_by_index(begin = var_5327_begin_0, end = var_5327_end_0, end_mask = var_5327_end_mask_0, x = q_435_cast_fp16)[name = tensor<string, []>("op_5327_cast_fp16")];
            tensor<int32, [4]> var_5331_begin_0 = const()[name = tensor<string, []>("op_5331_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5331_end_0 = const()[name = tensor<string, []>("op_5331_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5331_end_mask_0 = const()[name = tensor<string, []>("op_5331_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5331_cast_fp16 = slice_by_index(begin = var_5331_begin_0, end = var_5331_end_0, end_mask = var_5331_end_mask_0, x = k_261_cast_fp16)[name = tensor<string, []>("op_5331_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5333_cast_fp16 = mul(x = var_5327_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5333_cast_fp16")];
            tensor<int32, [5]> var_5341 = const()[name = tensor<string, []>("op_5341"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_587_cast_fp16 = reshape(shape = var_5341, x = var_5327_cast_fp16)[name = tensor<string, []>("x_587_cast_fp16")];
            tensor<int32, [2]> var_5343_split_sizes_0 = const()[name = tensor<string, []>("op_5343_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5343_axis_0 = const()[name = tensor<string, []>("op_5343_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5343_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5343_cast_fp16_1 = split(axis = var_5343_axis_0, split_sizes = var_5343_split_sizes_0, x = x_587_cast_fp16)[name = tensor<string, []>("op_5343_cast_fp16")];
            tensor<int32, [1]> squeeze_172_axes_0 = const()[name = tensor<string, []>("squeeze_172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_172_cast_fp16 = squeeze(axes = squeeze_172_axes_0, x = var_5343_cast_fp16_0)[name = tensor<string, []>("squeeze_172_cast_fp16")];
            tensor<int32, [1]> squeeze_173_axes_0 = const()[name = tensor<string, []>("squeeze_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_173_cast_fp16 = squeeze(axes = squeeze_173_axes_0, x = var_5343_cast_fp16_1)[name = tensor<string, []>("squeeze_173_cast_fp16")];
            tensor<fp16, []> const_707_promoted_to_fp16 = const()[name = tensor<string, []>("const_707_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5346_cast_fp16 = mul(x = squeeze_173_cast_fp16, y = const_707_promoted_to_fp16)[name = tensor<string, []>("op_5346_cast_fp16")];
            tensor<int32, []> x_589_axis_0 = const()[name = tensor<string, []>("x_589_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_589_cast_fp16 = stack(axis = x_589_axis_0, values = (var_5346_cast_fp16, squeeze_172_cast_fp16))[name = tensor<string, []>("x_589_cast_fp16")];
            tensor<int32, [4]> var_5352 = const()[name = tensor<string, []>("op_5352"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5353_cast_fp16 = reshape(shape = var_5352, x = x_589_cast_fp16)[name = tensor<string, []>("op_5353_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5354_cast_fp16 = mul(x = var_5353_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5354_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_87_cast_fp16 = add(x = var_5333_cast_fp16, y = var_5354_cast_fp16)[name = tensor<string, []>("q_patches_87_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5356_cast_fp16 = mul(x = var_5331_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5356_cast_fp16")];
            tensor<int32, [5]> var_5364 = const()[name = tensor<string, []>("op_5364"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_593_cast_fp16 = reshape(shape = var_5364, x = var_5331_cast_fp16)[name = tensor<string, []>("x_593_cast_fp16")];
            tensor<int32, [2]> var_5366_split_sizes_0 = const()[name = tensor<string, []>("op_5366_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5366_axis_0 = const()[name = tensor<string, []>("op_5366_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5366_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5366_cast_fp16_1 = split(axis = var_5366_axis_0, split_sizes = var_5366_split_sizes_0, x = x_593_cast_fp16)[name = tensor<string, []>("op_5366_cast_fp16")];
            tensor<int32, [1]> squeeze_174_axes_0 = const()[name = tensor<string, []>("squeeze_174_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_174_cast_fp16 = squeeze(axes = squeeze_174_axes_0, x = var_5366_cast_fp16_0)[name = tensor<string, []>("squeeze_174_cast_fp16")];
            tensor<int32, [1]> squeeze_175_axes_0 = const()[name = tensor<string, []>("squeeze_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_175_cast_fp16 = squeeze(axes = squeeze_175_axes_0, x = var_5366_cast_fp16_1)[name = tensor<string, []>("squeeze_175_cast_fp16")];
            tensor<fp16, []> const_715_promoted_to_fp16 = const()[name = tensor<string, []>("const_715_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5369_cast_fp16 = mul(x = squeeze_175_cast_fp16, y = const_715_promoted_to_fp16)[name = tensor<string, []>("op_5369_cast_fp16")];
            tensor<int32, []> x_595_axis_0 = const()[name = tensor<string, []>("x_595_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_595_cast_fp16 = stack(axis = x_595_axis_0, values = (var_5369_cast_fp16, squeeze_174_cast_fp16))[name = tensor<string, []>("x_595_cast_fp16")];
            tensor<int32, [4]> var_5375 = const()[name = tensor<string, []>("op_5375"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5376_cast_fp16 = reshape(shape = var_5375, x = x_595_cast_fp16)[name = tensor<string, []>("op_5376_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5377_cast_fp16 = mul(x = var_5376_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5377_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_87_cast_fp16 = add(x = var_5356_cast_fp16, y = var_5377_cast_fp16)[name = tensor<string, []>("k_patches_87_cast_fp16")];
            tensor<bool, []> var_5380_interleave_0 = const()[name = tensor<string, []>("op_5380_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5380_cast_fp16 = concat(axis = var_20, interleave = var_5380_interleave_0, values = (var_5319_cast_fp16, q_patches_87_cast_fp16))[name = tensor<string, []>("op_5380_cast_fp16")];
            tensor<bool, []> var_5383_interleave_0 = const()[name = tensor<string, []>("op_5383_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5383_cast_fp16 = concat(axis = var_20, interleave = var_5383_interleave_0, values = (var_5323_cast_fp16, k_patches_87_cast_fp16))[name = tensor<string, []>("op_5383_cast_fp16")];
            tensor<fp16, []> var_5385_to_fp16 = const()[name = tensor<string, []>("op_5385_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_439_cast_fp16 = mul(x = var_5380_cast_fp16, y = var_5385_to_fp16)[name = tensor<string, []>("q_439_cast_fp16")];
            tensor<bool, []> attn_173_transpose_x_1 = const()[name = tensor<string, []>("attn_173_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_173_transpose_y_1 = const()[name = tensor<string, []>("attn_173_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_173_cast_fp16 = matmul(transpose_x = attn_173_transpose_x_1, transpose_y = attn_173_transpose_y_1, x = q_439_cast_fp16, y = var_5383_cast_fp16)[name = tensor<string, []>("attn_173_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_175_cast_fp16 = softmax(axis = var_21, x = attn_173_cast_fp16)[name = tensor<string, []>("attn_175_cast_fp16")];
            tensor<bool, []> var_5390_transpose_x_0 = const()[name = tensor<string, []>("op_5390_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5390_transpose_y_0 = const()[name = tensor<string, []>("op_5390_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5390_cast_fp16 = matmul(transpose_x = var_5390_transpose_x_0, transpose_y = var_5390_transpose_y_0, x = attn_175_cast_fp16, y = v_263_cast_fp16)[name = tensor<string, []>("op_5390_cast_fp16")];
            tensor<int32, [1]> var_5391_axes_0 = const()[name = tensor<string, []>("op_5391_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5391_cast_fp16 = squeeze(axes = var_5391_axes_0, x = var_5390_cast_fp16)[name = tensor<string, []>("op_5391_cast_fp16")];
            tensor<string, []> var_5397_pad_type_0 = const()[name = tensor<string, []>("op_5397_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5397_strides_0 = const()[name = tensor<string, []>("op_5397_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5397_pad_0 = const()[name = tensor<string, []>("op_5397_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5397_dilations_0 = const()[name = tensor<string, []>("op_5397_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5397_groups_0 = const()[name = tensor<string, []>("op_5397_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46429888)))];
            tensor<fp16, [1, 64, 1, 197]> var_5397_cast_fp16 = conv(dilations = var_5397_dilations_0, groups = var_5397_groups_0, pad = var_5397_pad_0, pad_type = var_5397_pad_type_0, strides = var_5397_strides_0, weight = model_blocks_3_attn_q_projs_8_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5397_cast_fp16")];
            tensor<int32, [1]> var_5398_axes_0 = const()[name = tensor<string, []>("op_5398_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5398_cast_fp16 = squeeze(axes = var_5398_axes_0, x = var_5397_cast_fp16)[name = tensor<string, []>("op_5398_cast_fp16")];
            tensor<int32, [3]> q_441_perm_0 = const()[name = tensor<string, []>("q_441_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5405_pad_type_0 = const()[name = tensor<string, []>("op_5405_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5405_strides_0 = const()[name = tensor<string, []>("op_5405_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5405_pad_0 = const()[name = tensor<string, []>("op_5405_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5405_dilations_0 = const()[name = tensor<string, []>("op_5405_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5405_groups_0 = const()[name = tensor<string, []>("op_5405_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46528256)))];
            tensor<fp16, [1, 64, 1, 197]> var_5405_cast_fp16 = conv(dilations = var_5405_dilations_0, groups = var_5405_groups_0, pad = var_5405_pad_0, pad_type = var_5405_pad_type_0, strides = var_5405_strides_0, weight = model_blocks_3_attn_k_projs_8_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5405_cast_fp16")];
            tensor<int32, [1]> var_5406_axes_0 = const()[name = tensor<string, []>("op_5406_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5406_cast_fp16 = squeeze(axes = var_5406_axes_0, x = var_5405_cast_fp16)[name = tensor<string, []>("op_5406_cast_fp16")];
            tensor<int32, [3]> k_265_perm_0 = const()[name = tensor<string, []>("k_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5413_pad_type_0 = const()[name = tensor<string, []>("op_5413_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5413_strides_0 = const()[name = tensor<string, []>("op_5413_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5413_pad_0 = const()[name = tensor<string, []>("op_5413_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5413_dilations_0 = const()[name = tensor<string, []>("op_5413_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5413_groups_0 = const()[name = tensor<string, []>("op_5413_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46626624)))];
            tensor<fp16, [1, 64, 1, 197]> var_5413_cast_fp16 = conv(dilations = var_5413_dilations_0, groups = var_5413_groups_0, pad = var_5413_pad_0, pad_type = var_5413_pad_type_0, strides = var_5413_strides_0, weight = model_blocks_3_attn_v_projs_8_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5413_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = tensor<string, []>("op_5414_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5414_cast_fp16 = squeeze(axes = var_5414_axes_0, x = var_5413_cast_fp16)[name = tensor<string, []>("op_5414_cast_fp16")];
            tensor<int32, [3]> v_265_perm_0 = const()[name = tensor<string, []>("v_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46724992)))];
            tensor<fp16, [1, 197, 64]> q_441_cast_fp16 = transpose(perm = q_441_perm_0, x = var_5398_cast_fp16)[name = tensor<string, []>("transpose_361")];
            tensor<fp16, [1, 197, 64]> q_443_cast_fp16 = add(x = q_441_cast_fp16, y = model_blocks_3_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_443_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46725184)))];
            tensor<fp16, [1, 197, 64]> v_265_cast_fp16 = transpose(perm = v_265_perm_0, x = var_5414_cast_fp16)[name = tensor<string, []>("transpose_359")];
            tensor<fp16, [1, 197, 64]> v_267_cast_fp16 = add(x = v_265_cast_fp16, y = model_blocks_3_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_267_cast_fp16")];
            tensor<int32, [1]> q_445_axes_0 = const()[name = tensor<string, []>("q_445_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_445_cast_fp16 = expand_dims(axes = q_445_axes_0, x = q_443_cast_fp16)[name = tensor<string, []>("q_445_cast_fp16")];
            tensor<int32, [1]> k_267_axes_0 = const()[name = tensor<string, []>("k_267_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_265_cast_fp16 = transpose(perm = k_265_perm_0, x = var_5406_cast_fp16)[name = tensor<string, []>("transpose_360")];
            tensor<fp16, [1, 1, 197, 64]> k_267_cast_fp16 = expand_dims(axes = k_267_axes_0, x = k_265_cast_fp16)[name = tensor<string, []>("k_267_cast_fp16")];
            tensor<int32, [1]> v_269_axes_0 = const()[name = tensor<string, []>("v_269_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_269_cast_fp16 = expand_dims(axes = v_269_axes_0, x = v_267_cast_fp16)[name = tensor<string, []>("v_269_cast_fp16")];
            tensor<int32, [4]> var_5423_begin_0 = const()[name = tensor<string, []>("op_5423_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5423_end_0 = const()[name = tensor<string, []>("op_5423_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5423_end_mask_0 = const()[name = tensor<string, []>("op_5423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5423_cast_fp16 = slice_by_index(begin = var_5423_begin_0, end = var_5423_end_0, end_mask = var_5423_end_mask_0, x = q_445_cast_fp16)[name = tensor<string, []>("op_5423_cast_fp16")];
            tensor<int32, [4]> var_5427_begin_0 = const()[name = tensor<string, []>("op_5427_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5427_end_0 = const()[name = tensor<string, []>("op_5427_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5427_end_mask_0 = const()[name = tensor<string, []>("op_5427_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = k_267_cast_fp16)[name = tensor<string, []>("op_5427_cast_fp16")];
            tensor<int32, [4]> var_5431_begin_0 = const()[name = tensor<string, []>("op_5431_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5431_end_0 = const()[name = tensor<string, []>("op_5431_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5431_end_mask_0 = const()[name = tensor<string, []>("op_5431_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5431_cast_fp16 = slice_by_index(begin = var_5431_begin_0, end = var_5431_end_0, end_mask = var_5431_end_mask_0, x = q_445_cast_fp16)[name = tensor<string, []>("op_5431_cast_fp16")];
            tensor<int32, [4]> var_5435_begin_0 = const()[name = tensor<string, []>("op_5435_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5435_end_0 = const()[name = tensor<string, []>("op_5435_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5435_end_mask_0 = const()[name = tensor<string, []>("op_5435_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5435_cast_fp16 = slice_by_index(begin = var_5435_begin_0, end = var_5435_end_0, end_mask = var_5435_end_mask_0, x = k_267_cast_fp16)[name = tensor<string, []>("op_5435_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5437_cast_fp16 = mul(x = var_5431_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5437_cast_fp16")];
            tensor<int32, [5]> var_5445 = const()[name = tensor<string, []>("op_5445"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_599_cast_fp16 = reshape(shape = var_5445, x = var_5431_cast_fp16)[name = tensor<string, []>("x_599_cast_fp16")];
            tensor<int32, [2]> var_5447_split_sizes_0 = const()[name = tensor<string, []>("op_5447_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5447_axis_0 = const()[name = tensor<string, []>("op_5447_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5447_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5447_cast_fp16_1 = split(axis = var_5447_axis_0, split_sizes = var_5447_split_sizes_0, x = x_599_cast_fp16)[name = tensor<string, []>("op_5447_cast_fp16")];
            tensor<int32, [1]> squeeze_176_axes_0 = const()[name = tensor<string, []>("squeeze_176_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_176_cast_fp16 = squeeze(axes = squeeze_176_axes_0, x = var_5447_cast_fp16_0)[name = tensor<string, []>("squeeze_176_cast_fp16")];
            tensor<int32, [1]> squeeze_177_axes_0 = const()[name = tensor<string, []>("squeeze_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_177_cast_fp16 = squeeze(axes = squeeze_177_axes_0, x = var_5447_cast_fp16_1)[name = tensor<string, []>("squeeze_177_cast_fp16")];
            tensor<fp16, []> const_723_promoted_to_fp16 = const()[name = tensor<string, []>("const_723_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5450_cast_fp16 = mul(x = squeeze_177_cast_fp16, y = const_723_promoted_to_fp16)[name = tensor<string, []>("op_5450_cast_fp16")];
            tensor<int32, []> x_601_axis_0 = const()[name = tensor<string, []>("x_601_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_601_cast_fp16 = stack(axis = x_601_axis_0, values = (var_5450_cast_fp16, squeeze_176_cast_fp16))[name = tensor<string, []>("x_601_cast_fp16")];
            tensor<int32, [4]> var_5456 = const()[name = tensor<string, []>("op_5456"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5457_cast_fp16 = reshape(shape = var_5456, x = x_601_cast_fp16)[name = tensor<string, []>("op_5457_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5458_cast_fp16 = mul(x = var_5457_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5458_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_89_cast_fp16 = add(x = var_5437_cast_fp16, y = var_5458_cast_fp16)[name = tensor<string, []>("q_patches_89_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5460_cast_fp16 = mul(x = var_5435_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5460_cast_fp16")];
            tensor<int32, [5]> var_5468 = const()[name = tensor<string, []>("op_5468"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_605_cast_fp16 = reshape(shape = var_5468, x = var_5435_cast_fp16)[name = tensor<string, []>("x_605_cast_fp16")];
            tensor<int32, [2]> var_5470_split_sizes_0 = const()[name = tensor<string, []>("op_5470_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5470_axis_0 = const()[name = tensor<string, []>("op_5470_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5470_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5470_cast_fp16_1 = split(axis = var_5470_axis_0, split_sizes = var_5470_split_sizes_0, x = x_605_cast_fp16)[name = tensor<string, []>("op_5470_cast_fp16")];
            tensor<int32, [1]> squeeze_178_axes_0 = const()[name = tensor<string, []>("squeeze_178_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_178_cast_fp16 = squeeze(axes = squeeze_178_axes_0, x = var_5470_cast_fp16_0)[name = tensor<string, []>("squeeze_178_cast_fp16")];
            tensor<int32, [1]> squeeze_179_axes_0 = const()[name = tensor<string, []>("squeeze_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_179_cast_fp16 = squeeze(axes = squeeze_179_axes_0, x = var_5470_cast_fp16_1)[name = tensor<string, []>("squeeze_179_cast_fp16")];
            tensor<fp16, []> const_731_promoted_to_fp16 = const()[name = tensor<string, []>("const_731_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5473_cast_fp16 = mul(x = squeeze_179_cast_fp16, y = const_731_promoted_to_fp16)[name = tensor<string, []>("op_5473_cast_fp16")];
            tensor<int32, []> x_607_axis_0 = const()[name = tensor<string, []>("x_607_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_607_cast_fp16 = stack(axis = x_607_axis_0, values = (var_5473_cast_fp16, squeeze_178_cast_fp16))[name = tensor<string, []>("x_607_cast_fp16")];
            tensor<int32, [4]> var_5479 = const()[name = tensor<string, []>("op_5479"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5480_cast_fp16 = reshape(shape = var_5479, x = x_607_cast_fp16)[name = tensor<string, []>("op_5480_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5481_cast_fp16 = mul(x = var_5480_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5481_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_89_cast_fp16 = add(x = var_5460_cast_fp16, y = var_5481_cast_fp16)[name = tensor<string, []>("k_patches_89_cast_fp16")];
            tensor<bool, []> var_5484_interleave_0 = const()[name = tensor<string, []>("op_5484_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5484_cast_fp16 = concat(axis = var_20, interleave = var_5484_interleave_0, values = (var_5423_cast_fp16, q_patches_89_cast_fp16))[name = tensor<string, []>("op_5484_cast_fp16")];
            tensor<bool, []> var_5487_interleave_0 = const()[name = tensor<string, []>("op_5487_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5487_cast_fp16 = concat(axis = var_20, interleave = var_5487_interleave_0, values = (var_5427_cast_fp16, k_patches_89_cast_fp16))[name = tensor<string, []>("op_5487_cast_fp16")];
            tensor<fp16, []> var_5489_to_fp16 = const()[name = tensor<string, []>("op_5489_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_449_cast_fp16 = mul(x = var_5484_cast_fp16, y = var_5489_to_fp16)[name = tensor<string, []>("q_449_cast_fp16")];
            tensor<bool, []> attn_177_transpose_x_1 = const()[name = tensor<string, []>("attn_177_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_177_transpose_y_1 = const()[name = tensor<string, []>("attn_177_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_177_cast_fp16 = matmul(transpose_x = attn_177_transpose_x_1, transpose_y = attn_177_transpose_y_1, x = q_449_cast_fp16, y = var_5487_cast_fp16)[name = tensor<string, []>("attn_177_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_179_cast_fp16 = softmax(axis = var_21, x = attn_177_cast_fp16)[name = tensor<string, []>("attn_179_cast_fp16")];
            tensor<bool, []> var_5494_transpose_x_0 = const()[name = tensor<string, []>("op_5494_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5494_transpose_y_0 = const()[name = tensor<string, []>("op_5494_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5494_cast_fp16 = matmul(transpose_x = var_5494_transpose_x_0, transpose_y = var_5494_transpose_y_0, x = attn_179_cast_fp16, y = v_269_cast_fp16)[name = tensor<string, []>("op_5494_cast_fp16")];
            tensor<int32, [1]> var_5495_axes_0 = const()[name = tensor<string, []>("op_5495_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5495_cast_fp16 = squeeze(axes = var_5495_axes_0, x = var_5494_cast_fp16)[name = tensor<string, []>("op_5495_cast_fp16")];
            tensor<string, []> var_5501_pad_type_0 = const()[name = tensor<string, []>("op_5501_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5501_strides_0 = const()[name = tensor<string, []>("op_5501_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5501_pad_0 = const()[name = tensor<string, []>("op_5501_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5501_dilations_0 = const()[name = tensor<string, []>("op_5501_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5501_groups_0 = const()[name = tensor<string, []>("op_5501_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46725376)))];
            tensor<fp16, [1, 64, 1, 197]> var_5501_cast_fp16 = conv(dilations = var_5501_dilations_0, groups = var_5501_groups_0, pad = var_5501_pad_0, pad_type = var_5501_pad_type_0, strides = var_5501_strides_0, weight = model_blocks_3_attn_q_projs_9_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5501_cast_fp16")];
            tensor<int32, [1]> var_5502_axes_0 = const()[name = tensor<string, []>("op_5502_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5502_cast_fp16 = squeeze(axes = var_5502_axes_0, x = var_5501_cast_fp16)[name = tensor<string, []>("op_5502_cast_fp16")];
            tensor<int32, [3]> q_451_perm_0 = const()[name = tensor<string, []>("q_451_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5509_pad_type_0 = const()[name = tensor<string, []>("op_5509_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5509_strides_0 = const()[name = tensor<string, []>("op_5509_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5509_pad_0 = const()[name = tensor<string, []>("op_5509_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5509_dilations_0 = const()[name = tensor<string, []>("op_5509_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5509_groups_0 = const()[name = tensor<string, []>("op_5509_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46823744)))];
            tensor<fp16, [1, 64, 1, 197]> var_5509_cast_fp16 = conv(dilations = var_5509_dilations_0, groups = var_5509_groups_0, pad = var_5509_pad_0, pad_type = var_5509_pad_type_0, strides = var_5509_strides_0, weight = model_blocks_3_attn_k_projs_9_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5509_cast_fp16")];
            tensor<int32, [1]> var_5510_axes_0 = const()[name = tensor<string, []>("op_5510_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5510_cast_fp16 = squeeze(axes = var_5510_axes_0, x = var_5509_cast_fp16)[name = tensor<string, []>("op_5510_cast_fp16")];
            tensor<int32, [3]> k_271_perm_0 = const()[name = tensor<string, []>("k_271_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5517_pad_type_0 = const()[name = tensor<string, []>("op_5517_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5517_strides_0 = const()[name = tensor<string, []>("op_5517_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5517_pad_0 = const()[name = tensor<string, []>("op_5517_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5517_dilations_0 = const()[name = tensor<string, []>("op_5517_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5517_groups_0 = const()[name = tensor<string, []>("op_5517_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46922112)))];
            tensor<fp16, [1, 64, 1, 197]> var_5517_cast_fp16 = conv(dilations = var_5517_dilations_0, groups = var_5517_groups_0, pad = var_5517_pad_0, pad_type = var_5517_pad_type_0, strides = var_5517_strides_0, weight = model_blocks_3_attn_v_projs_9_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5517_cast_fp16")];
            tensor<int32, [1]> var_5518_axes_0 = const()[name = tensor<string, []>("op_5518_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5518_cast_fp16 = squeeze(axes = var_5518_axes_0, x = var_5517_cast_fp16)[name = tensor<string, []>("op_5518_cast_fp16")];
            tensor<int32, [3]> v_271_perm_0 = const()[name = tensor<string, []>("v_271_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47020480)))];
            tensor<fp16, [1, 197, 64]> q_451_cast_fp16 = transpose(perm = q_451_perm_0, x = var_5502_cast_fp16)[name = tensor<string, []>("transpose_358")];
            tensor<fp16, [1, 197, 64]> q_453_cast_fp16 = add(x = q_451_cast_fp16, y = model_blocks_3_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_453_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47020672)))];
            tensor<fp16, [1, 197, 64]> v_271_cast_fp16 = transpose(perm = v_271_perm_0, x = var_5518_cast_fp16)[name = tensor<string, []>("transpose_356")];
            tensor<fp16, [1, 197, 64]> v_273_cast_fp16 = add(x = v_271_cast_fp16, y = model_blocks_3_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_273_cast_fp16")];
            tensor<int32, [1]> q_455_axes_0 = const()[name = tensor<string, []>("q_455_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_455_cast_fp16 = expand_dims(axes = q_455_axes_0, x = q_453_cast_fp16)[name = tensor<string, []>("q_455_cast_fp16")];
            tensor<int32, [1]> k_273_axes_0 = const()[name = tensor<string, []>("k_273_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_271_cast_fp16 = transpose(perm = k_271_perm_0, x = var_5510_cast_fp16)[name = tensor<string, []>("transpose_357")];
            tensor<fp16, [1, 1, 197, 64]> k_273_cast_fp16 = expand_dims(axes = k_273_axes_0, x = k_271_cast_fp16)[name = tensor<string, []>("k_273_cast_fp16")];
            tensor<int32, [1]> v_275_axes_0 = const()[name = tensor<string, []>("v_275_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_275_cast_fp16 = expand_dims(axes = v_275_axes_0, x = v_273_cast_fp16)[name = tensor<string, []>("v_275_cast_fp16")];
            tensor<int32, [4]> var_5527_begin_0 = const()[name = tensor<string, []>("op_5527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5527_end_0 = const()[name = tensor<string, []>("op_5527_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5527_end_mask_0 = const()[name = tensor<string, []>("op_5527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5527_cast_fp16 = slice_by_index(begin = var_5527_begin_0, end = var_5527_end_0, end_mask = var_5527_end_mask_0, x = q_455_cast_fp16)[name = tensor<string, []>("op_5527_cast_fp16")];
            tensor<int32, [4]> var_5531_begin_0 = const()[name = tensor<string, []>("op_5531_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5531_end_0 = const()[name = tensor<string, []>("op_5531_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5531_end_mask_0 = const()[name = tensor<string, []>("op_5531_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5531_cast_fp16 = slice_by_index(begin = var_5531_begin_0, end = var_5531_end_0, end_mask = var_5531_end_mask_0, x = k_273_cast_fp16)[name = tensor<string, []>("op_5531_cast_fp16")];
            tensor<int32, [4]> var_5535_begin_0 = const()[name = tensor<string, []>("op_5535_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5535_end_0 = const()[name = tensor<string, []>("op_5535_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5535_end_mask_0 = const()[name = tensor<string, []>("op_5535_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5535_cast_fp16 = slice_by_index(begin = var_5535_begin_0, end = var_5535_end_0, end_mask = var_5535_end_mask_0, x = q_455_cast_fp16)[name = tensor<string, []>("op_5535_cast_fp16")];
            tensor<int32, [4]> var_5539_begin_0 = const()[name = tensor<string, []>("op_5539_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5539_end_0 = const()[name = tensor<string, []>("op_5539_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5539_end_mask_0 = const()[name = tensor<string, []>("op_5539_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5539_cast_fp16 = slice_by_index(begin = var_5539_begin_0, end = var_5539_end_0, end_mask = var_5539_end_mask_0, x = k_273_cast_fp16)[name = tensor<string, []>("op_5539_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5541_cast_fp16 = mul(x = var_5535_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5541_cast_fp16")];
            tensor<int32, [5]> var_5549 = const()[name = tensor<string, []>("op_5549"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_611_cast_fp16 = reshape(shape = var_5549, x = var_5535_cast_fp16)[name = tensor<string, []>("x_611_cast_fp16")];
            tensor<int32, [2]> var_5551_split_sizes_0 = const()[name = tensor<string, []>("op_5551_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5551_axis_0 = const()[name = tensor<string, []>("op_5551_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5551_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5551_cast_fp16_1 = split(axis = var_5551_axis_0, split_sizes = var_5551_split_sizes_0, x = x_611_cast_fp16)[name = tensor<string, []>("op_5551_cast_fp16")];
            tensor<int32, [1]> squeeze_180_axes_0 = const()[name = tensor<string, []>("squeeze_180_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_180_cast_fp16 = squeeze(axes = squeeze_180_axes_0, x = var_5551_cast_fp16_0)[name = tensor<string, []>("squeeze_180_cast_fp16")];
            tensor<int32, [1]> squeeze_181_axes_0 = const()[name = tensor<string, []>("squeeze_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_181_cast_fp16 = squeeze(axes = squeeze_181_axes_0, x = var_5551_cast_fp16_1)[name = tensor<string, []>("squeeze_181_cast_fp16")];
            tensor<fp16, []> const_739_promoted_to_fp16 = const()[name = tensor<string, []>("const_739_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5554_cast_fp16 = mul(x = squeeze_181_cast_fp16, y = const_739_promoted_to_fp16)[name = tensor<string, []>("op_5554_cast_fp16")];
            tensor<int32, []> x_613_axis_0 = const()[name = tensor<string, []>("x_613_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_613_cast_fp16 = stack(axis = x_613_axis_0, values = (var_5554_cast_fp16, squeeze_180_cast_fp16))[name = tensor<string, []>("x_613_cast_fp16")];
            tensor<int32, [4]> var_5560 = const()[name = tensor<string, []>("op_5560"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5561_cast_fp16 = reshape(shape = var_5560, x = x_613_cast_fp16)[name = tensor<string, []>("op_5561_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5562_cast_fp16 = mul(x = var_5561_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5562_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_91_cast_fp16 = add(x = var_5541_cast_fp16, y = var_5562_cast_fp16)[name = tensor<string, []>("q_patches_91_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5564_cast_fp16 = mul(x = var_5539_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5564_cast_fp16")];
            tensor<int32, [5]> var_5572 = const()[name = tensor<string, []>("op_5572"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_617_cast_fp16 = reshape(shape = var_5572, x = var_5539_cast_fp16)[name = tensor<string, []>("x_617_cast_fp16")];
            tensor<int32, [2]> var_5574_split_sizes_0 = const()[name = tensor<string, []>("op_5574_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5574_axis_0 = const()[name = tensor<string, []>("op_5574_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5574_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5574_cast_fp16_1 = split(axis = var_5574_axis_0, split_sizes = var_5574_split_sizes_0, x = x_617_cast_fp16)[name = tensor<string, []>("op_5574_cast_fp16")];
            tensor<int32, [1]> squeeze_182_axes_0 = const()[name = tensor<string, []>("squeeze_182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_182_cast_fp16 = squeeze(axes = squeeze_182_axes_0, x = var_5574_cast_fp16_0)[name = tensor<string, []>("squeeze_182_cast_fp16")];
            tensor<int32, [1]> squeeze_183_axes_0 = const()[name = tensor<string, []>("squeeze_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_183_cast_fp16 = squeeze(axes = squeeze_183_axes_0, x = var_5574_cast_fp16_1)[name = tensor<string, []>("squeeze_183_cast_fp16")];
            tensor<fp16, []> const_747_promoted_to_fp16 = const()[name = tensor<string, []>("const_747_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5577_cast_fp16 = mul(x = squeeze_183_cast_fp16, y = const_747_promoted_to_fp16)[name = tensor<string, []>("op_5577_cast_fp16")];
            tensor<int32, []> x_619_axis_0 = const()[name = tensor<string, []>("x_619_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_619_cast_fp16 = stack(axis = x_619_axis_0, values = (var_5577_cast_fp16, squeeze_182_cast_fp16))[name = tensor<string, []>("x_619_cast_fp16")];
            tensor<int32, [4]> var_5583 = const()[name = tensor<string, []>("op_5583"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5584_cast_fp16 = reshape(shape = var_5583, x = x_619_cast_fp16)[name = tensor<string, []>("op_5584_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5585_cast_fp16 = mul(x = var_5584_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5585_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_91_cast_fp16 = add(x = var_5564_cast_fp16, y = var_5585_cast_fp16)[name = tensor<string, []>("k_patches_91_cast_fp16")];
            tensor<bool, []> var_5588_interleave_0 = const()[name = tensor<string, []>("op_5588_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5588_cast_fp16 = concat(axis = var_20, interleave = var_5588_interleave_0, values = (var_5527_cast_fp16, q_patches_91_cast_fp16))[name = tensor<string, []>("op_5588_cast_fp16")];
            tensor<bool, []> var_5591_interleave_0 = const()[name = tensor<string, []>("op_5591_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5591_cast_fp16 = concat(axis = var_20, interleave = var_5591_interleave_0, values = (var_5531_cast_fp16, k_patches_91_cast_fp16))[name = tensor<string, []>("op_5591_cast_fp16")];
            tensor<fp16, []> var_5593_to_fp16 = const()[name = tensor<string, []>("op_5593_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_459_cast_fp16 = mul(x = var_5588_cast_fp16, y = var_5593_to_fp16)[name = tensor<string, []>("q_459_cast_fp16")];
            tensor<bool, []> attn_181_transpose_x_1 = const()[name = tensor<string, []>("attn_181_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_181_transpose_y_1 = const()[name = tensor<string, []>("attn_181_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_181_cast_fp16 = matmul(transpose_x = attn_181_transpose_x_1, transpose_y = attn_181_transpose_y_1, x = q_459_cast_fp16, y = var_5591_cast_fp16)[name = tensor<string, []>("attn_181_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_183_cast_fp16 = softmax(axis = var_21, x = attn_181_cast_fp16)[name = tensor<string, []>("attn_183_cast_fp16")];
            tensor<bool, []> var_5598_transpose_x_0 = const()[name = tensor<string, []>("op_5598_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5598_transpose_y_0 = const()[name = tensor<string, []>("op_5598_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5598_cast_fp16 = matmul(transpose_x = var_5598_transpose_x_0, transpose_y = var_5598_transpose_y_0, x = attn_183_cast_fp16, y = v_275_cast_fp16)[name = tensor<string, []>("op_5598_cast_fp16")];
            tensor<int32, [1]> var_5599_axes_0 = const()[name = tensor<string, []>("op_5599_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5599_cast_fp16 = squeeze(axes = var_5599_axes_0, x = var_5598_cast_fp16)[name = tensor<string, []>("op_5599_cast_fp16")];
            tensor<string, []> var_5605_pad_type_0 = const()[name = tensor<string, []>("op_5605_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5605_strides_0 = const()[name = tensor<string, []>("op_5605_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5605_pad_0 = const()[name = tensor<string, []>("op_5605_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5605_dilations_0 = const()[name = tensor<string, []>("op_5605_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5605_groups_0 = const()[name = tensor<string, []>("op_5605_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47020864)))];
            tensor<fp16, [1, 64, 1, 197]> var_5605_cast_fp16 = conv(dilations = var_5605_dilations_0, groups = var_5605_groups_0, pad = var_5605_pad_0, pad_type = var_5605_pad_type_0, strides = var_5605_strides_0, weight = model_blocks_3_attn_q_projs_10_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5605_cast_fp16")];
            tensor<int32, [1]> var_5606_axes_0 = const()[name = tensor<string, []>("op_5606_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5606_cast_fp16 = squeeze(axes = var_5606_axes_0, x = var_5605_cast_fp16)[name = tensor<string, []>("op_5606_cast_fp16")];
            tensor<int32, [3]> q_461_perm_0 = const()[name = tensor<string, []>("q_461_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5613_pad_type_0 = const()[name = tensor<string, []>("op_5613_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5613_strides_0 = const()[name = tensor<string, []>("op_5613_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5613_pad_0 = const()[name = tensor<string, []>("op_5613_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5613_dilations_0 = const()[name = tensor<string, []>("op_5613_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5613_groups_0 = const()[name = tensor<string, []>("op_5613_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47119232)))];
            tensor<fp16, [1, 64, 1, 197]> var_5613_cast_fp16 = conv(dilations = var_5613_dilations_0, groups = var_5613_groups_0, pad = var_5613_pad_0, pad_type = var_5613_pad_type_0, strides = var_5613_strides_0, weight = model_blocks_3_attn_k_projs_10_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5613_cast_fp16")];
            tensor<int32, [1]> var_5614_axes_0 = const()[name = tensor<string, []>("op_5614_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5614_cast_fp16 = squeeze(axes = var_5614_axes_0, x = var_5613_cast_fp16)[name = tensor<string, []>("op_5614_cast_fp16")];
            tensor<int32, [3]> k_277_perm_0 = const()[name = tensor<string, []>("k_277_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5621_pad_type_0 = const()[name = tensor<string, []>("op_5621_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5621_strides_0 = const()[name = tensor<string, []>("op_5621_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5621_pad_0 = const()[name = tensor<string, []>("op_5621_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5621_dilations_0 = const()[name = tensor<string, []>("op_5621_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5621_groups_0 = const()[name = tensor<string, []>("op_5621_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47217600)))];
            tensor<fp16, [1, 64, 1, 197]> var_5621_cast_fp16 = conv(dilations = var_5621_dilations_0, groups = var_5621_groups_0, pad = var_5621_pad_0, pad_type = var_5621_pad_type_0, strides = var_5621_strides_0, weight = model_blocks_3_attn_v_projs_10_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5621_cast_fp16")];
            tensor<int32, [1]> var_5622_axes_0 = const()[name = tensor<string, []>("op_5622_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5622_cast_fp16 = squeeze(axes = var_5622_axes_0, x = var_5621_cast_fp16)[name = tensor<string, []>("op_5622_cast_fp16")];
            tensor<int32, [3]> v_277_perm_0 = const()[name = tensor<string, []>("v_277_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47315968)))];
            tensor<fp16, [1, 197, 64]> q_461_cast_fp16 = transpose(perm = q_461_perm_0, x = var_5606_cast_fp16)[name = tensor<string, []>("transpose_355")];
            tensor<fp16, [1, 197, 64]> q_463_cast_fp16 = add(x = q_461_cast_fp16, y = model_blocks_3_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_463_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47316160)))];
            tensor<fp16, [1, 197, 64]> v_277_cast_fp16 = transpose(perm = v_277_perm_0, x = var_5622_cast_fp16)[name = tensor<string, []>("transpose_353")];
            tensor<fp16, [1, 197, 64]> v_279_cast_fp16 = add(x = v_277_cast_fp16, y = model_blocks_3_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_279_cast_fp16")];
            tensor<int32, [1]> q_465_axes_0 = const()[name = tensor<string, []>("q_465_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_465_cast_fp16 = expand_dims(axes = q_465_axes_0, x = q_463_cast_fp16)[name = tensor<string, []>("q_465_cast_fp16")];
            tensor<int32, [1]> k_279_axes_0 = const()[name = tensor<string, []>("k_279_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_277_cast_fp16 = transpose(perm = k_277_perm_0, x = var_5614_cast_fp16)[name = tensor<string, []>("transpose_354")];
            tensor<fp16, [1, 1, 197, 64]> k_279_cast_fp16 = expand_dims(axes = k_279_axes_0, x = k_277_cast_fp16)[name = tensor<string, []>("k_279_cast_fp16")];
            tensor<int32, [1]> v_281_axes_0 = const()[name = tensor<string, []>("v_281_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_281_cast_fp16 = expand_dims(axes = v_281_axes_0, x = v_279_cast_fp16)[name = tensor<string, []>("v_281_cast_fp16")];
            tensor<int32, [4]> var_5631_begin_0 = const()[name = tensor<string, []>("op_5631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5631_end_0 = const()[name = tensor<string, []>("op_5631_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5631_end_mask_0 = const()[name = tensor<string, []>("op_5631_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5631_cast_fp16 = slice_by_index(begin = var_5631_begin_0, end = var_5631_end_0, end_mask = var_5631_end_mask_0, x = q_465_cast_fp16)[name = tensor<string, []>("op_5631_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = tensor<string, []>("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = tensor<string, []>("op_5635_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = tensor<string, []>("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_279_cast_fp16)[name = tensor<string, []>("op_5635_cast_fp16")];
            tensor<int32, [4]> var_5639_begin_0 = const()[name = tensor<string, []>("op_5639_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5639_end_0 = const()[name = tensor<string, []>("op_5639_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5639_end_mask_0 = const()[name = tensor<string, []>("op_5639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5639_cast_fp16 = slice_by_index(begin = var_5639_begin_0, end = var_5639_end_0, end_mask = var_5639_end_mask_0, x = q_465_cast_fp16)[name = tensor<string, []>("op_5639_cast_fp16")];
            tensor<int32, [4]> var_5643_begin_0 = const()[name = tensor<string, []>("op_5643_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5643_end_0 = const()[name = tensor<string, []>("op_5643_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5643_end_mask_0 = const()[name = tensor<string, []>("op_5643_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5643_cast_fp16 = slice_by_index(begin = var_5643_begin_0, end = var_5643_end_0, end_mask = var_5643_end_mask_0, x = k_279_cast_fp16)[name = tensor<string, []>("op_5643_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5645_cast_fp16 = mul(x = var_5639_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5645_cast_fp16")];
            tensor<int32, [5]> var_5653 = const()[name = tensor<string, []>("op_5653"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_623_cast_fp16 = reshape(shape = var_5653, x = var_5639_cast_fp16)[name = tensor<string, []>("x_623_cast_fp16")];
            tensor<int32, [2]> var_5655_split_sizes_0 = const()[name = tensor<string, []>("op_5655_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5655_axis_0 = const()[name = tensor<string, []>("op_5655_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5655_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5655_cast_fp16_1 = split(axis = var_5655_axis_0, split_sizes = var_5655_split_sizes_0, x = x_623_cast_fp16)[name = tensor<string, []>("op_5655_cast_fp16")];
            tensor<int32, [1]> squeeze_184_axes_0 = const()[name = tensor<string, []>("squeeze_184_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_184_cast_fp16 = squeeze(axes = squeeze_184_axes_0, x = var_5655_cast_fp16_0)[name = tensor<string, []>("squeeze_184_cast_fp16")];
            tensor<int32, [1]> squeeze_185_axes_0 = const()[name = tensor<string, []>("squeeze_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_185_cast_fp16 = squeeze(axes = squeeze_185_axes_0, x = var_5655_cast_fp16_1)[name = tensor<string, []>("squeeze_185_cast_fp16")];
            tensor<fp16, []> const_755_promoted_to_fp16 = const()[name = tensor<string, []>("const_755_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5658_cast_fp16 = mul(x = squeeze_185_cast_fp16, y = const_755_promoted_to_fp16)[name = tensor<string, []>("op_5658_cast_fp16")];
            tensor<int32, []> x_625_axis_0 = const()[name = tensor<string, []>("x_625_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_625_cast_fp16 = stack(axis = x_625_axis_0, values = (var_5658_cast_fp16, squeeze_184_cast_fp16))[name = tensor<string, []>("x_625_cast_fp16")];
            tensor<int32, [4]> var_5664 = const()[name = tensor<string, []>("op_5664"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5665_cast_fp16 = reshape(shape = var_5664, x = x_625_cast_fp16)[name = tensor<string, []>("op_5665_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5666_cast_fp16 = mul(x = var_5665_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5666_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_93_cast_fp16 = add(x = var_5645_cast_fp16, y = var_5666_cast_fp16)[name = tensor<string, []>("q_patches_93_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5668_cast_fp16 = mul(x = var_5643_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5668_cast_fp16")];
            tensor<int32, [5]> var_5676 = const()[name = tensor<string, []>("op_5676"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_629_cast_fp16 = reshape(shape = var_5676, x = var_5643_cast_fp16)[name = tensor<string, []>("x_629_cast_fp16")];
            tensor<int32, [2]> var_5678_split_sizes_0 = const()[name = tensor<string, []>("op_5678_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5678_axis_0 = const()[name = tensor<string, []>("op_5678_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5678_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5678_cast_fp16_1 = split(axis = var_5678_axis_0, split_sizes = var_5678_split_sizes_0, x = x_629_cast_fp16)[name = tensor<string, []>("op_5678_cast_fp16")];
            tensor<int32, [1]> squeeze_186_axes_0 = const()[name = tensor<string, []>("squeeze_186_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_186_cast_fp16 = squeeze(axes = squeeze_186_axes_0, x = var_5678_cast_fp16_0)[name = tensor<string, []>("squeeze_186_cast_fp16")];
            tensor<int32, [1]> squeeze_187_axes_0 = const()[name = tensor<string, []>("squeeze_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_187_cast_fp16 = squeeze(axes = squeeze_187_axes_0, x = var_5678_cast_fp16_1)[name = tensor<string, []>("squeeze_187_cast_fp16")];
            tensor<fp16, []> const_763_promoted_to_fp16 = const()[name = tensor<string, []>("const_763_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5681_cast_fp16 = mul(x = squeeze_187_cast_fp16, y = const_763_promoted_to_fp16)[name = tensor<string, []>("op_5681_cast_fp16")];
            tensor<int32, []> x_631_axis_0 = const()[name = tensor<string, []>("x_631_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_631_cast_fp16 = stack(axis = x_631_axis_0, values = (var_5681_cast_fp16, squeeze_186_cast_fp16))[name = tensor<string, []>("x_631_cast_fp16")];
            tensor<int32, [4]> var_5687 = const()[name = tensor<string, []>("op_5687"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5688_cast_fp16 = reshape(shape = var_5687, x = x_631_cast_fp16)[name = tensor<string, []>("op_5688_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5689_cast_fp16 = mul(x = var_5688_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5689_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_93_cast_fp16 = add(x = var_5668_cast_fp16, y = var_5689_cast_fp16)[name = tensor<string, []>("k_patches_93_cast_fp16")];
            tensor<bool, []> var_5692_interleave_0 = const()[name = tensor<string, []>("op_5692_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5692_cast_fp16 = concat(axis = var_20, interleave = var_5692_interleave_0, values = (var_5631_cast_fp16, q_patches_93_cast_fp16))[name = tensor<string, []>("op_5692_cast_fp16")];
            tensor<bool, []> var_5695_interleave_0 = const()[name = tensor<string, []>("op_5695_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5695_cast_fp16 = concat(axis = var_20, interleave = var_5695_interleave_0, values = (var_5635_cast_fp16, k_patches_93_cast_fp16))[name = tensor<string, []>("op_5695_cast_fp16")];
            tensor<fp16, []> var_5697_to_fp16 = const()[name = tensor<string, []>("op_5697_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_469_cast_fp16 = mul(x = var_5692_cast_fp16, y = var_5697_to_fp16)[name = tensor<string, []>("q_469_cast_fp16")];
            tensor<bool, []> attn_185_transpose_x_1 = const()[name = tensor<string, []>("attn_185_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_185_transpose_y_1 = const()[name = tensor<string, []>("attn_185_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_185_cast_fp16 = matmul(transpose_x = attn_185_transpose_x_1, transpose_y = attn_185_transpose_y_1, x = q_469_cast_fp16, y = var_5695_cast_fp16)[name = tensor<string, []>("attn_185_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_187_cast_fp16 = softmax(axis = var_21, x = attn_185_cast_fp16)[name = tensor<string, []>("attn_187_cast_fp16")];
            tensor<bool, []> var_5702_transpose_x_0 = const()[name = tensor<string, []>("op_5702_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5702_transpose_y_0 = const()[name = tensor<string, []>("op_5702_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5702_cast_fp16 = matmul(transpose_x = var_5702_transpose_x_0, transpose_y = var_5702_transpose_y_0, x = attn_187_cast_fp16, y = v_281_cast_fp16)[name = tensor<string, []>("op_5702_cast_fp16")];
            tensor<int32, [1]> var_5703_axes_0 = const()[name = tensor<string, []>("op_5703_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_5703_cast_fp16 = squeeze(axes = var_5703_axes_0, x = var_5702_cast_fp16)[name = tensor<string, []>("op_5703_cast_fp16")];
            tensor<string, []> var_5709_pad_type_0 = const()[name = tensor<string, []>("op_5709_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5709_strides_0 = const()[name = tensor<string, []>("op_5709_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5709_pad_0 = const()[name = tensor<string, []>("op_5709_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5709_dilations_0 = const()[name = tensor<string, []>("op_5709_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5709_groups_0 = const()[name = tensor<string, []>("op_5709_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47316352)))];
            tensor<fp16, [1, 64, 1, 197]> var_5709_cast_fp16 = conv(dilations = var_5709_dilations_0, groups = var_5709_groups_0, pad = var_5709_pad_0, pad_type = var_5709_pad_type_0, strides = var_5709_strides_0, weight = model_blocks_3_attn_q_projs_11_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5709_cast_fp16")];
            tensor<int32, [1]> var_5710_axes_0 = const()[name = tensor<string, []>("op_5710_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5710_cast_fp16 = squeeze(axes = var_5710_axes_0, x = var_5709_cast_fp16)[name = tensor<string, []>("op_5710_cast_fp16")];
            tensor<int32, [3]> q_471_perm_0 = const()[name = tensor<string, []>("q_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5717_pad_type_0 = const()[name = tensor<string, []>("op_5717_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5717_strides_0 = const()[name = tensor<string, []>("op_5717_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5717_pad_0 = const()[name = tensor<string, []>("op_5717_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5717_dilations_0 = const()[name = tensor<string, []>("op_5717_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5717_groups_0 = const()[name = tensor<string, []>("op_5717_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47414720)))];
            tensor<fp16, [1, 64, 1, 197]> var_5717_cast_fp16 = conv(dilations = var_5717_dilations_0, groups = var_5717_groups_0, pad = var_5717_pad_0, pad_type = var_5717_pad_type_0, strides = var_5717_strides_0, weight = model_blocks_3_attn_k_projs_11_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5717_cast_fp16")];
            tensor<int32, [1]> var_5718_axes_0 = const()[name = tensor<string, []>("op_5718_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5718_cast_fp16 = squeeze(axes = var_5718_axes_0, x = var_5717_cast_fp16)[name = tensor<string, []>("op_5718_cast_fp16")];
            tensor<int32, [3]> k_283_perm_0 = const()[name = tensor<string, []>("k_283_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5725_pad_type_0 = const()[name = tensor<string, []>("op_5725_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5725_strides_0 = const()[name = tensor<string, []>("op_5725_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5725_pad_0 = const()[name = tensor<string, []>("op_5725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5725_dilations_0 = const()[name = tensor<string, []>("op_5725_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5725_groups_0 = const()[name = tensor<string, []>("op_5725_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_3_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47513088)))];
            tensor<fp16, [1, 64, 1, 197]> var_5725_cast_fp16 = conv(dilations = var_5725_dilations_0, groups = var_5725_groups_0, pad = var_5725_pad_0, pad_type = var_5725_pad_type_0, strides = var_5725_strides_0, weight = model_blocks_3_attn_v_projs_11_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("op_5725_cast_fp16")];
            tensor<int32, [1]> var_5726_axes_0 = const()[name = tensor<string, []>("op_5726_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_5726_cast_fp16 = squeeze(axes = var_5726_axes_0, x = var_5725_cast_fp16)[name = tensor<string, []>("op_5726_cast_fp16")];
            tensor<int32, [3]> v_283_perm_0 = const()[name = tensor<string, []>("v_283_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_3_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47611456)))];
            tensor<fp16, [1, 197, 64]> q_471_cast_fp16 = transpose(perm = q_471_perm_0, x = var_5710_cast_fp16)[name = tensor<string, []>("transpose_352")];
            tensor<fp16, [1, 197, 64]> q_473_cast_fp16 = add(x = q_471_cast_fp16, y = model_blocks_3_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_473_cast_fp16")];
            tensor<fp16, [64]> model_blocks_3_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47611648)))];
            tensor<fp16, [1, 197, 64]> v_283_cast_fp16 = transpose(perm = v_283_perm_0, x = var_5726_cast_fp16)[name = tensor<string, []>("transpose_350")];
            tensor<fp16, [1, 197, 64]> v_285_cast_fp16 = add(x = v_283_cast_fp16, y = model_blocks_3_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_285_cast_fp16")];
            tensor<int32, [1]> q_475_axes_0 = const()[name = tensor<string, []>("q_475_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_475_cast_fp16 = expand_dims(axes = q_475_axes_0, x = q_473_cast_fp16)[name = tensor<string, []>("q_475_cast_fp16")];
            tensor<int32, [1]> k_285_axes_0 = const()[name = tensor<string, []>("k_285_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_283_cast_fp16 = transpose(perm = k_283_perm_0, x = var_5718_cast_fp16)[name = tensor<string, []>("transpose_351")];
            tensor<fp16, [1, 1, 197, 64]> k_285_cast_fp16 = expand_dims(axes = k_285_axes_0, x = k_283_cast_fp16)[name = tensor<string, []>("k_285_cast_fp16")];
            tensor<int32, [1]> v_287_axes_0 = const()[name = tensor<string, []>("v_287_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_287_cast_fp16 = expand_dims(axes = v_287_axes_0, x = v_285_cast_fp16)[name = tensor<string, []>("v_287_cast_fp16")];
            tensor<int32, [4]> var_5735_begin_0 = const()[name = tensor<string, []>("op_5735_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5735_end_0 = const()[name = tensor<string, []>("op_5735_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5735_end_mask_0 = const()[name = tensor<string, []>("op_5735_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5735_cast_fp16 = slice_by_index(begin = var_5735_begin_0, end = var_5735_end_0, end_mask = var_5735_end_mask_0, x = q_475_cast_fp16)[name = tensor<string, []>("op_5735_cast_fp16")];
            tensor<int32, [4]> var_5739_begin_0 = const()[name = tensor<string, []>("op_5739_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5739_end_0 = const()[name = tensor<string, []>("op_5739_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_5739_end_mask_0 = const()[name = tensor<string, []>("op_5739_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_5739_cast_fp16 = slice_by_index(begin = var_5739_begin_0, end = var_5739_end_0, end_mask = var_5739_end_mask_0, x = k_285_cast_fp16)[name = tensor<string, []>("op_5739_cast_fp16")];
            tensor<int32, [4]> var_5743_begin_0 = const()[name = tensor<string, []>("op_5743_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5743_end_0 = const()[name = tensor<string, []>("op_5743_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5743_end_mask_0 = const()[name = tensor<string, []>("op_5743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5743_cast_fp16 = slice_by_index(begin = var_5743_begin_0, end = var_5743_end_0, end_mask = var_5743_end_mask_0, x = q_475_cast_fp16)[name = tensor<string, []>("op_5743_cast_fp16")];
            tensor<int32, [4]> var_5747_begin_0 = const()[name = tensor<string, []>("op_5747_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5747_end_0 = const()[name = tensor<string, []>("op_5747_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_5747_end_mask_0 = const()[name = tensor<string, []>("op_5747_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_5747_cast_fp16 = slice_by_index(begin = var_5747_begin_0, end = var_5747_end_0, end_mask = var_5747_end_mask_0, x = k_285_cast_fp16)[name = tensor<string, []>("op_5747_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5749_cast_fp16 = mul(x = var_5743_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5749_cast_fp16")];
            tensor<int32, [5]> var_5757 = const()[name = tensor<string, []>("op_5757"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_635_cast_fp16 = reshape(shape = var_5757, x = var_5743_cast_fp16)[name = tensor<string, []>("x_635_cast_fp16")];
            tensor<int32, [2]> var_5759_split_sizes_0 = const()[name = tensor<string, []>("op_5759_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5759_axis_0 = const()[name = tensor<string, []>("op_5759_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5759_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5759_cast_fp16_1 = split(axis = var_5759_axis_0, split_sizes = var_5759_split_sizes_0, x = x_635_cast_fp16)[name = tensor<string, []>("op_5759_cast_fp16")];
            tensor<int32, [1]> squeeze_188_axes_0 = const()[name = tensor<string, []>("squeeze_188_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_188_cast_fp16 = squeeze(axes = squeeze_188_axes_0, x = var_5759_cast_fp16_0)[name = tensor<string, []>("squeeze_188_cast_fp16")];
            tensor<int32, [1]> squeeze_189_axes_0 = const()[name = tensor<string, []>("squeeze_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_189_cast_fp16 = squeeze(axes = squeeze_189_axes_0, x = var_5759_cast_fp16_1)[name = tensor<string, []>("squeeze_189_cast_fp16")];
            tensor<fp16, []> const_771_promoted_to_fp16 = const()[name = tensor<string, []>("const_771_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5762_cast_fp16 = mul(x = squeeze_189_cast_fp16, y = const_771_promoted_to_fp16)[name = tensor<string, []>("op_5762_cast_fp16")];
            tensor<int32, []> x_637_axis_0 = const()[name = tensor<string, []>("x_637_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_637_cast_fp16 = stack(axis = x_637_axis_0, values = (var_5762_cast_fp16, squeeze_188_cast_fp16))[name = tensor<string, []>("x_637_cast_fp16")];
            tensor<int32, [4]> var_5768 = const()[name = tensor<string, []>("op_5768"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5769_cast_fp16 = reshape(shape = var_5768, x = x_637_cast_fp16)[name = tensor<string, []>("op_5769_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5770_cast_fp16 = mul(x = var_5769_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5770_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_95_cast_fp16 = add(x = var_5749_cast_fp16, y = var_5770_cast_fp16)[name = tensor<string, []>("q_patches_95_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5772_cast_fp16 = mul(x = var_5747_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_5772_cast_fp16")];
            tensor<int32, [5]> var_5780 = const()[name = tensor<string, []>("op_5780"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_641_cast_fp16 = reshape(shape = var_5780, x = var_5747_cast_fp16)[name = tensor<string, []>("x_641_cast_fp16")];
            tensor<int32, [2]> var_5782_split_sizes_0 = const()[name = tensor<string, []>("op_5782_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5782_axis_0 = const()[name = tensor<string, []>("op_5782_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_5782_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_5782_cast_fp16_1 = split(axis = var_5782_axis_0, split_sizes = var_5782_split_sizes_0, x = x_641_cast_fp16)[name = tensor<string, []>("op_5782_cast_fp16")];
            tensor<int32, [1]> squeeze_190_axes_0 = const()[name = tensor<string, []>("squeeze_190_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_190_cast_fp16 = squeeze(axes = squeeze_190_axes_0, x = var_5782_cast_fp16_0)[name = tensor<string, []>("squeeze_190_cast_fp16")];
            tensor<int32, [1]> squeeze_191_axes_0 = const()[name = tensor<string, []>("squeeze_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_191_cast_fp16 = squeeze(axes = squeeze_191_axes_0, x = var_5782_cast_fp16_1)[name = tensor<string, []>("squeeze_191_cast_fp16")];
            tensor<fp16, []> const_779_promoted_to_fp16 = const()[name = tensor<string, []>("const_779_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_5785_cast_fp16 = mul(x = squeeze_191_cast_fp16, y = const_779_promoted_to_fp16)[name = tensor<string, []>("op_5785_cast_fp16")];
            tensor<int32, []> x_643_axis_0 = const()[name = tensor<string, []>("x_643_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_643_cast_fp16 = stack(axis = x_643_axis_0, values = (var_5785_cast_fp16, squeeze_190_cast_fp16))[name = tensor<string, []>("x_643_cast_fp16")];
            tensor<int32, [4]> var_5791 = const()[name = tensor<string, []>("op_5791"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_5792_cast_fp16 = reshape(shape = var_5791, x = x_643_cast_fp16)[name = tensor<string, []>("op_5792_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_5793_cast_fp16 = mul(x = var_5792_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_5793_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_95_cast_fp16 = add(x = var_5772_cast_fp16, y = var_5793_cast_fp16)[name = tensor<string, []>("k_patches_95_cast_fp16")];
            tensor<bool, []> var_5796_interleave_0 = const()[name = tensor<string, []>("op_5796_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5796_cast_fp16 = concat(axis = var_20, interleave = var_5796_interleave_0, values = (var_5735_cast_fp16, q_patches_95_cast_fp16))[name = tensor<string, []>("op_5796_cast_fp16")];
            tensor<bool, []> var_5799_interleave_0 = const()[name = tensor<string, []>("op_5799_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5799_cast_fp16 = concat(axis = var_20, interleave = var_5799_interleave_0, values = (var_5739_cast_fp16, k_patches_95_cast_fp16))[name = tensor<string, []>("op_5799_cast_fp16")];
            tensor<fp16, []> var_5801_to_fp16 = const()[name = tensor<string, []>("op_5801_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_479_cast_fp16 = mul(x = var_5796_cast_fp16, y = var_5801_to_fp16)[name = tensor<string, []>("q_479_cast_fp16")];
            tensor<bool, []> attn_189_transpose_x_1 = const()[name = tensor<string, []>("attn_189_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_189_transpose_y_1 = const()[name = tensor<string, []>("attn_189_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_189_cast_fp16 = matmul(transpose_x = attn_189_transpose_x_1, transpose_y = attn_189_transpose_y_1, x = q_479_cast_fp16, y = var_5799_cast_fp16)[name = tensor<string, []>("attn_189_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_191_cast_fp16 = softmax(axis = var_21, x = attn_189_cast_fp16)[name = tensor<string, []>("attn_191_cast_fp16")];
            tensor<bool, []> var_5806_transpose_x_0 = const()[name = tensor<string, []>("op_5806_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_5806_transpose_y_0 = const()[name = tensor<string, []>("op_5806_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_5806_cast_fp16 = matmul(transpose_x = var_5806_transpose_x_0, transpose_y = var_5806_transpose_y_0, x = attn_191_cast_fp16, y = v_287_cast_fp16)[name = tensor<string, []>("op_5806_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_7_cast_fp16 = squeeze(axes = out_7_axes_0, x = var_5806_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<bool, []> input_59_interleave_0 = const()[name = tensor<string, []>("input_59_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_59_cast_fp16 = concat(axis = var_21, interleave = input_59_interleave_0, values = (var_4663_cast_fp16, var_4767_cast_fp16, var_4871_cast_fp16, var_4975_cast_fp16, var_5079_cast_fp16, var_5183_cast_fp16, var_5287_cast_fp16, var_5391_cast_fp16, var_5495_cast_fp16, var_5599_cast_fp16, var_5703_cast_fp16, out_7_cast_fp16))[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<int32, [1]> x_645_axes_0 = const()[name = tensor<string, []>("x_645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_3_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47611840)))];
            tensor<fp16, [768]> model_blocks_3_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47613440)))];
            tensor<fp16, [1, 197, 768]> x_645_cast_fp16 = layer_norm(axes = x_645_axes_0, beta = model_blocks_3_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_3_attn_inner_attn_ln_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_645_cast_fp16")];
            tensor<int32, [4]> var_5814 = const()[name = tensor<string, []>("op_5814"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_647_cast_fp16 = reshape(shape = var_5814, x = x_645_cast_fp16)[name = tensor<string, []>("x_647_cast_fp16")];
            tensor<int32, [4]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_649_pad_type_0 = const()[name = tensor<string, []>("x_649_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_649_strides_0 = const()[name = tensor<string, []>("x_649_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_649_pad_0 = const()[name = tensor<string, []>("x_649_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_649_dilations_0 = const()[name = tensor<string, []>("x_649_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_649_groups_0 = const()[name = tensor<string, []>("x_649_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_3_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47615040)))];
            tensor<fp16, [768]> model_blocks_3_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48794752)))];
            tensor<fp16, [1, 768, 1, 197]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_647_cast_fp16)[name = tensor<string, []>("transpose_349")];
            tensor<fp16, [1, 768, 1, 197]> x_649_cast_fp16 = conv(bias = model_blocks_3_attn_proj_bias_to_fp16, dilations = x_649_dilations_0, groups = x_649_groups_0, pad = x_649_pad_0, pad_type = x_649_pad_type_0, strides = x_649_strides_0, weight = model_blocks_3_attn_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("x_649_cast_fp16")];
            tensor<int32, [4]> x_651_perm_0 = const()[name = tensor<string, []>("x_651_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_5825 = const()[name = tensor<string, []>("op_5825"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_651_cast_fp16 = transpose(perm = x_651_perm_0, x = x_649_cast_fp16)[name = tensor<string, []>("transpose_348")];
            tensor<fp16, [1, 197, 768]> var_5826_cast_fp16 = reshape(shape = var_5825, x = x_651_cast_fp16)[name = tensor<string, []>("op_5826_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_63_cast_fp16 = add(x = input_55_cast_fp16, y = var_5826_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [1]> x_653_axes_0 = const()[name = tensor<string, []>("x_653_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48796352)))];
            tensor<fp16, [768]> model_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48797952)))];
            tensor<fp16, [1, 197, 768]> x_653_cast_fp16 = layer_norm(axes = x_653_axes_0, beta = model_blocks_3_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_3_norm2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("x_653_cast_fp16")];
            tensor<int32, [4]> var_5838 = const()[name = tensor<string, []>("op_5838"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_655_cast_fp16 = reshape(shape = var_5838, x = x_653_cast_fp16)[name = tensor<string, []>("x_655_cast_fp16")];
            tensor<int32, [4]> input_65_perm_0 = const()[name = tensor<string, []>("input_65_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_67_pad_type_0 = const()[name = tensor<string, []>("input_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = tensor<string, []>("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = tensor<string, []>("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = tensor<string, []>("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_67_groups_0 = const()[name = tensor<string, []>("input_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_3_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48799552)))];
            tensor<fp16, [2048]> model_blocks_3_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51945344)))];
            tensor<fp16, [1, 768, 1, 197]> input_65_cast_fp16 = transpose(perm = input_65_perm_0, x = x_655_cast_fp16)[name = tensor<string, []>("transpose_347")];
            tensor<fp16, [1, 2048, 1, 197]> input_67_cast_fp16 = conv(bias = model_blocks_3_mlp_w1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = model_blocks_3_mlp_w1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> x2_199_pad_type_0 = const()[name = tensor<string, []>("x2_199_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_199_strides_0 = const()[name = tensor<string, []>("x2_199_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_199_pad_0 = const()[name = tensor<string, []>("x2_199_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_199_dilations_0 = const()[name = tensor<string, []>("x2_199_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_199_groups_0 = const()[name = tensor<string, []>("x2_199_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_3_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51949504)))];
            tensor<fp16, [2048]> model_blocks_3_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55095296)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_199_cast_fp16 = conv(bias = model_blocks_3_mlp_w2_bias_to_fp16, dilations = x2_199_dilations_0, groups = x2_199_groups_0, pad = x2_199_pad_0, pad_type = x2_199_pad_type_0, strides = x2_199_strides_0, weight = model_blocks_3_mlp_w2_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("x2_199_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_5855_cast_fp16 = silu(x = input_67_cast_fp16)[name = tensor<string, []>("op_5855_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_25_cast_fp16 = mul(x = var_5855_cast_fp16, y = x2_199_cast_fp16)[name = tensor<string, []>("hidden_25_cast_fp16")];
            tensor<int32, [4]> hidden_27_perm_0 = const()[name = tensor<string, []>("hidden_27_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_5858 = const()[name = tensor<string, []>("op_5858"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_27_cast_fp16 = transpose(perm = hidden_27_perm_0, x = hidden_25_cast_fp16)[name = tensor<string, []>("transpose_346")];
            tensor<fp16, [1, 197, 2048]> input_69_cast_fp16 = reshape(shape = var_5858, x = hidden_27_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<int32, [1]> hidden_29_axes_0 = const()[name = tensor<string, []>("hidden_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_3_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55099456)))];
            tensor<fp16, [2048]> model_blocks_3_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55103616)))];
            tensor<fp16, [1, 197, 2048]> hidden_29_cast_fp16 = layer_norm(axes = hidden_29_axes_0, beta = model_blocks_3_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_3_mlp_ffn_ln_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("hidden_29_cast_fp16")];
            tensor<int32, [4]> var_5864 = const()[name = tensor<string, []>("op_5864"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_31_cast_fp16 = reshape(shape = var_5864, x = hidden_29_cast_fp16)[name = tensor<string, []>("hidden_31_cast_fp16")];
            tensor<int32, [4]> input_71_perm_0 = const()[name = tensor<string, []>("input_71_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_657_pad_type_0 = const()[name = tensor<string, []>("x_657_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_657_strides_0 = const()[name = tensor<string, []>("x_657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_657_pad_0 = const()[name = tensor<string, []>("x_657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_657_dilations_0 = const()[name = tensor<string, []>("x_657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_657_groups_0 = const()[name = tensor<string, []>("x_657_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_3_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55107776)))];
            tensor<fp16, [768]> model_blocks_3_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_3_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58253568)))];
            tensor<fp16, [1, 2048, 1, 197]> input_71_cast_fp16 = transpose(perm = input_71_perm_0, x = hidden_31_cast_fp16)[name = tensor<string, []>("transpose_345")];
            tensor<fp16, [1, 768, 1, 197]> x_657_cast_fp16 = conv(bias = model_blocks_3_mlp_w3_bias_to_fp16, dilations = x_657_dilations_0, groups = x_657_groups_0, pad = x_657_pad_0, pad_type = x_657_pad_type_0, strides = x_657_strides_0, weight = model_blocks_3_mlp_w3_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("x_657_cast_fp16")];
            tensor<int32, [4]> x_659_perm_0 = const()[name = tensor<string, []>("x_659_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_5875 = const()[name = tensor<string, []>("op_5875"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_659_cast_fp16 = transpose(perm = x_659_perm_0, x = x_657_cast_fp16)[name = tensor<string, []>("transpose_344")];
            tensor<fp16, [1, 197, 768]> var_5876_cast_fp16 = reshape(shape = var_5875, x = x_659_cast_fp16)[name = tensor<string, []>("op_5876_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_73_cast_fp16 = add(x = input_63_cast_fp16, y = var_5876_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<int32, [1]> x_661_axes_0 = const()[name = tensor<string, []>("x_661_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58255168)))];
            tensor<fp16, [768]> model_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58256768)))];
            tensor<fp16, [1, 197, 768]> x_661_cast_fp16 = layer_norm(axes = x_661_axes_0, beta = model_blocks_4_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_4_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("x_661_cast_fp16")];
            tensor<int32, [4]> var_6010 = const()[name = tensor<string, []>("op_6010"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_663_cast_fp16 = reshape(shape = var_6010, x = x_661_cast_fp16)[name = tensor<string, []>("x_663_cast_fp16")];
            tensor<int32, [4]> input_75_perm_0 = const()[name = tensor<string, []>("input_75_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_6018_pad_type_0 = const()[name = tensor<string, []>("op_6018_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6018_strides_0 = const()[name = tensor<string, []>("op_6018_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6018_pad_0 = const()[name = tensor<string, []>("op_6018_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6018_dilations_0 = const()[name = tensor<string, []>("op_6018_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6018_groups_0 = const()[name = tensor<string, []>("op_6018_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58258368)))];
            tensor<fp16, [1, 768, 1, 197]> input_75_cast_fp16 = transpose(perm = input_75_perm_0, x = x_663_cast_fp16)[name = tensor<string, []>("transpose_343")];
            tensor<fp16, [1, 64, 1, 197]> var_6018_cast_fp16 = conv(dilations = var_6018_dilations_0, groups = var_6018_groups_0, pad = var_6018_pad_0, pad_type = var_6018_pad_type_0, strides = var_6018_strides_0, weight = model_blocks_4_attn_q_projs_0_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6018_cast_fp16")];
            tensor<int32, [1]> var_6019_axes_0 = const()[name = tensor<string, []>("op_6019_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6019_cast_fp16 = squeeze(axes = var_6019_axes_0, x = var_6018_cast_fp16)[name = tensor<string, []>("op_6019_cast_fp16")];
            tensor<int32, [3]> q_481_perm_0 = const()[name = tensor<string, []>("q_481_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6026_pad_type_0 = const()[name = tensor<string, []>("op_6026_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6026_strides_0 = const()[name = tensor<string, []>("op_6026_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6026_pad_0 = const()[name = tensor<string, []>("op_6026_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6026_dilations_0 = const()[name = tensor<string, []>("op_6026_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6026_groups_0 = const()[name = tensor<string, []>("op_6026_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58356736)))];
            tensor<fp16, [1, 64, 1, 197]> var_6026_cast_fp16 = conv(dilations = var_6026_dilations_0, groups = var_6026_groups_0, pad = var_6026_pad_0, pad_type = var_6026_pad_type_0, strides = var_6026_strides_0, weight = model_blocks_4_attn_k_projs_0_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6026_cast_fp16")];
            tensor<int32, [1]> var_6027_axes_0 = const()[name = tensor<string, []>("op_6027_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6027_cast_fp16 = squeeze(axes = var_6027_axes_0, x = var_6026_cast_fp16)[name = tensor<string, []>("op_6027_cast_fp16")];
            tensor<int32, [3]> k_289_perm_0 = const()[name = tensor<string, []>("k_289_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6034_pad_type_0 = const()[name = tensor<string, []>("op_6034_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6034_strides_0 = const()[name = tensor<string, []>("op_6034_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6034_pad_0 = const()[name = tensor<string, []>("op_6034_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6034_dilations_0 = const()[name = tensor<string, []>("op_6034_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6034_groups_0 = const()[name = tensor<string, []>("op_6034_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58455104)))];
            tensor<fp16, [1, 64, 1, 197]> var_6034_cast_fp16 = conv(dilations = var_6034_dilations_0, groups = var_6034_groups_0, pad = var_6034_pad_0, pad_type = var_6034_pad_type_0, strides = var_6034_strides_0, weight = model_blocks_4_attn_v_projs_0_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = tensor<string, []>("op_6035_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6035_cast_fp16 = squeeze(axes = var_6035_axes_0, x = var_6034_cast_fp16)[name = tensor<string, []>("op_6035_cast_fp16")];
            tensor<int32, [3]> v_289_perm_0 = const()[name = tensor<string, []>("v_289_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58553472)))];
            tensor<fp16, [1, 197, 64]> q_481_cast_fp16 = transpose(perm = q_481_perm_0, x = var_6019_cast_fp16)[name = tensor<string, []>("transpose_342")];
            tensor<fp16, [1, 197, 64]> q_483_cast_fp16 = add(x = q_481_cast_fp16, y = model_blocks_4_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_483_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58553664)))];
            tensor<fp16, [1, 197, 64]> v_289_cast_fp16 = transpose(perm = v_289_perm_0, x = var_6035_cast_fp16)[name = tensor<string, []>("transpose_340")];
            tensor<fp16, [1, 197, 64]> v_291_cast_fp16 = add(x = v_289_cast_fp16, y = model_blocks_4_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_291_cast_fp16")];
            tensor<int32, [1]> q_485_axes_0 = const()[name = tensor<string, []>("q_485_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_485_cast_fp16 = expand_dims(axes = q_485_axes_0, x = q_483_cast_fp16)[name = tensor<string, []>("q_485_cast_fp16")];
            tensor<int32, [1]> k_291_axes_0 = const()[name = tensor<string, []>("k_291_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_289_cast_fp16 = transpose(perm = k_289_perm_0, x = var_6027_cast_fp16)[name = tensor<string, []>("transpose_341")];
            tensor<fp16, [1, 1, 197, 64]> k_291_cast_fp16 = expand_dims(axes = k_291_axes_0, x = k_289_cast_fp16)[name = tensor<string, []>("k_291_cast_fp16")];
            tensor<int32, [1]> v_293_axes_0 = const()[name = tensor<string, []>("v_293_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_293_cast_fp16 = expand_dims(axes = v_293_axes_0, x = v_291_cast_fp16)[name = tensor<string, []>("v_293_cast_fp16")];
            tensor<int32, [4]> var_6044_begin_0 = const()[name = tensor<string, []>("op_6044_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6044_end_0 = const()[name = tensor<string, []>("op_6044_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6044_end_mask_0 = const()[name = tensor<string, []>("op_6044_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, x = q_485_cast_fp16)[name = tensor<string, []>("op_6044_cast_fp16")];
            tensor<int32, [4]> var_6048_begin_0 = const()[name = tensor<string, []>("op_6048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6048_end_0 = const()[name = tensor<string, []>("op_6048_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6048_end_mask_0 = const()[name = tensor<string, []>("op_6048_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6048_cast_fp16 = slice_by_index(begin = var_6048_begin_0, end = var_6048_end_0, end_mask = var_6048_end_mask_0, x = k_291_cast_fp16)[name = tensor<string, []>("op_6048_cast_fp16")];
            tensor<int32, [4]> var_6052_begin_0 = const()[name = tensor<string, []>("op_6052_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6052_end_0 = const()[name = tensor<string, []>("op_6052_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6052_end_mask_0 = const()[name = tensor<string, []>("op_6052_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, x = q_485_cast_fp16)[name = tensor<string, []>("op_6052_cast_fp16")];
            tensor<int32, [4]> var_6056_begin_0 = const()[name = tensor<string, []>("op_6056_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6056_end_0 = const()[name = tensor<string, []>("op_6056_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6056_end_mask_0 = const()[name = tensor<string, []>("op_6056_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6056_cast_fp16 = slice_by_index(begin = var_6056_begin_0, end = var_6056_end_0, end_mask = var_6056_end_mask_0, x = k_291_cast_fp16)[name = tensor<string, []>("op_6056_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6058_cast_fp16 = mul(x = var_6052_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6058_cast_fp16")];
            tensor<int32, [5]> var_6066 = const()[name = tensor<string, []>("op_6066"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_667_cast_fp16 = reshape(shape = var_6066, x = var_6052_cast_fp16)[name = tensor<string, []>("x_667_cast_fp16")];
            tensor<int32, [2]> var_6068_split_sizes_0 = const()[name = tensor<string, []>("op_6068_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6068_axis_0 = const()[name = tensor<string, []>("op_6068_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6068_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6068_cast_fp16_1 = split(axis = var_6068_axis_0, split_sizes = var_6068_split_sizes_0, x = x_667_cast_fp16)[name = tensor<string, []>("op_6068_cast_fp16")];
            tensor<int32, [1]> squeeze_192_axes_0 = const()[name = tensor<string, []>("squeeze_192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_192_cast_fp16 = squeeze(axes = squeeze_192_axes_0, x = var_6068_cast_fp16_0)[name = tensor<string, []>("squeeze_192_cast_fp16")];
            tensor<int32, [1]> squeeze_193_axes_0 = const()[name = tensor<string, []>("squeeze_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_193_cast_fp16 = squeeze(axes = squeeze_193_axes_0, x = var_6068_cast_fp16_1)[name = tensor<string, []>("squeeze_193_cast_fp16")];
            tensor<fp16, []> const_791_promoted_to_fp16 = const()[name = tensor<string, []>("const_791_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6071_cast_fp16 = mul(x = squeeze_193_cast_fp16, y = const_791_promoted_to_fp16)[name = tensor<string, []>("op_6071_cast_fp16")];
            tensor<int32, []> x_669_axis_0 = const()[name = tensor<string, []>("x_669_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_669_cast_fp16 = stack(axis = x_669_axis_0, values = (var_6071_cast_fp16, squeeze_192_cast_fp16))[name = tensor<string, []>("x_669_cast_fp16")];
            tensor<int32, [4]> var_6077 = const()[name = tensor<string, []>("op_6077"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6078_cast_fp16 = reshape(shape = var_6077, x = x_669_cast_fp16)[name = tensor<string, []>("op_6078_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6079_cast_fp16 = mul(x = var_6078_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6079_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_97_cast_fp16 = add(x = var_6058_cast_fp16, y = var_6079_cast_fp16)[name = tensor<string, []>("q_patches_97_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6081_cast_fp16 = mul(x = var_6056_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6081_cast_fp16")];
            tensor<int32, [5]> var_6089 = const()[name = tensor<string, []>("op_6089"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_673_cast_fp16 = reshape(shape = var_6089, x = var_6056_cast_fp16)[name = tensor<string, []>("x_673_cast_fp16")];
            tensor<int32, [2]> var_6091_split_sizes_0 = const()[name = tensor<string, []>("op_6091_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6091_axis_0 = const()[name = tensor<string, []>("op_6091_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6091_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6091_cast_fp16_1 = split(axis = var_6091_axis_0, split_sizes = var_6091_split_sizes_0, x = x_673_cast_fp16)[name = tensor<string, []>("op_6091_cast_fp16")];
            tensor<int32, [1]> squeeze_194_axes_0 = const()[name = tensor<string, []>("squeeze_194_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_194_cast_fp16 = squeeze(axes = squeeze_194_axes_0, x = var_6091_cast_fp16_0)[name = tensor<string, []>("squeeze_194_cast_fp16")];
            tensor<int32, [1]> squeeze_195_axes_0 = const()[name = tensor<string, []>("squeeze_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_195_cast_fp16 = squeeze(axes = squeeze_195_axes_0, x = var_6091_cast_fp16_1)[name = tensor<string, []>("squeeze_195_cast_fp16")];
            tensor<fp16, []> const_799_promoted_to_fp16 = const()[name = tensor<string, []>("const_799_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6094_cast_fp16 = mul(x = squeeze_195_cast_fp16, y = const_799_promoted_to_fp16)[name = tensor<string, []>("op_6094_cast_fp16")];
            tensor<int32, []> x_675_axis_0 = const()[name = tensor<string, []>("x_675_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_675_cast_fp16 = stack(axis = x_675_axis_0, values = (var_6094_cast_fp16, squeeze_194_cast_fp16))[name = tensor<string, []>("x_675_cast_fp16")];
            tensor<int32, [4]> var_6100 = const()[name = tensor<string, []>("op_6100"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6101_cast_fp16 = reshape(shape = var_6100, x = x_675_cast_fp16)[name = tensor<string, []>("op_6101_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6102_cast_fp16 = mul(x = var_6101_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6102_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_97_cast_fp16 = add(x = var_6081_cast_fp16, y = var_6102_cast_fp16)[name = tensor<string, []>("k_patches_97_cast_fp16")];
            tensor<bool, []> var_6105_interleave_0 = const()[name = tensor<string, []>("op_6105_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6105_cast_fp16 = concat(axis = var_20, interleave = var_6105_interleave_0, values = (var_6044_cast_fp16, q_patches_97_cast_fp16))[name = tensor<string, []>("op_6105_cast_fp16")];
            tensor<bool, []> var_6108_interleave_0 = const()[name = tensor<string, []>("op_6108_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6108_cast_fp16 = concat(axis = var_20, interleave = var_6108_interleave_0, values = (var_6048_cast_fp16, k_patches_97_cast_fp16))[name = tensor<string, []>("op_6108_cast_fp16")];
            tensor<fp16, []> var_6110_to_fp16 = const()[name = tensor<string, []>("op_6110_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_489_cast_fp16 = mul(x = var_6105_cast_fp16, y = var_6110_to_fp16)[name = tensor<string, []>("q_489_cast_fp16")];
            tensor<bool, []> attn_193_transpose_x_1 = const()[name = tensor<string, []>("attn_193_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_193_transpose_y_1 = const()[name = tensor<string, []>("attn_193_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_193_cast_fp16 = matmul(transpose_x = attn_193_transpose_x_1, transpose_y = attn_193_transpose_y_1, x = q_489_cast_fp16, y = var_6108_cast_fp16)[name = tensor<string, []>("attn_193_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_195_cast_fp16 = softmax(axis = var_21, x = attn_193_cast_fp16)[name = tensor<string, []>("attn_195_cast_fp16")];
            tensor<bool, []> var_6115_transpose_x_0 = const()[name = tensor<string, []>("op_6115_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6115_transpose_y_0 = const()[name = tensor<string, []>("op_6115_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6115_cast_fp16 = matmul(transpose_x = var_6115_transpose_x_0, transpose_y = var_6115_transpose_y_0, x = attn_195_cast_fp16, y = v_293_cast_fp16)[name = tensor<string, []>("op_6115_cast_fp16")];
            tensor<int32, [1]> var_6116_axes_0 = const()[name = tensor<string, []>("op_6116_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6116_cast_fp16 = squeeze(axes = var_6116_axes_0, x = var_6115_cast_fp16)[name = tensor<string, []>("op_6116_cast_fp16")];
            tensor<string, []> var_6122_pad_type_0 = const()[name = tensor<string, []>("op_6122_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6122_strides_0 = const()[name = tensor<string, []>("op_6122_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6122_pad_0 = const()[name = tensor<string, []>("op_6122_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6122_dilations_0 = const()[name = tensor<string, []>("op_6122_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6122_groups_0 = const()[name = tensor<string, []>("op_6122_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58553856)))];
            tensor<fp16, [1, 64, 1, 197]> var_6122_cast_fp16 = conv(dilations = var_6122_dilations_0, groups = var_6122_groups_0, pad = var_6122_pad_0, pad_type = var_6122_pad_type_0, strides = var_6122_strides_0, weight = model_blocks_4_attn_q_projs_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6122_cast_fp16")];
            tensor<int32, [1]> var_6123_axes_0 = const()[name = tensor<string, []>("op_6123_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6123_cast_fp16 = squeeze(axes = var_6123_axes_0, x = var_6122_cast_fp16)[name = tensor<string, []>("op_6123_cast_fp16")];
            tensor<int32, [3]> q_491_perm_0 = const()[name = tensor<string, []>("q_491_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6130_pad_type_0 = const()[name = tensor<string, []>("op_6130_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6130_strides_0 = const()[name = tensor<string, []>("op_6130_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6130_pad_0 = const()[name = tensor<string, []>("op_6130_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6130_dilations_0 = const()[name = tensor<string, []>("op_6130_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6130_groups_0 = const()[name = tensor<string, []>("op_6130_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58652224)))];
            tensor<fp16, [1, 64, 1, 197]> var_6130_cast_fp16 = conv(dilations = var_6130_dilations_0, groups = var_6130_groups_0, pad = var_6130_pad_0, pad_type = var_6130_pad_type_0, strides = var_6130_strides_0, weight = model_blocks_4_attn_k_projs_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6130_cast_fp16")];
            tensor<int32, [1]> var_6131_axes_0 = const()[name = tensor<string, []>("op_6131_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6131_cast_fp16 = squeeze(axes = var_6131_axes_0, x = var_6130_cast_fp16)[name = tensor<string, []>("op_6131_cast_fp16")];
            tensor<int32, [3]> k_295_perm_0 = const()[name = tensor<string, []>("k_295_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6138_pad_type_0 = const()[name = tensor<string, []>("op_6138_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6138_strides_0 = const()[name = tensor<string, []>("op_6138_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6138_pad_0 = const()[name = tensor<string, []>("op_6138_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6138_dilations_0 = const()[name = tensor<string, []>("op_6138_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6138_groups_0 = const()[name = tensor<string, []>("op_6138_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58750592)))];
            tensor<fp16, [1, 64, 1, 197]> var_6138_cast_fp16 = conv(dilations = var_6138_dilations_0, groups = var_6138_groups_0, pad = var_6138_pad_0, pad_type = var_6138_pad_type_0, strides = var_6138_strides_0, weight = model_blocks_4_attn_v_projs_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6138_cast_fp16")];
            tensor<int32, [1]> var_6139_axes_0 = const()[name = tensor<string, []>("op_6139_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6139_cast_fp16 = squeeze(axes = var_6139_axes_0, x = var_6138_cast_fp16)[name = tensor<string, []>("op_6139_cast_fp16")];
            tensor<int32, [3]> v_295_perm_0 = const()[name = tensor<string, []>("v_295_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58848960)))];
            tensor<fp16, [1, 197, 64]> q_491_cast_fp16 = transpose(perm = q_491_perm_0, x = var_6123_cast_fp16)[name = tensor<string, []>("transpose_339")];
            tensor<fp16, [1, 197, 64]> q_493_cast_fp16 = add(x = q_491_cast_fp16, y = model_blocks_4_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_493_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58849152)))];
            tensor<fp16, [1, 197, 64]> v_295_cast_fp16 = transpose(perm = v_295_perm_0, x = var_6139_cast_fp16)[name = tensor<string, []>("transpose_337")];
            tensor<fp16, [1, 197, 64]> v_297_cast_fp16 = add(x = v_295_cast_fp16, y = model_blocks_4_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_297_cast_fp16")];
            tensor<int32, [1]> q_495_axes_0 = const()[name = tensor<string, []>("q_495_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_495_cast_fp16 = expand_dims(axes = q_495_axes_0, x = q_493_cast_fp16)[name = tensor<string, []>("q_495_cast_fp16")];
            tensor<int32, [1]> k_297_axes_0 = const()[name = tensor<string, []>("k_297_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_295_cast_fp16 = transpose(perm = k_295_perm_0, x = var_6131_cast_fp16)[name = tensor<string, []>("transpose_338")];
            tensor<fp16, [1, 1, 197, 64]> k_297_cast_fp16 = expand_dims(axes = k_297_axes_0, x = k_295_cast_fp16)[name = tensor<string, []>("k_297_cast_fp16")];
            tensor<int32, [1]> v_299_axes_0 = const()[name = tensor<string, []>("v_299_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_299_cast_fp16 = expand_dims(axes = v_299_axes_0, x = v_297_cast_fp16)[name = tensor<string, []>("v_299_cast_fp16")];
            tensor<int32, [4]> var_6148_begin_0 = const()[name = tensor<string, []>("op_6148_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6148_end_0 = const()[name = tensor<string, []>("op_6148_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6148_end_mask_0 = const()[name = tensor<string, []>("op_6148_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6148_cast_fp16 = slice_by_index(begin = var_6148_begin_0, end = var_6148_end_0, end_mask = var_6148_end_mask_0, x = q_495_cast_fp16)[name = tensor<string, []>("op_6148_cast_fp16")];
            tensor<int32, [4]> var_6152_begin_0 = const()[name = tensor<string, []>("op_6152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6152_end_0 = const()[name = tensor<string, []>("op_6152_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6152_end_mask_0 = const()[name = tensor<string, []>("op_6152_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6152_cast_fp16 = slice_by_index(begin = var_6152_begin_0, end = var_6152_end_0, end_mask = var_6152_end_mask_0, x = k_297_cast_fp16)[name = tensor<string, []>("op_6152_cast_fp16")];
            tensor<int32, [4]> var_6156_begin_0 = const()[name = tensor<string, []>("op_6156_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6156_end_0 = const()[name = tensor<string, []>("op_6156_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6156_end_mask_0 = const()[name = tensor<string, []>("op_6156_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6156_cast_fp16 = slice_by_index(begin = var_6156_begin_0, end = var_6156_end_0, end_mask = var_6156_end_mask_0, x = q_495_cast_fp16)[name = tensor<string, []>("op_6156_cast_fp16")];
            tensor<int32, [4]> var_6160_begin_0 = const()[name = tensor<string, []>("op_6160_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6160_end_0 = const()[name = tensor<string, []>("op_6160_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6160_end_mask_0 = const()[name = tensor<string, []>("op_6160_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6160_cast_fp16 = slice_by_index(begin = var_6160_begin_0, end = var_6160_end_0, end_mask = var_6160_end_mask_0, x = k_297_cast_fp16)[name = tensor<string, []>("op_6160_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6162_cast_fp16 = mul(x = var_6156_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6162_cast_fp16")];
            tensor<int32, [5]> var_6170 = const()[name = tensor<string, []>("op_6170"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_679_cast_fp16 = reshape(shape = var_6170, x = var_6156_cast_fp16)[name = tensor<string, []>("x_679_cast_fp16")];
            tensor<int32, [2]> var_6172_split_sizes_0 = const()[name = tensor<string, []>("op_6172_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6172_axis_0 = const()[name = tensor<string, []>("op_6172_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6172_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6172_cast_fp16_1 = split(axis = var_6172_axis_0, split_sizes = var_6172_split_sizes_0, x = x_679_cast_fp16)[name = tensor<string, []>("op_6172_cast_fp16")];
            tensor<int32, [1]> squeeze_196_axes_0 = const()[name = tensor<string, []>("squeeze_196_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_196_cast_fp16 = squeeze(axes = squeeze_196_axes_0, x = var_6172_cast_fp16_0)[name = tensor<string, []>("squeeze_196_cast_fp16")];
            tensor<int32, [1]> squeeze_197_axes_0 = const()[name = tensor<string, []>("squeeze_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_197_cast_fp16 = squeeze(axes = squeeze_197_axes_0, x = var_6172_cast_fp16_1)[name = tensor<string, []>("squeeze_197_cast_fp16")];
            tensor<fp16, []> const_807_promoted_to_fp16 = const()[name = tensor<string, []>("const_807_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6175_cast_fp16 = mul(x = squeeze_197_cast_fp16, y = const_807_promoted_to_fp16)[name = tensor<string, []>("op_6175_cast_fp16")];
            tensor<int32, []> x_681_axis_0 = const()[name = tensor<string, []>("x_681_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_681_cast_fp16 = stack(axis = x_681_axis_0, values = (var_6175_cast_fp16, squeeze_196_cast_fp16))[name = tensor<string, []>("x_681_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = tensor<string, []>("op_6181"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6182_cast_fp16 = reshape(shape = var_6181, x = x_681_cast_fp16)[name = tensor<string, []>("op_6182_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6183_cast_fp16 = mul(x = var_6182_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6183_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_99_cast_fp16 = add(x = var_6162_cast_fp16, y = var_6183_cast_fp16)[name = tensor<string, []>("q_patches_99_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6185_cast_fp16 = mul(x = var_6160_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6185_cast_fp16")];
            tensor<int32, [5]> var_6193 = const()[name = tensor<string, []>("op_6193"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_685_cast_fp16 = reshape(shape = var_6193, x = var_6160_cast_fp16)[name = tensor<string, []>("x_685_cast_fp16")];
            tensor<int32, [2]> var_6195_split_sizes_0 = const()[name = tensor<string, []>("op_6195_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6195_axis_0 = const()[name = tensor<string, []>("op_6195_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6195_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6195_cast_fp16_1 = split(axis = var_6195_axis_0, split_sizes = var_6195_split_sizes_0, x = x_685_cast_fp16)[name = tensor<string, []>("op_6195_cast_fp16")];
            tensor<int32, [1]> squeeze_198_axes_0 = const()[name = tensor<string, []>("squeeze_198_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_198_cast_fp16 = squeeze(axes = squeeze_198_axes_0, x = var_6195_cast_fp16_0)[name = tensor<string, []>("squeeze_198_cast_fp16")];
            tensor<int32, [1]> squeeze_199_axes_0 = const()[name = tensor<string, []>("squeeze_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_199_cast_fp16 = squeeze(axes = squeeze_199_axes_0, x = var_6195_cast_fp16_1)[name = tensor<string, []>("squeeze_199_cast_fp16")];
            tensor<fp16, []> const_815_promoted_to_fp16 = const()[name = tensor<string, []>("const_815_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6198_cast_fp16 = mul(x = squeeze_199_cast_fp16, y = const_815_promoted_to_fp16)[name = tensor<string, []>("op_6198_cast_fp16")];
            tensor<int32, []> x_687_axis_0 = const()[name = tensor<string, []>("x_687_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_687_cast_fp16 = stack(axis = x_687_axis_0, values = (var_6198_cast_fp16, squeeze_198_cast_fp16))[name = tensor<string, []>("x_687_cast_fp16")];
            tensor<int32, [4]> var_6204 = const()[name = tensor<string, []>("op_6204"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6205_cast_fp16 = reshape(shape = var_6204, x = x_687_cast_fp16)[name = tensor<string, []>("op_6205_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6206_cast_fp16 = mul(x = var_6205_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6206_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_99_cast_fp16 = add(x = var_6185_cast_fp16, y = var_6206_cast_fp16)[name = tensor<string, []>("k_patches_99_cast_fp16")];
            tensor<bool, []> var_6209_interleave_0 = const()[name = tensor<string, []>("op_6209_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6209_cast_fp16 = concat(axis = var_20, interleave = var_6209_interleave_0, values = (var_6148_cast_fp16, q_patches_99_cast_fp16))[name = tensor<string, []>("op_6209_cast_fp16")];
            tensor<bool, []> var_6212_interleave_0 = const()[name = tensor<string, []>("op_6212_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6212_cast_fp16 = concat(axis = var_20, interleave = var_6212_interleave_0, values = (var_6152_cast_fp16, k_patches_99_cast_fp16))[name = tensor<string, []>("op_6212_cast_fp16")];
            tensor<fp16, []> var_6214_to_fp16 = const()[name = tensor<string, []>("op_6214_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_499_cast_fp16 = mul(x = var_6209_cast_fp16, y = var_6214_to_fp16)[name = tensor<string, []>("q_499_cast_fp16")];
            tensor<bool, []> attn_197_transpose_x_1 = const()[name = tensor<string, []>("attn_197_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_197_transpose_y_1 = const()[name = tensor<string, []>("attn_197_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_197_cast_fp16 = matmul(transpose_x = attn_197_transpose_x_1, transpose_y = attn_197_transpose_y_1, x = q_499_cast_fp16, y = var_6212_cast_fp16)[name = tensor<string, []>("attn_197_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_199_cast_fp16 = softmax(axis = var_21, x = attn_197_cast_fp16)[name = tensor<string, []>("attn_199_cast_fp16")];
            tensor<bool, []> var_6219_transpose_x_0 = const()[name = tensor<string, []>("op_6219_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6219_transpose_y_0 = const()[name = tensor<string, []>("op_6219_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6219_cast_fp16 = matmul(transpose_x = var_6219_transpose_x_0, transpose_y = var_6219_transpose_y_0, x = attn_199_cast_fp16, y = v_299_cast_fp16)[name = tensor<string, []>("op_6219_cast_fp16")];
            tensor<int32, [1]> var_6220_axes_0 = const()[name = tensor<string, []>("op_6220_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6220_cast_fp16 = squeeze(axes = var_6220_axes_0, x = var_6219_cast_fp16)[name = tensor<string, []>("op_6220_cast_fp16")];
            tensor<string, []> var_6226_pad_type_0 = const()[name = tensor<string, []>("op_6226_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6226_strides_0 = const()[name = tensor<string, []>("op_6226_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6226_pad_0 = const()[name = tensor<string, []>("op_6226_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6226_dilations_0 = const()[name = tensor<string, []>("op_6226_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6226_groups_0 = const()[name = tensor<string, []>("op_6226_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58849344)))];
            tensor<fp16, [1, 64, 1, 197]> var_6226_cast_fp16 = conv(dilations = var_6226_dilations_0, groups = var_6226_groups_0, pad = var_6226_pad_0, pad_type = var_6226_pad_type_0, strides = var_6226_strides_0, weight = model_blocks_4_attn_q_projs_2_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6226_cast_fp16")];
            tensor<int32, [1]> var_6227_axes_0 = const()[name = tensor<string, []>("op_6227_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6227_cast_fp16 = squeeze(axes = var_6227_axes_0, x = var_6226_cast_fp16)[name = tensor<string, []>("op_6227_cast_fp16")];
            tensor<int32, [3]> q_501_perm_0 = const()[name = tensor<string, []>("q_501_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6234_pad_type_0 = const()[name = tensor<string, []>("op_6234_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6234_strides_0 = const()[name = tensor<string, []>("op_6234_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6234_pad_0 = const()[name = tensor<string, []>("op_6234_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6234_dilations_0 = const()[name = tensor<string, []>("op_6234_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6234_groups_0 = const()[name = tensor<string, []>("op_6234_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58947712)))];
            tensor<fp16, [1, 64, 1, 197]> var_6234_cast_fp16 = conv(dilations = var_6234_dilations_0, groups = var_6234_groups_0, pad = var_6234_pad_0, pad_type = var_6234_pad_type_0, strides = var_6234_strides_0, weight = model_blocks_4_attn_k_projs_2_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6234_cast_fp16")];
            tensor<int32, [1]> var_6235_axes_0 = const()[name = tensor<string, []>("op_6235_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6235_cast_fp16 = squeeze(axes = var_6235_axes_0, x = var_6234_cast_fp16)[name = tensor<string, []>("op_6235_cast_fp16")];
            tensor<int32, [3]> k_301_perm_0 = const()[name = tensor<string, []>("k_301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6242_pad_type_0 = const()[name = tensor<string, []>("op_6242_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6242_strides_0 = const()[name = tensor<string, []>("op_6242_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6242_pad_0 = const()[name = tensor<string, []>("op_6242_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6242_dilations_0 = const()[name = tensor<string, []>("op_6242_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6242_groups_0 = const()[name = tensor<string, []>("op_6242_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59046080)))];
            tensor<fp16, [1, 64, 1, 197]> var_6242_cast_fp16 = conv(dilations = var_6242_dilations_0, groups = var_6242_groups_0, pad = var_6242_pad_0, pad_type = var_6242_pad_type_0, strides = var_6242_strides_0, weight = model_blocks_4_attn_v_projs_2_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6242_cast_fp16")];
            tensor<int32, [1]> var_6243_axes_0 = const()[name = tensor<string, []>("op_6243_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6243_cast_fp16 = squeeze(axes = var_6243_axes_0, x = var_6242_cast_fp16)[name = tensor<string, []>("op_6243_cast_fp16")];
            tensor<int32, [3]> v_301_perm_0 = const()[name = tensor<string, []>("v_301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59144448)))];
            tensor<fp16, [1, 197, 64]> q_501_cast_fp16 = transpose(perm = q_501_perm_0, x = var_6227_cast_fp16)[name = tensor<string, []>("transpose_336")];
            tensor<fp16, [1, 197, 64]> q_503_cast_fp16 = add(x = q_501_cast_fp16, y = model_blocks_4_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_503_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59144640)))];
            tensor<fp16, [1, 197, 64]> v_301_cast_fp16 = transpose(perm = v_301_perm_0, x = var_6243_cast_fp16)[name = tensor<string, []>("transpose_334")];
            tensor<fp16, [1, 197, 64]> v_303_cast_fp16 = add(x = v_301_cast_fp16, y = model_blocks_4_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_303_cast_fp16")];
            tensor<int32, [1]> q_505_axes_0 = const()[name = tensor<string, []>("q_505_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_505_cast_fp16 = expand_dims(axes = q_505_axes_0, x = q_503_cast_fp16)[name = tensor<string, []>("q_505_cast_fp16")];
            tensor<int32, [1]> k_303_axes_0 = const()[name = tensor<string, []>("k_303_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_301_cast_fp16 = transpose(perm = k_301_perm_0, x = var_6235_cast_fp16)[name = tensor<string, []>("transpose_335")];
            tensor<fp16, [1, 1, 197, 64]> k_303_cast_fp16 = expand_dims(axes = k_303_axes_0, x = k_301_cast_fp16)[name = tensor<string, []>("k_303_cast_fp16")];
            tensor<int32, [1]> v_305_axes_0 = const()[name = tensor<string, []>("v_305_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_305_cast_fp16 = expand_dims(axes = v_305_axes_0, x = v_303_cast_fp16)[name = tensor<string, []>("v_305_cast_fp16")];
            tensor<int32, [4]> var_6252_begin_0 = const()[name = tensor<string, []>("op_6252_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6252_end_0 = const()[name = tensor<string, []>("op_6252_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6252_end_mask_0 = const()[name = tensor<string, []>("op_6252_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6252_cast_fp16 = slice_by_index(begin = var_6252_begin_0, end = var_6252_end_0, end_mask = var_6252_end_mask_0, x = q_505_cast_fp16)[name = tensor<string, []>("op_6252_cast_fp16")];
            tensor<int32, [4]> var_6256_begin_0 = const()[name = tensor<string, []>("op_6256_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6256_end_0 = const()[name = tensor<string, []>("op_6256_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6256_end_mask_0 = const()[name = tensor<string, []>("op_6256_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6256_cast_fp16 = slice_by_index(begin = var_6256_begin_0, end = var_6256_end_0, end_mask = var_6256_end_mask_0, x = k_303_cast_fp16)[name = tensor<string, []>("op_6256_cast_fp16")];
            tensor<int32, [4]> var_6260_begin_0 = const()[name = tensor<string, []>("op_6260_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6260_end_0 = const()[name = tensor<string, []>("op_6260_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6260_end_mask_0 = const()[name = tensor<string, []>("op_6260_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6260_cast_fp16 = slice_by_index(begin = var_6260_begin_0, end = var_6260_end_0, end_mask = var_6260_end_mask_0, x = q_505_cast_fp16)[name = tensor<string, []>("op_6260_cast_fp16")];
            tensor<int32, [4]> var_6264_begin_0 = const()[name = tensor<string, []>("op_6264_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6264_end_0 = const()[name = tensor<string, []>("op_6264_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6264_end_mask_0 = const()[name = tensor<string, []>("op_6264_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6264_cast_fp16 = slice_by_index(begin = var_6264_begin_0, end = var_6264_end_0, end_mask = var_6264_end_mask_0, x = k_303_cast_fp16)[name = tensor<string, []>("op_6264_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6266_cast_fp16 = mul(x = var_6260_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6266_cast_fp16")];
            tensor<int32, [5]> var_6274 = const()[name = tensor<string, []>("op_6274"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_691_cast_fp16 = reshape(shape = var_6274, x = var_6260_cast_fp16)[name = tensor<string, []>("x_691_cast_fp16")];
            tensor<int32, [2]> var_6276_split_sizes_0 = const()[name = tensor<string, []>("op_6276_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6276_axis_0 = const()[name = tensor<string, []>("op_6276_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6276_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6276_cast_fp16_1 = split(axis = var_6276_axis_0, split_sizes = var_6276_split_sizes_0, x = x_691_cast_fp16)[name = tensor<string, []>("op_6276_cast_fp16")];
            tensor<int32, [1]> squeeze_200_axes_0 = const()[name = tensor<string, []>("squeeze_200_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_200_cast_fp16 = squeeze(axes = squeeze_200_axes_0, x = var_6276_cast_fp16_0)[name = tensor<string, []>("squeeze_200_cast_fp16")];
            tensor<int32, [1]> squeeze_201_axes_0 = const()[name = tensor<string, []>("squeeze_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_201_cast_fp16 = squeeze(axes = squeeze_201_axes_0, x = var_6276_cast_fp16_1)[name = tensor<string, []>("squeeze_201_cast_fp16")];
            tensor<fp16, []> const_823_promoted_to_fp16 = const()[name = tensor<string, []>("const_823_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6279_cast_fp16 = mul(x = squeeze_201_cast_fp16, y = const_823_promoted_to_fp16)[name = tensor<string, []>("op_6279_cast_fp16")];
            tensor<int32, []> x_693_axis_0 = const()[name = tensor<string, []>("x_693_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_693_cast_fp16 = stack(axis = x_693_axis_0, values = (var_6279_cast_fp16, squeeze_200_cast_fp16))[name = tensor<string, []>("x_693_cast_fp16")];
            tensor<int32, [4]> var_6285 = const()[name = tensor<string, []>("op_6285"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6286_cast_fp16 = reshape(shape = var_6285, x = x_693_cast_fp16)[name = tensor<string, []>("op_6286_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6287_cast_fp16 = mul(x = var_6286_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6287_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_101_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6287_cast_fp16)[name = tensor<string, []>("q_patches_101_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6289_cast_fp16 = mul(x = var_6264_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6289_cast_fp16")];
            tensor<int32, [5]> var_6297 = const()[name = tensor<string, []>("op_6297"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_697_cast_fp16 = reshape(shape = var_6297, x = var_6264_cast_fp16)[name = tensor<string, []>("x_697_cast_fp16")];
            tensor<int32, [2]> var_6299_split_sizes_0 = const()[name = tensor<string, []>("op_6299_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6299_axis_0 = const()[name = tensor<string, []>("op_6299_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6299_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6299_cast_fp16_1 = split(axis = var_6299_axis_0, split_sizes = var_6299_split_sizes_0, x = x_697_cast_fp16)[name = tensor<string, []>("op_6299_cast_fp16")];
            tensor<int32, [1]> squeeze_202_axes_0 = const()[name = tensor<string, []>("squeeze_202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_202_cast_fp16 = squeeze(axes = squeeze_202_axes_0, x = var_6299_cast_fp16_0)[name = tensor<string, []>("squeeze_202_cast_fp16")];
            tensor<int32, [1]> squeeze_203_axes_0 = const()[name = tensor<string, []>("squeeze_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_203_cast_fp16 = squeeze(axes = squeeze_203_axes_0, x = var_6299_cast_fp16_1)[name = tensor<string, []>("squeeze_203_cast_fp16")];
            tensor<fp16, []> const_831_promoted_to_fp16 = const()[name = tensor<string, []>("const_831_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6302_cast_fp16 = mul(x = squeeze_203_cast_fp16, y = const_831_promoted_to_fp16)[name = tensor<string, []>("op_6302_cast_fp16")];
            tensor<int32, []> x_699_axis_0 = const()[name = tensor<string, []>("x_699_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_699_cast_fp16 = stack(axis = x_699_axis_0, values = (var_6302_cast_fp16, squeeze_202_cast_fp16))[name = tensor<string, []>("x_699_cast_fp16")];
            tensor<int32, [4]> var_6308 = const()[name = tensor<string, []>("op_6308"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6309_cast_fp16 = reshape(shape = var_6308, x = x_699_cast_fp16)[name = tensor<string, []>("op_6309_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6310_cast_fp16 = mul(x = var_6309_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6310_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_101_cast_fp16 = add(x = var_6289_cast_fp16, y = var_6310_cast_fp16)[name = tensor<string, []>("k_patches_101_cast_fp16")];
            tensor<bool, []> var_6313_interleave_0 = const()[name = tensor<string, []>("op_6313_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6313_cast_fp16 = concat(axis = var_20, interleave = var_6313_interleave_0, values = (var_6252_cast_fp16, q_patches_101_cast_fp16))[name = tensor<string, []>("op_6313_cast_fp16")];
            tensor<bool, []> var_6316_interleave_0 = const()[name = tensor<string, []>("op_6316_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6316_cast_fp16 = concat(axis = var_20, interleave = var_6316_interleave_0, values = (var_6256_cast_fp16, k_patches_101_cast_fp16))[name = tensor<string, []>("op_6316_cast_fp16")];
            tensor<fp16, []> var_6318_to_fp16 = const()[name = tensor<string, []>("op_6318_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_509_cast_fp16 = mul(x = var_6313_cast_fp16, y = var_6318_to_fp16)[name = tensor<string, []>("q_509_cast_fp16")];
            tensor<bool, []> attn_201_transpose_x_1 = const()[name = tensor<string, []>("attn_201_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_201_transpose_y_1 = const()[name = tensor<string, []>("attn_201_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_201_cast_fp16 = matmul(transpose_x = attn_201_transpose_x_1, transpose_y = attn_201_transpose_y_1, x = q_509_cast_fp16, y = var_6316_cast_fp16)[name = tensor<string, []>("attn_201_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_203_cast_fp16 = softmax(axis = var_21, x = attn_201_cast_fp16)[name = tensor<string, []>("attn_203_cast_fp16")];
            tensor<bool, []> var_6323_transpose_x_0 = const()[name = tensor<string, []>("op_6323_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6323_transpose_y_0 = const()[name = tensor<string, []>("op_6323_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6323_cast_fp16 = matmul(transpose_x = var_6323_transpose_x_0, transpose_y = var_6323_transpose_y_0, x = attn_203_cast_fp16, y = v_305_cast_fp16)[name = tensor<string, []>("op_6323_cast_fp16")];
            tensor<int32, [1]> var_6324_axes_0 = const()[name = tensor<string, []>("op_6324_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6324_cast_fp16 = squeeze(axes = var_6324_axes_0, x = var_6323_cast_fp16)[name = tensor<string, []>("op_6324_cast_fp16")];
            tensor<string, []> var_6330_pad_type_0 = const()[name = tensor<string, []>("op_6330_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6330_strides_0 = const()[name = tensor<string, []>("op_6330_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6330_pad_0 = const()[name = tensor<string, []>("op_6330_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6330_dilations_0 = const()[name = tensor<string, []>("op_6330_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6330_groups_0 = const()[name = tensor<string, []>("op_6330_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59144832)))];
            tensor<fp16, [1, 64, 1, 197]> var_6330_cast_fp16 = conv(dilations = var_6330_dilations_0, groups = var_6330_groups_0, pad = var_6330_pad_0, pad_type = var_6330_pad_type_0, strides = var_6330_strides_0, weight = model_blocks_4_attn_q_projs_3_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6330_cast_fp16")];
            tensor<int32, [1]> var_6331_axes_0 = const()[name = tensor<string, []>("op_6331_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6331_cast_fp16 = squeeze(axes = var_6331_axes_0, x = var_6330_cast_fp16)[name = tensor<string, []>("op_6331_cast_fp16")];
            tensor<int32, [3]> q_511_perm_0 = const()[name = tensor<string, []>("q_511_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6338_pad_type_0 = const()[name = tensor<string, []>("op_6338_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6338_strides_0 = const()[name = tensor<string, []>("op_6338_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6338_pad_0 = const()[name = tensor<string, []>("op_6338_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6338_dilations_0 = const()[name = tensor<string, []>("op_6338_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6338_groups_0 = const()[name = tensor<string, []>("op_6338_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59243200)))];
            tensor<fp16, [1, 64, 1, 197]> var_6338_cast_fp16 = conv(dilations = var_6338_dilations_0, groups = var_6338_groups_0, pad = var_6338_pad_0, pad_type = var_6338_pad_type_0, strides = var_6338_strides_0, weight = model_blocks_4_attn_k_projs_3_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6338_cast_fp16")];
            tensor<int32, [1]> var_6339_axes_0 = const()[name = tensor<string, []>("op_6339_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6339_cast_fp16 = squeeze(axes = var_6339_axes_0, x = var_6338_cast_fp16)[name = tensor<string, []>("op_6339_cast_fp16")];
            tensor<int32, [3]> k_307_perm_0 = const()[name = tensor<string, []>("k_307_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6346_pad_type_0 = const()[name = tensor<string, []>("op_6346_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6346_strides_0 = const()[name = tensor<string, []>("op_6346_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6346_pad_0 = const()[name = tensor<string, []>("op_6346_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6346_dilations_0 = const()[name = tensor<string, []>("op_6346_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6346_groups_0 = const()[name = tensor<string, []>("op_6346_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59341568)))];
            tensor<fp16, [1, 64, 1, 197]> var_6346_cast_fp16 = conv(dilations = var_6346_dilations_0, groups = var_6346_groups_0, pad = var_6346_pad_0, pad_type = var_6346_pad_type_0, strides = var_6346_strides_0, weight = model_blocks_4_attn_v_projs_3_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6346_cast_fp16")];
            tensor<int32, [1]> var_6347_axes_0 = const()[name = tensor<string, []>("op_6347_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6347_cast_fp16 = squeeze(axes = var_6347_axes_0, x = var_6346_cast_fp16)[name = tensor<string, []>("op_6347_cast_fp16")];
            tensor<int32, [3]> v_307_perm_0 = const()[name = tensor<string, []>("v_307_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59439936)))];
            tensor<fp16, [1, 197, 64]> q_511_cast_fp16 = transpose(perm = q_511_perm_0, x = var_6331_cast_fp16)[name = tensor<string, []>("transpose_333")];
            tensor<fp16, [1, 197, 64]> q_513_cast_fp16 = add(x = q_511_cast_fp16, y = model_blocks_4_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_513_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59440128)))];
            tensor<fp16, [1, 197, 64]> v_307_cast_fp16 = transpose(perm = v_307_perm_0, x = var_6347_cast_fp16)[name = tensor<string, []>("transpose_331")];
            tensor<fp16, [1, 197, 64]> v_309_cast_fp16 = add(x = v_307_cast_fp16, y = model_blocks_4_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_309_cast_fp16")];
            tensor<int32, [1]> q_515_axes_0 = const()[name = tensor<string, []>("q_515_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_515_cast_fp16 = expand_dims(axes = q_515_axes_0, x = q_513_cast_fp16)[name = tensor<string, []>("q_515_cast_fp16")];
            tensor<int32, [1]> k_309_axes_0 = const()[name = tensor<string, []>("k_309_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_307_cast_fp16 = transpose(perm = k_307_perm_0, x = var_6339_cast_fp16)[name = tensor<string, []>("transpose_332")];
            tensor<fp16, [1, 1, 197, 64]> k_309_cast_fp16 = expand_dims(axes = k_309_axes_0, x = k_307_cast_fp16)[name = tensor<string, []>("k_309_cast_fp16")];
            tensor<int32, [1]> v_311_axes_0 = const()[name = tensor<string, []>("v_311_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_311_cast_fp16 = expand_dims(axes = v_311_axes_0, x = v_309_cast_fp16)[name = tensor<string, []>("v_311_cast_fp16")];
            tensor<int32, [4]> var_6356_begin_0 = const()[name = tensor<string, []>("op_6356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6356_end_0 = const()[name = tensor<string, []>("op_6356_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6356_end_mask_0 = const()[name = tensor<string, []>("op_6356_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6356_cast_fp16 = slice_by_index(begin = var_6356_begin_0, end = var_6356_end_0, end_mask = var_6356_end_mask_0, x = q_515_cast_fp16)[name = tensor<string, []>("op_6356_cast_fp16")];
            tensor<int32, [4]> var_6360_begin_0 = const()[name = tensor<string, []>("op_6360_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6360_end_0 = const()[name = tensor<string, []>("op_6360_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6360_end_mask_0 = const()[name = tensor<string, []>("op_6360_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6360_cast_fp16 = slice_by_index(begin = var_6360_begin_0, end = var_6360_end_0, end_mask = var_6360_end_mask_0, x = k_309_cast_fp16)[name = tensor<string, []>("op_6360_cast_fp16")];
            tensor<int32, [4]> var_6364_begin_0 = const()[name = tensor<string, []>("op_6364_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6364_end_0 = const()[name = tensor<string, []>("op_6364_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6364_end_mask_0 = const()[name = tensor<string, []>("op_6364_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6364_cast_fp16 = slice_by_index(begin = var_6364_begin_0, end = var_6364_end_0, end_mask = var_6364_end_mask_0, x = q_515_cast_fp16)[name = tensor<string, []>("op_6364_cast_fp16")];
            tensor<int32, [4]> var_6368_begin_0 = const()[name = tensor<string, []>("op_6368_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6368_end_0 = const()[name = tensor<string, []>("op_6368_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6368_end_mask_0 = const()[name = tensor<string, []>("op_6368_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6368_cast_fp16 = slice_by_index(begin = var_6368_begin_0, end = var_6368_end_0, end_mask = var_6368_end_mask_0, x = k_309_cast_fp16)[name = tensor<string, []>("op_6368_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6370_cast_fp16 = mul(x = var_6364_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6370_cast_fp16")];
            tensor<int32, [5]> var_6378 = const()[name = tensor<string, []>("op_6378"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_703_cast_fp16 = reshape(shape = var_6378, x = var_6364_cast_fp16)[name = tensor<string, []>("x_703_cast_fp16")];
            tensor<int32, [2]> var_6380_split_sizes_0 = const()[name = tensor<string, []>("op_6380_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6380_axis_0 = const()[name = tensor<string, []>("op_6380_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6380_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6380_cast_fp16_1 = split(axis = var_6380_axis_0, split_sizes = var_6380_split_sizes_0, x = x_703_cast_fp16)[name = tensor<string, []>("op_6380_cast_fp16")];
            tensor<int32, [1]> squeeze_204_axes_0 = const()[name = tensor<string, []>("squeeze_204_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_204_cast_fp16 = squeeze(axes = squeeze_204_axes_0, x = var_6380_cast_fp16_0)[name = tensor<string, []>("squeeze_204_cast_fp16")];
            tensor<int32, [1]> squeeze_205_axes_0 = const()[name = tensor<string, []>("squeeze_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_205_cast_fp16 = squeeze(axes = squeeze_205_axes_0, x = var_6380_cast_fp16_1)[name = tensor<string, []>("squeeze_205_cast_fp16")];
            tensor<fp16, []> const_839_promoted_to_fp16 = const()[name = tensor<string, []>("const_839_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6383_cast_fp16 = mul(x = squeeze_205_cast_fp16, y = const_839_promoted_to_fp16)[name = tensor<string, []>("op_6383_cast_fp16")];
            tensor<int32, []> x_705_axis_0 = const()[name = tensor<string, []>("x_705_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_705_cast_fp16 = stack(axis = x_705_axis_0, values = (var_6383_cast_fp16, squeeze_204_cast_fp16))[name = tensor<string, []>("x_705_cast_fp16")];
            tensor<int32, [4]> var_6389 = const()[name = tensor<string, []>("op_6389"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6390_cast_fp16 = reshape(shape = var_6389, x = x_705_cast_fp16)[name = tensor<string, []>("op_6390_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6391_cast_fp16 = mul(x = var_6390_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6391_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_103_cast_fp16 = add(x = var_6370_cast_fp16, y = var_6391_cast_fp16)[name = tensor<string, []>("q_patches_103_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6393_cast_fp16 = mul(x = var_6368_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6393_cast_fp16")];
            tensor<int32, [5]> var_6401 = const()[name = tensor<string, []>("op_6401"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_709_cast_fp16 = reshape(shape = var_6401, x = var_6368_cast_fp16)[name = tensor<string, []>("x_709_cast_fp16")];
            tensor<int32, [2]> var_6403_split_sizes_0 = const()[name = tensor<string, []>("op_6403_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6403_axis_0 = const()[name = tensor<string, []>("op_6403_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6403_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6403_cast_fp16_1 = split(axis = var_6403_axis_0, split_sizes = var_6403_split_sizes_0, x = x_709_cast_fp16)[name = tensor<string, []>("op_6403_cast_fp16")];
            tensor<int32, [1]> squeeze_206_axes_0 = const()[name = tensor<string, []>("squeeze_206_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_206_cast_fp16 = squeeze(axes = squeeze_206_axes_0, x = var_6403_cast_fp16_0)[name = tensor<string, []>("squeeze_206_cast_fp16")];
            tensor<int32, [1]> squeeze_207_axes_0 = const()[name = tensor<string, []>("squeeze_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_207_cast_fp16 = squeeze(axes = squeeze_207_axes_0, x = var_6403_cast_fp16_1)[name = tensor<string, []>("squeeze_207_cast_fp16")];
            tensor<fp16, []> const_847_promoted_to_fp16 = const()[name = tensor<string, []>("const_847_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6406_cast_fp16 = mul(x = squeeze_207_cast_fp16, y = const_847_promoted_to_fp16)[name = tensor<string, []>("op_6406_cast_fp16")];
            tensor<int32, []> x_711_axis_0 = const()[name = tensor<string, []>("x_711_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_711_cast_fp16 = stack(axis = x_711_axis_0, values = (var_6406_cast_fp16, squeeze_206_cast_fp16))[name = tensor<string, []>("x_711_cast_fp16")];
            tensor<int32, [4]> var_6412 = const()[name = tensor<string, []>("op_6412"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6413_cast_fp16 = reshape(shape = var_6412, x = x_711_cast_fp16)[name = tensor<string, []>("op_6413_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6414_cast_fp16 = mul(x = var_6413_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6414_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_103_cast_fp16 = add(x = var_6393_cast_fp16, y = var_6414_cast_fp16)[name = tensor<string, []>("k_patches_103_cast_fp16")];
            tensor<bool, []> var_6417_interleave_0 = const()[name = tensor<string, []>("op_6417_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6417_cast_fp16 = concat(axis = var_20, interleave = var_6417_interleave_0, values = (var_6356_cast_fp16, q_patches_103_cast_fp16))[name = tensor<string, []>("op_6417_cast_fp16")];
            tensor<bool, []> var_6420_interleave_0 = const()[name = tensor<string, []>("op_6420_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6420_cast_fp16 = concat(axis = var_20, interleave = var_6420_interleave_0, values = (var_6360_cast_fp16, k_patches_103_cast_fp16))[name = tensor<string, []>("op_6420_cast_fp16")];
            tensor<fp16, []> var_6422_to_fp16 = const()[name = tensor<string, []>("op_6422_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_519_cast_fp16 = mul(x = var_6417_cast_fp16, y = var_6422_to_fp16)[name = tensor<string, []>("q_519_cast_fp16")];
            tensor<bool, []> attn_205_transpose_x_1 = const()[name = tensor<string, []>("attn_205_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_205_transpose_y_1 = const()[name = tensor<string, []>("attn_205_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_205_cast_fp16 = matmul(transpose_x = attn_205_transpose_x_1, transpose_y = attn_205_transpose_y_1, x = q_519_cast_fp16, y = var_6420_cast_fp16)[name = tensor<string, []>("attn_205_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_207_cast_fp16 = softmax(axis = var_21, x = attn_205_cast_fp16)[name = tensor<string, []>("attn_207_cast_fp16")];
            tensor<bool, []> var_6427_transpose_x_0 = const()[name = tensor<string, []>("op_6427_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6427_transpose_y_0 = const()[name = tensor<string, []>("op_6427_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6427_cast_fp16 = matmul(transpose_x = var_6427_transpose_x_0, transpose_y = var_6427_transpose_y_0, x = attn_207_cast_fp16, y = v_311_cast_fp16)[name = tensor<string, []>("op_6427_cast_fp16")];
            tensor<int32, [1]> var_6428_axes_0 = const()[name = tensor<string, []>("op_6428_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6428_cast_fp16 = squeeze(axes = var_6428_axes_0, x = var_6427_cast_fp16)[name = tensor<string, []>("op_6428_cast_fp16")];
            tensor<string, []> var_6434_pad_type_0 = const()[name = tensor<string, []>("op_6434_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6434_strides_0 = const()[name = tensor<string, []>("op_6434_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6434_pad_0 = const()[name = tensor<string, []>("op_6434_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6434_dilations_0 = const()[name = tensor<string, []>("op_6434_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6434_groups_0 = const()[name = tensor<string, []>("op_6434_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59440320)))];
            tensor<fp16, [1, 64, 1, 197]> var_6434_cast_fp16 = conv(dilations = var_6434_dilations_0, groups = var_6434_groups_0, pad = var_6434_pad_0, pad_type = var_6434_pad_type_0, strides = var_6434_strides_0, weight = model_blocks_4_attn_q_projs_4_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6434_cast_fp16")];
            tensor<int32, [1]> var_6435_axes_0 = const()[name = tensor<string, []>("op_6435_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6435_cast_fp16 = squeeze(axes = var_6435_axes_0, x = var_6434_cast_fp16)[name = tensor<string, []>("op_6435_cast_fp16")];
            tensor<int32, [3]> q_521_perm_0 = const()[name = tensor<string, []>("q_521_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6442_pad_type_0 = const()[name = tensor<string, []>("op_6442_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6442_strides_0 = const()[name = tensor<string, []>("op_6442_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6442_pad_0 = const()[name = tensor<string, []>("op_6442_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6442_dilations_0 = const()[name = tensor<string, []>("op_6442_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6442_groups_0 = const()[name = tensor<string, []>("op_6442_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59538688)))];
            tensor<fp16, [1, 64, 1, 197]> var_6442_cast_fp16 = conv(dilations = var_6442_dilations_0, groups = var_6442_groups_0, pad = var_6442_pad_0, pad_type = var_6442_pad_type_0, strides = var_6442_strides_0, weight = model_blocks_4_attn_k_projs_4_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6442_cast_fp16")];
            tensor<int32, [1]> var_6443_axes_0 = const()[name = tensor<string, []>("op_6443_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6443_cast_fp16 = squeeze(axes = var_6443_axes_0, x = var_6442_cast_fp16)[name = tensor<string, []>("op_6443_cast_fp16")];
            tensor<int32, [3]> k_313_perm_0 = const()[name = tensor<string, []>("k_313_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6450_pad_type_0 = const()[name = tensor<string, []>("op_6450_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6450_strides_0 = const()[name = tensor<string, []>("op_6450_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6450_pad_0 = const()[name = tensor<string, []>("op_6450_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6450_dilations_0 = const()[name = tensor<string, []>("op_6450_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6450_groups_0 = const()[name = tensor<string, []>("op_6450_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59637056)))];
            tensor<fp16, [1, 64, 1, 197]> var_6450_cast_fp16 = conv(dilations = var_6450_dilations_0, groups = var_6450_groups_0, pad = var_6450_pad_0, pad_type = var_6450_pad_type_0, strides = var_6450_strides_0, weight = model_blocks_4_attn_v_projs_4_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6450_cast_fp16")];
            tensor<int32, [1]> var_6451_axes_0 = const()[name = tensor<string, []>("op_6451_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6451_cast_fp16 = squeeze(axes = var_6451_axes_0, x = var_6450_cast_fp16)[name = tensor<string, []>("op_6451_cast_fp16")];
            tensor<int32, [3]> v_313_perm_0 = const()[name = tensor<string, []>("v_313_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59735424)))];
            tensor<fp16, [1, 197, 64]> q_521_cast_fp16 = transpose(perm = q_521_perm_0, x = var_6435_cast_fp16)[name = tensor<string, []>("transpose_330")];
            tensor<fp16, [1, 197, 64]> q_523_cast_fp16 = add(x = q_521_cast_fp16, y = model_blocks_4_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_523_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59735616)))];
            tensor<fp16, [1, 197, 64]> v_313_cast_fp16 = transpose(perm = v_313_perm_0, x = var_6451_cast_fp16)[name = tensor<string, []>("transpose_328")];
            tensor<fp16, [1, 197, 64]> v_315_cast_fp16 = add(x = v_313_cast_fp16, y = model_blocks_4_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_315_cast_fp16")];
            tensor<int32, [1]> q_525_axes_0 = const()[name = tensor<string, []>("q_525_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_525_cast_fp16 = expand_dims(axes = q_525_axes_0, x = q_523_cast_fp16)[name = tensor<string, []>("q_525_cast_fp16")];
            tensor<int32, [1]> k_315_axes_0 = const()[name = tensor<string, []>("k_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_313_cast_fp16 = transpose(perm = k_313_perm_0, x = var_6443_cast_fp16)[name = tensor<string, []>("transpose_329")];
            tensor<fp16, [1, 1, 197, 64]> k_315_cast_fp16 = expand_dims(axes = k_315_axes_0, x = k_313_cast_fp16)[name = tensor<string, []>("k_315_cast_fp16")];
            tensor<int32, [1]> v_317_axes_0 = const()[name = tensor<string, []>("v_317_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_317_cast_fp16 = expand_dims(axes = v_317_axes_0, x = v_315_cast_fp16)[name = tensor<string, []>("v_317_cast_fp16")];
            tensor<int32, [4]> var_6460_begin_0 = const()[name = tensor<string, []>("op_6460_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6460_end_0 = const()[name = tensor<string, []>("op_6460_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6460_end_mask_0 = const()[name = tensor<string, []>("op_6460_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6460_cast_fp16 = slice_by_index(begin = var_6460_begin_0, end = var_6460_end_0, end_mask = var_6460_end_mask_0, x = q_525_cast_fp16)[name = tensor<string, []>("op_6460_cast_fp16")];
            tensor<int32, [4]> var_6464_begin_0 = const()[name = tensor<string, []>("op_6464_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6464_end_0 = const()[name = tensor<string, []>("op_6464_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6464_end_mask_0 = const()[name = tensor<string, []>("op_6464_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6464_cast_fp16 = slice_by_index(begin = var_6464_begin_0, end = var_6464_end_0, end_mask = var_6464_end_mask_0, x = k_315_cast_fp16)[name = tensor<string, []>("op_6464_cast_fp16")];
            tensor<int32, [4]> var_6468_begin_0 = const()[name = tensor<string, []>("op_6468_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6468_end_0 = const()[name = tensor<string, []>("op_6468_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6468_end_mask_0 = const()[name = tensor<string, []>("op_6468_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6468_cast_fp16 = slice_by_index(begin = var_6468_begin_0, end = var_6468_end_0, end_mask = var_6468_end_mask_0, x = q_525_cast_fp16)[name = tensor<string, []>("op_6468_cast_fp16")];
            tensor<int32, [4]> var_6472_begin_0 = const()[name = tensor<string, []>("op_6472_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6472_end_0 = const()[name = tensor<string, []>("op_6472_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6472_end_mask_0 = const()[name = tensor<string, []>("op_6472_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6472_cast_fp16 = slice_by_index(begin = var_6472_begin_0, end = var_6472_end_0, end_mask = var_6472_end_mask_0, x = k_315_cast_fp16)[name = tensor<string, []>("op_6472_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6474_cast_fp16 = mul(x = var_6468_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6474_cast_fp16")];
            tensor<int32, [5]> var_6482 = const()[name = tensor<string, []>("op_6482"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_715_cast_fp16 = reshape(shape = var_6482, x = var_6468_cast_fp16)[name = tensor<string, []>("x_715_cast_fp16")];
            tensor<int32, [2]> var_6484_split_sizes_0 = const()[name = tensor<string, []>("op_6484_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6484_axis_0 = const()[name = tensor<string, []>("op_6484_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6484_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6484_cast_fp16_1 = split(axis = var_6484_axis_0, split_sizes = var_6484_split_sizes_0, x = x_715_cast_fp16)[name = tensor<string, []>("op_6484_cast_fp16")];
            tensor<int32, [1]> squeeze_208_axes_0 = const()[name = tensor<string, []>("squeeze_208_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_208_cast_fp16 = squeeze(axes = squeeze_208_axes_0, x = var_6484_cast_fp16_0)[name = tensor<string, []>("squeeze_208_cast_fp16")];
            tensor<int32, [1]> squeeze_209_axes_0 = const()[name = tensor<string, []>("squeeze_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_209_cast_fp16 = squeeze(axes = squeeze_209_axes_0, x = var_6484_cast_fp16_1)[name = tensor<string, []>("squeeze_209_cast_fp16")];
            tensor<fp16, []> const_855_promoted_to_fp16 = const()[name = tensor<string, []>("const_855_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6487_cast_fp16 = mul(x = squeeze_209_cast_fp16, y = const_855_promoted_to_fp16)[name = tensor<string, []>("op_6487_cast_fp16")];
            tensor<int32, []> x_717_axis_0 = const()[name = tensor<string, []>("x_717_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_717_cast_fp16 = stack(axis = x_717_axis_0, values = (var_6487_cast_fp16, squeeze_208_cast_fp16))[name = tensor<string, []>("x_717_cast_fp16")];
            tensor<int32, [4]> var_6493 = const()[name = tensor<string, []>("op_6493"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6494_cast_fp16 = reshape(shape = var_6493, x = x_717_cast_fp16)[name = tensor<string, []>("op_6494_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6495_cast_fp16 = mul(x = var_6494_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6495_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_105_cast_fp16 = add(x = var_6474_cast_fp16, y = var_6495_cast_fp16)[name = tensor<string, []>("q_patches_105_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6497_cast_fp16 = mul(x = var_6472_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6497_cast_fp16")];
            tensor<int32, [5]> var_6505 = const()[name = tensor<string, []>("op_6505"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_721_cast_fp16 = reshape(shape = var_6505, x = var_6472_cast_fp16)[name = tensor<string, []>("x_721_cast_fp16")];
            tensor<int32, [2]> var_6507_split_sizes_0 = const()[name = tensor<string, []>("op_6507_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6507_axis_0 = const()[name = tensor<string, []>("op_6507_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6507_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6507_cast_fp16_1 = split(axis = var_6507_axis_0, split_sizes = var_6507_split_sizes_0, x = x_721_cast_fp16)[name = tensor<string, []>("op_6507_cast_fp16")];
            tensor<int32, [1]> squeeze_210_axes_0 = const()[name = tensor<string, []>("squeeze_210_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_210_cast_fp16 = squeeze(axes = squeeze_210_axes_0, x = var_6507_cast_fp16_0)[name = tensor<string, []>("squeeze_210_cast_fp16")];
            tensor<int32, [1]> squeeze_211_axes_0 = const()[name = tensor<string, []>("squeeze_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_211_cast_fp16 = squeeze(axes = squeeze_211_axes_0, x = var_6507_cast_fp16_1)[name = tensor<string, []>("squeeze_211_cast_fp16")];
            tensor<fp16, []> const_863_promoted_to_fp16 = const()[name = tensor<string, []>("const_863_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6510_cast_fp16 = mul(x = squeeze_211_cast_fp16, y = const_863_promoted_to_fp16)[name = tensor<string, []>("op_6510_cast_fp16")];
            tensor<int32, []> x_723_axis_0 = const()[name = tensor<string, []>("x_723_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_723_cast_fp16 = stack(axis = x_723_axis_0, values = (var_6510_cast_fp16, squeeze_210_cast_fp16))[name = tensor<string, []>("x_723_cast_fp16")];
            tensor<int32, [4]> var_6516 = const()[name = tensor<string, []>("op_6516"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6517_cast_fp16 = reshape(shape = var_6516, x = x_723_cast_fp16)[name = tensor<string, []>("op_6517_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6518_cast_fp16 = mul(x = var_6517_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6518_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_105_cast_fp16 = add(x = var_6497_cast_fp16, y = var_6518_cast_fp16)[name = tensor<string, []>("k_patches_105_cast_fp16")];
            tensor<bool, []> var_6521_interleave_0 = const()[name = tensor<string, []>("op_6521_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6521_cast_fp16 = concat(axis = var_20, interleave = var_6521_interleave_0, values = (var_6460_cast_fp16, q_patches_105_cast_fp16))[name = tensor<string, []>("op_6521_cast_fp16")];
            tensor<bool, []> var_6524_interleave_0 = const()[name = tensor<string, []>("op_6524_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6524_cast_fp16 = concat(axis = var_20, interleave = var_6524_interleave_0, values = (var_6464_cast_fp16, k_patches_105_cast_fp16))[name = tensor<string, []>("op_6524_cast_fp16")];
            tensor<fp16, []> var_6526_to_fp16 = const()[name = tensor<string, []>("op_6526_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_529_cast_fp16 = mul(x = var_6521_cast_fp16, y = var_6526_to_fp16)[name = tensor<string, []>("q_529_cast_fp16")];
            tensor<bool, []> attn_209_transpose_x_1 = const()[name = tensor<string, []>("attn_209_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_209_transpose_y_1 = const()[name = tensor<string, []>("attn_209_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_209_cast_fp16 = matmul(transpose_x = attn_209_transpose_x_1, transpose_y = attn_209_transpose_y_1, x = q_529_cast_fp16, y = var_6524_cast_fp16)[name = tensor<string, []>("attn_209_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_211_cast_fp16 = softmax(axis = var_21, x = attn_209_cast_fp16)[name = tensor<string, []>("attn_211_cast_fp16")];
            tensor<bool, []> var_6531_transpose_x_0 = const()[name = tensor<string, []>("op_6531_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6531_transpose_y_0 = const()[name = tensor<string, []>("op_6531_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6531_cast_fp16 = matmul(transpose_x = var_6531_transpose_x_0, transpose_y = var_6531_transpose_y_0, x = attn_211_cast_fp16, y = v_317_cast_fp16)[name = tensor<string, []>("op_6531_cast_fp16")];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = tensor<string, []>("op_6532_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6532_cast_fp16 = squeeze(axes = var_6532_axes_0, x = var_6531_cast_fp16)[name = tensor<string, []>("op_6532_cast_fp16")];
            tensor<string, []> var_6538_pad_type_0 = const()[name = tensor<string, []>("op_6538_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6538_strides_0 = const()[name = tensor<string, []>("op_6538_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6538_pad_0 = const()[name = tensor<string, []>("op_6538_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6538_dilations_0 = const()[name = tensor<string, []>("op_6538_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6538_groups_0 = const()[name = tensor<string, []>("op_6538_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59735808)))];
            tensor<fp16, [1, 64, 1, 197]> var_6538_cast_fp16 = conv(dilations = var_6538_dilations_0, groups = var_6538_groups_0, pad = var_6538_pad_0, pad_type = var_6538_pad_type_0, strides = var_6538_strides_0, weight = model_blocks_4_attn_q_projs_5_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6538_cast_fp16")];
            tensor<int32, [1]> var_6539_axes_0 = const()[name = tensor<string, []>("op_6539_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6539_cast_fp16 = squeeze(axes = var_6539_axes_0, x = var_6538_cast_fp16)[name = tensor<string, []>("op_6539_cast_fp16")];
            tensor<int32, [3]> q_531_perm_0 = const()[name = tensor<string, []>("q_531_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6546_pad_type_0 = const()[name = tensor<string, []>("op_6546_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6546_strides_0 = const()[name = tensor<string, []>("op_6546_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6546_pad_0 = const()[name = tensor<string, []>("op_6546_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6546_dilations_0 = const()[name = tensor<string, []>("op_6546_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6546_groups_0 = const()[name = tensor<string, []>("op_6546_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59834176)))];
            tensor<fp16, [1, 64, 1, 197]> var_6546_cast_fp16 = conv(dilations = var_6546_dilations_0, groups = var_6546_groups_0, pad = var_6546_pad_0, pad_type = var_6546_pad_type_0, strides = var_6546_strides_0, weight = model_blocks_4_attn_k_projs_5_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6546_cast_fp16")];
            tensor<int32, [1]> var_6547_axes_0 = const()[name = tensor<string, []>("op_6547_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6547_cast_fp16 = squeeze(axes = var_6547_axes_0, x = var_6546_cast_fp16)[name = tensor<string, []>("op_6547_cast_fp16")];
            tensor<int32, [3]> k_319_perm_0 = const()[name = tensor<string, []>("k_319_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6554_pad_type_0 = const()[name = tensor<string, []>("op_6554_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6554_strides_0 = const()[name = tensor<string, []>("op_6554_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6554_pad_0 = const()[name = tensor<string, []>("op_6554_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6554_dilations_0 = const()[name = tensor<string, []>("op_6554_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6554_groups_0 = const()[name = tensor<string, []>("op_6554_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59932544)))];
            tensor<fp16, [1, 64, 1, 197]> var_6554_cast_fp16 = conv(dilations = var_6554_dilations_0, groups = var_6554_groups_0, pad = var_6554_pad_0, pad_type = var_6554_pad_type_0, strides = var_6554_strides_0, weight = model_blocks_4_attn_v_projs_5_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6554_cast_fp16")];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = tensor<string, []>("op_6555_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6555_cast_fp16 = squeeze(axes = var_6555_axes_0, x = var_6554_cast_fp16)[name = tensor<string, []>("op_6555_cast_fp16")];
            tensor<int32, [3]> v_319_perm_0 = const()[name = tensor<string, []>("v_319_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60030912)))];
            tensor<fp16, [1, 197, 64]> q_531_cast_fp16 = transpose(perm = q_531_perm_0, x = var_6539_cast_fp16)[name = tensor<string, []>("transpose_327")];
            tensor<fp16, [1, 197, 64]> q_533_cast_fp16 = add(x = q_531_cast_fp16, y = model_blocks_4_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_533_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60031104)))];
            tensor<fp16, [1, 197, 64]> v_319_cast_fp16 = transpose(perm = v_319_perm_0, x = var_6555_cast_fp16)[name = tensor<string, []>("transpose_325")];
            tensor<fp16, [1, 197, 64]> v_321_cast_fp16 = add(x = v_319_cast_fp16, y = model_blocks_4_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_321_cast_fp16")];
            tensor<int32, [1]> q_535_axes_0 = const()[name = tensor<string, []>("q_535_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_535_cast_fp16 = expand_dims(axes = q_535_axes_0, x = q_533_cast_fp16)[name = tensor<string, []>("q_535_cast_fp16")];
            tensor<int32, [1]> k_321_axes_0 = const()[name = tensor<string, []>("k_321_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_319_cast_fp16 = transpose(perm = k_319_perm_0, x = var_6547_cast_fp16)[name = tensor<string, []>("transpose_326")];
            tensor<fp16, [1, 1, 197, 64]> k_321_cast_fp16 = expand_dims(axes = k_321_axes_0, x = k_319_cast_fp16)[name = tensor<string, []>("k_321_cast_fp16")];
            tensor<int32, [1]> v_323_axes_0 = const()[name = tensor<string, []>("v_323_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_323_cast_fp16 = expand_dims(axes = v_323_axes_0, x = v_321_cast_fp16)[name = tensor<string, []>("v_323_cast_fp16")];
            tensor<int32, [4]> var_6564_begin_0 = const()[name = tensor<string, []>("op_6564_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6564_end_0 = const()[name = tensor<string, []>("op_6564_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6564_end_mask_0 = const()[name = tensor<string, []>("op_6564_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6564_cast_fp16 = slice_by_index(begin = var_6564_begin_0, end = var_6564_end_0, end_mask = var_6564_end_mask_0, x = q_535_cast_fp16)[name = tensor<string, []>("op_6564_cast_fp16")];
            tensor<int32, [4]> var_6568_begin_0 = const()[name = tensor<string, []>("op_6568_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6568_end_0 = const()[name = tensor<string, []>("op_6568_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6568_end_mask_0 = const()[name = tensor<string, []>("op_6568_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6568_cast_fp16 = slice_by_index(begin = var_6568_begin_0, end = var_6568_end_0, end_mask = var_6568_end_mask_0, x = k_321_cast_fp16)[name = tensor<string, []>("op_6568_cast_fp16")];
            tensor<int32, [4]> var_6572_begin_0 = const()[name = tensor<string, []>("op_6572_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6572_end_0 = const()[name = tensor<string, []>("op_6572_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6572_end_mask_0 = const()[name = tensor<string, []>("op_6572_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6572_cast_fp16 = slice_by_index(begin = var_6572_begin_0, end = var_6572_end_0, end_mask = var_6572_end_mask_0, x = q_535_cast_fp16)[name = tensor<string, []>("op_6572_cast_fp16")];
            tensor<int32, [4]> var_6576_begin_0 = const()[name = tensor<string, []>("op_6576_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6576_end_0 = const()[name = tensor<string, []>("op_6576_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6576_end_mask_0 = const()[name = tensor<string, []>("op_6576_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6576_cast_fp16 = slice_by_index(begin = var_6576_begin_0, end = var_6576_end_0, end_mask = var_6576_end_mask_0, x = k_321_cast_fp16)[name = tensor<string, []>("op_6576_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6578_cast_fp16 = mul(x = var_6572_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6578_cast_fp16")];
            tensor<int32, [5]> var_6586 = const()[name = tensor<string, []>("op_6586"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_727_cast_fp16 = reshape(shape = var_6586, x = var_6572_cast_fp16)[name = tensor<string, []>("x_727_cast_fp16")];
            tensor<int32, [2]> var_6588_split_sizes_0 = const()[name = tensor<string, []>("op_6588_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6588_axis_0 = const()[name = tensor<string, []>("op_6588_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6588_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6588_cast_fp16_1 = split(axis = var_6588_axis_0, split_sizes = var_6588_split_sizes_0, x = x_727_cast_fp16)[name = tensor<string, []>("op_6588_cast_fp16")];
            tensor<int32, [1]> squeeze_212_axes_0 = const()[name = tensor<string, []>("squeeze_212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_212_cast_fp16 = squeeze(axes = squeeze_212_axes_0, x = var_6588_cast_fp16_0)[name = tensor<string, []>("squeeze_212_cast_fp16")];
            tensor<int32, [1]> squeeze_213_axes_0 = const()[name = tensor<string, []>("squeeze_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_213_cast_fp16 = squeeze(axes = squeeze_213_axes_0, x = var_6588_cast_fp16_1)[name = tensor<string, []>("squeeze_213_cast_fp16")];
            tensor<fp16, []> const_871_promoted_to_fp16 = const()[name = tensor<string, []>("const_871_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6591_cast_fp16 = mul(x = squeeze_213_cast_fp16, y = const_871_promoted_to_fp16)[name = tensor<string, []>("op_6591_cast_fp16")];
            tensor<int32, []> x_729_axis_0 = const()[name = tensor<string, []>("x_729_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_729_cast_fp16 = stack(axis = x_729_axis_0, values = (var_6591_cast_fp16, squeeze_212_cast_fp16))[name = tensor<string, []>("x_729_cast_fp16")];
            tensor<int32, [4]> var_6597 = const()[name = tensor<string, []>("op_6597"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6598_cast_fp16 = reshape(shape = var_6597, x = x_729_cast_fp16)[name = tensor<string, []>("op_6598_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6599_cast_fp16 = mul(x = var_6598_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6599_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_107_cast_fp16 = add(x = var_6578_cast_fp16, y = var_6599_cast_fp16)[name = tensor<string, []>("q_patches_107_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6601_cast_fp16 = mul(x = var_6576_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6601_cast_fp16")];
            tensor<int32, [5]> var_6609 = const()[name = tensor<string, []>("op_6609"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_733_cast_fp16 = reshape(shape = var_6609, x = var_6576_cast_fp16)[name = tensor<string, []>("x_733_cast_fp16")];
            tensor<int32, [2]> var_6611_split_sizes_0 = const()[name = tensor<string, []>("op_6611_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6611_axis_0 = const()[name = tensor<string, []>("op_6611_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6611_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6611_cast_fp16_1 = split(axis = var_6611_axis_0, split_sizes = var_6611_split_sizes_0, x = x_733_cast_fp16)[name = tensor<string, []>("op_6611_cast_fp16")];
            tensor<int32, [1]> squeeze_214_axes_0 = const()[name = tensor<string, []>("squeeze_214_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_214_cast_fp16 = squeeze(axes = squeeze_214_axes_0, x = var_6611_cast_fp16_0)[name = tensor<string, []>("squeeze_214_cast_fp16")];
            tensor<int32, [1]> squeeze_215_axes_0 = const()[name = tensor<string, []>("squeeze_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_215_cast_fp16 = squeeze(axes = squeeze_215_axes_0, x = var_6611_cast_fp16_1)[name = tensor<string, []>("squeeze_215_cast_fp16")];
            tensor<fp16, []> const_879_promoted_to_fp16 = const()[name = tensor<string, []>("const_879_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6614_cast_fp16 = mul(x = squeeze_215_cast_fp16, y = const_879_promoted_to_fp16)[name = tensor<string, []>("op_6614_cast_fp16")];
            tensor<int32, []> x_735_axis_0 = const()[name = tensor<string, []>("x_735_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_735_cast_fp16 = stack(axis = x_735_axis_0, values = (var_6614_cast_fp16, squeeze_214_cast_fp16))[name = tensor<string, []>("x_735_cast_fp16")];
            tensor<int32, [4]> var_6620 = const()[name = tensor<string, []>("op_6620"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6621_cast_fp16 = reshape(shape = var_6620, x = x_735_cast_fp16)[name = tensor<string, []>("op_6621_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6622_cast_fp16 = mul(x = var_6621_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6622_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_107_cast_fp16 = add(x = var_6601_cast_fp16, y = var_6622_cast_fp16)[name = tensor<string, []>("k_patches_107_cast_fp16")];
            tensor<bool, []> var_6625_interleave_0 = const()[name = tensor<string, []>("op_6625_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6625_cast_fp16 = concat(axis = var_20, interleave = var_6625_interleave_0, values = (var_6564_cast_fp16, q_patches_107_cast_fp16))[name = tensor<string, []>("op_6625_cast_fp16")];
            tensor<bool, []> var_6628_interleave_0 = const()[name = tensor<string, []>("op_6628_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6628_cast_fp16 = concat(axis = var_20, interleave = var_6628_interleave_0, values = (var_6568_cast_fp16, k_patches_107_cast_fp16))[name = tensor<string, []>("op_6628_cast_fp16")];
            tensor<fp16, []> var_6630_to_fp16 = const()[name = tensor<string, []>("op_6630_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_539_cast_fp16 = mul(x = var_6625_cast_fp16, y = var_6630_to_fp16)[name = tensor<string, []>("q_539_cast_fp16")];
            tensor<bool, []> attn_213_transpose_x_1 = const()[name = tensor<string, []>("attn_213_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_213_transpose_y_1 = const()[name = tensor<string, []>("attn_213_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_213_cast_fp16 = matmul(transpose_x = attn_213_transpose_x_1, transpose_y = attn_213_transpose_y_1, x = q_539_cast_fp16, y = var_6628_cast_fp16)[name = tensor<string, []>("attn_213_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_215_cast_fp16 = softmax(axis = var_21, x = attn_213_cast_fp16)[name = tensor<string, []>("attn_215_cast_fp16")];
            tensor<bool, []> var_6635_transpose_x_0 = const()[name = tensor<string, []>("op_6635_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6635_transpose_y_0 = const()[name = tensor<string, []>("op_6635_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6635_cast_fp16 = matmul(transpose_x = var_6635_transpose_x_0, transpose_y = var_6635_transpose_y_0, x = attn_215_cast_fp16, y = v_323_cast_fp16)[name = tensor<string, []>("op_6635_cast_fp16")];
            tensor<int32, [1]> var_6636_axes_0 = const()[name = tensor<string, []>("op_6636_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6636_cast_fp16 = squeeze(axes = var_6636_axes_0, x = var_6635_cast_fp16)[name = tensor<string, []>("op_6636_cast_fp16")];
            tensor<string, []> var_6642_pad_type_0 = const()[name = tensor<string, []>("op_6642_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6642_strides_0 = const()[name = tensor<string, []>("op_6642_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6642_pad_0 = const()[name = tensor<string, []>("op_6642_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6642_dilations_0 = const()[name = tensor<string, []>("op_6642_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6642_groups_0 = const()[name = tensor<string, []>("op_6642_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60031296)))];
            tensor<fp16, [1, 64, 1, 197]> var_6642_cast_fp16 = conv(dilations = var_6642_dilations_0, groups = var_6642_groups_0, pad = var_6642_pad_0, pad_type = var_6642_pad_type_0, strides = var_6642_strides_0, weight = model_blocks_4_attn_q_projs_6_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6642_cast_fp16")];
            tensor<int32, [1]> var_6643_axes_0 = const()[name = tensor<string, []>("op_6643_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6643_cast_fp16 = squeeze(axes = var_6643_axes_0, x = var_6642_cast_fp16)[name = tensor<string, []>("op_6643_cast_fp16")];
            tensor<int32, [3]> q_541_perm_0 = const()[name = tensor<string, []>("q_541_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6650_pad_type_0 = const()[name = tensor<string, []>("op_6650_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6650_strides_0 = const()[name = tensor<string, []>("op_6650_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6650_pad_0 = const()[name = tensor<string, []>("op_6650_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6650_dilations_0 = const()[name = tensor<string, []>("op_6650_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6650_groups_0 = const()[name = tensor<string, []>("op_6650_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60129664)))];
            tensor<fp16, [1, 64, 1, 197]> var_6650_cast_fp16 = conv(dilations = var_6650_dilations_0, groups = var_6650_groups_0, pad = var_6650_pad_0, pad_type = var_6650_pad_type_0, strides = var_6650_strides_0, weight = model_blocks_4_attn_k_projs_6_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6650_cast_fp16")];
            tensor<int32, [1]> var_6651_axes_0 = const()[name = tensor<string, []>("op_6651_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6651_cast_fp16 = squeeze(axes = var_6651_axes_0, x = var_6650_cast_fp16)[name = tensor<string, []>("op_6651_cast_fp16")];
            tensor<int32, [3]> k_325_perm_0 = const()[name = tensor<string, []>("k_325_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6658_pad_type_0 = const()[name = tensor<string, []>("op_6658_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6658_strides_0 = const()[name = tensor<string, []>("op_6658_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6658_pad_0 = const()[name = tensor<string, []>("op_6658_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6658_dilations_0 = const()[name = tensor<string, []>("op_6658_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6658_groups_0 = const()[name = tensor<string, []>("op_6658_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60228032)))];
            tensor<fp16, [1, 64, 1, 197]> var_6658_cast_fp16 = conv(dilations = var_6658_dilations_0, groups = var_6658_groups_0, pad = var_6658_pad_0, pad_type = var_6658_pad_type_0, strides = var_6658_strides_0, weight = model_blocks_4_attn_v_projs_6_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6658_cast_fp16")];
            tensor<int32, [1]> var_6659_axes_0 = const()[name = tensor<string, []>("op_6659_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6659_cast_fp16 = squeeze(axes = var_6659_axes_0, x = var_6658_cast_fp16)[name = tensor<string, []>("op_6659_cast_fp16")];
            tensor<int32, [3]> v_325_perm_0 = const()[name = tensor<string, []>("v_325_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60326400)))];
            tensor<fp16, [1, 197, 64]> q_541_cast_fp16 = transpose(perm = q_541_perm_0, x = var_6643_cast_fp16)[name = tensor<string, []>("transpose_324")];
            tensor<fp16, [1, 197, 64]> q_543_cast_fp16 = add(x = q_541_cast_fp16, y = model_blocks_4_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_543_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60326592)))];
            tensor<fp16, [1, 197, 64]> v_325_cast_fp16 = transpose(perm = v_325_perm_0, x = var_6659_cast_fp16)[name = tensor<string, []>("transpose_322")];
            tensor<fp16, [1, 197, 64]> v_327_cast_fp16 = add(x = v_325_cast_fp16, y = model_blocks_4_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_327_cast_fp16")];
            tensor<int32, [1]> q_545_axes_0 = const()[name = tensor<string, []>("q_545_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_545_cast_fp16 = expand_dims(axes = q_545_axes_0, x = q_543_cast_fp16)[name = tensor<string, []>("q_545_cast_fp16")];
            tensor<int32, [1]> k_327_axes_0 = const()[name = tensor<string, []>("k_327_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_325_cast_fp16 = transpose(perm = k_325_perm_0, x = var_6651_cast_fp16)[name = tensor<string, []>("transpose_323")];
            tensor<fp16, [1, 1, 197, 64]> k_327_cast_fp16 = expand_dims(axes = k_327_axes_0, x = k_325_cast_fp16)[name = tensor<string, []>("k_327_cast_fp16")];
            tensor<int32, [1]> v_329_axes_0 = const()[name = tensor<string, []>("v_329_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_329_cast_fp16 = expand_dims(axes = v_329_axes_0, x = v_327_cast_fp16)[name = tensor<string, []>("v_329_cast_fp16")];
            tensor<int32, [4]> var_6668_begin_0 = const()[name = tensor<string, []>("op_6668_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6668_end_0 = const()[name = tensor<string, []>("op_6668_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6668_end_mask_0 = const()[name = tensor<string, []>("op_6668_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6668_cast_fp16 = slice_by_index(begin = var_6668_begin_0, end = var_6668_end_0, end_mask = var_6668_end_mask_0, x = q_545_cast_fp16)[name = tensor<string, []>("op_6668_cast_fp16")];
            tensor<int32, [4]> var_6672_begin_0 = const()[name = tensor<string, []>("op_6672_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6672_end_0 = const()[name = tensor<string, []>("op_6672_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6672_end_mask_0 = const()[name = tensor<string, []>("op_6672_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6672_cast_fp16 = slice_by_index(begin = var_6672_begin_0, end = var_6672_end_0, end_mask = var_6672_end_mask_0, x = k_327_cast_fp16)[name = tensor<string, []>("op_6672_cast_fp16")];
            tensor<int32, [4]> var_6676_begin_0 = const()[name = tensor<string, []>("op_6676_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6676_end_0 = const()[name = tensor<string, []>("op_6676_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6676_end_mask_0 = const()[name = tensor<string, []>("op_6676_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6676_cast_fp16 = slice_by_index(begin = var_6676_begin_0, end = var_6676_end_0, end_mask = var_6676_end_mask_0, x = q_545_cast_fp16)[name = tensor<string, []>("op_6676_cast_fp16")];
            tensor<int32, [4]> var_6680_begin_0 = const()[name = tensor<string, []>("op_6680_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6680_end_0 = const()[name = tensor<string, []>("op_6680_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6680_end_mask_0 = const()[name = tensor<string, []>("op_6680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6680_cast_fp16 = slice_by_index(begin = var_6680_begin_0, end = var_6680_end_0, end_mask = var_6680_end_mask_0, x = k_327_cast_fp16)[name = tensor<string, []>("op_6680_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6682_cast_fp16 = mul(x = var_6676_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6682_cast_fp16")];
            tensor<int32, [5]> var_6690 = const()[name = tensor<string, []>("op_6690"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_739_cast_fp16 = reshape(shape = var_6690, x = var_6676_cast_fp16)[name = tensor<string, []>("x_739_cast_fp16")];
            tensor<int32, [2]> var_6692_split_sizes_0 = const()[name = tensor<string, []>("op_6692_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6692_axis_0 = const()[name = tensor<string, []>("op_6692_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6692_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6692_cast_fp16_1 = split(axis = var_6692_axis_0, split_sizes = var_6692_split_sizes_0, x = x_739_cast_fp16)[name = tensor<string, []>("op_6692_cast_fp16")];
            tensor<int32, [1]> squeeze_216_axes_0 = const()[name = tensor<string, []>("squeeze_216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_216_cast_fp16 = squeeze(axes = squeeze_216_axes_0, x = var_6692_cast_fp16_0)[name = tensor<string, []>("squeeze_216_cast_fp16")];
            tensor<int32, [1]> squeeze_217_axes_0 = const()[name = tensor<string, []>("squeeze_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_217_cast_fp16 = squeeze(axes = squeeze_217_axes_0, x = var_6692_cast_fp16_1)[name = tensor<string, []>("squeeze_217_cast_fp16")];
            tensor<fp16, []> const_887_promoted_to_fp16 = const()[name = tensor<string, []>("const_887_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6695_cast_fp16 = mul(x = squeeze_217_cast_fp16, y = const_887_promoted_to_fp16)[name = tensor<string, []>("op_6695_cast_fp16")];
            tensor<int32, []> x_741_axis_0 = const()[name = tensor<string, []>("x_741_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_741_cast_fp16 = stack(axis = x_741_axis_0, values = (var_6695_cast_fp16, squeeze_216_cast_fp16))[name = tensor<string, []>("x_741_cast_fp16")];
            tensor<int32, [4]> var_6701 = const()[name = tensor<string, []>("op_6701"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6702_cast_fp16 = reshape(shape = var_6701, x = x_741_cast_fp16)[name = tensor<string, []>("op_6702_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6703_cast_fp16 = mul(x = var_6702_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6703_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_109_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6703_cast_fp16)[name = tensor<string, []>("q_patches_109_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6705_cast_fp16 = mul(x = var_6680_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6705_cast_fp16")];
            tensor<int32, [5]> var_6713 = const()[name = tensor<string, []>("op_6713"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_745_cast_fp16 = reshape(shape = var_6713, x = var_6680_cast_fp16)[name = tensor<string, []>("x_745_cast_fp16")];
            tensor<int32, [2]> var_6715_split_sizes_0 = const()[name = tensor<string, []>("op_6715_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6715_axis_0 = const()[name = tensor<string, []>("op_6715_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6715_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6715_cast_fp16_1 = split(axis = var_6715_axis_0, split_sizes = var_6715_split_sizes_0, x = x_745_cast_fp16)[name = tensor<string, []>("op_6715_cast_fp16")];
            tensor<int32, [1]> squeeze_218_axes_0 = const()[name = tensor<string, []>("squeeze_218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_218_cast_fp16 = squeeze(axes = squeeze_218_axes_0, x = var_6715_cast_fp16_0)[name = tensor<string, []>("squeeze_218_cast_fp16")];
            tensor<int32, [1]> squeeze_219_axes_0 = const()[name = tensor<string, []>("squeeze_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_219_cast_fp16 = squeeze(axes = squeeze_219_axes_0, x = var_6715_cast_fp16_1)[name = tensor<string, []>("squeeze_219_cast_fp16")];
            tensor<fp16, []> const_895_promoted_to_fp16 = const()[name = tensor<string, []>("const_895_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6718_cast_fp16 = mul(x = squeeze_219_cast_fp16, y = const_895_promoted_to_fp16)[name = tensor<string, []>("op_6718_cast_fp16")];
            tensor<int32, []> x_747_axis_0 = const()[name = tensor<string, []>("x_747_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_747_cast_fp16 = stack(axis = x_747_axis_0, values = (var_6718_cast_fp16, squeeze_218_cast_fp16))[name = tensor<string, []>("x_747_cast_fp16")];
            tensor<int32, [4]> var_6724 = const()[name = tensor<string, []>("op_6724"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6725_cast_fp16 = reshape(shape = var_6724, x = x_747_cast_fp16)[name = tensor<string, []>("op_6725_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6726_cast_fp16 = mul(x = var_6725_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6726_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_109_cast_fp16 = add(x = var_6705_cast_fp16, y = var_6726_cast_fp16)[name = tensor<string, []>("k_patches_109_cast_fp16")];
            tensor<bool, []> var_6729_interleave_0 = const()[name = tensor<string, []>("op_6729_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6729_cast_fp16 = concat(axis = var_20, interleave = var_6729_interleave_0, values = (var_6668_cast_fp16, q_patches_109_cast_fp16))[name = tensor<string, []>("op_6729_cast_fp16")];
            tensor<bool, []> var_6732_interleave_0 = const()[name = tensor<string, []>("op_6732_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6732_cast_fp16 = concat(axis = var_20, interleave = var_6732_interleave_0, values = (var_6672_cast_fp16, k_patches_109_cast_fp16))[name = tensor<string, []>("op_6732_cast_fp16")];
            tensor<fp16, []> var_6734_to_fp16 = const()[name = tensor<string, []>("op_6734_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_549_cast_fp16 = mul(x = var_6729_cast_fp16, y = var_6734_to_fp16)[name = tensor<string, []>("q_549_cast_fp16")];
            tensor<bool, []> attn_217_transpose_x_1 = const()[name = tensor<string, []>("attn_217_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_217_transpose_y_1 = const()[name = tensor<string, []>("attn_217_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_217_cast_fp16 = matmul(transpose_x = attn_217_transpose_x_1, transpose_y = attn_217_transpose_y_1, x = q_549_cast_fp16, y = var_6732_cast_fp16)[name = tensor<string, []>("attn_217_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_219_cast_fp16 = softmax(axis = var_21, x = attn_217_cast_fp16)[name = tensor<string, []>("attn_219_cast_fp16")];
            tensor<bool, []> var_6739_transpose_x_0 = const()[name = tensor<string, []>("op_6739_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6739_transpose_y_0 = const()[name = tensor<string, []>("op_6739_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6739_cast_fp16 = matmul(transpose_x = var_6739_transpose_x_0, transpose_y = var_6739_transpose_y_0, x = attn_219_cast_fp16, y = v_329_cast_fp16)[name = tensor<string, []>("op_6739_cast_fp16")];
            tensor<int32, [1]> var_6740_axes_0 = const()[name = tensor<string, []>("op_6740_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6740_cast_fp16 = squeeze(axes = var_6740_axes_0, x = var_6739_cast_fp16)[name = tensor<string, []>("op_6740_cast_fp16")];
            tensor<string, []> var_6746_pad_type_0 = const()[name = tensor<string, []>("op_6746_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6746_strides_0 = const()[name = tensor<string, []>("op_6746_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6746_pad_0 = const()[name = tensor<string, []>("op_6746_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6746_dilations_0 = const()[name = tensor<string, []>("op_6746_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6746_groups_0 = const()[name = tensor<string, []>("op_6746_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60326784)))];
            tensor<fp16, [1, 64, 1, 197]> var_6746_cast_fp16 = conv(dilations = var_6746_dilations_0, groups = var_6746_groups_0, pad = var_6746_pad_0, pad_type = var_6746_pad_type_0, strides = var_6746_strides_0, weight = model_blocks_4_attn_q_projs_7_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6746_cast_fp16")];
            tensor<int32, [1]> var_6747_axes_0 = const()[name = tensor<string, []>("op_6747_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6747_cast_fp16 = squeeze(axes = var_6747_axes_0, x = var_6746_cast_fp16)[name = tensor<string, []>("op_6747_cast_fp16")];
            tensor<int32, [3]> q_551_perm_0 = const()[name = tensor<string, []>("q_551_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6754_pad_type_0 = const()[name = tensor<string, []>("op_6754_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6754_strides_0 = const()[name = tensor<string, []>("op_6754_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6754_pad_0 = const()[name = tensor<string, []>("op_6754_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6754_dilations_0 = const()[name = tensor<string, []>("op_6754_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6754_groups_0 = const()[name = tensor<string, []>("op_6754_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60425152)))];
            tensor<fp16, [1, 64, 1, 197]> var_6754_cast_fp16 = conv(dilations = var_6754_dilations_0, groups = var_6754_groups_0, pad = var_6754_pad_0, pad_type = var_6754_pad_type_0, strides = var_6754_strides_0, weight = model_blocks_4_attn_k_projs_7_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6754_cast_fp16")];
            tensor<int32, [1]> var_6755_axes_0 = const()[name = tensor<string, []>("op_6755_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6755_cast_fp16 = squeeze(axes = var_6755_axes_0, x = var_6754_cast_fp16)[name = tensor<string, []>("op_6755_cast_fp16")];
            tensor<int32, [3]> k_331_perm_0 = const()[name = tensor<string, []>("k_331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6762_pad_type_0 = const()[name = tensor<string, []>("op_6762_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6762_strides_0 = const()[name = tensor<string, []>("op_6762_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6762_pad_0 = const()[name = tensor<string, []>("op_6762_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6762_dilations_0 = const()[name = tensor<string, []>("op_6762_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6762_groups_0 = const()[name = tensor<string, []>("op_6762_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60523520)))];
            tensor<fp16, [1, 64, 1, 197]> var_6762_cast_fp16 = conv(dilations = var_6762_dilations_0, groups = var_6762_groups_0, pad = var_6762_pad_0, pad_type = var_6762_pad_type_0, strides = var_6762_strides_0, weight = model_blocks_4_attn_v_projs_7_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6762_cast_fp16")];
            tensor<int32, [1]> var_6763_axes_0 = const()[name = tensor<string, []>("op_6763_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6763_cast_fp16 = squeeze(axes = var_6763_axes_0, x = var_6762_cast_fp16)[name = tensor<string, []>("op_6763_cast_fp16")];
            tensor<int32, [3]> v_331_perm_0 = const()[name = tensor<string, []>("v_331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60621888)))];
            tensor<fp16, [1, 197, 64]> q_551_cast_fp16 = transpose(perm = q_551_perm_0, x = var_6747_cast_fp16)[name = tensor<string, []>("transpose_321")];
            tensor<fp16, [1, 197, 64]> q_553_cast_fp16 = add(x = q_551_cast_fp16, y = model_blocks_4_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_553_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60622080)))];
            tensor<fp16, [1, 197, 64]> v_331_cast_fp16 = transpose(perm = v_331_perm_0, x = var_6763_cast_fp16)[name = tensor<string, []>("transpose_319")];
            tensor<fp16, [1, 197, 64]> v_333_cast_fp16 = add(x = v_331_cast_fp16, y = model_blocks_4_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_333_cast_fp16")];
            tensor<int32, [1]> q_555_axes_0 = const()[name = tensor<string, []>("q_555_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_555_cast_fp16 = expand_dims(axes = q_555_axes_0, x = q_553_cast_fp16)[name = tensor<string, []>("q_555_cast_fp16")];
            tensor<int32, [1]> k_333_axes_0 = const()[name = tensor<string, []>("k_333_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_331_cast_fp16 = transpose(perm = k_331_perm_0, x = var_6755_cast_fp16)[name = tensor<string, []>("transpose_320")];
            tensor<fp16, [1, 1, 197, 64]> k_333_cast_fp16 = expand_dims(axes = k_333_axes_0, x = k_331_cast_fp16)[name = tensor<string, []>("k_333_cast_fp16")];
            tensor<int32, [1]> v_335_axes_0 = const()[name = tensor<string, []>("v_335_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_335_cast_fp16 = expand_dims(axes = v_335_axes_0, x = v_333_cast_fp16)[name = tensor<string, []>("v_335_cast_fp16")];
            tensor<int32, [4]> var_6772_begin_0 = const()[name = tensor<string, []>("op_6772_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6772_end_0 = const()[name = tensor<string, []>("op_6772_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6772_end_mask_0 = const()[name = tensor<string, []>("op_6772_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6772_cast_fp16 = slice_by_index(begin = var_6772_begin_0, end = var_6772_end_0, end_mask = var_6772_end_mask_0, x = q_555_cast_fp16)[name = tensor<string, []>("op_6772_cast_fp16")];
            tensor<int32, [4]> var_6776_begin_0 = const()[name = tensor<string, []>("op_6776_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6776_end_0 = const()[name = tensor<string, []>("op_6776_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6776_end_mask_0 = const()[name = tensor<string, []>("op_6776_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6776_cast_fp16 = slice_by_index(begin = var_6776_begin_0, end = var_6776_end_0, end_mask = var_6776_end_mask_0, x = k_333_cast_fp16)[name = tensor<string, []>("op_6776_cast_fp16")];
            tensor<int32, [4]> var_6780_begin_0 = const()[name = tensor<string, []>("op_6780_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6780_end_0 = const()[name = tensor<string, []>("op_6780_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6780_end_mask_0 = const()[name = tensor<string, []>("op_6780_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6780_cast_fp16 = slice_by_index(begin = var_6780_begin_0, end = var_6780_end_0, end_mask = var_6780_end_mask_0, x = q_555_cast_fp16)[name = tensor<string, []>("op_6780_cast_fp16")];
            tensor<int32, [4]> var_6784_begin_0 = const()[name = tensor<string, []>("op_6784_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6784_end_0 = const()[name = tensor<string, []>("op_6784_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6784_end_mask_0 = const()[name = tensor<string, []>("op_6784_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6784_cast_fp16 = slice_by_index(begin = var_6784_begin_0, end = var_6784_end_0, end_mask = var_6784_end_mask_0, x = k_333_cast_fp16)[name = tensor<string, []>("op_6784_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6786_cast_fp16 = mul(x = var_6780_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6786_cast_fp16")];
            tensor<int32, [5]> var_6794 = const()[name = tensor<string, []>("op_6794"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_751_cast_fp16 = reshape(shape = var_6794, x = var_6780_cast_fp16)[name = tensor<string, []>("x_751_cast_fp16")];
            tensor<int32, [2]> var_6796_split_sizes_0 = const()[name = tensor<string, []>("op_6796_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6796_axis_0 = const()[name = tensor<string, []>("op_6796_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6796_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6796_cast_fp16_1 = split(axis = var_6796_axis_0, split_sizes = var_6796_split_sizes_0, x = x_751_cast_fp16)[name = tensor<string, []>("op_6796_cast_fp16")];
            tensor<int32, [1]> squeeze_220_axes_0 = const()[name = tensor<string, []>("squeeze_220_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_220_cast_fp16 = squeeze(axes = squeeze_220_axes_0, x = var_6796_cast_fp16_0)[name = tensor<string, []>("squeeze_220_cast_fp16")];
            tensor<int32, [1]> squeeze_221_axes_0 = const()[name = tensor<string, []>("squeeze_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_221_cast_fp16 = squeeze(axes = squeeze_221_axes_0, x = var_6796_cast_fp16_1)[name = tensor<string, []>("squeeze_221_cast_fp16")];
            tensor<fp16, []> const_903_promoted_to_fp16 = const()[name = tensor<string, []>("const_903_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6799_cast_fp16 = mul(x = squeeze_221_cast_fp16, y = const_903_promoted_to_fp16)[name = tensor<string, []>("op_6799_cast_fp16")];
            tensor<int32, []> x_753_axis_0 = const()[name = tensor<string, []>("x_753_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_753_cast_fp16 = stack(axis = x_753_axis_0, values = (var_6799_cast_fp16, squeeze_220_cast_fp16))[name = tensor<string, []>("x_753_cast_fp16")];
            tensor<int32, [4]> var_6805 = const()[name = tensor<string, []>("op_6805"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6806_cast_fp16 = reshape(shape = var_6805, x = x_753_cast_fp16)[name = tensor<string, []>("op_6806_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6807_cast_fp16 = mul(x = var_6806_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6807_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_111_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6807_cast_fp16)[name = tensor<string, []>("q_patches_111_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6809_cast_fp16 = mul(x = var_6784_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6809_cast_fp16")];
            tensor<int32, [5]> var_6817 = const()[name = tensor<string, []>("op_6817"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_757_cast_fp16 = reshape(shape = var_6817, x = var_6784_cast_fp16)[name = tensor<string, []>("x_757_cast_fp16")];
            tensor<int32, [2]> var_6819_split_sizes_0 = const()[name = tensor<string, []>("op_6819_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6819_axis_0 = const()[name = tensor<string, []>("op_6819_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6819_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6819_cast_fp16_1 = split(axis = var_6819_axis_0, split_sizes = var_6819_split_sizes_0, x = x_757_cast_fp16)[name = tensor<string, []>("op_6819_cast_fp16")];
            tensor<int32, [1]> squeeze_222_axes_0 = const()[name = tensor<string, []>("squeeze_222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_222_cast_fp16 = squeeze(axes = squeeze_222_axes_0, x = var_6819_cast_fp16_0)[name = tensor<string, []>("squeeze_222_cast_fp16")];
            tensor<int32, [1]> squeeze_223_axes_0 = const()[name = tensor<string, []>("squeeze_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_223_cast_fp16 = squeeze(axes = squeeze_223_axes_0, x = var_6819_cast_fp16_1)[name = tensor<string, []>("squeeze_223_cast_fp16")];
            tensor<fp16, []> const_911_promoted_to_fp16 = const()[name = tensor<string, []>("const_911_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6822_cast_fp16 = mul(x = squeeze_223_cast_fp16, y = const_911_promoted_to_fp16)[name = tensor<string, []>("op_6822_cast_fp16")];
            tensor<int32, []> x_759_axis_0 = const()[name = tensor<string, []>("x_759_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_759_cast_fp16 = stack(axis = x_759_axis_0, values = (var_6822_cast_fp16, squeeze_222_cast_fp16))[name = tensor<string, []>("x_759_cast_fp16")];
            tensor<int32, [4]> var_6828 = const()[name = tensor<string, []>("op_6828"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6829_cast_fp16 = reshape(shape = var_6828, x = x_759_cast_fp16)[name = tensor<string, []>("op_6829_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6830_cast_fp16 = mul(x = var_6829_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6830_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_111_cast_fp16 = add(x = var_6809_cast_fp16, y = var_6830_cast_fp16)[name = tensor<string, []>("k_patches_111_cast_fp16")];
            tensor<bool, []> var_6833_interleave_0 = const()[name = tensor<string, []>("op_6833_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6833_cast_fp16 = concat(axis = var_20, interleave = var_6833_interleave_0, values = (var_6772_cast_fp16, q_patches_111_cast_fp16))[name = tensor<string, []>("op_6833_cast_fp16")];
            tensor<bool, []> var_6836_interleave_0 = const()[name = tensor<string, []>("op_6836_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6836_cast_fp16 = concat(axis = var_20, interleave = var_6836_interleave_0, values = (var_6776_cast_fp16, k_patches_111_cast_fp16))[name = tensor<string, []>("op_6836_cast_fp16")];
            tensor<fp16, []> var_6838_to_fp16 = const()[name = tensor<string, []>("op_6838_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_559_cast_fp16 = mul(x = var_6833_cast_fp16, y = var_6838_to_fp16)[name = tensor<string, []>("q_559_cast_fp16")];
            tensor<bool, []> attn_221_transpose_x_1 = const()[name = tensor<string, []>("attn_221_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_221_transpose_y_1 = const()[name = tensor<string, []>("attn_221_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_221_cast_fp16 = matmul(transpose_x = attn_221_transpose_x_1, transpose_y = attn_221_transpose_y_1, x = q_559_cast_fp16, y = var_6836_cast_fp16)[name = tensor<string, []>("attn_221_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_223_cast_fp16 = softmax(axis = var_21, x = attn_221_cast_fp16)[name = tensor<string, []>("attn_223_cast_fp16")];
            tensor<bool, []> var_6843_transpose_x_0 = const()[name = tensor<string, []>("op_6843_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6843_transpose_y_0 = const()[name = tensor<string, []>("op_6843_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6843_cast_fp16 = matmul(transpose_x = var_6843_transpose_x_0, transpose_y = var_6843_transpose_y_0, x = attn_223_cast_fp16, y = v_335_cast_fp16)[name = tensor<string, []>("op_6843_cast_fp16")];
            tensor<int32, [1]> var_6844_axes_0 = const()[name = tensor<string, []>("op_6844_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6844_cast_fp16 = squeeze(axes = var_6844_axes_0, x = var_6843_cast_fp16)[name = tensor<string, []>("op_6844_cast_fp16")];
            tensor<string, []> var_6850_pad_type_0 = const()[name = tensor<string, []>("op_6850_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6850_strides_0 = const()[name = tensor<string, []>("op_6850_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6850_pad_0 = const()[name = tensor<string, []>("op_6850_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6850_dilations_0 = const()[name = tensor<string, []>("op_6850_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6850_groups_0 = const()[name = tensor<string, []>("op_6850_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60622272)))];
            tensor<fp16, [1, 64, 1, 197]> var_6850_cast_fp16 = conv(dilations = var_6850_dilations_0, groups = var_6850_groups_0, pad = var_6850_pad_0, pad_type = var_6850_pad_type_0, strides = var_6850_strides_0, weight = model_blocks_4_attn_q_projs_8_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6850_cast_fp16")];
            tensor<int32, [1]> var_6851_axes_0 = const()[name = tensor<string, []>("op_6851_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6851_cast_fp16 = squeeze(axes = var_6851_axes_0, x = var_6850_cast_fp16)[name = tensor<string, []>("op_6851_cast_fp16")];
            tensor<int32, [3]> q_561_perm_0 = const()[name = tensor<string, []>("q_561_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6858_pad_type_0 = const()[name = tensor<string, []>("op_6858_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6858_strides_0 = const()[name = tensor<string, []>("op_6858_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6858_pad_0 = const()[name = tensor<string, []>("op_6858_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6858_dilations_0 = const()[name = tensor<string, []>("op_6858_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6858_groups_0 = const()[name = tensor<string, []>("op_6858_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60720640)))];
            tensor<fp16, [1, 64, 1, 197]> var_6858_cast_fp16 = conv(dilations = var_6858_dilations_0, groups = var_6858_groups_0, pad = var_6858_pad_0, pad_type = var_6858_pad_type_0, strides = var_6858_strides_0, weight = model_blocks_4_attn_k_projs_8_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6858_cast_fp16")];
            tensor<int32, [1]> var_6859_axes_0 = const()[name = tensor<string, []>("op_6859_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6859_cast_fp16 = squeeze(axes = var_6859_axes_0, x = var_6858_cast_fp16)[name = tensor<string, []>("op_6859_cast_fp16")];
            tensor<int32, [3]> k_337_perm_0 = const()[name = tensor<string, []>("k_337_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6866_pad_type_0 = const()[name = tensor<string, []>("op_6866_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6866_strides_0 = const()[name = tensor<string, []>("op_6866_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6866_pad_0 = const()[name = tensor<string, []>("op_6866_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6866_dilations_0 = const()[name = tensor<string, []>("op_6866_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6866_groups_0 = const()[name = tensor<string, []>("op_6866_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60819008)))];
            tensor<fp16, [1, 64, 1, 197]> var_6866_cast_fp16 = conv(dilations = var_6866_dilations_0, groups = var_6866_groups_0, pad = var_6866_pad_0, pad_type = var_6866_pad_type_0, strides = var_6866_strides_0, weight = model_blocks_4_attn_v_projs_8_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6867_axes_0 = const()[name = tensor<string, []>("op_6867_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6867_cast_fp16 = squeeze(axes = var_6867_axes_0, x = var_6866_cast_fp16)[name = tensor<string, []>("op_6867_cast_fp16")];
            tensor<int32, [3]> v_337_perm_0 = const()[name = tensor<string, []>("v_337_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60917376)))];
            tensor<fp16, [1, 197, 64]> q_561_cast_fp16 = transpose(perm = q_561_perm_0, x = var_6851_cast_fp16)[name = tensor<string, []>("transpose_318")];
            tensor<fp16, [1, 197, 64]> q_563_cast_fp16 = add(x = q_561_cast_fp16, y = model_blocks_4_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_563_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60917568)))];
            tensor<fp16, [1, 197, 64]> v_337_cast_fp16 = transpose(perm = v_337_perm_0, x = var_6867_cast_fp16)[name = tensor<string, []>("transpose_316")];
            tensor<fp16, [1, 197, 64]> v_339_cast_fp16 = add(x = v_337_cast_fp16, y = model_blocks_4_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_339_cast_fp16")];
            tensor<int32, [1]> q_565_axes_0 = const()[name = tensor<string, []>("q_565_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_565_cast_fp16 = expand_dims(axes = q_565_axes_0, x = q_563_cast_fp16)[name = tensor<string, []>("q_565_cast_fp16")];
            tensor<int32, [1]> k_339_axes_0 = const()[name = tensor<string, []>("k_339_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_337_cast_fp16 = transpose(perm = k_337_perm_0, x = var_6859_cast_fp16)[name = tensor<string, []>("transpose_317")];
            tensor<fp16, [1, 1, 197, 64]> k_339_cast_fp16 = expand_dims(axes = k_339_axes_0, x = k_337_cast_fp16)[name = tensor<string, []>("k_339_cast_fp16")];
            tensor<int32, [1]> v_341_axes_0 = const()[name = tensor<string, []>("v_341_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_341_cast_fp16 = expand_dims(axes = v_341_axes_0, x = v_339_cast_fp16)[name = tensor<string, []>("v_341_cast_fp16")];
            tensor<int32, [4]> var_6876_begin_0 = const()[name = tensor<string, []>("op_6876_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6876_end_0 = const()[name = tensor<string, []>("op_6876_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6876_end_mask_0 = const()[name = tensor<string, []>("op_6876_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6876_cast_fp16 = slice_by_index(begin = var_6876_begin_0, end = var_6876_end_0, end_mask = var_6876_end_mask_0, x = q_565_cast_fp16)[name = tensor<string, []>("op_6876_cast_fp16")];
            tensor<int32, [4]> var_6880_begin_0 = const()[name = tensor<string, []>("op_6880_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6880_end_0 = const()[name = tensor<string, []>("op_6880_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6880_end_mask_0 = const()[name = tensor<string, []>("op_6880_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, x = k_339_cast_fp16)[name = tensor<string, []>("op_6880_cast_fp16")];
            tensor<int32, [4]> var_6884_begin_0 = const()[name = tensor<string, []>("op_6884_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6884_end_0 = const()[name = tensor<string, []>("op_6884_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6884_end_mask_0 = const()[name = tensor<string, []>("op_6884_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6884_cast_fp16 = slice_by_index(begin = var_6884_begin_0, end = var_6884_end_0, end_mask = var_6884_end_mask_0, x = q_565_cast_fp16)[name = tensor<string, []>("op_6884_cast_fp16")];
            tensor<int32, [4]> var_6888_begin_0 = const()[name = tensor<string, []>("op_6888_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6888_end_0 = const()[name = tensor<string, []>("op_6888_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6888_end_mask_0 = const()[name = tensor<string, []>("op_6888_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, x = k_339_cast_fp16)[name = tensor<string, []>("op_6888_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6890_cast_fp16 = mul(x = var_6884_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6890_cast_fp16")];
            tensor<int32, [5]> var_6898 = const()[name = tensor<string, []>("op_6898"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_763_cast_fp16 = reshape(shape = var_6898, x = var_6884_cast_fp16)[name = tensor<string, []>("x_763_cast_fp16")];
            tensor<int32, [2]> var_6900_split_sizes_0 = const()[name = tensor<string, []>("op_6900_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6900_axis_0 = const()[name = tensor<string, []>("op_6900_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6900_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6900_cast_fp16_1 = split(axis = var_6900_axis_0, split_sizes = var_6900_split_sizes_0, x = x_763_cast_fp16)[name = tensor<string, []>("op_6900_cast_fp16")];
            tensor<int32, [1]> squeeze_224_axes_0 = const()[name = tensor<string, []>("squeeze_224_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_224_cast_fp16 = squeeze(axes = squeeze_224_axes_0, x = var_6900_cast_fp16_0)[name = tensor<string, []>("squeeze_224_cast_fp16")];
            tensor<int32, [1]> squeeze_225_axes_0 = const()[name = tensor<string, []>("squeeze_225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_225_cast_fp16 = squeeze(axes = squeeze_225_axes_0, x = var_6900_cast_fp16_1)[name = tensor<string, []>("squeeze_225_cast_fp16")];
            tensor<fp16, []> const_919_promoted_to_fp16 = const()[name = tensor<string, []>("const_919_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6903_cast_fp16 = mul(x = squeeze_225_cast_fp16, y = const_919_promoted_to_fp16)[name = tensor<string, []>("op_6903_cast_fp16")];
            tensor<int32, []> x_765_axis_0 = const()[name = tensor<string, []>("x_765_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_765_cast_fp16 = stack(axis = x_765_axis_0, values = (var_6903_cast_fp16, squeeze_224_cast_fp16))[name = tensor<string, []>("x_765_cast_fp16")];
            tensor<int32, [4]> var_6909 = const()[name = tensor<string, []>("op_6909"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6910_cast_fp16 = reshape(shape = var_6909, x = x_765_cast_fp16)[name = tensor<string, []>("op_6910_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6911_cast_fp16 = mul(x = var_6910_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6911_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_113_cast_fp16 = add(x = var_6890_cast_fp16, y = var_6911_cast_fp16)[name = tensor<string, []>("q_patches_113_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6913_cast_fp16 = mul(x = var_6888_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6913_cast_fp16")];
            tensor<int32, [5]> var_6921 = const()[name = tensor<string, []>("op_6921"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_769_cast_fp16 = reshape(shape = var_6921, x = var_6888_cast_fp16)[name = tensor<string, []>("x_769_cast_fp16")];
            tensor<int32, [2]> var_6923_split_sizes_0 = const()[name = tensor<string, []>("op_6923_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6923_axis_0 = const()[name = tensor<string, []>("op_6923_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_6923_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_6923_cast_fp16_1 = split(axis = var_6923_axis_0, split_sizes = var_6923_split_sizes_0, x = x_769_cast_fp16)[name = tensor<string, []>("op_6923_cast_fp16")];
            tensor<int32, [1]> squeeze_226_axes_0 = const()[name = tensor<string, []>("squeeze_226_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_226_cast_fp16 = squeeze(axes = squeeze_226_axes_0, x = var_6923_cast_fp16_0)[name = tensor<string, []>("squeeze_226_cast_fp16")];
            tensor<int32, [1]> squeeze_227_axes_0 = const()[name = tensor<string, []>("squeeze_227_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_227_cast_fp16 = squeeze(axes = squeeze_227_axes_0, x = var_6923_cast_fp16_1)[name = tensor<string, []>("squeeze_227_cast_fp16")];
            tensor<fp16, []> const_927_promoted_to_fp16 = const()[name = tensor<string, []>("const_927_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_6926_cast_fp16 = mul(x = squeeze_227_cast_fp16, y = const_927_promoted_to_fp16)[name = tensor<string, []>("op_6926_cast_fp16")];
            tensor<int32, []> x_771_axis_0 = const()[name = tensor<string, []>("x_771_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_771_cast_fp16 = stack(axis = x_771_axis_0, values = (var_6926_cast_fp16, squeeze_226_cast_fp16))[name = tensor<string, []>("x_771_cast_fp16")];
            tensor<int32, [4]> var_6932 = const()[name = tensor<string, []>("op_6932"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_6933_cast_fp16 = reshape(shape = var_6932, x = x_771_cast_fp16)[name = tensor<string, []>("op_6933_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6934_cast_fp16 = mul(x = var_6933_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_6934_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_113_cast_fp16 = add(x = var_6913_cast_fp16, y = var_6934_cast_fp16)[name = tensor<string, []>("k_patches_113_cast_fp16")];
            tensor<bool, []> var_6937_interleave_0 = const()[name = tensor<string, []>("op_6937_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6937_cast_fp16 = concat(axis = var_20, interleave = var_6937_interleave_0, values = (var_6876_cast_fp16, q_patches_113_cast_fp16))[name = tensor<string, []>("op_6937_cast_fp16")];
            tensor<bool, []> var_6940_interleave_0 = const()[name = tensor<string, []>("op_6940_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6940_cast_fp16 = concat(axis = var_20, interleave = var_6940_interleave_0, values = (var_6880_cast_fp16, k_patches_113_cast_fp16))[name = tensor<string, []>("op_6940_cast_fp16")];
            tensor<fp16, []> var_6942_to_fp16 = const()[name = tensor<string, []>("op_6942_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_569_cast_fp16 = mul(x = var_6937_cast_fp16, y = var_6942_to_fp16)[name = tensor<string, []>("q_569_cast_fp16")];
            tensor<bool, []> attn_225_transpose_x_1 = const()[name = tensor<string, []>("attn_225_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_225_transpose_y_1 = const()[name = tensor<string, []>("attn_225_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_225_cast_fp16 = matmul(transpose_x = attn_225_transpose_x_1, transpose_y = attn_225_transpose_y_1, x = q_569_cast_fp16, y = var_6940_cast_fp16)[name = tensor<string, []>("attn_225_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_227_cast_fp16 = softmax(axis = var_21, x = attn_225_cast_fp16)[name = tensor<string, []>("attn_227_cast_fp16")];
            tensor<bool, []> var_6947_transpose_x_0 = const()[name = tensor<string, []>("op_6947_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_6947_transpose_y_0 = const()[name = tensor<string, []>("op_6947_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_6947_cast_fp16 = matmul(transpose_x = var_6947_transpose_x_0, transpose_y = var_6947_transpose_y_0, x = attn_227_cast_fp16, y = v_341_cast_fp16)[name = tensor<string, []>("op_6947_cast_fp16")];
            tensor<int32, [1]> var_6948_axes_0 = const()[name = tensor<string, []>("op_6948_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_6948_cast_fp16 = squeeze(axes = var_6948_axes_0, x = var_6947_cast_fp16)[name = tensor<string, []>("op_6948_cast_fp16")];
            tensor<string, []> var_6954_pad_type_0 = const()[name = tensor<string, []>("op_6954_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6954_strides_0 = const()[name = tensor<string, []>("op_6954_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6954_pad_0 = const()[name = tensor<string, []>("op_6954_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6954_dilations_0 = const()[name = tensor<string, []>("op_6954_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6954_groups_0 = const()[name = tensor<string, []>("op_6954_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60917760)))];
            tensor<fp16, [1, 64, 1, 197]> var_6954_cast_fp16 = conv(dilations = var_6954_dilations_0, groups = var_6954_groups_0, pad = var_6954_pad_0, pad_type = var_6954_pad_type_0, strides = var_6954_strides_0, weight = model_blocks_4_attn_q_projs_9_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6954_cast_fp16")];
            tensor<int32, [1]> var_6955_axes_0 = const()[name = tensor<string, []>("op_6955_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6955_cast_fp16 = squeeze(axes = var_6955_axes_0, x = var_6954_cast_fp16)[name = tensor<string, []>("op_6955_cast_fp16")];
            tensor<int32, [3]> q_571_perm_0 = const()[name = tensor<string, []>("q_571_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6962_pad_type_0 = const()[name = tensor<string, []>("op_6962_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6962_strides_0 = const()[name = tensor<string, []>("op_6962_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6962_pad_0 = const()[name = tensor<string, []>("op_6962_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6962_dilations_0 = const()[name = tensor<string, []>("op_6962_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6962_groups_0 = const()[name = tensor<string, []>("op_6962_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61016128)))];
            tensor<fp16, [1, 64, 1, 197]> var_6962_cast_fp16 = conv(dilations = var_6962_dilations_0, groups = var_6962_groups_0, pad = var_6962_pad_0, pad_type = var_6962_pad_type_0, strides = var_6962_strides_0, weight = model_blocks_4_attn_k_projs_9_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6962_cast_fp16")];
            tensor<int32, [1]> var_6963_axes_0 = const()[name = tensor<string, []>("op_6963_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6963_cast_fp16 = squeeze(axes = var_6963_axes_0, x = var_6962_cast_fp16)[name = tensor<string, []>("op_6963_cast_fp16")];
            tensor<int32, [3]> k_343_perm_0 = const()[name = tensor<string, []>("k_343_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_6970_pad_type_0 = const()[name = tensor<string, []>("op_6970_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_6970_strides_0 = const()[name = tensor<string, []>("op_6970_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6970_pad_0 = const()[name = tensor<string, []>("op_6970_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6970_dilations_0 = const()[name = tensor<string, []>("op_6970_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_6970_groups_0 = const()[name = tensor<string, []>("op_6970_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61114496)))];
            tensor<fp16, [1, 64, 1, 197]> var_6970_cast_fp16 = conv(dilations = var_6970_dilations_0, groups = var_6970_groups_0, pad = var_6970_pad_0, pad_type = var_6970_pad_type_0, strides = var_6970_strides_0, weight = model_blocks_4_attn_v_projs_9_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_6970_cast_fp16")];
            tensor<int32, [1]> var_6971_axes_0 = const()[name = tensor<string, []>("op_6971_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_6971_cast_fp16 = squeeze(axes = var_6971_axes_0, x = var_6970_cast_fp16)[name = tensor<string, []>("op_6971_cast_fp16")];
            tensor<int32, [3]> v_343_perm_0 = const()[name = tensor<string, []>("v_343_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61212864)))];
            tensor<fp16, [1, 197, 64]> q_571_cast_fp16 = transpose(perm = q_571_perm_0, x = var_6955_cast_fp16)[name = tensor<string, []>("transpose_315")];
            tensor<fp16, [1, 197, 64]> q_573_cast_fp16 = add(x = q_571_cast_fp16, y = model_blocks_4_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_573_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61213056)))];
            tensor<fp16, [1, 197, 64]> v_343_cast_fp16 = transpose(perm = v_343_perm_0, x = var_6971_cast_fp16)[name = tensor<string, []>("transpose_313")];
            tensor<fp16, [1, 197, 64]> v_345_cast_fp16 = add(x = v_343_cast_fp16, y = model_blocks_4_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_345_cast_fp16")];
            tensor<int32, [1]> q_575_axes_0 = const()[name = tensor<string, []>("q_575_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_575_cast_fp16 = expand_dims(axes = q_575_axes_0, x = q_573_cast_fp16)[name = tensor<string, []>("q_575_cast_fp16")];
            tensor<int32, [1]> k_345_axes_0 = const()[name = tensor<string, []>("k_345_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_343_cast_fp16 = transpose(perm = k_343_perm_0, x = var_6963_cast_fp16)[name = tensor<string, []>("transpose_314")];
            tensor<fp16, [1, 1, 197, 64]> k_345_cast_fp16 = expand_dims(axes = k_345_axes_0, x = k_343_cast_fp16)[name = tensor<string, []>("k_345_cast_fp16")];
            tensor<int32, [1]> v_347_axes_0 = const()[name = tensor<string, []>("v_347_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_347_cast_fp16 = expand_dims(axes = v_347_axes_0, x = v_345_cast_fp16)[name = tensor<string, []>("v_347_cast_fp16")];
            tensor<int32, [4]> var_6980_begin_0 = const()[name = tensor<string, []>("op_6980_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6980_end_0 = const()[name = tensor<string, []>("op_6980_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6980_end_mask_0 = const()[name = tensor<string, []>("op_6980_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6980_cast_fp16 = slice_by_index(begin = var_6980_begin_0, end = var_6980_end_0, end_mask = var_6980_end_mask_0, x = q_575_cast_fp16)[name = tensor<string, []>("op_6980_cast_fp16")];
            tensor<int32, [4]> var_6984_begin_0 = const()[name = tensor<string, []>("op_6984_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6984_end_0 = const()[name = tensor<string, []>("op_6984_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_6984_end_mask_0 = const()[name = tensor<string, []>("op_6984_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_6984_cast_fp16 = slice_by_index(begin = var_6984_begin_0, end = var_6984_end_0, end_mask = var_6984_end_mask_0, x = k_345_cast_fp16)[name = tensor<string, []>("op_6984_cast_fp16")];
            tensor<int32, [4]> var_6988_begin_0 = const()[name = tensor<string, []>("op_6988_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6988_end_0 = const()[name = tensor<string, []>("op_6988_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6988_end_mask_0 = const()[name = tensor<string, []>("op_6988_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6988_cast_fp16 = slice_by_index(begin = var_6988_begin_0, end = var_6988_end_0, end_mask = var_6988_end_mask_0, x = q_575_cast_fp16)[name = tensor<string, []>("op_6988_cast_fp16")];
            tensor<int32, [4]> var_6992_begin_0 = const()[name = tensor<string, []>("op_6992_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6992_end_0 = const()[name = tensor<string, []>("op_6992_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_6992_end_mask_0 = const()[name = tensor<string, []>("op_6992_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_6992_cast_fp16 = slice_by_index(begin = var_6992_begin_0, end = var_6992_end_0, end_mask = var_6992_end_mask_0, x = k_345_cast_fp16)[name = tensor<string, []>("op_6992_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_6994_cast_fp16 = mul(x = var_6988_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_6994_cast_fp16")];
            tensor<int32, [5]> var_7002 = const()[name = tensor<string, []>("op_7002"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_775_cast_fp16 = reshape(shape = var_7002, x = var_6988_cast_fp16)[name = tensor<string, []>("x_775_cast_fp16")];
            tensor<int32, [2]> var_7004_split_sizes_0 = const()[name = tensor<string, []>("op_7004_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7004_axis_0 = const()[name = tensor<string, []>("op_7004_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7004_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7004_cast_fp16_1 = split(axis = var_7004_axis_0, split_sizes = var_7004_split_sizes_0, x = x_775_cast_fp16)[name = tensor<string, []>("op_7004_cast_fp16")];
            tensor<int32, [1]> squeeze_228_axes_0 = const()[name = tensor<string, []>("squeeze_228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_228_cast_fp16 = squeeze(axes = squeeze_228_axes_0, x = var_7004_cast_fp16_0)[name = tensor<string, []>("squeeze_228_cast_fp16")];
            tensor<int32, [1]> squeeze_229_axes_0 = const()[name = tensor<string, []>("squeeze_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_229_cast_fp16 = squeeze(axes = squeeze_229_axes_0, x = var_7004_cast_fp16_1)[name = tensor<string, []>("squeeze_229_cast_fp16")];
            tensor<fp16, []> const_935_promoted_to_fp16 = const()[name = tensor<string, []>("const_935_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7007_cast_fp16 = mul(x = squeeze_229_cast_fp16, y = const_935_promoted_to_fp16)[name = tensor<string, []>("op_7007_cast_fp16")];
            tensor<int32, []> x_777_axis_0 = const()[name = tensor<string, []>("x_777_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_777_cast_fp16 = stack(axis = x_777_axis_0, values = (var_7007_cast_fp16, squeeze_228_cast_fp16))[name = tensor<string, []>("x_777_cast_fp16")];
            tensor<int32, [4]> var_7013 = const()[name = tensor<string, []>("op_7013"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7014_cast_fp16 = reshape(shape = var_7013, x = x_777_cast_fp16)[name = tensor<string, []>("op_7014_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7015_cast_fp16 = mul(x = var_7014_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7015_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_115_cast_fp16 = add(x = var_6994_cast_fp16, y = var_7015_cast_fp16)[name = tensor<string, []>("q_patches_115_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7017_cast_fp16 = mul(x = var_6992_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7017_cast_fp16")];
            tensor<int32, [5]> var_7025 = const()[name = tensor<string, []>("op_7025"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_781_cast_fp16 = reshape(shape = var_7025, x = var_6992_cast_fp16)[name = tensor<string, []>("x_781_cast_fp16")];
            tensor<int32, [2]> var_7027_split_sizes_0 = const()[name = tensor<string, []>("op_7027_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7027_axis_0 = const()[name = tensor<string, []>("op_7027_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7027_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7027_cast_fp16_1 = split(axis = var_7027_axis_0, split_sizes = var_7027_split_sizes_0, x = x_781_cast_fp16)[name = tensor<string, []>("op_7027_cast_fp16")];
            tensor<int32, [1]> squeeze_230_axes_0 = const()[name = tensor<string, []>("squeeze_230_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_230_cast_fp16 = squeeze(axes = squeeze_230_axes_0, x = var_7027_cast_fp16_0)[name = tensor<string, []>("squeeze_230_cast_fp16")];
            tensor<int32, [1]> squeeze_231_axes_0 = const()[name = tensor<string, []>("squeeze_231_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_231_cast_fp16 = squeeze(axes = squeeze_231_axes_0, x = var_7027_cast_fp16_1)[name = tensor<string, []>("squeeze_231_cast_fp16")];
            tensor<fp16, []> const_943_promoted_to_fp16 = const()[name = tensor<string, []>("const_943_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7030_cast_fp16 = mul(x = squeeze_231_cast_fp16, y = const_943_promoted_to_fp16)[name = tensor<string, []>("op_7030_cast_fp16")];
            tensor<int32, []> x_783_axis_0 = const()[name = tensor<string, []>("x_783_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_783_cast_fp16 = stack(axis = x_783_axis_0, values = (var_7030_cast_fp16, squeeze_230_cast_fp16))[name = tensor<string, []>("x_783_cast_fp16")];
            tensor<int32, [4]> var_7036 = const()[name = tensor<string, []>("op_7036"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7037_cast_fp16 = reshape(shape = var_7036, x = x_783_cast_fp16)[name = tensor<string, []>("op_7037_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7038_cast_fp16 = mul(x = var_7037_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7038_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_115_cast_fp16 = add(x = var_7017_cast_fp16, y = var_7038_cast_fp16)[name = tensor<string, []>("k_patches_115_cast_fp16")];
            tensor<bool, []> var_7041_interleave_0 = const()[name = tensor<string, []>("op_7041_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7041_cast_fp16 = concat(axis = var_20, interleave = var_7041_interleave_0, values = (var_6980_cast_fp16, q_patches_115_cast_fp16))[name = tensor<string, []>("op_7041_cast_fp16")];
            tensor<bool, []> var_7044_interleave_0 = const()[name = tensor<string, []>("op_7044_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7044_cast_fp16 = concat(axis = var_20, interleave = var_7044_interleave_0, values = (var_6984_cast_fp16, k_patches_115_cast_fp16))[name = tensor<string, []>("op_7044_cast_fp16")];
            tensor<fp16, []> var_7046_to_fp16 = const()[name = tensor<string, []>("op_7046_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_579_cast_fp16 = mul(x = var_7041_cast_fp16, y = var_7046_to_fp16)[name = tensor<string, []>("q_579_cast_fp16")];
            tensor<bool, []> attn_229_transpose_x_1 = const()[name = tensor<string, []>("attn_229_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_229_transpose_y_1 = const()[name = tensor<string, []>("attn_229_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_229_cast_fp16 = matmul(transpose_x = attn_229_transpose_x_1, transpose_y = attn_229_transpose_y_1, x = q_579_cast_fp16, y = var_7044_cast_fp16)[name = tensor<string, []>("attn_229_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_231_cast_fp16 = softmax(axis = var_21, x = attn_229_cast_fp16)[name = tensor<string, []>("attn_231_cast_fp16")];
            tensor<bool, []> var_7051_transpose_x_0 = const()[name = tensor<string, []>("op_7051_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7051_transpose_y_0 = const()[name = tensor<string, []>("op_7051_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7051_cast_fp16 = matmul(transpose_x = var_7051_transpose_x_0, transpose_y = var_7051_transpose_y_0, x = attn_231_cast_fp16, y = v_347_cast_fp16)[name = tensor<string, []>("op_7051_cast_fp16")];
            tensor<int32, [1]> var_7052_axes_0 = const()[name = tensor<string, []>("op_7052_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7052_cast_fp16 = squeeze(axes = var_7052_axes_0, x = var_7051_cast_fp16)[name = tensor<string, []>("op_7052_cast_fp16")];
            tensor<string, []> var_7058_pad_type_0 = const()[name = tensor<string, []>("op_7058_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7058_strides_0 = const()[name = tensor<string, []>("op_7058_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7058_pad_0 = const()[name = tensor<string, []>("op_7058_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7058_dilations_0 = const()[name = tensor<string, []>("op_7058_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7058_groups_0 = const()[name = tensor<string, []>("op_7058_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61213248)))];
            tensor<fp16, [1, 64, 1, 197]> var_7058_cast_fp16 = conv(dilations = var_7058_dilations_0, groups = var_7058_groups_0, pad = var_7058_pad_0, pad_type = var_7058_pad_type_0, strides = var_7058_strides_0, weight = model_blocks_4_attn_q_projs_10_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7058_cast_fp16")];
            tensor<int32, [1]> var_7059_axes_0 = const()[name = tensor<string, []>("op_7059_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7059_cast_fp16 = squeeze(axes = var_7059_axes_0, x = var_7058_cast_fp16)[name = tensor<string, []>("op_7059_cast_fp16")];
            tensor<int32, [3]> q_581_perm_0 = const()[name = tensor<string, []>("q_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7066_pad_type_0 = const()[name = tensor<string, []>("op_7066_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7066_strides_0 = const()[name = tensor<string, []>("op_7066_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7066_pad_0 = const()[name = tensor<string, []>("op_7066_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7066_dilations_0 = const()[name = tensor<string, []>("op_7066_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7066_groups_0 = const()[name = tensor<string, []>("op_7066_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61311616)))];
            tensor<fp16, [1, 64, 1, 197]> var_7066_cast_fp16 = conv(dilations = var_7066_dilations_0, groups = var_7066_groups_0, pad = var_7066_pad_0, pad_type = var_7066_pad_type_0, strides = var_7066_strides_0, weight = model_blocks_4_attn_k_projs_10_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7066_cast_fp16")];
            tensor<int32, [1]> var_7067_axes_0 = const()[name = tensor<string, []>("op_7067_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7067_cast_fp16 = squeeze(axes = var_7067_axes_0, x = var_7066_cast_fp16)[name = tensor<string, []>("op_7067_cast_fp16")];
            tensor<int32, [3]> k_349_perm_0 = const()[name = tensor<string, []>("k_349_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7074_pad_type_0 = const()[name = tensor<string, []>("op_7074_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7074_strides_0 = const()[name = tensor<string, []>("op_7074_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7074_pad_0 = const()[name = tensor<string, []>("op_7074_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7074_dilations_0 = const()[name = tensor<string, []>("op_7074_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7074_groups_0 = const()[name = tensor<string, []>("op_7074_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61409984)))];
            tensor<fp16, [1, 64, 1, 197]> var_7074_cast_fp16 = conv(dilations = var_7074_dilations_0, groups = var_7074_groups_0, pad = var_7074_pad_0, pad_type = var_7074_pad_type_0, strides = var_7074_strides_0, weight = model_blocks_4_attn_v_projs_10_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7074_cast_fp16")];
            tensor<int32, [1]> var_7075_axes_0 = const()[name = tensor<string, []>("op_7075_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7075_cast_fp16 = squeeze(axes = var_7075_axes_0, x = var_7074_cast_fp16)[name = tensor<string, []>("op_7075_cast_fp16")];
            tensor<int32, [3]> v_349_perm_0 = const()[name = tensor<string, []>("v_349_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61508352)))];
            tensor<fp16, [1, 197, 64]> q_581_cast_fp16 = transpose(perm = q_581_perm_0, x = var_7059_cast_fp16)[name = tensor<string, []>("transpose_312")];
            tensor<fp16, [1, 197, 64]> q_583_cast_fp16 = add(x = q_581_cast_fp16, y = model_blocks_4_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_583_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61508544)))];
            tensor<fp16, [1, 197, 64]> v_349_cast_fp16 = transpose(perm = v_349_perm_0, x = var_7075_cast_fp16)[name = tensor<string, []>("transpose_310")];
            tensor<fp16, [1, 197, 64]> v_351_cast_fp16 = add(x = v_349_cast_fp16, y = model_blocks_4_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_351_cast_fp16")];
            tensor<int32, [1]> q_585_axes_0 = const()[name = tensor<string, []>("q_585_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_585_cast_fp16 = expand_dims(axes = q_585_axes_0, x = q_583_cast_fp16)[name = tensor<string, []>("q_585_cast_fp16")];
            tensor<int32, [1]> k_351_axes_0 = const()[name = tensor<string, []>("k_351_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_349_cast_fp16 = transpose(perm = k_349_perm_0, x = var_7067_cast_fp16)[name = tensor<string, []>("transpose_311")];
            tensor<fp16, [1, 1, 197, 64]> k_351_cast_fp16 = expand_dims(axes = k_351_axes_0, x = k_349_cast_fp16)[name = tensor<string, []>("k_351_cast_fp16")];
            tensor<int32, [1]> v_353_axes_0 = const()[name = tensor<string, []>("v_353_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_353_cast_fp16 = expand_dims(axes = v_353_axes_0, x = v_351_cast_fp16)[name = tensor<string, []>("v_353_cast_fp16")];
            tensor<int32, [4]> var_7084_begin_0 = const()[name = tensor<string, []>("op_7084_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7084_end_0 = const()[name = tensor<string, []>("op_7084_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7084_end_mask_0 = const()[name = tensor<string, []>("op_7084_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, x = q_585_cast_fp16)[name = tensor<string, []>("op_7084_cast_fp16")];
            tensor<int32, [4]> var_7088_begin_0 = const()[name = tensor<string, []>("op_7088_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7088_end_0 = const()[name = tensor<string, []>("op_7088_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7088_end_mask_0 = const()[name = tensor<string, []>("op_7088_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7088_cast_fp16 = slice_by_index(begin = var_7088_begin_0, end = var_7088_end_0, end_mask = var_7088_end_mask_0, x = k_351_cast_fp16)[name = tensor<string, []>("op_7088_cast_fp16")];
            tensor<int32, [4]> var_7092_begin_0 = const()[name = tensor<string, []>("op_7092_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7092_end_0 = const()[name = tensor<string, []>("op_7092_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7092_end_mask_0 = const()[name = tensor<string, []>("op_7092_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7092_cast_fp16 = slice_by_index(begin = var_7092_begin_0, end = var_7092_end_0, end_mask = var_7092_end_mask_0, x = q_585_cast_fp16)[name = tensor<string, []>("op_7092_cast_fp16")];
            tensor<int32, [4]> var_7096_begin_0 = const()[name = tensor<string, []>("op_7096_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7096_end_0 = const()[name = tensor<string, []>("op_7096_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7096_end_mask_0 = const()[name = tensor<string, []>("op_7096_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7096_cast_fp16 = slice_by_index(begin = var_7096_begin_0, end = var_7096_end_0, end_mask = var_7096_end_mask_0, x = k_351_cast_fp16)[name = tensor<string, []>("op_7096_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7098_cast_fp16 = mul(x = var_7092_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7098_cast_fp16")];
            tensor<int32, [5]> var_7106 = const()[name = tensor<string, []>("op_7106"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_787_cast_fp16 = reshape(shape = var_7106, x = var_7092_cast_fp16)[name = tensor<string, []>("x_787_cast_fp16")];
            tensor<int32, [2]> var_7108_split_sizes_0 = const()[name = tensor<string, []>("op_7108_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7108_axis_0 = const()[name = tensor<string, []>("op_7108_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7108_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7108_cast_fp16_1 = split(axis = var_7108_axis_0, split_sizes = var_7108_split_sizes_0, x = x_787_cast_fp16)[name = tensor<string, []>("op_7108_cast_fp16")];
            tensor<int32, [1]> squeeze_232_axes_0 = const()[name = tensor<string, []>("squeeze_232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_232_cast_fp16 = squeeze(axes = squeeze_232_axes_0, x = var_7108_cast_fp16_0)[name = tensor<string, []>("squeeze_232_cast_fp16")];
            tensor<int32, [1]> squeeze_233_axes_0 = const()[name = tensor<string, []>("squeeze_233_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_233_cast_fp16 = squeeze(axes = squeeze_233_axes_0, x = var_7108_cast_fp16_1)[name = tensor<string, []>("squeeze_233_cast_fp16")];
            tensor<fp16, []> const_951_promoted_to_fp16 = const()[name = tensor<string, []>("const_951_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7111_cast_fp16 = mul(x = squeeze_233_cast_fp16, y = const_951_promoted_to_fp16)[name = tensor<string, []>("op_7111_cast_fp16")];
            tensor<int32, []> x_789_axis_0 = const()[name = tensor<string, []>("x_789_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_789_cast_fp16 = stack(axis = x_789_axis_0, values = (var_7111_cast_fp16, squeeze_232_cast_fp16))[name = tensor<string, []>("x_789_cast_fp16")];
            tensor<int32, [4]> var_7117 = const()[name = tensor<string, []>("op_7117"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7118_cast_fp16 = reshape(shape = var_7117, x = x_789_cast_fp16)[name = tensor<string, []>("op_7118_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7119_cast_fp16 = mul(x = var_7118_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7119_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_117_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7119_cast_fp16)[name = tensor<string, []>("q_patches_117_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7121_cast_fp16 = mul(x = var_7096_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7121_cast_fp16")];
            tensor<int32, [5]> var_7129 = const()[name = tensor<string, []>("op_7129"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_793_cast_fp16 = reshape(shape = var_7129, x = var_7096_cast_fp16)[name = tensor<string, []>("x_793_cast_fp16")];
            tensor<int32, [2]> var_7131_split_sizes_0 = const()[name = tensor<string, []>("op_7131_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7131_axis_0 = const()[name = tensor<string, []>("op_7131_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7131_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7131_cast_fp16_1 = split(axis = var_7131_axis_0, split_sizes = var_7131_split_sizes_0, x = x_793_cast_fp16)[name = tensor<string, []>("op_7131_cast_fp16")];
            tensor<int32, [1]> squeeze_234_axes_0 = const()[name = tensor<string, []>("squeeze_234_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_234_cast_fp16 = squeeze(axes = squeeze_234_axes_0, x = var_7131_cast_fp16_0)[name = tensor<string, []>("squeeze_234_cast_fp16")];
            tensor<int32, [1]> squeeze_235_axes_0 = const()[name = tensor<string, []>("squeeze_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_235_cast_fp16 = squeeze(axes = squeeze_235_axes_0, x = var_7131_cast_fp16_1)[name = tensor<string, []>("squeeze_235_cast_fp16")];
            tensor<fp16, []> const_959_promoted_to_fp16 = const()[name = tensor<string, []>("const_959_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7134_cast_fp16 = mul(x = squeeze_235_cast_fp16, y = const_959_promoted_to_fp16)[name = tensor<string, []>("op_7134_cast_fp16")];
            tensor<int32, []> x_795_axis_0 = const()[name = tensor<string, []>("x_795_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_795_cast_fp16 = stack(axis = x_795_axis_0, values = (var_7134_cast_fp16, squeeze_234_cast_fp16))[name = tensor<string, []>("x_795_cast_fp16")];
            tensor<int32, [4]> var_7140 = const()[name = tensor<string, []>("op_7140"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7141_cast_fp16 = reshape(shape = var_7140, x = x_795_cast_fp16)[name = tensor<string, []>("op_7141_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7142_cast_fp16 = mul(x = var_7141_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7142_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_117_cast_fp16 = add(x = var_7121_cast_fp16, y = var_7142_cast_fp16)[name = tensor<string, []>("k_patches_117_cast_fp16")];
            tensor<bool, []> var_7145_interleave_0 = const()[name = tensor<string, []>("op_7145_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7145_cast_fp16 = concat(axis = var_20, interleave = var_7145_interleave_0, values = (var_7084_cast_fp16, q_patches_117_cast_fp16))[name = tensor<string, []>("op_7145_cast_fp16")];
            tensor<bool, []> var_7148_interleave_0 = const()[name = tensor<string, []>("op_7148_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7148_cast_fp16 = concat(axis = var_20, interleave = var_7148_interleave_0, values = (var_7088_cast_fp16, k_patches_117_cast_fp16))[name = tensor<string, []>("op_7148_cast_fp16")];
            tensor<fp16, []> var_7150_to_fp16 = const()[name = tensor<string, []>("op_7150_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_589_cast_fp16 = mul(x = var_7145_cast_fp16, y = var_7150_to_fp16)[name = tensor<string, []>("q_589_cast_fp16")];
            tensor<bool, []> attn_233_transpose_x_1 = const()[name = tensor<string, []>("attn_233_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_233_transpose_y_1 = const()[name = tensor<string, []>("attn_233_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_233_cast_fp16 = matmul(transpose_x = attn_233_transpose_x_1, transpose_y = attn_233_transpose_y_1, x = q_589_cast_fp16, y = var_7148_cast_fp16)[name = tensor<string, []>("attn_233_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_235_cast_fp16 = softmax(axis = var_21, x = attn_233_cast_fp16)[name = tensor<string, []>("attn_235_cast_fp16")];
            tensor<bool, []> var_7155_transpose_x_0 = const()[name = tensor<string, []>("op_7155_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7155_transpose_y_0 = const()[name = tensor<string, []>("op_7155_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7155_cast_fp16 = matmul(transpose_x = var_7155_transpose_x_0, transpose_y = var_7155_transpose_y_0, x = attn_235_cast_fp16, y = v_353_cast_fp16)[name = tensor<string, []>("op_7155_cast_fp16")];
            tensor<int32, [1]> var_7156_axes_0 = const()[name = tensor<string, []>("op_7156_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7156_cast_fp16 = squeeze(axes = var_7156_axes_0, x = var_7155_cast_fp16)[name = tensor<string, []>("op_7156_cast_fp16")];
            tensor<string, []> var_7162_pad_type_0 = const()[name = tensor<string, []>("op_7162_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7162_strides_0 = const()[name = tensor<string, []>("op_7162_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7162_pad_0 = const()[name = tensor<string, []>("op_7162_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7162_dilations_0 = const()[name = tensor<string, []>("op_7162_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7162_groups_0 = const()[name = tensor<string, []>("op_7162_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61508736)))];
            tensor<fp16, [1, 64, 1, 197]> var_7162_cast_fp16 = conv(dilations = var_7162_dilations_0, groups = var_7162_groups_0, pad = var_7162_pad_0, pad_type = var_7162_pad_type_0, strides = var_7162_strides_0, weight = model_blocks_4_attn_q_projs_11_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7162_cast_fp16")];
            tensor<int32, [1]> var_7163_axes_0 = const()[name = tensor<string, []>("op_7163_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7163_cast_fp16 = squeeze(axes = var_7163_axes_0, x = var_7162_cast_fp16)[name = tensor<string, []>("op_7163_cast_fp16")];
            tensor<int32, [3]> q_591_perm_0 = const()[name = tensor<string, []>("q_591_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7170_pad_type_0 = const()[name = tensor<string, []>("op_7170_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7170_strides_0 = const()[name = tensor<string, []>("op_7170_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7170_pad_0 = const()[name = tensor<string, []>("op_7170_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7170_dilations_0 = const()[name = tensor<string, []>("op_7170_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7170_groups_0 = const()[name = tensor<string, []>("op_7170_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61607104)))];
            tensor<fp16, [1, 64, 1, 197]> var_7170_cast_fp16 = conv(dilations = var_7170_dilations_0, groups = var_7170_groups_0, pad = var_7170_pad_0, pad_type = var_7170_pad_type_0, strides = var_7170_strides_0, weight = model_blocks_4_attn_k_projs_11_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7170_cast_fp16")];
            tensor<int32, [1]> var_7171_axes_0 = const()[name = tensor<string, []>("op_7171_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7171_cast_fp16 = squeeze(axes = var_7171_axes_0, x = var_7170_cast_fp16)[name = tensor<string, []>("op_7171_cast_fp16")];
            tensor<int32, [3]> k_355_perm_0 = const()[name = tensor<string, []>("k_355_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7178_pad_type_0 = const()[name = tensor<string, []>("op_7178_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7178_strides_0 = const()[name = tensor<string, []>("op_7178_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7178_pad_0 = const()[name = tensor<string, []>("op_7178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7178_dilations_0 = const()[name = tensor<string, []>("op_7178_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7178_groups_0 = const()[name = tensor<string, []>("op_7178_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_4_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61705472)))];
            tensor<fp16, [1, 64, 1, 197]> var_7178_cast_fp16 = conv(dilations = var_7178_dilations_0, groups = var_7178_groups_0, pad = var_7178_pad_0, pad_type = var_7178_pad_type_0, strides = var_7178_strides_0, weight = model_blocks_4_attn_v_projs_11_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("op_7178_cast_fp16")];
            tensor<int32, [1]> var_7179_axes_0 = const()[name = tensor<string, []>("op_7179_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7179_cast_fp16 = squeeze(axes = var_7179_axes_0, x = var_7178_cast_fp16)[name = tensor<string, []>("op_7179_cast_fp16")];
            tensor<int32, [3]> v_355_perm_0 = const()[name = tensor<string, []>("v_355_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_4_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61803840)))];
            tensor<fp16, [1, 197, 64]> q_591_cast_fp16 = transpose(perm = q_591_perm_0, x = var_7163_cast_fp16)[name = tensor<string, []>("transpose_309")];
            tensor<fp16, [1, 197, 64]> q_593_cast_fp16 = add(x = q_591_cast_fp16, y = model_blocks_4_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_593_cast_fp16")];
            tensor<fp16, [64]> model_blocks_4_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61804032)))];
            tensor<fp16, [1, 197, 64]> v_355_cast_fp16 = transpose(perm = v_355_perm_0, x = var_7179_cast_fp16)[name = tensor<string, []>("transpose_307")];
            tensor<fp16, [1, 197, 64]> v_357_cast_fp16 = add(x = v_355_cast_fp16, y = model_blocks_4_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_357_cast_fp16")];
            tensor<int32, [1]> q_595_axes_0 = const()[name = tensor<string, []>("q_595_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_595_cast_fp16 = expand_dims(axes = q_595_axes_0, x = q_593_cast_fp16)[name = tensor<string, []>("q_595_cast_fp16")];
            tensor<int32, [1]> k_357_axes_0 = const()[name = tensor<string, []>("k_357_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_355_cast_fp16 = transpose(perm = k_355_perm_0, x = var_7171_cast_fp16)[name = tensor<string, []>("transpose_308")];
            tensor<fp16, [1, 1, 197, 64]> k_357_cast_fp16 = expand_dims(axes = k_357_axes_0, x = k_355_cast_fp16)[name = tensor<string, []>("k_357_cast_fp16")];
            tensor<int32, [1]> v_359_axes_0 = const()[name = tensor<string, []>("v_359_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_359_cast_fp16 = expand_dims(axes = v_359_axes_0, x = v_357_cast_fp16)[name = tensor<string, []>("v_359_cast_fp16")];
            tensor<int32, [4]> var_7188_begin_0 = const()[name = tensor<string, []>("op_7188_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7188_end_0 = const()[name = tensor<string, []>("op_7188_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7188_end_mask_0 = const()[name = tensor<string, []>("op_7188_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, x = q_595_cast_fp16)[name = tensor<string, []>("op_7188_cast_fp16")];
            tensor<int32, [4]> var_7192_begin_0 = const()[name = tensor<string, []>("op_7192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7192_end_0 = const()[name = tensor<string, []>("op_7192_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7192_end_mask_0 = const()[name = tensor<string, []>("op_7192_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7192_cast_fp16 = slice_by_index(begin = var_7192_begin_0, end = var_7192_end_0, end_mask = var_7192_end_mask_0, x = k_357_cast_fp16)[name = tensor<string, []>("op_7192_cast_fp16")];
            tensor<int32, [4]> var_7196_begin_0 = const()[name = tensor<string, []>("op_7196_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7196_end_0 = const()[name = tensor<string, []>("op_7196_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7196_end_mask_0 = const()[name = tensor<string, []>("op_7196_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7196_cast_fp16 = slice_by_index(begin = var_7196_begin_0, end = var_7196_end_0, end_mask = var_7196_end_mask_0, x = q_595_cast_fp16)[name = tensor<string, []>("op_7196_cast_fp16")];
            tensor<int32, [4]> var_7200_begin_0 = const()[name = tensor<string, []>("op_7200_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7200_end_0 = const()[name = tensor<string, []>("op_7200_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7200_end_mask_0 = const()[name = tensor<string, []>("op_7200_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7200_cast_fp16 = slice_by_index(begin = var_7200_begin_0, end = var_7200_end_0, end_mask = var_7200_end_mask_0, x = k_357_cast_fp16)[name = tensor<string, []>("op_7200_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7202_cast_fp16 = mul(x = var_7196_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7202_cast_fp16")];
            tensor<int32, [5]> var_7210 = const()[name = tensor<string, []>("op_7210"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_799_cast_fp16 = reshape(shape = var_7210, x = var_7196_cast_fp16)[name = tensor<string, []>("x_799_cast_fp16")];
            tensor<int32, [2]> var_7212_split_sizes_0 = const()[name = tensor<string, []>("op_7212_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7212_axis_0 = const()[name = tensor<string, []>("op_7212_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7212_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7212_cast_fp16_1 = split(axis = var_7212_axis_0, split_sizes = var_7212_split_sizes_0, x = x_799_cast_fp16)[name = tensor<string, []>("op_7212_cast_fp16")];
            tensor<int32, [1]> squeeze_236_axes_0 = const()[name = tensor<string, []>("squeeze_236_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_236_cast_fp16 = squeeze(axes = squeeze_236_axes_0, x = var_7212_cast_fp16_0)[name = tensor<string, []>("squeeze_236_cast_fp16")];
            tensor<int32, [1]> squeeze_237_axes_0 = const()[name = tensor<string, []>("squeeze_237_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_237_cast_fp16 = squeeze(axes = squeeze_237_axes_0, x = var_7212_cast_fp16_1)[name = tensor<string, []>("squeeze_237_cast_fp16")];
            tensor<fp16, []> const_967_promoted_to_fp16 = const()[name = tensor<string, []>("const_967_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7215_cast_fp16 = mul(x = squeeze_237_cast_fp16, y = const_967_promoted_to_fp16)[name = tensor<string, []>("op_7215_cast_fp16")];
            tensor<int32, []> x_801_axis_0 = const()[name = tensor<string, []>("x_801_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_801_cast_fp16 = stack(axis = x_801_axis_0, values = (var_7215_cast_fp16, squeeze_236_cast_fp16))[name = tensor<string, []>("x_801_cast_fp16")];
            tensor<int32, [4]> var_7221 = const()[name = tensor<string, []>("op_7221"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7222_cast_fp16 = reshape(shape = var_7221, x = x_801_cast_fp16)[name = tensor<string, []>("op_7222_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7223_cast_fp16 = mul(x = var_7222_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7223_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_119_cast_fp16 = add(x = var_7202_cast_fp16, y = var_7223_cast_fp16)[name = tensor<string, []>("q_patches_119_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7225_cast_fp16 = mul(x = var_7200_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7225_cast_fp16")];
            tensor<int32, [5]> var_7233 = const()[name = tensor<string, []>("op_7233"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_805_cast_fp16 = reshape(shape = var_7233, x = var_7200_cast_fp16)[name = tensor<string, []>("x_805_cast_fp16")];
            tensor<int32, [2]> var_7235_split_sizes_0 = const()[name = tensor<string, []>("op_7235_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7235_axis_0 = const()[name = tensor<string, []>("op_7235_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7235_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7235_cast_fp16_1 = split(axis = var_7235_axis_0, split_sizes = var_7235_split_sizes_0, x = x_805_cast_fp16)[name = tensor<string, []>("op_7235_cast_fp16")];
            tensor<int32, [1]> squeeze_238_axes_0 = const()[name = tensor<string, []>("squeeze_238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_238_cast_fp16 = squeeze(axes = squeeze_238_axes_0, x = var_7235_cast_fp16_0)[name = tensor<string, []>("squeeze_238_cast_fp16")];
            tensor<int32, [1]> squeeze_239_axes_0 = const()[name = tensor<string, []>("squeeze_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_239_cast_fp16 = squeeze(axes = squeeze_239_axes_0, x = var_7235_cast_fp16_1)[name = tensor<string, []>("squeeze_239_cast_fp16")];
            tensor<fp16, []> const_975_promoted_to_fp16 = const()[name = tensor<string, []>("const_975_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7238_cast_fp16 = mul(x = squeeze_239_cast_fp16, y = const_975_promoted_to_fp16)[name = tensor<string, []>("op_7238_cast_fp16")];
            tensor<int32, []> x_807_axis_0 = const()[name = tensor<string, []>("x_807_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_807_cast_fp16 = stack(axis = x_807_axis_0, values = (var_7238_cast_fp16, squeeze_238_cast_fp16))[name = tensor<string, []>("x_807_cast_fp16")];
            tensor<int32, [4]> var_7244 = const()[name = tensor<string, []>("op_7244"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7245_cast_fp16 = reshape(shape = var_7244, x = x_807_cast_fp16)[name = tensor<string, []>("op_7245_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7246_cast_fp16 = mul(x = var_7245_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7246_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_119_cast_fp16 = add(x = var_7225_cast_fp16, y = var_7246_cast_fp16)[name = tensor<string, []>("k_patches_119_cast_fp16")];
            tensor<bool, []> var_7249_interleave_0 = const()[name = tensor<string, []>("op_7249_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7249_cast_fp16 = concat(axis = var_20, interleave = var_7249_interleave_0, values = (var_7188_cast_fp16, q_patches_119_cast_fp16))[name = tensor<string, []>("op_7249_cast_fp16")];
            tensor<bool, []> var_7252_interleave_0 = const()[name = tensor<string, []>("op_7252_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7252_cast_fp16 = concat(axis = var_20, interleave = var_7252_interleave_0, values = (var_7192_cast_fp16, k_patches_119_cast_fp16))[name = tensor<string, []>("op_7252_cast_fp16")];
            tensor<fp16, []> var_7254_to_fp16 = const()[name = tensor<string, []>("op_7254_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_599_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7254_to_fp16)[name = tensor<string, []>("q_599_cast_fp16")];
            tensor<bool, []> attn_237_transpose_x_1 = const()[name = tensor<string, []>("attn_237_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_237_transpose_y_1 = const()[name = tensor<string, []>("attn_237_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_237_cast_fp16 = matmul(transpose_x = attn_237_transpose_x_1, transpose_y = attn_237_transpose_y_1, x = q_599_cast_fp16, y = var_7252_cast_fp16)[name = tensor<string, []>("attn_237_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_239_cast_fp16 = softmax(axis = var_21, x = attn_237_cast_fp16)[name = tensor<string, []>("attn_239_cast_fp16")];
            tensor<bool, []> var_7259_transpose_x_0 = const()[name = tensor<string, []>("op_7259_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7259_transpose_y_0 = const()[name = tensor<string, []>("op_7259_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7259_cast_fp16 = matmul(transpose_x = var_7259_transpose_x_0, transpose_y = var_7259_transpose_y_0, x = attn_239_cast_fp16, y = v_359_cast_fp16)[name = tensor<string, []>("op_7259_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_9_cast_fp16 = squeeze(axes = out_9_axes_0, x = var_7259_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<bool, []> input_77_interleave_0 = const()[name = tensor<string, []>("input_77_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_77_cast_fp16 = concat(axis = var_21, interleave = input_77_interleave_0, values = (var_6116_cast_fp16, var_6220_cast_fp16, var_6324_cast_fp16, var_6428_cast_fp16, var_6532_cast_fp16, var_6636_cast_fp16, var_6740_cast_fp16, var_6844_cast_fp16, var_6948_cast_fp16, var_7052_cast_fp16, var_7156_cast_fp16, out_9_cast_fp16))[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<int32, [1]> x_809_axes_0 = const()[name = tensor<string, []>("x_809_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_4_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61804224)))];
            tensor<fp16, [768]> model_blocks_4_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61805824)))];
            tensor<fp16, [1, 197, 768]> x_809_cast_fp16 = layer_norm(axes = x_809_axes_0, beta = model_blocks_4_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_4_attn_inner_attn_ln_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("x_809_cast_fp16")];
            tensor<int32, [4]> var_7267 = const()[name = tensor<string, []>("op_7267"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_811_cast_fp16 = reshape(shape = var_7267, x = x_809_cast_fp16)[name = tensor<string, []>("x_811_cast_fp16")];
            tensor<int32, [4]> input_79_perm_0 = const()[name = tensor<string, []>("input_79_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_813_pad_type_0 = const()[name = tensor<string, []>("x_813_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_813_strides_0 = const()[name = tensor<string, []>("x_813_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_813_pad_0 = const()[name = tensor<string, []>("x_813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_813_dilations_0 = const()[name = tensor<string, []>("x_813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_813_groups_0 = const()[name = tensor<string, []>("x_813_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_4_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61807424)))];
            tensor<fp16, [768]> model_blocks_4_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62987136)))];
            tensor<fp16, [1, 768, 1, 197]> input_79_cast_fp16 = transpose(perm = input_79_perm_0, x = x_811_cast_fp16)[name = tensor<string, []>("transpose_306")];
            tensor<fp16, [1, 768, 1, 197]> x_813_cast_fp16 = conv(bias = model_blocks_4_attn_proj_bias_to_fp16, dilations = x_813_dilations_0, groups = x_813_groups_0, pad = x_813_pad_0, pad_type = x_813_pad_type_0, strides = x_813_strides_0, weight = model_blocks_4_attn_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("x_813_cast_fp16")];
            tensor<int32, [4]> x_815_perm_0 = const()[name = tensor<string, []>("x_815_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_7278 = const()[name = tensor<string, []>("op_7278"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_815_cast_fp16 = transpose(perm = x_815_perm_0, x = x_813_cast_fp16)[name = tensor<string, []>("transpose_305")];
            tensor<fp16, [1, 197, 768]> var_7279_cast_fp16 = reshape(shape = var_7278, x = x_815_cast_fp16)[name = tensor<string, []>("op_7279_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_81_cast_fp16 = add(x = input_73_cast_fp16, y = var_7279_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<int32, [1]> x_817_axes_0 = const()[name = tensor<string, []>("x_817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62988736)))];
            tensor<fp16, [768]> model_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62990336)))];
            tensor<fp16, [1, 197, 768]> x_817_cast_fp16 = layer_norm(axes = x_817_axes_0, beta = model_blocks_4_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_4_norm2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("x_817_cast_fp16")];
            tensor<int32, [4]> var_7291 = const()[name = tensor<string, []>("op_7291"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_819_cast_fp16 = reshape(shape = var_7291, x = x_817_cast_fp16)[name = tensor<string, []>("x_819_cast_fp16")];
            tensor<int32, [4]> input_83_perm_0 = const()[name = tensor<string, []>("input_83_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_85_pad_type_0 = const()[name = tensor<string, []>("input_85_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_85_strides_0 = const()[name = tensor<string, []>("input_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_85_pad_0 = const()[name = tensor<string, []>("input_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_85_dilations_0 = const()[name = tensor<string, []>("input_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_85_groups_0 = const()[name = tensor<string, []>("input_85_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_4_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62991936)))];
            tensor<fp16, [2048]> model_blocks_4_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66137728)))];
            tensor<fp16, [1, 768, 1, 197]> input_83_cast_fp16 = transpose(perm = input_83_perm_0, x = x_819_cast_fp16)[name = tensor<string, []>("transpose_304")];
            tensor<fp16, [1, 2048, 1, 197]> input_85_cast_fp16 = conv(bias = model_blocks_4_mlp_w1_bias_to_fp16, dilations = input_85_dilations_0, groups = input_85_groups_0, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = input_85_strides_0, weight = model_blocks_4_mlp_w1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<string, []> x2_249_pad_type_0 = const()[name = tensor<string, []>("x2_249_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_249_strides_0 = const()[name = tensor<string, []>("x2_249_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_249_pad_0 = const()[name = tensor<string, []>("x2_249_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_249_dilations_0 = const()[name = tensor<string, []>("x2_249_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_249_groups_0 = const()[name = tensor<string, []>("x2_249_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_4_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66141888)))];
            tensor<fp16, [2048]> model_blocks_4_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69287680)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_249_cast_fp16 = conv(bias = model_blocks_4_mlp_w2_bias_to_fp16, dilations = x2_249_dilations_0, groups = x2_249_groups_0, pad = x2_249_pad_0, pad_type = x2_249_pad_type_0, strides = x2_249_strides_0, weight = model_blocks_4_mlp_w2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("x2_249_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_7308_cast_fp16 = silu(x = input_85_cast_fp16)[name = tensor<string, []>("op_7308_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_33_cast_fp16 = mul(x = var_7308_cast_fp16, y = x2_249_cast_fp16)[name = tensor<string, []>("hidden_33_cast_fp16")];
            tensor<int32, [4]> hidden_35_perm_0 = const()[name = tensor<string, []>("hidden_35_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_7311 = const()[name = tensor<string, []>("op_7311"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_35_cast_fp16 = transpose(perm = hidden_35_perm_0, x = hidden_33_cast_fp16)[name = tensor<string, []>("transpose_303")];
            tensor<fp16, [1, 197, 2048]> input_87_cast_fp16 = reshape(shape = var_7311, x = hidden_35_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<int32, [1]> hidden_37_axes_0 = const()[name = tensor<string, []>("hidden_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_4_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69291840)))];
            tensor<fp16, [2048]> model_blocks_4_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69296000)))];
            tensor<fp16, [1, 197, 2048]> hidden_37_cast_fp16 = layer_norm(axes = hidden_37_axes_0, beta = model_blocks_4_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_4_mlp_ffn_ln_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_37_cast_fp16")];
            tensor<int32, [4]> var_7317 = const()[name = tensor<string, []>("op_7317"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_39_cast_fp16 = reshape(shape = var_7317, x = hidden_37_cast_fp16)[name = tensor<string, []>("hidden_39_cast_fp16")];
            tensor<int32, [4]> input_89_perm_0 = const()[name = tensor<string, []>("input_89_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_821_pad_type_0 = const()[name = tensor<string, []>("x_821_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_821_strides_0 = const()[name = tensor<string, []>("x_821_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_821_pad_0 = const()[name = tensor<string, []>("x_821_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_821_dilations_0 = const()[name = tensor<string, []>("x_821_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_821_groups_0 = const()[name = tensor<string, []>("x_821_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_4_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69300160)))];
            tensor<fp16, [768]> model_blocks_4_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_4_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72445952)))];
            tensor<fp16, [1, 2048, 1, 197]> input_89_cast_fp16 = transpose(perm = input_89_perm_0, x = hidden_39_cast_fp16)[name = tensor<string, []>("transpose_302")];
            tensor<fp16, [1, 768, 1, 197]> x_821_cast_fp16 = conv(bias = model_blocks_4_mlp_w3_bias_to_fp16, dilations = x_821_dilations_0, groups = x_821_groups_0, pad = x_821_pad_0, pad_type = x_821_pad_type_0, strides = x_821_strides_0, weight = model_blocks_4_mlp_w3_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("x_821_cast_fp16")];
            tensor<int32, [4]> x_823_perm_0 = const()[name = tensor<string, []>("x_823_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_7328 = const()[name = tensor<string, []>("op_7328"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_823_cast_fp16 = transpose(perm = x_823_perm_0, x = x_821_cast_fp16)[name = tensor<string, []>("transpose_301")];
            tensor<fp16, [1, 197, 768]> var_7329_cast_fp16 = reshape(shape = var_7328, x = x_823_cast_fp16)[name = tensor<string, []>("op_7329_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_91_cast_fp16 = add(x = input_81_cast_fp16, y = var_7329_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<int32, [1]> x_825_axes_0 = const()[name = tensor<string, []>("x_825_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72447552)))];
            tensor<fp16, [768]> model_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72449152)))];
            tensor<fp16, [1, 197, 768]> x_825_cast_fp16 = layer_norm(axes = x_825_axes_0, beta = model_blocks_5_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_5_norm1_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("x_825_cast_fp16")];
            tensor<int32, [4]> var_7463 = const()[name = tensor<string, []>("op_7463"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_827_cast_fp16 = reshape(shape = var_7463, x = x_825_cast_fp16)[name = tensor<string, []>("x_827_cast_fp16")];
            tensor<int32, [4]> input_93_perm_0 = const()[name = tensor<string, []>("input_93_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_7471_pad_type_0 = const()[name = tensor<string, []>("op_7471_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7471_strides_0 = const()[name = tensor<string, []>("op_7471_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7471_pad_0 = const()[name = tensor<string, []>("op_7471_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7471_dilations_0 = const()[name = tensor<string, []>("op_7471_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7471_groups_0 = const()[name = tensor<string, []>("op_7471_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72450752)))];
            tensor<fp16, [1, 768, 1, 197]> input_93_cast_fp16 = transpose(perm = input_93_perm_0, x = x_827_cast_fp16)[name = tensor<string, []>("transpose_300")];
            tensor<fp16, [1, 64, 1, 197]> var_7471_cast_fp16 = conv(dilations = var_7471_dilations_0, groups = var_7471_groups_0, pad = var_7471_pad_0, pad_type = var_7471_pad_type_0, strides = var_7471_strides_0, weight = model_blocks_5_attn_q_projs_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7471_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = tensor<string, []>("op_7472_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7472_cast_fp16 = squeeze(axes = var_7472_axes_0, x = var_7471_cast_fp16)[name = tensor<string, []>("op_7472_cast_fp16")];
            tensor<int32, [3]> q_601_perm_0 = const()[name = tensor<string, []>("q_601_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7479_pad_type_0 = const()[name = tensor<string, []>("op_7479_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7479_strides_0 = const()[name = tensor<string, []>("op_7479_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7479_pad_0 = const()[name = tensor<string, []>("op_7479_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7479_dilations_0 = const()[name = tensor<string, []>("op_7479_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7479_groups_0 = const()[name = tensor<string, []>("op_7479_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72549120)))];
            tensor<fp16, [1, 64, 1, 197]> var_7479_cast_fp16 = conv(dilations = var_7479_dilations_0, groups = var_7479_groups_0, pad = var_7479_pad_0, pad_type = var_7479_pad_type_0, strides = var_7479_strides_0, weight = model_blocks_5_attn_k_projs_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7479_cast_fp16")];
            tensor<int32, [1]> var_7480_axes_0 = const()[name = tensor<string, []>("op_7480_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7480_cast_fp16 = squeeze(axes = var_7480_axes_0, x = var_7479_cast_fp16)[name = tensor<string, []>("op_7480_cast_fp16")];
            tensor<int32, [3]> k_361_perm_0 = const()[name = tensor<string, []>("k_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7487_pad_type_0 = const()[name = tensor<string, []>("op_7487_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7487_strides_0 = const()[name = tensor<string, []>("op_7487_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7487_pad_0 = const()[name = tensor<string, []>("op_7487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7487_dilations_0 = const()[name = tensor<string, []>("op_7487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7487_groups_0 = const()[name = tensor<string, []>("op_7487_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72647488)))];
            tensor<fp16, [1, 64, 1, 197]> var_7487_cast_fp16 = conv(dilations = var_7487_dilations_0, groups = var_7487_groups_0, pad = var_7487_pad_0, pad_type = var_7487_pad_type_0, strides = var_7487_strides_0, weight = model_blocks_5_attn_v_projs_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7487_cast_fp16")];
            tensor<int32, [1]> var_7488_axes_0 = const()[name = tensor<string, []>("op_7488_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7488_cast_fp16 = squeeze(axes = var_7488_axes_0, x = var_7487_cast_fp16)[name = tensor<string, []>("op_7488_cast_fp16")];
            tensor<int32, [3]> v_361_perm_0 = const()[name = tensor<string, []>("v_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72745856)))];
            tensor<fp16, [1, 197, 64]> q_601_cast_fp16 = transpose(perm = q_601_perm_0, x = var_7472_cast_fp16)[name = tensor<string, []>("transpose_299")];
            tensor<fp16, [1, 197, 64]> q_603_cast_fp16 = add(x = q_601_cast_fp16, y = model_blocks_5_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_603_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72746048)))];
            tensor<fp16, [1, 197, 64]> v_361_cast_fp16 = transpose(perm = v_361_perm_0, x = var_7488_cast_fp16)[name = tensor<string, []>("transpose_297")];
            tensor<fp16, [1, 197, 64]> v_363_cast_fp16 = add(x = v_361_cast_fp16, y = model_blocks_5_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_363_cast_fp16")];
            tensor<int32, [1]> q_605_axes_0 = const()[name = tensor<string, []>("q_605_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_605_cast_fp16 = expand_dims(axes = q_605_axes_0, x = q_603_cast_fp16)[name = tensor<string, []>("q_605_cast_fp16")];
            tensor<int32, [1]> k_363_axes_0 = const()[name = tensor<string, []>("k_363_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_361_cast_fp16 = transpose(perm = k_361_perm_0, x = var_7480_cast_fp16)[name = tensor<string, []>("transpose_298")];
            tensor<fp16, [1, 1, 197, 64]> k_363_cast_fp16 = expand_dims(axes = k_363_axes_0, x = k_361_cast_fp16)[name = tensor<string, []>("k_363_cast_fp16")];
            tensor<int32, [1]> v_365_axes_0 = const()[name = tensor<string, []>("v_365_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_365_cast_fp16 = expand_dims(axes = v_365_axes_0, x = v_363_cast_fp16)[name = tensor<string, []>("v_365_cast_fp16")];
            tensor<int32, [4]> var_7497_begin_0 = const()[name = tensor<string, []>("op_7497_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7497_end_0 = const()[name = tensor<string, []>("op_7497_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7497_end_mask_0 = const()[name = tensor<string, []>("op_7497_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7497_cast_fp16 = slice_by_index(begin = var_7497_begin_0, end = var_7497_end_0, end_mask = var_7497_end_mask_0, x = q_605_cast_fp16)[name = tensor<string, []>("op_7497_cast_fp16")];
            tensor<int32, [4]> var_7501_begin_0 = const()[name = tensor<string, []>("op_7501_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7501_end_0 = const()[name = tensor<string, []>("op_7501_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7501_end_mask_0 = const()[name = tensor<string, []>("op_7501_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7501_cast_fp16 = slice_by_index(begin = var_7501_begin_0, end = var_7501_end_0, end_mask = var_7501_end_mask_0, x = k_363_cast_fp16)[name = tensor<string, []>("op_7501_cast_fp16")];
            tensor<int32, [4]> var_7505_begin_0 = const()[name = tensor<string, []>("op_7505_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7505_end_0 = const()[name = tensor<string, []>("op_7505_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7505_end_mask_0 = const()[name = tensor<string, []>("op_7505_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7505_cast_fp16 = slice_by_index(begin = var_7505_begin_0, end = var_7505_end_0, end_mask = var_7505_end_mask_0, x = q_605_cast_fp16)[name = tensor<string, []>("op_7505_cast_fp16")];
            tensor<int32, [4]> var_7509_begin_0 = const()[name = tensor<string, []>("op_7509_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7509_end_0 = const()[name = tensor<string, []>("op_7509_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7509_end_mask_0 = const()[name = tensor<string, []>("op_7509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7509_cast_fp16 = slice_by_index(begin = var_7509_begin_0, end = var_7509_end_0, end_mask = var_7509_end_mask_0, x = k_363_cast_fp16)[name = tensor<string, []>("op_7509_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7511_cast_fp16 = mul(x = var_7505_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7511_cast_fp16")];
            tensor<int32, [5]> var_7519 = const()[name = tensor<string, []>("op_7519"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_831_cast_fp16 = reshape(shape = var_7519, x = var_7505_cast_fp16)[name = tensor<string, []>("x_831_cast_fp16")];
            tensor<int32, [2]> var_7521_split_sizes_0 = const()[name = tensor<string, []>("op_7521_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7521_axis_0 = const()[name = tensor<string, []>("op_7521_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7521_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7521_cast_fp16_1 = split(axis = var_7521_axis_0, split_sizes = var_7521_split_sizes_0, x = x_831_cast_fp16)[name = tensor<string, []>("op_7521_cast_fp16")];
            tensor<int32, [1]> squeeze_240_axes_0 = const()[name = tensor<string, []>("squeeze_240_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_240_cast_fp16 = squeeze(axes = squeeze_240_axes_0, x = var_7521_cast_fp16_0)[name = tensor<string, []>("squeeze_240_cast_fp16")];
            tensor<int32, [1]> squeeze_241_axes_0 = const()[name = tensor<string, []>("squeeze_241_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_241_cast_fp16 = squeeze(axes = squeeze_241_axes_0, x = var_7521_cast_fp16_1)[name = tensor<string, []>("squeeze_241_cast_fp16")];
            tensor<fp16, []> const_987_promoted_to_fp16 = const()[name = tensor<string, []>("const_987_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7524_cast_fp16 = mul(x = squeeze_241_cast_fp16, y = const_987_promoted_to_fp16)[name = tensor<string, []>("op_7524_cast_fp16")];
            tensor<int32, []> x_833_axis_0 = const()[name = tensor<string, []>("x_833_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_833_cast_fp16 = stack(axis = x_833_axis_0, values = (var_7524_cast_fp16, squeeze_240_cast_fp16))[name = tensor<string, []>("x_833_cast_fp16")];
            tensor<int32, [4]> var_7530 = const()[name = tensor<string, []>("op_7530"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7531_cast_fp16 = reshape(shape = var_7530, x = x_833_cast_fp16)[name = tensor<string, []>("op_7531_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7532_cast_fp16 = mul(x = var_7531_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7532_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_121_cast_fp16 = add(x = var_7511_cast_fp16, y = var_7532_cast_fp16)[name = tensor<string, []>("q_patches_121_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7534_cast_fp16 = mul(x = var_7509_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7534_cast_fp16")];
            tensor<int32, [5]> var_7542 = const()[name = tensor<string, []>("op_7542"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_837_cast_fp16 = reshape(shape = var_7542, x = var_7509_cast_fp16)[name = tensor<string, []>("x_837_cast_fp16")];
            tensor<int32, [2]> var_7544_split_sizes_0 = const()[name = tensor<string, []>("op_7544_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7544_axis_0 = const()[name = tensor<string, []>("op_7544_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7544_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7544_cast_fp16_1 = split(axis = var_7544_axis_0, split_sizes = var_7544_split_sizes_0, x = x_837_cast_fp16)[name = tensor<string, []>("op_7544_cast_fp16")];
            tensor<int32, [1]> squeeze_242_axes_0 = const()[name = tensor<string, []>("squeeze_242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_242_cast_fp16 = squeeze(axes = squeeze_242_axes_0, x = var_7544_cast_fp16_0)[name = tensor<string, []>("squeeze_242_cast_fp16")];
            tensor<int32, [1]> squeeze_243_axes_0 = const()[name = tensor<string, []>("squeeze_243_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_243_cast_fp16 = squeeze(axes = squeeze_243_axes_0, x = var_7544_cast_fp16_1)[name = tensor<string, []>("squeeze_243_cast_fp16")];
            tensor<fp16, []> const_995_promoted_to_fp16 = const()[name = tensor<string, []>("const_995_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7547_cast_fp16 = mul(x = squeeze_243_cast_fp16, y = const_995_promoted_to_fp16)[name = tensor<string, []>("op_7547_cast_fp16")];
            tensor<int32, []> x_839_axis_0 = const()[name = tensor<string, []>("x_839_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_839_cast_fp16 = stack(axis = x_839_axis_0, values = (var_7547_cast_fp16, squeeze_242_cast_fp16))[name = tensor<string, []>("x_839_cast_fp16")];
            tensor<int32, [4]> var_7553 = const()[name = tensor<string, []>("op_7553"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7554_cast_fp16 = reshape(shape = var_7553, x = x_839_cast_fp16)[name = tensor<string, []>("op_7554_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7555_cast_fp16 = mul(x = var_7554_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7555_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_121_cast_fp16 = add(x = var_7534_cast_fp16, y = var_7555_cast_fp16)[name = tensor<string, []>("k_patches_121_cast_fp16")];
            tensor<bool, []> var_7558_interleave_0 = const()[name = tensor<string, []>("op_7558_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7558_cast_fp16 = concat(axis = var_20, interleave = var_7558_interleave_0, values = (var_7497_cast_fp16, q_patches_121_cast_fp16))[name = tensor<string, []>("op_7558_cast_fp16")];
            tensor<bool, []> var_7561_interleave_0 = const()[name = tensor<string, []>("op_7561_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7561_cast_fp16 = concat(axis = var_20, interleave = var_7561_interleave_0, values = (var_7501_cast_fp16, k_patches_121_cast_fp16))[name = tensor<string, []>("op_7561_cast_fp16")];
            tensor<fp16, []> var_7563_to_fp16 = const()[name = tensor<string, []>("op_7563_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_609_cast_fp16 = mul(x = var_7558_cast_fp16, y = var_7563_to_fp16)[name = tensor<string, []>("q_609_cast_fp16")];
            tensor<bool, []> attn_241_transpose_x_1 = const()[name = tensor<string, []>("attn_241_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_241_transpose_y_1 = const()[name = tensor<string, []>("attn_241_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_241_cast_fp16 = matmul(transpose_x = attn_241_transpose_x_1, transpose_y = attn_241_transpose_y_1, x = q_609_cast_fp16, y = var_7561_cast_fp16)[name = tensor<string, []>("attn_241_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_243_cast_fp16 = softmax(axis = var_21, x = attn_241_cast_fp16)[name = tensor<string, []>("attn_243_cast_fp16")];
            tensor<bool, []> var_7568_transpose_x_0 = const()[name = tensor<string, []>("op_7568_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7568_transpose_y_0 = const()[name = tensor<string, []>("op_7568_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7568_cast_fp16 = matmul(transpose_x = var_7568_transpose_x_0, transpose_y = var_7568_transpose_y_0, x = attn_243_cast_fp16, y = v_365_cast_fp16)[name = tensor<string, []>("op_7568_cast_fp16")];
            tensor<int32, [1]> var_7569_axes_0 = const()[name = tensor<string, []>("op_7569_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7569_cast_fp16 = squeeze(axes = var_7569_axes_0, x = var_7568_cast_fp16)[name = tensor<string, []>("op_7569_cast_fp16")];
            tensor<string, []> var_7575_pad_type_0 = const()[name = tensor<string, []>("op_7575_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7575_strides_0 = const()[name = tensor<string, []>("op_7575_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7575_pad_0 = const()[name = tensor<string, []>("op_7575_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7575_dilations_0 = const()[name = tensor<string, []>("op_7575_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7575_groups_0 = const()[name = tensor<string, []>("op_7575_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72746240)))];
            tensor<fp16, [1, 64, 1, 197]> var_7575_cast_fp16 = conv(dilations = var_7575_dilations_0, groups = var_7575_groups_0, pad = var_7575_pad_0, pad_type = var_7575_pad_type_0, strides = var_7575_strides_0, weight = model_blocks_5_attn_q_projs_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7575_cast_fp16")];
            tensor<int32, [1]> var_7576_axes_0 = const()[name = tensor<string, []>("op_7576_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7576_cast_fp16 = squeeze(axes = var_7576_axes_0, x = var_7575_cast_fp16)[name = tensor<string, []>("op_7576_cast_fp16")];
            tensor<int32, [3]> q_611_perm_0 = const()[name = tensor<string, []>("q_611_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7583_pad_type_0 = const()[name = tensor<string, []>("op_7583_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7583_strides_0 = const()[name = tensor<string, []>("op_7583_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7583_pad_0 = const()[name = tensor<string, []>("op_7583_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7583_dilations_0 = const()[name = tensor<string, []>("op_7583_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7583_groups_0 = const()[name = tensor<string, []>("op_7583_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72844608)))];
            tensor<fp16, [1, 64, 1, 197]> var_7583_cast_fp16 = conv(dilations = var_7583_dilations_0, groups = var_7583_groups_0, pad = var_7583_pad_0, pad_type = var_7583_pad_type_0, strides = var_7583_strides_0, weight = model_blocks_5_attn_k_projs_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7583_cast_fp16")];
            tensor<int32, [1]> var_7584_axes_0 = const()[name = tensor<string, []>("op_7584_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7584_cast_fp16 = squeeze(axes = var_7584_axes_0, x = var_7583_cast_fp16)[name = tensor<string, []>("op_7584_cast_fp16")];
            tensor<int32, [3]> k_367_perm_0 = const()[name = tensor<string, []>("k_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7591_pad_type_0 = const()[name = tensor<string, []>("op_7591_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7591_strides_0 = const()[name = tensor<string, []>("op_7591_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7591_pad_0 = const()[name = tensor<string, []>("op_7591_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7591_dilations_0 = const()[name = tensor<string, []>("op_7591_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7591_groups_0 = const()[name = tensor<string, []>("op_7591_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72942976)))];
            tensor<fp16, [1, 64, 1, 197]> var_7591_cast_fp16 = conv(dilations = var_7591_dilations_0, groups = var_7591_groups_0, pad = var_7591_pad_0, pad_type = var_7591_pad_type_0, strides = var_7591_strides_0, weight = model_blocks_5_attn_v_projs_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7591_cast_fp16")];
            tensor<int32, [1]> var_7592_axes_0 = const()[name = tensor<string, []>("op_7592_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7592_cast_fp16 = squeeze(axes = var_7592_axes_0, x = var_7591_cast_fp16)[name = tensor<string, []>("op_7592_cast_fp16")];
            tensor<int32, [3]> v_367_perm_0 = const()[name = tensor<string, []>("v_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73041344)))];
            tensor<fp16, [1, 197, 64]> q_611_cast_fp16 = transpose(perm = q_611_perm_0, x = var_7576_cast_fp16)[name = tensor<string, []>("transpose_296")];
            tensor<fp16, [1, 197, 64]> q_613_cast_fp16 = add(x = q_611_cast_fp16, y = model_blocks_5_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_613_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73041536)))];
            tensor<fp16, [1, 197, 64]> v_367_cast_fp16 = transpose(perm = v_367_perm_0, x = var_7592_cast_fp16)[name = tensor<string, []>("transpose_294")];
            tensor<fp16, [1, 197, 64]> v_369_cast_fp16 = add(x = v_367_cast_fp16, y = model_blocks_5_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_369_cast_fp16")];
            tensor<int32, [1]> q_615_axes_0 = const()[name = tensor<string, []>("q_615_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_615_cast_fp16 = expand_dims(axes = q_615_axes_0, x = q_613_cast_fp16)[name = tensor<string, []>("q_615_cast_fp16")];
            tensor<int32, [1]> k_369_axes_0 = const()[name = tensor<string, []>("k_369_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_367_cast_fp16 = transpose(perm = k_367_perm_0, x = var_7584_cast_fp16)[name = tensor<string, []>("transpose_295")];
            tensor<fp16, [1, 1, 197, 64]> k_369_cast_fp16 = expand_dims(axes = k_369_axes_0, x = k_367_cast_fp16)[name = tensor<string, []>("k_369_cast_fp16")];
            tensor<int32, [1]> v_371_axes_0 = const()[name = tensor<string, []>("v_371_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_371_cast_fp16 = expand_dims(axes = v_371_axes_0, x = v_369_cast_fp16)[name = tensor<string, []>("v_371_cast_fp16")];
            tensor<int32, [4]> var_7601_begin_0 = const()[name = tensor<string, []>("op_7601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7601_end_0 = const()[name = tensor<string, []>("op_7601_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7601_end_mask_0 = const()[name = tensor<string, []>("op_7601_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7601_cast_fp16 = slice_by_index(begin = var_7601_begin_0, end = var_7601_end_0, end_mask = var_7601_end_mask_0, x = q_615_cast_fp16)[name = tensor<string, []>("op_7601_cast_fp16")];
            tensor<int32, [4]> var_7605_begin_0 = const()[name = tensor<string, []>("op_7605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7605_end_0 = const()[name = tensor<string, []>("op_7605_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7605_end_mask_0 = const()[name = tensor<string, []>("op_7605_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7605_cast_fp16 = slice_by_index(begin = var_7605_begin_0, end = var_7605_end_0, end_mask = var_7605_end_mask_0, x = k_369_cast_fp16)[name = tensor<string, []>("op_7605_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = tensor<string, []>("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = tensor<string, []>("op_7609_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = tensor<string, []>("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = q_615_cast_fp16)[name = tensor<string, []>("op_7609_cast_fp16")];
            tensor<int32, [4]> var_7613_begin_0 = const()[name = tensor<string, []>("op_7613_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7613_end_0 = const()[name = tensor<string, []>("op_7613_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7613_end_mask_0 = const()[name = tensor<string, []>("op_7613_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7613_cast_fp16 = slice_by_index(begin = var_7613_begin_0, end = var_7613_end_0, end_mask = var_7613_end_mask_0, x = k_369_cast_fp16)[name = tensor<string, []>("op_7613_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7615_cast_fp16 = mul(x = var_7609_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7615_cast_fp16")];
            tensor<int32, [5]> var_7623 = const()[name = tensor<string, []>("op_7623"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_843_cast_fp16 = reshape(shape = var_7623, x = var_7609_cast_fp16)[name = tensor<string, []>("x_843_cast_fp16")];
            tensor<int32, [2]> var_7625_split_sizes_0 = const()[name = tensor<string, []>("op_7625_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7625_axis_0 = const()[name = tensor<string, []>("op_7625_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7625_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7625_cast_fp16_1 = split(axis = var_7625_axis_0, split_sizes = var_7625_split_sizes_0, x = x_843_cast_fp16)[name = tensor<string, []>("op_7625_cast_fp16")];
            tensor<int32, [1]> squeeze_244_axes_0 = const()[name = tensor<string, []>("squeeze_244_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_244_cast_fp16 = squeeze(axes = squeeze_244_axes_0, x = var_7625_cast_fp16_0)[name = tensor<string, []>("squeeze_244_cast_fp16")];
            tensor<int32, [1]> squeeze_245_axes_0 = const()[name = tensor<string, []>("squeeze_245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_245_cast_fp16 = squeeze(axes = squeeze_245_axes_0, x = var_7625_cast_fp16_1)[name = tensor<string, []>("squeeze_245_cast_fp16")];
            tensor<fp16, []> const_1003_promoted_to_fp16 = const()[name = tensor<string, []>("const_1003_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7628_cast_fp16 = mul(x = squeeze_245_cast_fp16, y = const_1003_promoted_to_fp16)[name = tensor<string, []>("op_7628_cast_fp16")];
            tensor<int32, []> x_845_axis_0 = const()[name = tensor<string, []>("x_845_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_845_cast_fp16 = stack(axis = x_845_axis_0, values = (var_7628_cast_fp16, squeeze_244_cast_fp16))[name = tensor<string, []>("x_845_cast_fp16")];
            tensor<int32, [4]> var_7634 = const()[name = tensor<string, []>("op_7634"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7635_cast_fp16 = reshape(shape = var_7634, x = x_845_cast_fp16)[name = tensor<string, []>("op_7635_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7636_cast_fp16 = mul(x = var_7635_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7636_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_123_cast_fp16 = add(x = var_7615_cast_fp16, y = var_7636_cast_fp16)[name = tensor<string, []>("q_patches_123_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7638_cast_fp16 = mul(x = var_7613_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7638_cast_fp16")];
            tensor<int32, [5]> var_7646 = const()[name = tensor<string, []>("op_7646"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_849_cast_fp16 = reshape(shape = var_7646, x = var_7613_cast_fp16)[name = tensor<string, []>("x_849_cast_fp16")];
            tensor<int32, [2]> var_7648_split_sizes_0 = const()[name = tensor<string, []>("op_7648_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7648_axis_0 = const()[name = tensor<string, []>("op_7648_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7648_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7648_cast_fp16_1 = split(axis = var_7648_axis_0, split_sizes = var_7648_split_sizes_0, x = x_849_cast_fp16)[name = tensor<string, []>("op_7648_cast_fp16")];
            tensor<int32, [1]> squeeze_246_axes_0 = const()[name = tensor<string, []>("squeeze_246_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_246_cast_fp16 = squeeze(axes = squeeze_246_axes_0, x = var_7648_cast_fp16_0)[name = tensor<string, []>("squeeze_246_cast_fp16")];
            tensor<int32, [1]> squeeze_247_axes_0 = const()[name = tensor<string, []>("squeeze_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_247_cast_fp16 = squeeze(axes = squeeze_247_axes_0, x = var_7648_cast_fp16_1)[name = tensor<string, []>("squeeze_247_cast_fp16")];
            tensor<fp16, []> const_1011_promoted_to_fp16 = const()[name = tensor<string, []>("const_1011_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7651_cast_fp16 = mul(x = squeeze_247_cast_fp16, y = const_1011_promoted_to_fp16)[name = tensor<string, []>("op_7651_cast_fp16")];
            tensor<int32, []> x_851_axis_0 = const()[name = tensor<string, []>("x_851_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_851_cast_fp16 = stack(axis = x_851_axis_0, values = (var_7651_cast_fp16, squeeze_246_cast_fp16))[name = tensor<string, []>("x_851_cast_fp16")];
            tensor<int32, [4]> var_7657 = const()[name = tensor<string, []>("op_7657"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7658_cast_fp16 = reshape(shape = var_7657, x = x_851_cast_fp16)[name = tensor<string, []>("op_7658_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7659_cast_fp16 = mul(x = var_7658_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7659_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_123_cast_fp16 = add(x = var_7638_cast_fp16, y = var_7659_cast_fp16)[name = tensor<string, []>("k_patches_123_cast_fp16")];
            tensor<bool, []> var_7662_interleave_0 = const()[name = tensor<string, []>("op_7662_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7662_cast_fp16 = concat(axis = var_20, interleave = var_7662_interleave_0, values = (var_7601_cast_fp16, q_patches_123_cast_fp16))[name = tensor<string, []>("op_7662_cast_fp16")];
            tensor<bool, []> var_7665_interleave_0 = const()[name = tensor<string, []>("op_7665_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7665_cast_fp16 = concat(axis = var_20, interleave = var_7665_interleave_0, values = (var_7605_cast_fp16, k_patches_123_cast_fp16))[name = tensor<string, []>("op_7665_cast_fp16")];
            tensor<fp16, []> var_7667_to_fp16 = const()[name = tensor<string, []>("op_7667_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_619_cast_fp16 = mul(x = var_7662_cast_fp16, y = var_7667_to_fp16)[name = tensor<string, []>("q_619_cast_fp16")];
            tensor<bool, []> attn_245_transpose_x_1 = const()[name = tensor<string, []>("attn_245_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_245_transpose_y_1 = const()[name = tensor<string, []>("attn_245_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_245_cast_fp16 = matmul(transpose_x = attn_245_transpose_x_1, transpose_y = attn_245_transpose_y_1, x = q_619_cast_fp16, y = var_7665_cast_fp16)[name = tensor<string, []>("attn_245_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_247_cast_fp16 = softmax(axis = var_21, x = attn_245_cast_fp16)[name = tensor<string, []>("attn_247_cast_fp16")];
            tensor<bool, []> var_7672_transpose_x_0 = const()[name = tensor<string, []>("op_7672_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7672_transpose_y_0 = const()[name = tensor<string, []>("op_7672_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7672_cast_fp16 = matmul(transpose_x = var_7672_transpose_x_0, transpose_y = var_7672_transpose_y_0, x = attn_247_cast_fp16, y = v_371_cast_fp16)[name = tensor<string, []>("op_7672_cast_fp16")];
            tensor<int32, [1]> var_7673_axes_0 = const()[name = tensor<string, []>("op_7673_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7673_cast_fp16 = squeeze(axes = var_7673_axes_0, x = var_7672_cast_fp16)[name = tensor<string, []>("op_7673_cast_fp16")];
            tensor<string, []> var_7679_pad_type_0 = const()[name = tensor<string, []>("op_7679_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7679_strides_0 = const()[name = tensor<string, []>("op_7679_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7679_pad_0 = const()[name = tensor<string, []>("op_7679_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7679_dilations_0 = const()[name = tensor<string, []>("op_7679_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7679_groups_0 = const()[name = tensor<string, []>("op_7679_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73041728)))];
            tensor<fp16, [1, 64, 1, 197]> var_7679_cast_fp16 = conv(dilations = var_7679_dilations_0, groups = var_7679_groups_0, pad = var_7679_pad_0, pad_type = var_7679_pad_type_0, strides = var_7679_strides_0, weight = model_blocks_5_attn_q_projs_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7679_cast_fp16")];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = tensor<string, []>("op_7680_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7680_cast_fp16 = squeeze(axes = var_7680_axes_0, x = var_7679_cast_fp16)[name = tensor<string, []>("op_7680_cast_fp16")];
            tensor<int32, [3]> q_621_perm_0 = const()[name = tensor<string, []>("q_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7687_pad_type_0 = const()[name = tensor<string, []>("op_7687_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7687_strides_0 = const()[name = tensor<string, []>("op_7687_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7687_pad_0 = const()[name = tensor<string, []>("op_7687_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7687_dilations_0 = const()[name = tensor<string, []>("op_7687_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7687_groups_0 = const()[name = tensor<string, []>("op_7687_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73140096)))];
            tensor<fp16, [1, 64, 1, 197]> var_7687_cast_fp16 = conv(dilations = var_7687_dilations_0, groups = var_7687_groups_0, pad = var_7687_pad_0, pad_type = var_7687_pad_type_0, strides = var_7687_strides_0, weight = model_blocks_5_attn_k_projs_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7687_cast_fp16")];
            tensor<int32, [1]> var_7688_axes_0 = const()[name = tensor<string, []>("op_7688_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7688_cast_fp16 = squeeze(axes = var_7688_axes_0, x = var_7687_cast_fp16)[name = tensor<string, []>("op_7688_cast_fp16")];
            tensor<int32, [3]> k_373_perm_0 = const()[name = tensor<string, []>("k_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7695_pad_type_0 = const()[name = tensor<string, []>("op_7695_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7695_strides_0 = const()[name = tensor<string, []>("op_7695_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7695_pad_0 = const()[name = tensor<string, []>("op_7695_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7695_dilations_0 = const()[name = tensor<string, []>("op_7695_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7695_groups_0 = const()[name = tensor<string, []>("op_7695_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73238464)))];
            tensor<fp16, [1, 64, 1, 197]> var_7695_cast_fp16 = conv(dilations = var_7695_dilations_0, groups = var_7695_groups_0, pad = var_7695_pad_0, pad_type = var_7695_pad_type_0, strides = var_7695_strides_0, weight = model_blocks_5_attn_v_projs_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7695_cast_fp16")];
            tensor<int32, [1]> var_7696_axes_0 = const()[name = tensor<string, []>("op_7696_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7696_cast_fp16 = squeeze(axes = var_7696_axes_0, x = var_7695_cast_fp16)[name = tensor<string, []>("op_7696_cast_fp16")];
            tensor<int32, [3]> v_373_perm_0 = const()[name = tensor<string, []>("v_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73336832)))];
            tensor<fp16, [1, 197, 64]> q_621_cast_fp16 = transpose(perm = q_621_perm_0, x = var_7680_cast_fp16)[name = tensor<string, []>("transpose_293")];
            tensor<fp16, [1, 197, 64]> q_623_cast_fp16 = add(x = q_621_cast_fp16, y = model_blocks_5_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_623_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73337024)))];
            tensor<fp16, [1, 197, 64]> v_373_cast_fp16 = transpose(perm = v_373_perm_0, x = var_7696_cast_fp16)[name = tensor<string, []>("transpose_291")];
            tensor<fp16, [1, 197, 64]> v_375_cast_fp16 = add(x = v_373_cast_fp16, y = model_blocks_5_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_375_cast_fp16")];
            tensor<int32, [1]> q_625_axes_0 = const()[name = tensor<string, []>("q_625_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_625_cast_fp16 = expand_dims(axes = q_625_axes_0, x = q_623_cast_fp16)[name = tensor<string, []>("q_625_cast_fp16")];
            tensor<int32, [1]> k_375_axes_0 = const()[name = tensor<string, []>("k_375_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_373_cast_fp16 = transpose(perm = k_373_perm_0, x = var_7688_cast_fp16)[name = tensor<string, []>("transpose_292")];
            tensor<fp16, [1, 1, 197, 64]> k_375_cast_fp16 = expand_dims(axes = k_375_axes_0, x = k_373_cast_fp16)[name = tensor<string, []>("k_375_cast_fp16")];
            tensor<int32, [1]> v_377_axes_0 = const()[name = tensor<string, []>("v_377_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_377_cast_fp16 = expand_dims(axes = v_377_axes_0, x = v_375_cast_fp16)[name = tensor<string, []>("v_377_cast_fp16")];
            tensor<int32, [4]> var_7705_begin_0 = const()[name = tensor<string, []>("op_7705_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7705_end_0 = const()[name = tensor<string, []>("op_7705_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7705_end_mask_0 = const()[name = tensor<string, []>("op_7705_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7705_cast_fp16 = slice_by_index(begin = var_7705_begin_0, end = var_7705_end_0, end_mask = var_7705_end_mask_0, x = q_625_cast_fp16)[name = tensor<string, []>("op_7705_cast_fp16")];
            tensor<int32, [4]> var_7709_begin_0 = const()[name = tensor<string, []>("op_7709_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7709_end_0 = const()[name = tensor<string, []>("op_7709_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7709_end_mask_0 = const()[name = tensor<string, []>("op_7709_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7709_cast_fp16 = slice_by_index(begin = var_7709_begin_0, end = var_7709_end_0, end_mask = var_7709_end_mask_0, x = k_375_cast_fp16)[name = tensor<string, []>("op_7709_cast_fp16")];
            tensor<int32, [4]> var_7713_begin_0 = const()[name = tensor<string, []>("op_7713_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7713_end_0 = const()[name = tensor<string, []>("op_7713_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7713_end_mask_0 = const()[name = tensor<string, []>("op_7713_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7713_cast_fp16 = slice_by_index(begin = var_7713_begin_0, end = var_7713_end_0, end_mask = var_7713_end_mask_0, x = q_625_cast_fp16)[name = tensor<string, []>("op_7713_cast_fp16")];
            tensor<int32, [4]> var_7717_begin_0 = const()[name = tensor<string, []>("op_7717_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7717_end_0 = const()[name = tensor<string, []>("op_7717_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7717_end_mask_0 = const()[name = tensor<string, []>("op_7717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7717_cast_fp16 = slice_by_index(begin = var_7717_begin_0, end = var_7717_end_0, end_mask = var_7717_end_mask_0, x = k_375_cast_fp16)[name = tensor<string, []>("op_7717_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7719_cast_fp16 = mul(x = var_7713_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7719_cast_fp16")];
            tensor<int32, [5]> var_7727 = const()[name = tensor<string, []>("op_7727"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_855_cast_fp16 = reshape(shape = var_7727, x = var_7713_cast_fp16)[name = tensor<string, []>("x_855_cast_fp16")];
            tensor<int32, [2]> var_7729_split_sizes_0 = const()[name = tensor<string, []>("op_7729_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7729_axis_0 = const()[name = tensor<string, []>("op_7729_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7729_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7729_cast_fp16_1 = split(axis = var_7729_axis_0, split_sizes = var_7729_split_sizes_0, x = x_855_cast_fp16)[name = tensor<string, []>("op_7729_cast_fp16")];
            tensor<int32, [1]> squeeze_248_axes_0 = const()[name = tensor<string, []>("squeeze_248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_248_cast_fp16 = squeeze(axes = squeeze_248_axes_0, x = var_7729_cast_fp16_0)[name = tensor<string, []>("squeeze_248_cast_fp16")];
            tensor<int32, [1]> squeeze_249_axes_0 = const()[name = tensor<string, []>("squeeze_249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_249_cast_fp16 = squeeze(axes = squeeze_249_axes_0, x = var_7729_cast_fp16_1)[name = tensor<string, []>("squeeze_249_cast_fp16")];
            tensor<fp16, []> const_1019_promoted_to_fp16 = const()[name = tensor<string, []>("const_1019_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7732_cast_fp16 = mul(x = squeeze_249_cast_fp16, y = const_1019_promoted_to_fp16)[name = tensor<string, []>("op_7732_cast_fp16")];
            tensor<int32, []> x_857_axis_0 = const()[name = tensor<string, []>("x_857_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_857_cast_fp16 = stack(axis = x_857_axis_0, values = (var_7732_cast_fp16, squeeze_248_cast_fp16))[name = tensor<string, []>("x_857_cast_fp16")];
            tensor<int32, [4]> var_7738 = const()[name = tensor<string, []>("op_7738"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7739_cast_fp16 = reshape(shape = var_7738, x = x_857_cast_fp16)[name = tensor<string, []>("op_7739_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7740_cast_fp16 = mul(x = var_7739_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7740_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_125_cast_fp16 = add(x = var_7719_cast_fp16, y = var_7740_cast_fp16)[name = tensor<string, []>("q_patches_125_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7742_cast_fp16 = mul(x = var_7717_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7742_cast_fp16")];
            tensor<int32, [5]> var_7750 = const()[name = tensor<string, []>("op_7750"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_861_cast_fp16 = reshape(shape = var_7750, x = var_7717_cast_fp16)[name = tensor<string, []>("x_861_cast_fp16")];
            tensor<int32, [2]> var_7752_split_sizes_0 = const()[name = tensor<string, []>("op_7752_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7752_axis_0 = const()[name = tensor<string, []>("op_7752_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7752_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7752_cast_fp16_1 = split(axis = var_7752_axis_0, split_sizes = var_7752_split_sizes_0, x = x_861_cast_fp16)[name = tensor<string, []>("op_7752_cast_fp16")];
            tensor<int32, [1]> squeeze_250_axes_0 = const()[name = tensor<string, []>("squeeze_250_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_250_cast_fp16 = squeeze(axes = squeeze_250_axes_0, x = var_7752_cast_fp16_0)[name = tensor<string, []>("squeeze_250_cast_fp16")];
            tensor<int32, [1]> squeeze_251_axes_0 = const()[name = tensor<string, []>("squeeze_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_251_cast_fp16 = squeeze(axes = squeeze_251_axes_0, x = var_7752_cast_fp16_1)[name = tensor<string, []>("squeeze_251_cast_fp16")];
            tensor<fp16, []> const_1027_promoted_to_fp16 = const()[name = tensor<string, []>("const_1027_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7755_cast_fp16 = mul(x = squeeze_251_cast_fp16, y = const_1027_promoted_to_fp16)[name = tensor<string, []>("op_7755_cast_fp16")];
            tensor<int32, []> x_863_axis_0 = const()[name = tensor<string, []>("x_863_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_863_cast_fp16 = stack(axis = x_863_axis_0, values = (var_7755_cast_fp16, squeeze_250_cast_fp16))[name = tensor<string, []>("x_863_cast_fp16")];
            tensor<int32, [4]> var_7761 = const()[name = tensor<string, []>("op_7761"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7762_cast_fp16 = reshape(shape = var_7761, x = x_863_cast_fp16)[name = tensor<string, []>("op_7762_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7763_cast_fp16 = mul(x = var_7762_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7763_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_125_cast_fp16 = add(x = var_7742_cast_fp16, y = var_7763_cast_fp16)[name = tensor<string, []>("k_patches_125_cast_fp16")];
            tensor<bool, []> var_7766_interleave_0 = const()[name = tensor<string, []>("op_7766_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7766_cast_fp16 = concat(axis = var_20, interleave = var_7766_interleave_0, values = (var_7705_cast_fp16, q_patches_125_cast_fp16))[name = tensor<string, []>("op_7766_cast_fp16")];
            tensor<bool, []> var_7769_interleave_0 = const()[name = tensor<string, []>("op_7769_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7769_cast_fp16 = concat(axis = var_20, interleave = var_7769_interleave_0, values = (var_7709_cast_fp16, k_patches_125_cast_fp16))[name = tensor<string, []>("op_7769_cast_fp16")];
            tensor<fp16, []> var_7771_to_fp16 = const()[name = tensor<string, []>("op_7771_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_629_cast_fp16 = mul(x = var_7766_cast_fp16, y = var_7771_to_fp16)[name = tensor<string, []>("q_629_cast_fp16")];
            tensor<bool, []> attn_249_transpose_x_1 = const()[name = tensor<string, []>("attn_249_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_249_transpose_y_1 = const()[name = tensor<string, []>("attn_249_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_249_cast_fp16 = matmul(transpose_x = attn_249_transpose_x_1, transpose_y = attn_249_transpose_y_1, x = q_629_cast_fp16, y = var_7769_cast_fp16)[name = tensor<string, []>("attn_249_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_251_cast_fp16 = softmax(axis = var_21, x = attn_249_cast_fp16)[name = tensor<string, []>("attn_251_cast_fp16")];
            tensor<bool, []> var_7776_transpose_x_0 = const()[name = tensor<string, []>("op_7776_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7776_transpose_y_0 = const()[name = tensor<string, []>("op_7776_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7776_cast_fp16 = matmul(transpose_x = var_7776_transpose_x_0, transpose_y = var_7776_transpose_y_0, x = attn_251_cast_fp16, y = v_377_cast_fp16)[name = tensor<string, []>("op_7776_cast_fp16")];
            tensor<int32, [1]> var_7777_axes_0 = const()[name = tensor<string, []>("op_7777_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7777_cast_fp16 = squeeze(axes = var_7777_axes_0, x = var_7776_cast_fp16)[name = tensor<string, []>("op_7777_cast_fp16")];
            tensor<string, []> var_7783_pad_type_0 = const()[name = tensor<string, []>("op_7783_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7783_strides_0 = const()[name = tensor<string, []>("op_7783_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7783_pad_0 = const()[name = tensor<string, []>("op_7783_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7783_dilations_0 = const()[name = tensor<string, []>("op_7783_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7783_groups_0 = const()[name = tensor<string, []>("op_7783_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73337216)))];
            tensor<fp16, [1, 64, 1, 197]> var_7783_cast_fp16 = conv(dilations = var_7783_dilations_0, groups = var_7783_groups_0, pad = var_7783_pad_0, pad_type = var_7783_pad_type_0, strides = var_7783_strides_0, weight = model_blocks_5_attn_q_projs_3_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7783_cast_fp16")];
            tensor<int32, [1]> var_7784_axes_0 = const()[name = tensor<string, []>("op_7784_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7784_cast_fp16 = squeeze(axes = var_7784_axes_0, x = var_7783_cast_fp16)[name = tensor<string, []>("op_7784_cast_fp16")];
            tensor<int32, [3]> q_631_perm_0 = const()[name = tensor<string, []>("q_631_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7791_pad_type_0 = const()[name = tensor<string, []>("op_7791_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7791_strides_0 = const()[name = tensor<string, []>("op_7791_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7791_pad_0 = const()[name = tensor<string, []>("op_7791_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7791_dilations_0 = const()[name = tensor<string, []>("op_7791_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7791_groups_0 = const()[name = tensor<string, []>("op_7791_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73435584)))];
            tensor<fp16, [1, 64, 1, 197]> var_7791_cast_fp16 = conv(dilations = var_7791_dilations_0, groups = var_7791_groups_0, pad = var_7791_pad_0, pad_type = var_7791_pad_type_0, strides = var_7791_strides_0, weight = model_blocks_5_attn_k_projs_3_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7791_cast_fp16")];
            tensor<int32, [1]> var_7792_axes_0 = const()[name = tensor<string, []>("op_7792_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7792_cast_fp16 = squeeze(axes = var_7792_axes_0, x = var_7791_cast_fp16)[name = tensor<string, []>("op_7792_cast_fp16")];
            tensor<int32, [3]> k_379_perm_0 = const()[name = tensor<string, []>("k_379_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7799_pad_type_0 = const()[name = tensor<string, []>("op_7799_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7799_strides_0 = const()[name = tensor<string, []>("op_7799_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7799_pad_0 = const()[name = tensor<string, []>("op_7799_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7799_dilations_0 = const()[name = tensor<string, []>("op_7799_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7799_groups_0 = const()[name = tensor<string, []>("op_7799_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73533952)))];
            tensor<fp16, [1, 64, 1, 197]> var_7799_cast_fp16 = conv(dilations = var_7799_dilations_0, groups = var_7799_groups_0, pad = var_7799_pad_0, pad_type = var_7799_pad_type_0, strides = var_7799_strides_0, weight = model_blocks_5_attn_v_projs_3_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7799_cast_fp16")];
            tensor<int32, [1]> var_7800_axes_0 = const()[name = tensor<string, []>("op_7800_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7800_cast_fp16 = squeeze(axes = var_7800_axes_0, x = var_7799_cast_fp16)[name = tensor<string, []>("op_7800_cast_fp16")];
            tensor<int32, [3]> v_379_perm_0 = const()[name = tensor<string, []>("v_379_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73632320)))];
            tensor<fp16, [1, 197, 64]> q_631_cast_fp16 = transpose(perm = q_631_perm_0, x = var_7784_cast_fp16)[name = tensor<string, []>("transpose_290")];
            tensor<fp16, [1, 197, 64]> q_633_cast_fp16 = add(x = q_631_cast_fp16, y = model_blocks_5_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_633_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73632512)))];
            tensor<fp16, [1, 197, 64]> v_379_cast_fp16 = transpose(perm = v_379_perm_0, x = var_7800_cast_fp16)[name = tensor<string, []>("transpose_288")];
            tensor<fp16, [1, 197, 64]> v_381_cast_fp16 = add(x = v_379_cast_fp16, y = model_blocks_5_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_381_cast_fp16")];
            tensor<int32, [1]> q_635_axes_0 = const()[name = tensor<string, []>("q_635_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_635_cast_fp16 = expand_dims(axes = q_635_axes_0, x = q_633_cast_fp16)[name = tensor<string, []>("q_635_cast_fp16")];
            tensor<int32, [1]> k_381_axes_0 = const()[name = tensor<string, []>("k_381_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_379_cast_fp16 = transpose(perm = k_379_perm_0, x = var_7792_cast_fp16)[name = tensor<string, []>("transpose_289")];
            tensor<fp16, [1, 1, 197, 64]> k_381_cast_fp16 = expand_dims(axes = k_381_axes_0, x = k_379_cast_fp16)[name = tensor<string, []>("k_381_cast_fp16")];
            tensor<int32, [1]> v_383_axes_0 = const()[name = tensor<string, []>("v_383_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_383_cast_fp16 = expand_dims(axes = v_383_axes_0, x = v_381_cast_fp16)[name = tensor<string, []>("v_383_cast_fp16")];
            tensor<int32, [4]> var_7809_begin_0 = const()[name = tensor<string, []>("op_7809_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7809_end_0 = const()[name = tensor<string, []>("op_7809_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7809_end_mask_0 = const()[name = tensor<string, []>("op_7809_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7809_cast_fp16 = slice_by_index(begin = var_7809_begin_0, end = var_7809_end_0, end_mask = var_7809_end_mask_0, x = q_635_cast_fp16)[name = tensor<string, []>("op_7809_cast_fp16")];
            tensor<int32, [4]> var_7813_begin_0 = const()[name = tensor<string, []>("op_7813_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7813_end_0 = const()[name = tensor<string, []>("op_7813_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7813_end_mask_0 = const()[name = tensor<string, []>("op_7813_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7813_cast_fp16 = slice_by_index(begin = var_7813_begin_0, end = var_7813_end_0, end_mask = var_7813_end_mask_0, x = k_381_cast_fp16)[name = tensor<string, []>("op_7813_cast_fp16")];
            tensor<int32, [4]> var_7817_begin_0 = const()[name = tensor<string, []>("op_7817_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7817_end_0 = const()[name = tensor<string, []>("op_7817_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7817_end_mask_0 = const()[name = tensor<string, []>("op_7817_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7817_cast_fp16 = slice_by_index(begin = var_7817_begin_0, end = var_7817_end_0, end_mask = var_7817_end_mask_0, x = q_635_cast_fp16)[name = tensor<string, []>("op_7817_cast_fp16")];
            tensor<int32, [4]> var_7821_begin_0 = const()[name = tensor<string, []>("op_7821_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7821_end_0 = const()[name = tensor<string, []>("op_7821_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7821_end_mask_0 = const()[name = tensor<string, []>("op_7821_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7821_cast_fp16 = slice_by_index(begin = var_7821_begin_0, end = var_7821_end_0, end_mask = var_7821_end_mask_0, x = k_381_cast_fp16)[name = tensor<string, []>("op_7821_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7823_cast_fp16 = mul(x = var_7817_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7823_cast_fp16")];
            tensor<int32, [5]> var_7831 = const()[name = tensor<string, []>("op_7831"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_867_cast_fp16 = reshape(shape = var_7831, x = var_7817_cast_fp16)[name = tensor<string, []>("x_867_cast_fp16")];
            tensor<int32, [2]> var_7833_split_sizes_0 = const()[name = tensor<string, []>("op_7833_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7833_axis_0 = const()[name = tensor<string, []>("op_7833_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7833_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7833_cast_fp16_1 = split(axis = var_7833_axis_0, split_sizes = var_7833_split_sizes_0, x = x_867_cast_fp16)[name = tensor<string, []>("op_7833_cast_fp16")];
            tensor<int32, [1]> squeeze_252_axes_0 = const()[name = tensor<string, []>("squeeze_252_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_252_cast_fp16 = squeeze(axes = squeeze_252_axes_0, x = var_7833_cast_fp16_0)[name = tensor<string, []>("squeeze_252_cast_fp16")];
            tensor<int32, [1]> squeeze_253_axes_0 = const()[name = tensor<string, []>("squeeze_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_253_cast_fp16 = squeeze(axes = squeeze_253_axes_0, x = var_7833_cast_fp16_1)[name = tensor<string, []>("squeeze_253_cast_fp16")];
            tensor<fp16, []> const_1035_promoted_to_fp16 = const()[name = tensor<string, []>("const_1035_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7836_cast_fp16 = mul(x = squeeze_253_cast_fp16, y = const_1035_promoted_to_fp16)[name = tensor<string, []>("op_7836_cast_fp16")];
            tensor<int32, []> x_869_axis_0 = const()[name = tensor<string, []>("x_869_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_869_cast_fp16 = stack(axis = x_869_axis_0, values = (var_7836_cast_fp16, squeeze_252_cast_fp16))[name = tensor<string, []>("x_869_cast_fp16")];
            tensor<int32, [4]> var_7842 = const()[name = tensor<string, []>("op_7842"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7843_cast_fp16 = reshape(shape = var_7842, x = x_869_cast_fp16)[name = tensor<string, []>("op_7843_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7844_cast_fp16 = mul(x = var_7843_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7844_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_127_cast_fp16 = add(x = var_7823_cast_fp16, y = var_7844_cast_fp16)[name = tensor<string, []>("q_patches_127_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7846_cast_fp16 = mul(x = var_7821_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7846_cast_fp16")];
            tensor<int32, [5]> var_7854 = const()[name = tensor<string, []>("op_7854"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_873_cast_fp16 = reshape(shape = var_7854, x = var_7821_cast_fp16)[name = tensor<string, []>("x_873_cast_fp16")];
            tensor<int32, [2]> var_7856_split_sizes_0 = const()[name = tensor<string, []>("op_7856_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7856_axis_0 = const()[name = tensor<string, []>("op_7856_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7856_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7856_cast_fp16_1 = split(axis = var_7856_axis_0, split_sizes = var_7856_split_sizes_0, x = x_873_cast_fp16)[name = tensor<string, []>("op_7856_cast_fp16")];
            tensor<int32, [1]> squeeze_254_axes_0 = const()[name = tensor<string, []>("squeeze_254_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_254_cast_fp16 = squeeze(axes = squeeze_254_axes_0, x = var_7856_cast_fp16_0)[name = tensor<string, []>("squeeze_254_cast_fp16")];
            tensor<int32, [1]> squeeze_255_axes_0 = const()[name = tensor<string, []>("squeeze_255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_255_cast_fp16 = squeeze(axes = squeeze_255_axes_0, x = var_7856_cast_fp16_1)[name = tensor<string, []>("squeeze_255_cast_fp16")];
            tensor<fp16, []> const_1043_promoted_to_fp16 = const()[name = tensor<string, []>("const_1043_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7859_cast_fp16 = mul(x = squeeze_255_cast_fp16, y = const_1043_promoted_to_fp16)[name = tensor<string, []>("op_7859_cast_fp16")];
            tensor<int32, []> x_875_axis_0 = const()[name = tensor<string, []>("x_875_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_875_cast_fp16 = stack(axis = x_875_axis_0, values = (var_7859_cast_fp16, squeeze_254_cast_fp16))[name = tensor<string, []>("x_875_cast_fp16")];
            tensor<int32, [4]> var_7865 = const()[name = tensor<string, []>("op_7865"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7866_cast_fp16 = reshape(shape = var_7865, x = x_875_cast_fp16)[name = tensor<string, []>("op_7866_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7867_cast_fp16 = mul(x = var_7866_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7867_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_127_cast_fp16 = add(x = var_7846_cast_fp16, y = var_7867_cast_fp16)[name = tensor<string, []>("k_patches_127_cast_fp16")];
            tensor<bool, []> var_7870_interleave_0 = const()[name = tensor<string, []>("op_7870_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7870_cast_fp16 = concat(axis = var_20, interleave = var_7870_interleave_0, values = (var_7809_cast_fp16, q_patches_127_cast_fp16))[name = tensor<string, []>("op_7870_cast_fp16")];
            tensor<bool, []> var_7873_interleave_0 = const()[name = tensor<string, []>("op_7873_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7873_cast_fp16 = concat(axis = var_20, interleave = var_7873_interleave_0, values = (var_7813_cast_fp16, k_patches_127_cast_fp16))[name = tensor<string, []>("op_7873_cast_fp16")];
            tensor<fp16, []> var_7875_to_fp16 = const()[name = tensor<string, []>("op_7875_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_639_cast_fp16 = mul(x = var_7870_cast_fp16, y = var_7875_to_fp16)[name = tensor<string, []>("q_639_cast_fp16")];
            tensor<bool, []> attn_253_transpose_x_1 = const()[name = tensor<string, []>("attn_253_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_253_transpose_y_1 = const()[name = tensor<string, []>("attn_253_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_253_cast_fp16 = matmul(transpose_x = attn_253_transpose_x_1, transpose_y = attn_253_transpose_y_1, x = q_639_cast_fp16, y = var_7873_cast_fp16)[name = tensor<string, []>("attn_253_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_255_cast_fp16 = softmax(axis = var_21, x = attn_253_cast_fp16)[name = tensor<string, []>("attn_255_cast_fp16")];
            tensor<bool, []> var_7880_transpose_x_0 = const()[name = tensor<string, []>("op_7880_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7880_transpose_y_0 = const()[name = tensor<string, []>("op_7880_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7880_cast_fp16 = matmul(transpose_x = var_7880_transpose_x_0, transpose_y = var_7880_transpose_y_0, x = attn_255_cast_fp16, y = v_383_cast_fp16)[name = tensor<string, []>("op_7880_cast_fp16")];
            tensor<int32, [1]> var_7881_axes_0 = const()[name = tensor<string, []>("op_7881_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7881_cast_fp16 = squeeze(axes = var_7881_axes_0, x = var_7880_cast_fp16)[name = tensor<string, []>("op_7881_cast_fp16")];
            tensor<string, []> var_7887_pad_type_0 = const()[name = tensor<string, []>("op_7887_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7887_strides_0 = const()[name = tensor<string, []>("op_7887_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7887_pad_0 = const()[name = tensor<string, []>("op_7887_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7887_dilations_0 = const()[name = tensor<string, []>("op_7887_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7887_groups_0 = const()[name = tensor<string, []>("op_7887_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73632704)))];
            tensor<fp16, [1, 64, 1, 197]> var_7887_cast_fp16 = conv(dilations = var_7887_dilations_0, groups = var_7887_groups_0, pad = var_7887_pad_0, pad_type = var_7887_pad_type_0, strides = var_7887_strides_0, weight = model_blocks_5_attn_q_projs_4_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7887_cast_fp16")];
            tensor<int32, [1]> var_7888_axes_0 = const()[name = tensor<string, []>("op_7888_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7888_cast_fp16 = squeeze(axes = var_7888_axes_0, x = var_7887_cast_fp16)[name = tensor<string, []>("op_7888_cast_fp16")];
            tensor<int32, [3]> q_641_perm_0 = const()[name = tensor<string, []>("q_641_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7895_pad_type_0 = const()[name = tensor<string, []>("op_7895_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7895_strides_0 = const()[name = tensor<string, []>("op_7895_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7895_pad_0 = const()[name = tensor<string, []>("op_7895_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7895_dilations_0 = const()[name = tensor<string, []>("op_7895_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7895_groups_0 = const()[name = tensor<string, []>("op_7895_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73731072)))];
            tensor<fp16, [1, 64, 1, 197]> var_7895_cast_fp16 = conv(dilations = var_7895_dilations_0, groups = var_7895_groups_0, pad = var_7895_pad_0, pad_type = var_7895_pad_type_0, strides = var_7895_strides_0, weight = model_blocks_5_attn_k_projs_4_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7895_cast_fp16")];
            tensor<int32, [1]> var_7896_axes_0 = const()[name = tensor<string, []>("op_7896_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7896_cast_fp16 = squeeze(axes = var_7896_axes_0, x = var_7895_cast_fp16)[name = tensor<string, []>("op_7896_cast_fp16")];
            tensor<int32, [3]> k_385_perm_0 = const()[name = tensor<string, []>("k_385_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7903_pad_type_0 = const()[name = tensor<string, []>("op_7903_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7903_strides_0 = const()[name = tensor<string, []>("op_7903_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7903_pad_0 = const()[name = tensor<string, []>("op_7903_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7903_dilations_0 = const()[name = tensor<string, []>("op_7903_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7903_groups_0 = const()[name = tensor<string, []>("op_7903_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73829440)))];
            tensor<fp16, [1, 64, 1, 197]> var_7903_cast_fp16 = conv(dilations = var_7903_dilations_0, groups = var_7903_groups_0, pad = var_7903_pad_0, pad_type = var_7903_pad_type_0, strides = var_7903_strides_0, weight = model_blocks_5_attn_v_projs_4_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7903_cast_fp16")];
            tensor<int32, [1]> var_7904_axes_0 = const()[name = tensor<string, []>("op_7904_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7904_cast_fp16 = squeeze(axes = var_7904_axes_0, x = var_7903_cast_fp16)[name = tensor<string, []>("op_7904_cast_fp16")];
            tensor<int32, [3]> v_385_perm_0 = const()[name = tensor<string, []>("v_385_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73927808)))];
            tensor<fp16, [1, 197, 64]> q_641_cast_fp16 = transpose(perm = q_641_perm_0, x = var_7888_cast_fp16)[name = tensor<string, []>("transpose_287")];
            tensor<fp16, [1, 197, 64]> q_643_cast_fp16 = add(x = q_641_cast_fp16, y = model_blocks_5_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_643_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73928000)))];
            tensor<fp16, [1, 197, 64]> v_385_cast_fp16 = transpose(perm = v_385_perm_0, x = var_7904_cast_fp16)[name = tensor<string, []>("transpose_285")];
            tensor<fp16, [1, 197, 64]> v_387_cast_fp16 = add(x = v_385_cast_fp16, y = model_blocks_5_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_387_cast_fp16")];
            tensor<int32, [1]> q_645_axes_0 = const()[name = tensor<string, []>("q_645_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_645_cast_fp16 = expand_dims(axes = q_645_axes_0, x = q_643_cast_fp16)[name = tensor<string, []>("q_645_cast_fp16")];
            tensor<int32, [1]> k_387_axes_0 = const()[name = tensor<string, []>("k_387_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_385_cast_fp16 = transpose(perm = k_385_perm_0, x = var_7896_cast_fp16)[name = tensor<string, []>("transpose_286")];
            tensor<fp16, [1, 1, 197, 64]> k_387_cast_fp16 = expand_dims(axes = k_387_axes_0, x = k_385_cast_fp16)[name = tensor<string, []>("k_387_cast_fp16")];
            tensor<int32, [1]> v_389_axes_0 = const()[name = tensor<string, []>("v_389_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_389_cast_fp16 = expand_dims(axes = v_389_axes_0, x = v_387_cast_fp16)[name = tensor<string, []>("v_389_cast_fp16")];
            tensor<int32, [4]> var_7913_begin_0 = const()[name = tensor<string, []>("op_7913_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7913_end_0 = const()[name = tensor<string, []>("op_7913_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7913_end_mask_0 = const()[name = tensor<string, []>("op_7913_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7913_cast_fp16 = slice_by_index(begin = var_7913_begin_0, end = var_7913_end_0, end_mask = var_7913_end_mask_0, x = q_645_cast_fp16)[name = tensor<string, []>("op_7913_cast_fp16")];
            tensor<int32, [4]> var_7917_begin_0 = const()[name = tensor<string, []>("op_7917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7917_end_0 = const()[name = tensor<string, []>("op_7917_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_7917_end_mask_0 = const()[name = tensor<string, []>("op_7917_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_7917_cast_fp16 = slice_by_index(begin = var_7917_begin_0, end = var_7917_end_0, end_mask = var_7917_end_mask_0, x = k_387_cast_fp16)[name = tensor<string, []>("op_7917_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = tensor<string, []>("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = tensor<string, []>("op_7921_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = tensor<string, []>("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_645_cast_fp16)[name = tensor<string, []>("op_7921_cast_fp16")];
            tensor<int32, [4]> var_7925_begin_0 = const()[name = tensor<string, []>("op_7925_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7925_end_0 = const()[name = tensor<string, []>("op_7925_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_7925_end_mask_0 = const()[name = tensor<string, []>("op_7925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_7925_cast_fp16 = slice_by_index(begin = var_7925_begin_0, end = var_7925_end_0, end_mask = var_7925_end_mask_0, x = k_387_cast_fp16)[name = tensor<string, []>("op_7925_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7927_cast_fp16 = mul(x = var_7921_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7927_cast_fp16")];
            tensor<int32, [5]> var_7935 = const()[name = tensor<string, []>("op_7935"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_879_cast_fp16 = reshape(shape = var_7935, x = var_7921_cast_fp16)[name = tensor<string, []>("x_879_cast_fp16")];
            tensor<int32, [2]> var_7937_split_sizes_0 = const()[name = tensor<string, []>("op_7937_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7937_axis_0 = const()[name = tensor<string, []>("op_7937_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7937_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7937_cast_fp16_1 = split(axis = var_7937_axis_0, split_sizes = var_7937_split_sizes_0, x = x_879_cast_fp16)[name = tensor<string, []>("op_7937_cast_fp16")];
            tensor<int32, [1]> squeeze_256_axes_0 = const()[name = tensor<string, []>("squeeze_256_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_256_cast_fp16 = squeeze(axes = squeeze_256_axes_0, x = var_7937_cast_fp16_0)[name = tensor<string, []>("squeeze_256_cast_fp16")];
            tensor<int32, [1]> squeeze_257_axes_0 = const()[name = tensor<string, []>("squeeze_257_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_257_cast_fp16 = squeeze(axes = squeeze_257_axes_0, x = var_7937_cast_fp16_1)[name = tensor<string, []>("squeeze_257_cast_fp16")];
            tensor<fp16, []> const_1051_promoted_to_fp16 = const()[name = tensor<string, []>("const_1051_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7940_cast_fp16 = mul(x = squeeze_257_cast_fp16, y = const_1051_promoted_to_fp16)[name = tensor<string, []>("op_7940_cast_fp16")];
            tensor<int32, []> x_881_axis_0 = const()[name = tensor<string, []>("x_881_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_881_cast_fp16 = stack(axis = x_881_axis_0, values = (var_7940_cast_fp16, squeeze_256_cast_fp16))[name = tensor<string, []>("x_881_cast_fp16")];
            tensor<int32, [4]> var_7946 = const()[name = tensor<string, []>("op_7946"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7947_cast_fp16 = reshape(shape = var_7946, x = x_881_cast_fp16)[name = tensor<string, []>("op_7947_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7948_cast_fp16 = mul(x = var_7947_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_129_cast_fp16 = add(x = var_7927_cast_fp16, y = var_7948_cast_fp16)[name = tensor<string, []>("q_patches_129_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7950_cast_fp16 = mul(x = var_7925_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_7950_cast_fp16")];
            tensor<int32, [5]> var_7958 = const()[name = tensor<string, []>("op_7958"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_885_cast_fp16 = reshape(shape = var_7958, x = var_7925_cast_fp16)[name = tensor<string, []>("x_885_cast_fp16")];
            tensor<int32, [2]> var_7960_split_sizes_0 = const()[name = tensor<string, []>("op_7960_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7960_axis_0 = const()[name = tensor<string, []>("op_7960_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_7960_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_7960_cast_fp16_1 = split(axis = var_7960_axis_0, split_sizes = var_7960_split_sizes_0, x = x_885_cast_fp16)[name = tensor<string, []>("op_7960_cast_fp16")];
            tensor<int32, [1]> squeeze_258_axes_0 = const()[name = tensor<string, []>("squeeze_258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_258_cast_fp16 = squeeze(axes = squeeze_258_axes_0, x = var_7960_cast_fp16_0)[name = tensor<string, []>("squeeze_258_cast_fp16")];
            tensor<int32, [1]> squeeze_259_axes_0 = const()[name = tensor<string, []>("squeeze_259_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_259_cast_fp16 = squeeze(axes = squeeze_259_axes_0, x = var_7960_cast_fp16_1)[name = tensor<string, []>("squeeze_259_cast_fp16")];
            tensor<fp16, []> const_1059_promoted_to_fp16 = const()[name = tensor<string, []>("const_1059_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_7963_cast_fp16 = mul(x = squeeze_259_cast_fp16, y = const_1059_promoted_to_fp16)[name = tensor<string, []>("op_7963_cast_fp16")];
            tensor<int32, []> x_887_axis_0 = const()[name = tensor<string, []>("x_887_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_887_cast_fp16 = stack(axis = x_887_axis_0, values = (var_7963_cast_fp16, squeeze_258_cast_fp16))[name = tensor<string, []>("x_887_cast_fp16")];
            tensor<int32, [4]> var_7969 = const()[name = tensor<string, []>("op_7969"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_7970_cast_fp16 = reshape(shape = var_7969, x = x_887_cast_fp16)[name = tensor<string, []>("op_7970_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_7971_cast_fp16 = mul(x = var_7970_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_7971_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_129_cast_fp16 = add(x = var_7950_cast_fp16, y = var_7971_cast_fp16)[name = tensor<string, []>("k_patches_129_cast_fp16")];
            tensor<bool, []> var_7974_interleave_0 = const()[name = tensor<string, []>("op_7974_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7974_cast_fp16 = concat(axis = var_20, interleave = var_7974_interleave_0, values = (var_7913_cast_fp16, q_patches_129_cast_fp16))[name = tensor<string, []>("op_7974_cast_fp16")];
            tensor<bool, []> var_7977_interleave_0 = const()[name = tensor<string, []>("op_7977_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7977_cast_fp16 = concat(axis = var_20, interleave = var_7977_interleave_0, values = (var_7917_cast_fp16, k_patches_129_cast_fp16))[name = tensor<string, []>("op_7977_cast_fp16")];
            tensor<fp16, []> var_7979_to_fp16 = const()[name = tensor<string, []>("op_7979_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_649_cast_fp16 = mul(x = var_7974_cast_fp16, y = var_7979_to_fp16)[name = tensor<string, []>("q_649_cast_fp16")];
            tensor<bool, []> attn_257_transpose_x_1 = const()[name = tensor<string, []>("attn_257_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_257_transpose_y_1 = const()[name = tensor<string, []>("attn_257_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_257_cast_fp16 = matmul(transpose_x = attn_257_transpose_x_1, transpose_y = attn_257_transpose_y_1, x = q_649_cast_fp16, y = var_7977_cast_fp16)[name = tensor<string, []>("attn_257_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_259_cast_fp16 = softmax(axis = var_21, x = attn_257_cast_fp16)[name = tensor<string, []>("attn_259_cast_fp16")];
            tensor<bool, []> var_7984_transpose_x_0 = const()[name = tensor<string, []>("op_7984_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_7984_transpose_y_0 = const()[name = tensor<string, []>("op_7984_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_7984_cast_fp16 = matmul(transpose_x = var_7984_transpose_x_0, transpose_y = var_7984_transpose_y_0, x = attn_259_cast_fp16, y = v_389_cast_fp16)[name = tensor<string, []>("op_7984_cast_fp16")];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = tensor<string, []>("op_7985_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_7985_cast_fp16 = squeeze(axes = var_7985_axes_0, x = var_7984_cast_fp16)[name = tensor<string, []>("op_7985_cast_fp16")];
            tensor<string, []> var_7991_pad_type_0 = const()[name = tensor<string, []>("op_7991_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7991_strides_0 = const()[name = tensor<string, []>("op_7991_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7991_pad_0 = const()[name = tensor<string, []>("op_7991_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7991_dilations_0 = const()[name = tensor<string, []>("op_7991_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7991_groups_0 = const()[name = tensor<string, []>("op_7991_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73928192)))];
            tensor<fp16, [1, 64, 1, 197]> var_7991_cast_fp16 = conv(dilations = var_7991_dilations_0, groups = var_7991_groups_0, pad = var_7991_pad_0, pad_type = var_7991_pad_type_0, strides = var_7991_strides_0, weight = model_blocks_5_attn_q_projs_5_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7991_cast_fp16")];
            tensor<int32, [1]> var_7992_axes_0 = const()[name = tensor<string, []>("op_7992_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_7992_cast_fp16 = squeeze(axes = var_7992_axes_0, x = var_7991_cast_fp16)[name = tensor<string, []>("op_7992_cast_fp16")];
            tensor<int32, [3]> q_651_perm_0 = const()[name = tensor<string, []>("q_651_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_7999_pad_type_0 = const()[name = tensor<string, []>("op_7999_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_7999_strides_0 = const()[name = tensor<string, []>("op_7999_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7999_pad_0 = const()[name = tensor<string, []>("op_7999_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7999_dilations_0 = const()[name = tensor<string, []>("op_7999_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_7999_groups_0 = const()[name = tensor<string, []>("op_7999_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74026560)))];
            tensor<fp16, [1, 64, 1, 197]> var_7999_cast_fp16 = conv(dilations = var_7999_dilations_0, groups = var_7999_groups_0, pad = var_7999_pad_0, pad_type = var_7999_pad_type_0, strides = var_7999_strides_0, weight = model_blocks_5_attn_k_projs_5_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_7999_cast_fp16")];
            tensor<int32, [1]> var_8000_axes_0 = const()[name = tensor<string, []>("op_8000_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8000_cast_fp16 = squeeze(axes = var_8000_axes_0, x = var_7999_cast_fp16)[name = tensor<string, []>("op_8000_cast_fp16")];
            tensor<int32, [3]> k_391_perm_0 = const()[name = tensor<string, []>("k_391_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8007_pad_type_0 = const()[name = tensor<string, []>("op_8007_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8007_strides_0 = const()[name = tensor<string, []>("op_8007_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8007_pad_0 = const()[name = tensor<string, []>("op_8007_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8007_dilations_0 = const()[name = tensor<string, []>("op_8007_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8007_groups_0 = const()[name = tensor<string, []>("op_8007_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74124928)))];
            tensor<fp16, [1, 64, 1, 197]> var_8007_cast_fp16 = conv(dilations = var_8007_dilations_0, groups = var_8007_groups_0, pad = var_8007_pad_0, pad_type = var_8007_pad_type_0, strides = var_8007_strides_0, weight = model_blocks_5_attn_v_projs_5_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8007_cast_fp16")];
            tensor<int32, [1]> var_8008_axes_0 = const()[name = tensor<string, []>("op_8008_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8008_cast_fp16 = squeeze(axes = var_8008_axes_0, x = var_8007_cast_fp16)[name = tensor<string, []>("op_8008_cast_fp16")];
            tensor<int32, [3]> v_391_perm_0 = const()[name = tensor<string, []>("v_391_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74223296)))];
            tensor<fp16, [1, 197, 64]> q_651_cast_fp16 = transpose(perm = q_651_perm_0, x = var_7992_cast_fp16)[name = tensor<string, []>("transpose_284")];
            tensor<fp16, [1, 197, 64]> q_653_cast_fp16 = add(x = q_651_cast_fp16, y = model_blocks_5_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_653_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74223488)))];
            tensor<fp16, [1, 197, 64]> v_391_cast_fp16 = transpose(perm = v_391_perm_0, x = var_8008_cast_fp16)[name = tensor<string, []>("transpose_282")];
            tensor<fp16, [1, 197, 64]> v_393_cast_fp16 = add(x = v_391_cast_fp16, y = model_blocks_5_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_393_cast_fp16")];
            tensor<int32, [1]> q_655_axes_0 = const()[name = tensor<string, []>("q_655_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_655_cast_fp16 = expand_dims(axes = q_655_axes_0, x = q_653_cast_fp16)[name = tensor<string, []>("q_655_cast_fp16")];
            tensor<int32, [1]> k_393_axes_0 = const()[name = tensor<string, []>("k_393_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_391_cast_fp16 = transpose(perm = k_391_perm_0, x = var_8000_cast_fp16)[name = tensor<string, []>("transpose_283")];
            tensor<fp16, [1, 1, 197, 64]> k_393_cast_fp16 = expand_dims(axes = k_393_axes_0, x = k_391_cast_fp16)[name = tensor<string, []>("k_393_cast_fp16")];
            tensor<int32, [1]> v_395_axes_0 = const()[name = tensor<string, []>("v_395_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_395_cast_fp16 = expand_dims(axes = v_395_axes_0, x = v_393_cast_fp16)[name = tensor<string, []>("v_395_cast_fp16")];
            tensor<int32, [4]> var_8017_begin_0 = const()[name = tensor<string, []>("op_8017_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8017_end_0 = const()[name = tensor<string, []>("op_8017_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8017_end_mask_0 = const()[name = tensor<string, []>("op_8017_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8017_cast_fp16 = slice_by_index(begin = var_8017_begin_0, end = var_8017_end_0, end_mask = var_8017_end_mask_0, x = q_655_cast_fp16)[name = tensor<string, []>("op_8017_cast_fp16")];
            tensor<int32, [4]> var_8021_begin_0 = const()[name = tensor<string, []>("op_8021_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8021_end_0 = const()[name = tensor<string, []>("op_8021_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8021_end_mask_0 = const()[name = tensor<string, []>("op_8021_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8021_cast_fp16 = slice_by_index(begin = var_8021_begin_0, end = var_8021_end_0, end_mask = var_8021_end_mask_0, x = k_393_cast_fp16)[name = tensor<string, []>("op_8021_cast_fp16")];
            tensor<int32, [4]> var_8025_begin_0 = const()[name = tensor<string, []>("op_8025_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8025_end_0 = const()[name = tensor<string, []>("op_8025_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8025_end_mask_0 = const()[name = tensor<string, []>("op_8025_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8025_cast_fp16 = slice_by_index(begin = var_8025_begin_0, end = var_8025_end_0, end_mask = var_8025_end_mask_0, x = q_655_cast_fp16)[name = tensor<string, []>("op_8025_cast_fp16")];
            tensor<int32, [4]> var_8029_begin_0 = const()[name = tensor<string, []>("op_8029_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8029_end_0 = const()[name = tensor<string, []>("op_8029_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8029_end_mask_0 = const()[name = tensor<string, []>("op_8029_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8029_cast_fp16 = slice_by_index(begin = var_8029_begin_0, end = var_8029_end_0, end_mask = var_8029_end_mask_0, x = k_393_cast_fp16)[name = tensor<string, []>("op_8029_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8031_cast_fp16 = mul(x = var_8025_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8031_cast_fp16")];
            tensor<int32, [5]> var_8039 = const()[name = tensor<string, []>("op_8039"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_891_cast_fp16 = reshape(shape = var_8039, x = var_8025_cast_fp16)[name = tensor<string, []>("x_891_cast_fp16")];
            tensor<int32, [2]> var_8041_split_sizes_0 = const()[name = tensor<string, []>("op_8041_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8041_axis_0 = const()[name = tensor<string, []>("op_8041_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8041_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8041_cast_fp16_1 = split(axis = var_8041_axis_0, split_sizes = var_8041_split_sizes_0, x = x_891_cast_fp16)[name = tensor<string, []>("op_8041_cast_fp16")];
            tensor<int32, [1]> squeeze_260_axes_0 = const()[name = tensor<string, []>("squeeze_260_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_260_cast_fp16 = squeeze(axes = squeeze_260_axes_0, x = var_8041_cast_fp16_0)[name = tensor<string, []>("squeeze_260_cast_fp16")];
            tensor<int32, [1]> squeeze_261_axes_0 = const()[name = tensor<string, []>("squeeze_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_261_cast_fp16 = squeeze(axes = squeeze_261_axes_0, x = var_8041_cast_fp16_1)[name = tensor<string, []>("squeeze_261_cast_fp16")];
            tensor<fp16, []> const_1067_promoted_to_fp16 = const()[name = tensor<string, []>("const_1067_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8044_cast_fp16 = mul(x = squeeze_261_cast_fp16, y = const_1067_promoted_to_fp16)[name = tensor<string, []>("op_8044_cast_fp16")];
            tensor<int32, []> x_893_axis_0 = const()[name = tensor<string, []>("x_893_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_893_cast_fp16 = stack(axis = x_893_axis_0, values = (var_8044_cast_fp16, squeeze_260_cast_fp16))[name = tensor<string, []>("x_893_cast_fp16")];
            tensor<int32, [4]> var_8050 = const()[name = tensor<string, []>("op_8050"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8051_cast_fp16 = reshape(shape = var_8050, x = x_893_cast_fp16)[name = tensor<string, []>("op_8051_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8052_cast_fp16 = mul(x = var_8051_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8052_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_131_cast_fp16 = add(x = var_8031_cast_fp16, y = var_8052_cast_fp16)[name = tensor<string, []>("q_patches_131_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8054_cast_fp16 = mul(x = var_8029_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8054_cast_fp16")];
            tensor<int32, [5]> var_8062 = const()[name = tensor<string, []>("op_8062"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_897_cast_fp16 = reshape(shape = var_8062, x = var_8029_cast_fp16)[name = tensor<string, []>("x_897_cast_fp16")];
            tensor<int32, [2]> var_8064_split_sizes_0 = const()[name = tensor<string, []>("op_8064_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8064_axis_0 = const()[name = tensor<string, []>("op_8064_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8064_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8064_cast_fp16_1 = split(axis = var_8064_axis_0, split_sizes = var_8064_split_sizes_0, x = x_897_cast_fp16)[name = tensor<string, []>("op_8064_cast_fp16")];
            tensor<int32, [1]> squeeze_262_axes_0 = const()[name = tensor<string, []>("squeeze_262_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_262_cast_fp16 = squeeze(axes = squeeze_262_axes_0, x = var_8064_cast_fp16_0)[name = tensor<string, []>("squeeze_262_cast_fp16")];
            tensor<int32, [1]> squeeze_263_axes_0 = const()[name = tensor<string, []>("squeeze_263_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_263_cast_fp16 = squeeze(axes = squeeze_263_axes_0, x = var_8064_cast_fp16_1)[name = tensor<string, []>("squeeze_263_cast_fp16")];
            tensor<fp16, []> const_1075_promoted_to_fp16 = const()[name = tensor<string, []>("const_1075_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8067_cast_fp16 = mul(x = squeeze_263_cast_fp16, y = const_1075_promoted_to_fp16)[name = tensor<string, []>("op_8067_cast_fp16")];
            tensor<int32, []> x_899_axis_0 = const()[name = tensor<string, []>("x_899_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_899_cast_fp16 = stack(axis = x_899_axis_0, values = (var_8067_cast_fp16, squeeze_262_cast_fp16))[name = tensor<string, []>("x_899_cast_fp16")];
            tensor<int32, [4]> var_8073 = const()[name = tensor<string, []>("op_8073"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8074_cast_fp16 = reshape(shape = var_8073, x = x_899_cast_fp16)[name = tensor<string, []>("op_8074_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8075_cast_fp16 = mul(x = var_8074_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8075_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_131_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8075_cast_fp16)[name = tensor<string, []>("k_patches_131_cast_fp16")];
            tensor<bool, []> var_8078_interleave_0 = const()[name = tensor<string, []>("op_8078_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8078_cast_fp16 = concat(axis = var_20, interleave = var_8078_interleave_0, values = (var_8017_cast_fp16, q_patches_131_cast_fp16))[name = tensor<string, []>("op_8078_cast_fp16")];
            tensor<bool, []> var_8081_interleave_0 = const()[name = tensor<string, []>("op_8081_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8081_cast_fp16 = concat(axis = var_20, interleave = var_8081_interleave_0, values = (var_8021_cast_fp16, k_patches_131_cast_fp16))[name = tensor<string, []>("op_8081_cast_fp16")];
            tensor<fp16, []> var_8083_to_fp16 = const()[name = tensor<string, []>("op_8083_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_659_cast_fp16 = mul(x = var_8078_cast_fp16, y = var_8083_to_fp16)[name = tensor<string, []>("q_659_cast_fp16")];
            tensor<bool, []> attn_261_transpose_x_1 = const()[name = tensor<string, []>("attn_261_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_261_transpose_y_1 = const()[name = tensor<string, []>("attn_261_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_261_cast_fp16 = matmul(transpose_x = attn_261_transpose_x_1, transpose_y = attn_261_transpose_y_1, x = q_659_cast_fp16, y = var_8081_cast_fp16)[name = tensor<string, []>("attn_261_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_263_cast_fp16 = softmax(axis = var_21, x = attn_261_cast_fp16)[name = tensor<string, []>("attn_263_cast_fp16")];
            tensor<bool, []> var_8088_transpose_x_0 = const()[name = tensor<string, []>("op_8088_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8088_transpose_y_0 = const()[name = tensor<string, []>("op_8088_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8088_cast_fp16 = matmul(transpose_x = var_8088_transpose_x_0, transpose_y = var_8088_transpose_y_0, x = attn_263_cast_fp16, y = v_395_cast_fp16)[name = tensor<string, []>("op_8088_cast_fp16")];
            tensor<int32, [1]> var_8089_axes_0 = const()[name = tensor<string, []>("op_8089_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8089_cast_fp16 = squeeze(axes = var_8089_axes_0, x = var_8088_cast_fp16)[name = tensor<string, []>("op_8089_cast_fp16")];
            tensor<string, []> var_8095_pad_type_0 = const()[name = tensor<string, []>("op_8095_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8095_strides_0 = const()[name = tensor<string, []>("op_8095_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8095_pad_0 = const()[name = tensor<string, []>("op_8095_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8095_dilations_0 = const()[name = tensor<string, []>("op_8095_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8095_groups_0 = const()[name = tensor<string, []>("op_8095_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74223680)))];
            tensor<fp16, [1, 64, 1, 197]> var_8095_cast_fp16 = conv(dilations = var_8095_dilations_0, groups = var_8095_groups_0, pad = var_8095_pad_0, pad_type = var_8095_pad_type_0, strides = var_8095_strides_0, weight = model_blocks_5_attn_q_projs_6_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8095_cast_fp16")];
            tensor<int32, [1]> var_8096_axes_0 = const()[name = tensor<string, []>("op_8096_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8096_cast_fp16 = squeeze(axes = var_8096_axes_0, x = var_8095_cast_fp16)[name = tensor<string, []>("op_8096_cast_fp16")];
            tensor<int32, [3]> q_661_perm_0 = const()[name = tensor<string, []>("q_661_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8103_pad_type_0 = const()[name = tensor<string, []>("op_8103_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8103_strides_0 = const()[name = tensor<string, []>("op_8103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8103_pad_0 = const()[name = tensor<string, []>("op_8103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8103_dilations_0 = const()[name = tensor<string, []>("op_8103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8103_groups_0 = const()[name = tensor<string, []>("op_8103_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74322048)))];
            tensor<fp16, [1, 64, 1, 197]> var_8103_cast_fp16 = conv(dilations = var_8103_dilations_0, groups = var_8103_groups_0, pad = var_8103_pad_0, pad_type = var_8103_pad_type_0, strides = var_8103_strides_0, weight = model_blocks_5_attn_k_projs_6_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8103_cast_fp16")];
            tensor<int32, [1]> var_8104_axes_0 = const()[name = tensor<string, []>("op_8104_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8104_cast_fp16 = squeeze(axes = var_8104_axes_0, x = var_8103_cast_fp16)[name = tensor<string, []>("op_8104_cast_fp16")];
            tensor<int32, [3]> k_397_perm_0 = const()[name = tensor<string, []>("k_397_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8111_pad_type_0 = const()[name = tensor<string, []>("op_8111_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8111_strides_0 = const()[name = tensor<string, []>("op_8111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8111_pad_0 = const()[name = tensor<string, []>("op_8111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8111_dilations_0 = const()[name = tensor<string, []>("op_8111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8111_groups_0 = const()[name = tensor<string, []>("op_8111_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74420416)))];
            tensor<fp16, [1, 64, 1, 197]> var_8111_cast_fp16 = conv(dilations = var_8111_dilations_0, groups = var_8111_groups_0, pad = var_8111_pad_0, pad_type = var_8111_pad_type_0, strides = var_8111_strides_0, weight = model_blocks_5_attn_v_projs_6_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8111_cast_fp16")];
            tensor<int32, [1]> var_8112_axes_0 = const()[name = tensor<string, []>("op_8112_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8112_cast_fp16 = squeeze(axes = var_8112_axes_0, x = var_8111_cast_fp16)[name = tensor<string, []>("op_8112_cast_fp16")];
            tensor<int32, [3]> v_397_perm_0 = const()[name = tensor<string, []>("v_397_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74518784)))];
            tensor<fp16, [1, 197, 64]> q_661_cast_fp16 = transpose(perm = q_661_perm_0, x = var_8096_cast_fp16)[name = tensor<string, []>("transpose_281")];
            tensor<fp16, [1, 197, 64]> q_663_cast_fp16 = add(x = q_661_cast_fp16, y = model_blocks_5_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_663_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74518976)))];
            tensor<fp16, [1, 197, 64]> v_397_cast_fp16 = transpose(perm = v_397_perm_0, x = var_8112_cast_fp16)[name = tensor<string, []>("transpose_279")];
            tensor<fp16, [1, 197, 64]> v_399_cast_fp16 = add(x = v_397_cast_fp16, y = model_blocks_5_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_399_cast_fp16")];
            tensor<int32, [1]> q_665_axes_0 = const()[name = tensor<string, []>("q_665_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_665_cast_fp16 = expand_dims(axes = q_665_axes_0, x = q_663_cast_fp16)[name = tensor<string, []>("q_665_cast_fp16")];
            tensor<int32, [1]> k_399_axes_0 = const()[name = tensor<string, []>("k_399_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_397_cast_fp16 = transpose(perm = k_397_perm_0, x = var_8104_cast_fp16)[name = tensor<string, []>("transpose_280")];
            tensor<fp16, [1, 1, 197, 64]> k_399_cast_fp16 = expand_dims(axes = k_399_axes_0, x = k_397_cast_fp16)[name = tensor<string, []>("k_399_cast_fp16")];
            tensor<int32, [1]> v_401_axes_0 = const()[name = tensor<string, []>("v_401_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_401_cast_fp16 = expand_dims(axes = v_401_axes_0, x = v_399_cast_fp16)[name = tensor<string, []>("v_401_cast_fp16")];
            tensor<int32, [4]> var_8121_begin_0 = const()[name = tensor<string, []>("op_8121_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8121_end_0 = const()[name = tensor<string, []>("op_8121_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8121_end_mask_0 = const()[name = tensor<string, []>("op_8121_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8121_cast_fp16 = slice_by_index(begin = var_8121_begin_0, end = var_8121_end_0, end_mask = var_8121_end_mask_0, x = q_665_cast_fp16)[name = tensor<string, []>("op_8121_cast_fp16")];
            tensor<int32, [4]> var_8125_begin_0 = const()[name = tensor<string, []>("op_8125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8125_end_0 = const()[name = tensor<string, []>("op_8125_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8125_end_mask_0 = const()[name = tensor<string, []>("op_8125_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8125_cast_fp16 = slice_by_index(begin = var_8125_begin_0, end = var_8125_end_0, end_mask = var_8125_end_mask_0, x = k_399_cast_fp16)[name = tensor<string, []>("op_8125_cast_fp16")];
            tensor<int32, [4]> var_8129_begin_0 = const()[name = tensor<string, []>("op_8129_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8129_end_0 = const()[name = tensor<string, []>("op_8129_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8129_end_mask_0 = const()[name = tensor<string, []>("op_8129_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8129_cast_fp16 = slice_by_index(begin = var_8129_begin_0, end = var_8129_end_0, end_mask = var_8129_end_mask_0, x = q_665_cast_fp16)[name = tensor<string, []>("op_8129_cast_fp16")];
            tensor<int32, [4]> var_8133_begin_0 = const()[name = tensor<string, []>("op_8133_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8133_end_0 = const()[name = tensor<string, []>("op_8133_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8133_end_mask_0 = const()[name = tensor<string, []>("op_8133_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8133_cast_fp16 = slice_by_index(begin = var_8133_begin_0, end = var_8133_end_0, end_mask = var_8133_end_mask_0, x = k_399_cast_fp16)[name = tensor<string, []>("op_8133_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8135_cast_fp16 = mul(x = var_8129_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8135_cast_fp16")];
            tensor<int32, [5]> var_8143 = const()[name = tensor<string, []>("op_8143"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_903_cast_fp16 = reshape(shape = var_8143, x = var_8129_cast_fp16)[name = tensor<string, []>("x_903_cast_fp16")];
            tensor<int32, [2]> var_8145_split_sizes_0 = const()[name = tensor<string, []>("op_8145_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8145_axis_0 = const()[name = tensor<string, []>("op_8145_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8145_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8145_cast_fp16_1 = split(axis = var_8145_axis_0, split_sizes = var_8145_split_sizes_0, x = x_903_cast_fp16)[name = tensor<string, []>("op_8145_cast_fp16")];
            tensor<int32, [1]> squeeze_264_axes_0 = const()[name = tensor<string, []>("squeeze_264_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_264_cast_fp16 = squeeze(axes = squeeze_264_axes_0, x = var_8145_cast_fp16_0)[name = tensor<string, []>("squeeze_264_cast_fp16")];
            tensor<int32, [1]> squeeze_265_axes_0 = const()[name = tensor<string, []>("squeeze_265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_265_cast_fp16 = squeeze(axes = squeeze_265_axes_0, x = var_8145_cast_fp16_1)[name = tensor<string, []>("squeeze_265_cast_fp16")];
            tensor<fp16, []> const_1083_promoted_to_fp16 = const()[name = tensor<string, []>("const_1083_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8148_cast_fp16 = mul(x = squeeze_265_cast_fp16, y = const_1083_promoted_to_fp16)[name = tensor<string, []>("op_8148_cast_fp16")];
            tensor<int32, []> x_905_axis_0 = const()[name = tensor<string, []>("x_905_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_905_cast_fp16 = stack(axis = x_905_axis_0, values = (var_8148_cast_fp16, squeeze_264_cast_fp16))[name = tensor<string, []>("x_905_cast_fp16")];
            tensor<int32, [4]> var_8154 = const()[name = tensor<string, []>("op_8154"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8155_cast_fp16 = reshape(shape = var_8154, x = x_905_cast_fp16)[name = tensor<string, []>("op_8155_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8156_cast_fp16 = mul(x = var_8155_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8156_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_133_cast_fp16 = add(x = var_8135_cast_fp16, y = var_8156_cast_fp16)[name = tensor<string, []>("q_patches_133_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8158_cast_fp16 = mul(x = var_8133_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8158_cast_fp16")];
            tensor<int32, [5]> var_8166 = const()[name = tensor<string, []>("op_8166"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_909_cast_fp16 = reshape(shape = var_8166, x = var_8133_cast_fp16)[name = tensor<string, []>("x_909_cast_fp16")];
            tensor<int32, [2]> var_8168_split_sizes_0 = const()[name = tensor<string, []>("op_8168_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8168_axis_0 = const()[name = tensor<string, []>("op_8168_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8168_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8168_cast_fp16_1 = split(axis = var_8168_axis_0, split_sizes = var_8168_split_sizes_0, x = x_909_cast_fp16)[name = tensor<string, []>("op_8168_cast_fp16")];
            tensor<int32, [1]> squeeze_266_axes_0 = const()[name = tensor<string, []>("squeeze_266_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_266_cast_fp16 = squeeze(axes = squeeze_266_axes_0, x = var_8168_cast_fp16_0)[name = tensor<string, []>("squeeze_266_cast_fp16")];
            tensor<int32, [1]> squeeze_267_axes_0 = const()[name = tensor<string, []>("squeeze_267_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_267_cast_fp16 = squeeze(axes = squeeze_267_axes_0, x = var_8168_cast_fp16_1)[name = tensor<string, []>("squeeze_267_cast_fp16")];
            tensor<fp16, []> const_1091_promoted_to_fp16 = const()[name = tensor<string, []>("const_1091_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8171_cast_fp16 = mul(x = squeeze_267_cast_fp16, y = const_1091_promoted_to_fp16)[name = tensor<string, []>("op_8171_cast_fp16")];
            tensor<int32, []> x_911_axis_0 = const()[name = tensor<string, []>("x_911_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_911_cast_fp16 = stack(axis = x_911_axis_0, values = (var_8171_cast_fp16, squeeze_266_cast_fp16))[name = tensor<string, []>("x_911_cast_fp16")];
            tensor<int32, [4]> var_8177 = const()[name = tensor<string, []>("op_8177"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8178_cast_fp16 = reshape(shape = var_8177, x = x_911_cast_fp16)[name = tensor<string, []>("op_8178_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8179_cast_fp16 = mul(x = var_8178_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8179_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_133_cast_fp16 = add(x = var_8158_cast_fp16, y = var_8179_cast_fp16)[name = tensor<string, []>("k_patches_133_cast_fp16")];
            tensor<bool, []> var_8182_interleave_0 = const()[name = tensor<string, []>("op_8182_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8182_cast_fp16 = concat(axis = var_20, interleave = var_8182_interleave_0, values = (var_8121_cast_fp16, q_patches_133_cast_fp16))[name = tensor<string, []>("op_8182_cast_fp16")];
            tensor<bool, []> var_8185_interleave_0 = const()[name = tensor<string, []>("op_8185_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8185_cast_fp16 = concat(axis = var_20, interleave = var_8185_interleave_0, values = (var_8125_cast_fp16, k_patches_133_cast_fp16))[name = tensor<string, []>("op_8185_cast_fp16")];
            tensor<fp16, []> var_8187_to_fp16 = const()[name = tensor<string, []>("op_8187_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_669_cast_fp16 = mul(x = var_8182_cast_fp16, y = var_8187_to_fp16)[name = tensor<string, []>("q_669_cast_fp16")];
            tensor<bool, []> attn_265_transpose_x_1 = const()[name = tensor<string, []>("attn_265_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_265_transpose_y_1 = const()[name = tensor<string, []>("attn_265_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_265_cast_fp16 = matmul(transpose_x = attn_265_transpose_x_1, transpose_y = attn_265_transpose_y_1, x = q_669_cast_fp16, y = var_8185_cast_fp16)[name = tensor<string, []>("attn_265_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_267_cast_fp16 = softmax(axis = var_21, x = attn_265_cast_fp16)[name = tensor<string, []>("attn_267_cast_fp16")];
            tensor<bool, []> var_8192_transpose_x_0 = const()[name = tensor<string, []>("op_8192_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8192_transpose_y_0 = const()[name = tensor<string, []>("op_8192_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8192_cast_fp16 = matmul(transpose_x = var_8192_transpose_x_0, transpose_y = var_8192_transpose_y_0, x = attn_267_cast_fp16, y = v_401_cast_fp16)[name = tensor<string, []>("op_8192_cast_fp16")];
            tensor<int32, [1]> var_8193_axes_0 = const()[name = tensor<string, []>("op_8193_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8193_cast_fp16 = squeeze(axes = var_8193_axes_0, x = var_8192_cast_fp16)[name = tensor<string, []>("op_8193_cast_fp16")];
            tensor<string, []> var_8199_pad_type_0 = const()[name = tensor<string, []>("op_8199_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8199_strides_0 = const()[name = tensor<string, []>("op_8199_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8199_pad_0 = const()[name = tensor<string, []>("op_8199_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8199_dilations_0 = const()[name = tensor<string, []>("op_8199_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8199_groups_0 = const()[name = tensor<string, []>("op_8199_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74519168)))];
            tensor<fp16, [1, 64, 1, 197]> var_8199_cast_fp16 = conv(dilations = var_8199_dilations_0, groups = var_8199_groups_0, pad = var_8199_pad_0, pad_type = var_8199_pad_type_0, strides = var_8199_strides_0, weight = model_blocks_5_attn_q_projs_7_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8199_cast_fp16")];
            tensor<int32, [1]> var_8200_axes_0 = const()[name = tensor<string, []>("op_8200_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8200_cast_fp16 = squeeze(axes = var_8200_axes_0, x = var_8199_cast_fp16)[name = tensor<string, []>("op_8200_cast_fp16")];
            tensor<int32, [3]> q_671_perm_0 = const()[name = tensor<string, []>("q_671_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8207_pad_type_0 = const()[name = tensor<string, []>("op_8207_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8207_strides_0 = const()[name = tensor<string, []>("op_8207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8207_pad_0 = const()[name = tensor<string, []>("op_8207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8207_dilations_0 = const()[name = tensor<string, []>("op_8207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8207_groups_0 = const()[name = tensor<string, []>("op_8207_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74617536)))];
            tensor<fp16, [1, 64, 1, 197]> var_8207_cast_fp16 = conv(dilations = var_8207_dilations_0, groups = var_8207_groups_0, pad = var_8207_pad_0, pad_type = var_8207_pad_type_0, strides = var_8207_strides_0, weight = model_blocks_5_attn_k_projs_7_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8207_cast_fp16")];
            tensor<int32, [1]> var_8208_axes_0 = const()[name = tensor<string, []>("op_8208_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8208_cast_fp16 = squeeze(axes = var_8208_axes_0, x = var_8207_cast_fp16)[name = tensor<string, []>("op_8208_cast_fp16")];
            tensor<int32, [3]> k_403_perm_0 = const()[name = tensor<string, []>("k_403_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8215_pad_type_0 = const()[name = tensor<string, []>("op_8215_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8215_strides_0 = const()[name = tensor<string, []>("op_8215_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8215_pad_0 = const()[name = tensor<string, []>("op_8215_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8215_dilations_0 = const()[name = tensor<string, []>("op_8215_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8215_groups_0 = const()[name = tensor<string, []>("op_8215_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74715904)))];
            tensor<fp16, [1, 64, 1, 197]> var_8215_cast_fp16 = conv(dilations = var_8215_dilations_0, groups = var_8215_groups_0, pad = var_8215_pad_0, pad_type = var_8215_pad_type_0, strides = var_8215_strides_0, weight = model_blocks_5_attn_v_projs_7_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8215_cast_fp16")];
            tensor<int32, [1]> var_8216_axes_0 = const()[name = tensor<string, []>("op_8216_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8216_cast_fp16 = squeeze(axes = var_8216_axes_0, x = var_8215_cast_fp16)[name = tensor<string, []>("op_8216_cast_fp16")];
            tensor<int32, [3]> v_403_perm_0 = const()[name = tensor<string, []>("v_403_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74814272)))];
            tensor<fp16, [1, 197, 64]> q_671_cast_fp16 = transpose(perm = q_671_perm_0, x = var_8200_cast_fp16)[name = tensor<string, []>("transpose_278")];
            tensor<fp16, [1, 197, 64]> q_673_cast_fp16 = add(x = q_671_cast_fp16, y = model_blocks_5_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_673_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74814464)))];
            tensor<fp16, [1, 197, 64]> v_403_cast_fp16 = transpose(perm = v_403_perm_0, x = var_8216_cast_fp16)[name = tensor<string, []>("transpose_276")];
            tensor<fp16, [1, 197, 64]> v_405_cast_fp16 = add(x = v_403_cast_fp16, y = model_blocks_5_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_405_cast_fp16")];
            tensor<int32, [1]> q_675_axes_0 = const()[name = tensor<string, []>("q_675_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_675_cast_fp16 = expand_dims(axes = q_675_axes_0, x = q_673_cast_fp16)[name = tensor<string, []>("q_675_cast_fp16")];
            tensor<int32, [1]> k_405_axes_0 = const()[name = tensor<string, []>("k_405_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_403_cast_fp16 = transpose(perm = k_403_perm_0, x = var_8208_cast_fp16)[name = tensor<string, []>("transpose_277")];
            tensor<fp16, [1, 1, 197, 64]> k_405_cast_fp16 = expand_dims(axes = k_405_axes_0, x = k_403_cast_fp16)[name = tensor<string, []>("k_405_cast_fp16")];
            tensor<int32, [1]> v_407_axes_0 = const()[name = tensor<string, []>("v_407_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_407_cast_fp16 = expand_dims(axes = v_407_axes_0, x = v_405_cast_fp16)[name = tensor<string, []>("v_407_cast_fp16")];
            tensor<int32, [4]> var_8225_begin_0 = const()[name = tensor<string, []>("op_8225_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8225_end_0 = const()[name = tensor<string, []>("op_8225_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8225_end_mask_0 = const()[name = tensor<string, []>("op_8225_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8225_cast_fp16 = slice_by_index(begin = var_8225_begin_0, end = var_8225_end_0, end_mask = var_8225_end_mask_0, x = q_675_cast_fp16)[name = tensor<string, []>("op_8225_cast_fp16")];
            tensor<int32, [4]> var_8229_begin_0 = const()[name = tensor<string, []>("op_8229_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8229_end_0 = const()[name = tensor<string, []>("op_8229_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8229_end_mask_0 = const()[name = tensor<string, []>("op_8229_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8229_cast_fp16 = slice_by_index(begin = var_8229_begin_0, end = var_8229_end_0, end_mask = var_8229_end_mask_0, x = k_405_cast_fp16)[name = tensor<string, []>("op_8229_cast_fp16")];
            tensor<int32, [4]> var_8233_begin_0 = const()[name = tensor<string, []>("op_8233_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8233_end_0 = const()[name = tensor<string, []>("op_8233_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8233_end_mask_0 = const()[name = tensor<string, []>("op_8233_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8233_cast_fp16 = slice_by_index(begin = var_8233_begin_0, end = var_8233_end_0, end_mask = var_8233_end_mask_0, x = q_675_cast_fp16)[name = tensor<string, []>("op_8233_cast_fp16")];
            tensor<int32, [4]> var_8237_begin_0 = const()[name = tensor<string, []>("op_8237_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8237_end_0 = const()[name = tensor<string, []>("op_8237_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8237_end_mask_0 = const()[name = tensor<string, []>("op_8237_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8237_cast_fp16 = slice_by_index(begin = var_8237_begin_0, end = var_8237_end_0, end_mask = var_8237_end_mask_0, x = k_405_cast_fp16)[name = tensor<string, []>("op_8237_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8239_cast_fp16 = mul(x = var_8233_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8239_cast_fp16")];
            tensor<int32, [5]> var_8247 = const()[name = tensor<string, []>("op_8247"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_915_cast_fp16 = reshape(shape = var_8247, x = var_8233_cast_fp16)[name = tensor<string, []>("x_915_cast_fp16")];
            tensor<int32, [2]> var_8249_split_sizes_0 = const()[name = tensor<string, []>("op_8249_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8249_axis_0 = const()[name = tensor<string, []>("op_8249_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8249_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8249_cast_fp16_1 = split(axis = var_8249_axis_0, split_sizes = var_8249_split_sizes_0, x = x_915_cast_fp16)[name = tensor<string, []>("op_8249_cast_fp16")];
            tensor<int32, [1]> squeeze_268_axes_0 = const()[name = tensor<string, []>("squeeze_268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_268_cast_fp16 = squeeze(axes = squeeze_268_axes_0, x = var_8249_cast_fp16_0)[name = tensor<string, []>("squeeze_268_cast_fp16")];
            tensor<int32, [1]> squeeze_269_axes_0 = const()[name = tensor<string, []>("squeeze_269_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_269_cast_fp16 = squeeze(axes = squeeze_269_axes_0, x = var_8249_cast_fp16_1)[name = tensor<string, []>("squeeze_269_cast_fp16")];
            tensor<fp16, []> const_1099_promoted_to_fp16 = const()[name = tensor<string, []>("const_1099_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8252_cast_fp16 = mul(x = squeeze_269_cast_fp16, y = const_1099_promoted_to_fp16)[name = tensor<string, []>("op_8252_cast_fp16")];
            tensor<int32, []> x_917_axis_0 = const()[name = tensor<string, []>("x_917_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_917_cast_fp16 = stack(axis = x_917_axis_0, values = (var_8252_cast_fp16, squeeze_268_cast_fp16))[name = tensor<string, []>("x_917_cast_fp16")];
            tensor<int32, [4]> var_8258 = const()[name = tensor<string, []>("op_8258"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8259_cast_fp16 = reshape(shape = var_8258, x = x_917_cast_fp16)[name = tensor<string, []>("op_8259_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8260_cast_fp16 = mul(x = var_8259_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8260_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_135_cast_fp16 = add(x = var_8239_cast_fp16, y = var_8260_cast_fp16)[name = tensor<string, []>("q_patches_135_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8262_cast_fp16 = mul(x = var_8237_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8262_cast_fp16")];
            tensor<int32, [5]> var_8270 = const()[name = tensor<string, []>("op_8270"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_921_cast_fp16 = reshape(shape = var_8270, x = var_8237_cast_fp16)[name = tensor<string, []>("x_921_cast_fp16")];
            tensor<int32, [2]> var_8272_split_sizes_0 = const()[name = tensor<string, []>("op_8272_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8272_axis_0 = const()[name = tensor<string, []>("op_8272_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8272_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8272_cast_fp16_1 = split(axis = var_8272_axis_0, split_sizes = var_8272_split_sizes_0, x = x_921_cast_fp16)[name = tensor<string, []>("op_8272_cast_fp16")];
            tensor<int32, [1]> squeeze_270_axes_0 = const()[name = tensor<string, []>("squeeze_270_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_270_cast_fp16 = squeeze(axes = squeeze_270_axes_0, x = var_8272_cast_fp16_0)[name = tensor<string, []>("squeeze_270_cast_fp16")];
            tensor<int32, [1]> squeeze_271_axes_0 = const()[name = tensor<string, []>("squeeze_271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_271_cast_fp16 = squeeze(axes = squeeze_271_axes_0, x = var_8272_cast_fp16_1)[name = tensor<string, []>("squeeze_271_cast_fp16")];
            tensor<fp16, []> const_1107_promoted_to_fp16 = const()[name = tensor<string, []>("const_1107_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8275_cast_fp16 = mul(x = squeeze_271_cast_fp16, y = const_1107_promoted_to_fp16)[name = tensor<string, []>("op_8275_cast_fp16")];
            tensor<int32, []> x_923_axis_0 = const()[name = tensor<string, []>("x_923_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_923_cast_fp16 = stack(axis = x_923_axis_0, values = (var_8275_cast_fp16, squeeze_270_cast_fp16))[name = tensor<string, []>("x_923_cast_fp16")];
            tensor<int32, [4]> var_8281 = const()[name = tensor<string, []>("op_8281"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8282_cast_fp16 = reshape(shape = var_8281, x = x_923_cast_fp16)[name = tensor<string, []>("op_8282_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8283_cast_fp16 = mul(x = var_8282_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8283_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_135_cast_fp16 = add(x = var_8262_cast_fp16, y = var_8283_cast_fp16)[name = tensor<string, []>("k_patches_135_cast_fp16")];
            tensor<bool, []> var_8286_interleave_0 = const()[name = tensor<string, []>("op_8286_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8286_cast_fp16 = concat(axis = var_20, interleave = var_8286_interleave_0, values = (var_8225_cast_fp16, q_patches_135_cast_fp16))[name = tensor<string, []>("op_8286_cast_fp16")];
            tensor<bool, []> var_8289_interleave_0 = const()[name = tensor<string, []>("op_8289_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8289_cast_fp16 = concat(axis = var_20, interleave = var_8289_interleave_0, values = (var_8229_cast_fp16, k_patches_135_cast_fp16))[name = tensor<string, []>("op_8289_cast_fp16")];
            tensor<fp16, []> var_8291_to_fp16 = const()[name = tensor<string, []>("op_8291_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_679_cast_fp16 = mul(x = var_8286_cast_fp16, y = var_8291_to_fp16)[name = tensor<string, []>("q_679_cast_fp16")];
            tensor<bool, []> attn_269_transpose_x_1 = const()[name = tensor<string, []>("attn_269_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_269_transpose_y_1 = const()[name = tensor<string, []>("attn_269_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_269_cast_fp16 = matmul(transpose_x = attn_269_transpose_x_1, transpose_y = attn_269_transpose_y_1, x = q_679_cast_fp16, y = var_8289_cast_fp16)[name = tensor<string, []>("attn_269_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_271_cast_fp16 = softmax(axis = var_21, x = attn_269_cast_fp16)[name = tensor<string, []>("attn_271_cast_fp16")];
            tensor<bool, []> var_8296_transpose_x_0 = const()[name = tensor<string, []>("op_8296_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8296_transpose_y_0 = const()[name = tensor<string, []>("op_8296_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8296_cast_fp16 = matmul(transpose_x = var_8296_transpose_x_0, transpose_y = var_8296_transpose_y_0, x = attn_271_cast_fp16, y = v_407_cast_fp16)[name = tensor<string, []>("op_8296_cast_fp16")];
            tensor<int32, [1]> var_8297_axes_0 = const()[name = tensor<string, []>("op_8297_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8297_cast_fp16 = squeeze(axes = var_8297_axes_0, x = var_8296_cast_fp16)[name = tensor<string, []>("op_8297_cast_fp16")];
            tensor<string, []> var_8303_pad_type_0 = const()[name = tensor<string, []>("op_8303_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8303_strides_0 = const()[name = tensor<string, []>("op_8303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8303_pad_0 = const()[name = tensor<string, []>("op_8303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8303_dilations_0 = const()[name = tensor<string, []>("op_8303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8303_groups_0 = const()[name = tensor<string, []>("op_8303_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74814656)))];
            tensor<fp16, [1, 64, 1, 197]> var_8303_cast_fp16 = conv(dilations = var_8303_dilations_0, groups = var_8303_groups_0, pad = var_8303_pad_0, pad_type = var_8303_pad_type_0, strides = var_8303_strides_0, weight = model_blocks_5_attn_q_projs_8_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8303_cast_fp16")];
            tensor<int32, [1]> var_8304_axes_0 = const()[name = tensor<string, []>("op_8304_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8304_cast_fp16 = squeeze(axes = var_8304_axes_0, x = var_8303_cast_fp16)[name = tensor<string, []>("op_8304_cast_fp16")];
            tensor<int32, [3]> q_681_perm_0 = const()[name = tensor<string, []>("q_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8311_pad_type_0 = const()[name = tensor<string, []>("op_8311_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8311_strides_0 = const()[name = tensor<string, []>("op_8311_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8311_pad_0 = const()[name = tensor<string, []>("op_8311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8311_dilations_0 = const()[name = tensor<string, []>("op_8311_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8311_groups_0 = const()[name = tensor<string, []>("op_8311_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74913024)))];
            tensor<fp16, [1, 64, 1, 197]> var_8311_cast_fp16 = conv(dilations = var_8311_dilations_0, groups = var_8311_groups_0, pad = var_8311_pad_0, pad_type = var_8311_pad_type_0, strides = var_8311_strides_0, weight = model_blocks_5_attn_k_projs_8_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8311_cast_fp16")];
            tensor<int32, [1]> var_8312_axes_0 = const()[name = tensor<string, []>("op_8312_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8312_cast_fp16 = squeeze(axes = var_8312_axes_0, x = var_8311_cast_fp16)[name = tensor<string, []>("op_8312_cast_fp16")];
            tensor<int32, [3]> k_409_perm_0 = const()[name = tensor<string, []>("k_409_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8319_pad_type_0 = const()[name = tensor<string, []>("op_8319_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8319_strides_0 = const()[name = tensor<string, []>("op_8319_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8319_pad_0 = const()[name = tensor<string, []>("op_8319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8319_dilations_0 = const()[name = tensor<string, []>("op_8319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8319_groups_0 = const()[name = tensor<string, []>("op_8319_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75011392)))];
            tensor<fp16, [1, 64, 1, 197]> var_8319_cast_fp16 = conv(dilations = var_8319_dilations_0, groups = var_8319_groups_0, pad = var_8319_pad_0, pad_type = var_8319_pad_type_0, strides = var_8319_strides_0, weight = model_blocks_5_attn_v_projs_8_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8320_axes_0 = const()[name = tensor<string, []>("op_8320_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8320_cast_fp16 = squeeze(axes = var_8320_axes_0, x = var_8319_cast_fp16)[name = tensor<string, []>("op_8320_cast_fp16")];
            tensor<int32, [3]> v_409_perm_0 = const()[name = tensor<string, []>("v_409_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75109760)))];
            tensor<fp16, [1, 197, 64]> q_681_cast_fp16 = transpose(perm = q_681_perm_0, x = var_8304_cast_fp16)[name = tensor<string, []>("transpose_275")];
            tensor<fp16, [1, 197, 64]> q_683_cast_fp16 = add(x = q_681_cast_fp16, y = model_blocks_5_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_683_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75109952)))];
            tensor<fp16, [1, 197, 64]> v_409_cast_fp16 = transpose(perm = v_409_perm_0, x = var_8320_cast_fp16)[name = tensor<string, []>("transpose_273")];
            tensor<fp16, [1, 197, 64]> v_411_cast_fp16 = add(x = v_409_cast_fp16, y = model_blocks_5_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_411_cast_fp16")];
            tensor<int32, [1]> q_685_axes_0 = const()[name = tensor<string, []>("q_685_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_685_cast_fp16 = expand_dims(axes = q_685_axes_0, x = q_683_cast_fp16)[name = tensor<string, []>("q_685_cast_fp16")];
            tensor<int32, [1]> k_411_axes_0 = const()[name = tensor<string, []>("k_411_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_409_cast_fp16 = transpose(perm = k_409_perm_0, x = var_8312_cast_fp16)[name = tensor<string, []>("transpose_274")];
            tensor<fp16, [1, 1, 197, 64]> k_411_cast_fp16 = expand_dims(axes = k_411_axes_0, x = k_409_cast_fp16)[name = tensor<string, []>("k_411_cast_fp16")];
            tensor<int32, [1]> v_413_axes_0 = const()[name = tensor<string, []>("v_413_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_413_cast_fp16 = expand_dims(axes = v_413_axes_0, x = v_411_cast_fp16)[name = tensor<string, []>("v_413_cast_fp16")];
            tensor<int32, [4]> var_8329_begin_0 = const()[name = tensor<string, []>("op_8329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8329_end_0 = const()[name = tensor<string, []>("op_8329_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8329_end_mask_0 = const()[name = tensor<string, []>("op_8329_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8329_cast_fp16 = slice_by_index(begin = var_8329_begin_0, end = var_8329_end_0, end_mask = var_8329_end_mask_0, x = q_685_cast_fp16)[name = tensor<string, []>("op_8329_cast_fp16")];
            tensor<int32, [4]> var_8333_begin_0 = const()[name = tensor<string, []>("op_8333_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8333_end_0 = const()[name = tensor<string, []>("op_8333_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8333_end_mask_0 = const()[name = tensor<string, []>("op_8333_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, x = k_411_cast_fp16)[name = tensor<string, []>("op_8333_cast_fp16")];
            tensor<int32, [4]> var_8337_begin_0 = const()[name = tensor<string, []>("op_8337_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8337_end_0 = const()[name = tensor<string, []>("op_8337_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8337_end_mask_0 = const()[name = tensor<string, []>("op_8337_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8337_cast_fp16 = slice_by_index(begin = var_8337_begin_0, end = var_8337_end_0, end_mask = var_8337_end_mask_0, x = q_685_cast_fp16)[name = tensor<string, []>("op_8337_cast_fp16")];
            tensor<int32, [4]> var_8341_begin_0 = const()[name = tensor<string, []>("op_8341_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8341_end_0 = const()[name = tensor<string, []>("op_8341_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8341_end_mask_0 = const()[name = tensor<string, []>("op_8341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, x = k_411_cast_fp16)[name = tensor<string, []>("op_8341_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8343_cast_fp16 = mul(x = var_8337_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8343_cast_fp16")];
            tensor<int32, [5]> var_8351 = const()[name = tensor<string, []>("op_8351"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_927_cast_fp16 = reshape(shape = var_8351, x = var_8337_cast_fp16)[name = tensor<string, []>("x_927_cast_fp16")];
            tensor<int32, [2]> var_8353_split_sizes_0 = const()[name = tensor<string, []>("op_8353_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8353_axis_0 = const()[name = tensor<string, []>("op_8353_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8353_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8353_cast_fp16_1 = split(axis = var_8353_axis_0, split_sizes = var_8353_split_sizes_0, x = x_927_cast_fp16)[name = tensor<string, []>("op_8353_cast_fp16")];
            tensor<int32, [1]> squeeze_272_axes_0 = const()[name = tensor<string, []>("squeeze_272_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_272_cast_fp16 = squeeze(axes = squeeze_272_axes_0, x = var_8353_cast_fp16_0)[name = tensor<string, []>("squeeze_272_cast_fp16")];
            tensor<int32, [1]> squeeze_273_axes_0 = const()[name = tensor<string, []>("squeeze_273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_273_cast_fp16 = squeeze(axes = squeeze_273_axes_0, x = var_8353_cast_fp16_1)[name = tensor<string, []>("squeeze_273_cast_fp16")];
            tensor<fp16, []> const_1115_promoted_to_fp16 = const()[name = tensor<string, []>("const_1115_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8356_cast_fp16 = mul(x = squeeze_273_cast_fp16, y = const_1115_promoted_to_fp16)[name = tensor<string, []>("op_8356_cast_fp16")];
            tensor<int32, []> x_929_axis_0 = const()[name = tensor<string, []>("x_929_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_929_cast_fp16 = stack(axis = x_929_axis_0, values = (var_8356_cast_fp16, squeeze_272_cast_fp16))[name = tensor<string, []>("x_929_cast_fp16")];
            tensor<int32, [4]> var_8362 = const()[name = tensor<string, []>("op_8362"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8363_cast_fp16 = reshape(shape = var_8362, x = x_929_cast_fp16)[name = tensor<string, []>("op_8363_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8364_cast_fp16 = mul(x = var_8363_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8364_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_137_cast_fp16 = add(x = var_8343_cast_fp16, y = var_8364_cast_fp16)[name = tensor<string, []>("q_patches_137_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8366_cast_fp16 = mul(x = var_8341_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8366_cast_fp16")];
            tensor<int32, [5]> var_8374 = const()[name = tensor<string, []>("op_8374"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_933_cast_fp16 = reshape(shape = var_8374, x = var_8341_cast_fp16)[name = tensor<string, []>("x_933_cast_fp16")];
            tensor<int32, [2]> var_8376_split_sizes_0 = const()[name = tensor<string, []>("op_8376_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8376_axis_0 = const()[name = tensor<string, []>("op_8376_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8376_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8376_cast_fp16_1 = split(axis = var_8376_axis_0, split_sizes = var_8376_split_sizes_0, x = x_933_cast_fp16)[name = tensor<string, []>("op_8376_cast_fp16")];
            tensor<int32, [1]> squeeze_274_axes_0 = const()[name = tensor<string, []>("squeeze_274_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_274_cast_fp16 = squeeze(axes = squeeze_274_axes_0, x = var_8376_cast_fp16_0)[name = tensor<string, []>("squeeze_274_cast_fp16")];
            tensor<int32, [1]> squeeze_275_axes_0 = const()[name = tensor<string, []>("squeeze_275_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_275_cast_fp16 = squeeze(axes = squeeze_275_axes_0, x = var_8376_cast_fp16_1)[name = tensor<string, []>("squeeze_275_cast_fp16")];
            tensor<fp16, []> const_1123_promoted_to_fp16 = const()[name = tensor<string, []>("const_1123_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8379_cast_fp16 = mul(x = squeeze_275_cast_fp16, y = const_1123_promoted_to_fp16)[name = tensor<string, []>("op_8379_cast_fp16")];
            tensor<int32, []> x_935_axis_0 = const()[name = tensor<string, []>("x_935_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_935_cast_fp16 = stack(axis = x_935_axis_0, values = (var_8379_cast_fp16, squeeze_274_cast_fp16))[name = tensor<string, []>("x_935_cast_fp16")];
            tensor<int32, [4]> var_8385 = const()[name = tensor<string, []>("op_8385"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8386_cast_fp16 = reshape(shape = var_8385, x = x_935_cast_fp16)[name = tensor<string, []>("op_8386_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8387_cast_fp16 = mul(x = var_8386_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8387_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_137_cast_fp16 = add(x = var_8366_cast_fp16, y = var_8387_cast_fp16)[name = tensor<string, []>("k_patches_137_cast_fp16")];
            tensor<bool, []> var_8390_interleave_0 = const()[name = tensor<string, []>("op_8390_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8390_cast_fp16 = concat(axis = var_20, interleave = var_8390_interleave_0, values = (var_8329_cast_fp16, q_patches_137_cast_fp16))[name = tensor<string, []>("op_8390_cast_fp16")];
            tensor<bool, []> var_8393_interleave_0 = const()[name = tensor<string, []>("op_8393_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8393_cast_fp16 = concat(axis = var_20, interleave = var_8393_interleave_0, values = (var_8333_cast_fp16, k_patches_137_cast_fp16))[name = tensor<string, []>("op_8393_cast_fp16")];
            tensor<fp16, []> var_8395_to_fp16 = const()[name = tensor<string, []>("op_8395_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_689_cast_fp16 = mul(x = var_8390_cast_fp16, y = var_8395_to_fp16)[name = tensor<string, []>("q_689_cast_fp16")];
            tensor<bool, []> attn_273_transpose_x_1 = const()[name = tensor<string, []>("attn_273_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_273_transpose_y_1 = const()[name = tensor<string, []>("attn_273_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_273_cast_fp16 = matmul(transpose_x = attn_273_transpose_x_1, transpose_y = attn_273_transpose_y_1, x = q_689_cast_fp16, y = var_8393_cast_fp16)[name = tensor<string, []>("attn_273_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_275_cast_fp16 = softmax(axis = var_21, x = attn_273_cast_fp16)[name = tensor<string, []>("attn_275_cast_fp16")];
            tensor<bool, []> var_8400_transpose_x_0 = const()[name = tensor<string, []>("op_8400_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8400_transpose_y_0 = const()[name = tensor<string, []>("op_8400_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8400_cast_fp16 = matmul(transpose_x = var_8400_transpose_x_0, transpose_y = var_8400_transpose_y_0, x = attn_275_cast_fp16, y = v_413_cast_fp16)[name = tensor<string, []>("op_8400_cast_fp16")];
            tensor<int32, [1]> var_8401_axes_0 = const()[name = tensor<string, []>("op_8401_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8401_cast_fp16 = squeeze(axes = var_8401_axes_0, x = var_8400_cast_fp16)[name = tensor<string, []>("op_8401_cast_fp16")];
            tensor<string, []> var_8407_pad_type_0 = const()[name = tensor<string, []>("op_8407_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8407_strides_0 = const()[name = tensor<string, []>("op_8407_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8407_pad_0 = const()[name = tensor<string, []>("op_8407_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8407_dilations_0 = const()[name = tensor<string, []>("op_8407_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8407_groups_0 = const()[name = tensor<string, []>("op_8407_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75110144)))];
            tensor<fp16, [1, 64, 1, 197]> var_8407_cast_fp16 = conv(dilations = var_8407_dilations_0, groups = var_8407_groups_0, pad = var_8407_pad_0, pad_type = var_8407_pad_type_0, strides = var_8407_strides_0, weight = model_blocks_5_attn_q_projs_9_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8407_cast_fp16")];
            tensor<int32, [1]> var_8408_axes_0 = const()[name = tensor<string, []>("op_8408_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8408_cast_fp16 = squeeze(axes = var_8408_axes_0, x = var_8407_cast_fp16)[name = tensor<string, []>("op_8408_cast_fp16")];
            tensor<int32, [3]> q_691_perm_0 = const()[name = tensor<string, []>("q_691_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8415_pad_type_0 = const()[name = tensor<string, []>("op_8415_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8415_strides_0 = const()[name = tensor<string, []>("op_8415_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8415_pad_0 = const()[name = tensor<string, []>("op_8415_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8415_dilations_0 = const()[name = tensor<string, []>("op_8415_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8415_groups_0 = const()[name = tensor<string, []>("op_8415_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75208512)))];
            tensor<fp16, [1, 64, 1, 197]> var_8415_cast_fp16 = conv(dilations = var_8415_dilations_0, groups = var_8415_groups_0, pad = var_8415_pad_0, pad_type = var_8415_pad_type_0, strides = var_8415_strides_0, weight = model_blocks_5_attn_k_projs_9_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8415_cast_fp16")];
            tensor<int32, [1]> var_8416_axes_0 = const()[name = tensor<string, []>("op_8416_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8416_cast_fp16 = squeeze(axes = var_8416_axes_0, x = var_8415_cast_fp16)[name = tensor<string, []>("op_8416_cast_fp16")];
            tensor<int32, [3]> k_415_perm_0 = const()[name = tensor<string, []>("k_415_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8423_pad_type_0 = const()[name = tensor<string, []>("op_8423_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8423_strides_0 = const()[name = tensor<string, []>("op_8423_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8423_pad_0 = const()[name = tensor<string, []>("op_8423_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8423_dilations_0 = const()[name = tensor<string, []>("op_8423_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8423_groups_0 = const()[name = tensor<string, []>("op_8423_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75306880)))];
            tensor<fp16, [1, 64, 1, 197]> var_8423_cast_fp16 = conv(dilations = var_8423_dilations_0, groups = var_8423_groups_0, pad = var_8423_pad_0, pad_type = var_8423_pad_type_0, strides = var_8423_strides_0, weight = model_blocks_5_attn_v_projs_9_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8423_cast_fp16")];
            tensor<int32, [1]> var_8424_axes_0 = const()[name = tensor<string, []>("op_8424_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8424_cast_fp16 = squeeze(axes = var_8424_axes_0, x = var_8423_cast_fp16)[name = tensor<string, []>("op_8424_cast_fp16")];
            tensor<int32, [3]> v_415_perm_0 = const()[name = tensor<string, []>("v_415_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75405248)))];
            tensor<fp16, [1, 197, 64]> q_691_cast_fp16 = transpose(perm = q_691_perm_0, x = var_8408_cast_fp16)[name = tensor<string, []>("transpose_272")];
            tensor<fp16, [1, 197, 64]> q_693_cast_fp16 = add(x = q_691_cast_fp16, y = model_blocks_5_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_693_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75405440)))];
            tensor<fp16, [1, 197, 64]> v_415_cast_fp16 = transpose(perm = v_415_perm_0, x = var_8424_cast_fp16)[name = tensor<string, []>("transpose_270")];
            tensor<fp16, [1, 197, 64]> v_417_cast_fp16 = add(x = v_415_cast_fp16, y = model_blocks_5_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_417_cast_fp16")];
            tensor<int32, [1]> q_695_axes_0 = const()[name = tensor<string, []>("q_695_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_695_cast_fp16 = expand_dims(axes = q_695_axes_0, x = q_693_cast_fp16)[name = tensor<string, []>("q_695_cast_fp16")];
            tensor<int32, [1]> k_417_axes_0 = const()[name = tensor<string, []>("k_417_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_415_cast_fp16 = transpose(perm = k_415_perm_0, x = var_8416_cast_fp16)[name = tensor<string, []>("transpose_271")];
            tensor<fp16, [1, 1, 197, 64]> k_417_cast_fp16 = expand_dims(axes = k_417_axes_0, x = k_415_cast_fp16)[name = tensor<string, []>("k_417_cast_fp16")];
            tensor<int32, [1]> v_419_axes_0 = const()[name = tensor<string, []>("v_419_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_419_cast_fp16 = expand_dims(axes = v_419_axes_0, x = v_417_cast_fp16)[name = tensor<string, []>("v_419_cast_fp16")];
            tensor<int32, [4]> var_8433_begin_0 = const()[name = tensor<string, []>("op_8433_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8433_end_0 = const()[name = tensor<string, []>("op_8433_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8433_end_mask_0 = const()[name = tensor<string, []>("op_8433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8433_cast_fp16 = slice_by_index(begin = var_8433_begin_0, end = var_8433_end_0, end_mask = var_8433_end_mask_0, x = q_695_cast_fp16)[name = tensor<string, []>("op_8433_cast_fp16")];
            tensor<int32, [4]> var_8437_begin_0 = const()[name = tensor<string, []>("op_8437_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8437_end_0 = const()[name = tensor<string, []>("op_8437_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8437_end_mask_0 = const()[name = tensor<string, []>("op_8437_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8437_cast_fp16 = slice_by_index(begin = var_8437_begin_0, end = var_8437_end_0, end_mask = var_8437_end_mask_0, x = k_417_cast_fp16)[name = tensor<string, []>("op_8437_cast_fp16")];
            tensor<int32, [4]> var_8441_begin_0 = const()[name = tensor<string, []>("op_8441_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8441_end_0 = const()[name = tensor<string, []>("op_8441_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8441_end_mask_0 = const()[name = tensor<string, []>("op_8441_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8441_cast_fp16 = slice_by_index(begin = var_8441_begin_0, end = var_8441_end_0, end_mask = var_8441_end_mask_0, x = q_695_cast_fp16)[name = tensor<string, []>("op_8441_cast_fp16")];
            tensor<int32, [4]> var_8445_begin_0 = const()[name = tensor<string, []>("op_8445_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8445_end_0 = const()[name = tensor<string, []>("op_8445_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8445_end_mask_0 = const()[name = tensor<string, []>("op_8445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, x = k_417_cast_fp16)[name = tensor<string, []>("op_8445_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8447_cast_fp16 = mul(x = var_8441_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8447_cast_fp16")];
            tensor<int32, [5]> var_8455 = const()[name = tensor<string, []>("op_8455"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_939_cast_fp16 = reshape(shape = var_8455, x = var_8441_cast_fp16)[name = tensor<string, []>("x_939_cast_fp16")];
            tensor<int32, [2]> var_8457_split_sizes_0 = const()[name = tensor<string, []>("op_8457_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8457_axis_0 = const()[name = tensor<string, []>("op_8457_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8457_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8457_cast_fp16_1 = split(axis = var_8457_axis_0, split_sizes = var_8457_split_sizes_0, x = x_939_cast_fp16)[name = tensor<string, []>("op_8457_cast_fp16")];
            tensor<int32, [1]> squeeze_276_axes_0 = const()[name = tensor<string, []>("squeeze_276_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_276_cast_fp16 = squeeze(axes = squeeze_276_axes_0, x = var_8457_cast_fp16_0)[name = tensor<string, []>("squeeze_276_cast_fp16")];
            tensor<int32, [1]> squeeze_277_axes_0 = const()[name = tensor<string, []>("squeeze_277_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_277_cast_fp16 = squeeze(axes = squeeze_277_axes_0, x = var_8457_cast_fp16_1)[name = tensor<string, []>("squeeze_277_cast_fp16")];
            tensor<fp16, []> const_1131_promoted_to_fp16 = const()[name = tensor<string, []>("const_1131_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8460_cast_fp16 = mul(x = squeeze_277_cast_fp16, y = const_1131_promoted_to_fp16)[name = tensor<string, []>("op_8460_cast_fp16")];
            tensor<int32, []> x_941_axis_0 = const()[name = tensor<string, []>("x_941_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_941_cast_fp16 = stack(axis = x_941_axis_0, values = (var_8460_cast_fp16, squeeze_276_cast_fp16))[name = tensor<string, []>("x_941_cast_fp16")];
            tensor<int32, [4]> var_8466 = const()[name = tensor<string, []>("op_8466"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8467_cast_fp16 = reshape(shape = var_8466, x = x_941_cast_fp16)[name = tensor<string, []>("op_8467_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8468_cast_fp16 = mul(x = var_8467_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8468_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_139_cast_fp16 = add(x = var_8447_cast_fp16, y = var_8468_cast_fp16)[name = tensor<string, []>("q_patches_139_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8470_cast_fp16 = mul(x = var_8445_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8470_cast_fp16")];
            tensor<int32, [5]> var_8478 = const()[name = tensor<string, []>("op_8478"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_945_cast_fp16 = reshape(shape = var_8478, x = var_8445_cast_fp16)[name = tensor<string, []>("x_945_cast_fp16")];
            tensor<int32, [2]> var_8480_split_sizes_0 = const()[name = tensor<string, []>("op_8480_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8480_axis_0 = const()[name = tensor<string, []>("op_8480_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8480_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8480_cast_fp16_1 = split(axis = var_8480_axis_0, split_sizes = var_8480_split_sizes_0, x = x_945_cast_fp16)[name = tensor<string, []>("op_8480_cast_fp16")];
            tensor<int32, [1]> squeeze_278_axes_0 = const()[name = tensor<string, []>("squeeze_278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_278_cast_fp16 = squeeze(axes = squeeze_278_axes_0, x = var_8480_cast_fp16_0)[name = tensor<string, []>("squeeze_278_cast_fp16")];
            tensor<int32, [1]> squeeze_279_axes_0 = const()[name = tensor<string, []>("squeeze_279_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_279_cast_fp16 = squeeze(axes = squeeze_279_axes_0, x = var_8480_cast_fp16_1)[name = tensor<string, []>("squeeze_279_cast_fp16")];
            tensor<fp16, []> const_1139_promoted_to_fp16 = const()[name = tensor<string, []>("const_1139_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8483_cast_fp16 = mul(x = squeeze_279_cast_fp16, y = const_1139_promoted_to_fp16)[name = tensor<string, []>("op_8483_cast_fp16")];
            tensor<int32, []> x_947_axis_0 = const()[name = tensor<string, []>("x_947_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_947_cast_fp16 = stack(axis = x_947_axis_0, values = (var_8483_cast_fp16, squeeze_278_cast_fp16))[name = tensor<string, []>("x_947_cast_fp16")];
            tensor<int32, [4]> var_8489 = const()[name = tensor<string, []>("op_8489"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8490_cast_fp16 = reshape(shape = var_8489, x = x_947_cast_fp16)[name = tensor<string, []>("op_8490_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8491_cast_fp16 = mul(x = var_8490_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8491_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_139_cast_fp16 = add(x = var_8470_cast_fp16, y = var_8491_cast_fp16)[name = tensor<string, []>("k_patches_139_cast_fp16")];
            tensor<bool, []> var_8494_interleave_0 = const()[name = tensor<string, []>("op_8494_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8494_cast_fp16 = concat(axis = var_20, interleave = var_8494_interleave_0, values = (var_8433_cast_fp16, q_patches_139_cast_fp16))[name = tensor<string, []>("op_8494_cast_fp16")];
            tensor<bool, []> var_8497_interleave_0 = const()[name = tensor<string, []>("op_8497_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8497_cast_fp16 = concat(axis = var_20, interleave = var_8497_interleave_0, values = (var_8437_cast_fp16, k_patches_139_cast_fp16))[name = tensor<string, []>("op_8497_cast_fp16")];
            tensor<fp16, []> var_8499_to_fp16 = const()[name = tensor<string, []>("op_8499_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_699_cast_fp16 = mul(x = var_8494_cast_fp16, y = var_8499_to_fp16)[name = tensor<string, []>("q_699_cast_fp16")];
            tensor<bool, []> attn_277_transpose_x_1 = const()[name = tensor<string, []>("attn_277_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_277_transpose_y_1 = const()[name = tensor<string, []>("attn_277_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_277_cast_fp16 = matmul(transpose_x = attn_277_transpose_x_1, transpose_y = attn_277_transpose_y_1, x = q_699_cast_fp16, y = var_8497_cast_fp16)[name = tensor<string, []>("attn_277_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_279_cast_fp16 = softmax(axis = var_21, x = attn_277_cast_fp16)[name = tensor<string, []>("attn_279_cast_fp16")];
            tensor<bool, []> var_8504_transpose_x_0 = const()[name = tensor<string, []>("op_8504_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8504_transpose_y_0 = const()[name = tensor<string, []>("op_8504_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8504_cast_fp16 = matmul(transpose_x = var_8504_transpose_x_0, transpose_y = var_8504_transpose_y_0, x = attn_279_cast_fp16, y = v_419_cast_fp16)[name = tensor<string, []>("op_8504_cast_fp16")];
            tensor<int32, [1]> var_8505_axes_0 = const()[name = tensor<string, []>("op_8505_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8505_cast_fp16 = squeeze(axes = var_8505_axes_0, x = var_8504_cast_fp16)[name = tensor<string, []>("op_8505_cast_fp16")];
            tensor<string, []> var_8511_pad_type_0 = const()[name = tensor<string, []>("op_8511_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8511_strides_0 = const()[name = tensor<string, []>("op_8511_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8511_pad_0 = const()[name = tensor<string, []>("op_8511_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8511_dilations_0 = const()[name = tensor<string, []>("op_8511_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8511_groups_0 = const()[name = tensor<string, []>("op_8511_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75405632)))];
            tensor<fp16, [1, 64, 1, 197]> var_8511_cast_fp16 = conv(dilations = var_8511_dilations_0, groups = var_8511_groups_0, pad = var_8511_pad_0, pad_type = var_8511_pad_type_0, strides = var_8511_strides_0, weight = model_blocks_5_attn_q_projs_10_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8511_cast_fp16")];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = tensor<string, []>("op_8512_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8512_cast_fp16 = squeeze(axes = var_8512_axes_0, x = var_8511_cast_fp16)[name = tensor<string, []>("op_8512_cast_fp16")];
            tensor<int32, [3]> q_701_perm_0 = const()[name = tensor<string, []>("q_701_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8519_pad_type_0 = const()[name = tensor<string, []>("op_8519_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8519_strides_0 = const()[name = tensor<string, []>("op_8519_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8519_pad_0 = const()[name = tensor<string, []>("op_8519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8519_dilations_0 = const()[name = tensor<string, []>("op_8519_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8519_groups_0 = const()[name = tensor<string, []>("op_8519_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75504000)))];
            tensor<fp16, [1, 64, 1, 197]> var_8519_cast_fp16 = conv(dilations = var_8519_dilations_0, groups = var_8519_groups_0, pad = var_8519_pad_0, pad_type = var_8519_pad_type_0, strides = var_8519_strides_0, weight = model_blocks_5_attn_k_projs_10_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8519_cast_fp16")];
            tensor<int32, [1]> var_8520_axes_0 = const()[name = tensor<string, []>("op_8520_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8520_cast_fp16 = squeeze(axes = var_8520_axes_0, x = var_8519_cast_fp16)[name = tensor<string, []>("op_8520_cast_fp16")];
            tensor<int32, [3]> k_421_perm_0 = const()[name = tensor<string, []>("k_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8527_pad_type_0 = const()[name = tensor<string, []>("op_8527_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8527_strides_0 = const()[name = tensor<string, []>("op_8527_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8527_pad_0 = const()[name = tensor<string, []>("op_8527_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8527_dilations_0 = const()[name = tensor<string, []>("op_8527_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8527_groups_0 = const()[name = tensor<string, []>("op_8527_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75602368)))];
            tensor<fp16, [1, 64, 1, 197]> var_8527_cast_fp16 = conv(dilations = var_8527_dilations_0, groups = var_8527_groups_0, pad = var_8527_pad_0, pad_type = var_8527_pad_type_0, strides = var_8527_strides_0, weight = model_blocks_5_attn_v_projs_10_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8527_cast_fp16")];
            tensor<int32, [1]> var_8528_axes_0 = const()[name = tensor<string, []>("op_8528_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8528_cast_fp16 = squeeze(axes = var_8528_axes_0, x = var_8527_cast_fp16)[name = tensor<string, []>("op_8528_cast_fp16")];
            tensor<int32, [3]> v_421_perm_0 = const()[name = tensor<string, []>("v_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75700736)))];
            tensor<fp16, [1, 197, 64]> q_701_cast_fp16 = transpose(perm = q_701_perm_0, x = var_8512_cast_fp16)[name = tensor<string, []>("transpose_269")];
            tensor<fp16, [1, 197, 64]> q_703_cast_fp16 = add(x = q_701_cast_fp16, y = model_blocks_5_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_703_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75700928)))];
            tensor<fp16, [1, 197, 64]> v_421_cast_fp16 = transpose(perm = v_421_perm_0, x = var_8528_cast_fp16)[name = tensor<string, []>("transpose_267")];
            tensor<fp16, [1, 197, 64]> v_423_cast_fp16 = add(x = v_421_cast_fp16, y = model_blocks_5_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_423_cast_fp16")];
            tensor<int32, [1]> q_705_axes_0 = const()[name = tensor<string, []>("q_705_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_705_cast_fp16 = expand_dims(axes = q_705_axes_0, x = q_703_cast_fp16)[name = tensor<string, []>("q_705_cast_fp16")];
            tensor<int32, [1]> k_423_axes_0 = const()[name = tensor<string, []>("k_423_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_421_cast_fp16 = transpose(perm = k_421_perm_0, x = var_8520_cast_fp16)[name = tensor<string, []>("transpose_268")];
            tensor<fp16, [1, 1, 197, 64]> k_423_cast_fp16 = expand_dims(axes = k_423_axes_0, x = k_421_cast_fp16)[name = tensor<string, []>("k_423_cast_fp16")];
            tensor<int32, [1]> v_425_axes_0 = const()[name = tensor<string, []>("v_425_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_425_cast_fp16 = expand_dims(axes = v_425_axes_0, x = v_423_cast_fp16)[name = tensor<string, []>("v_425_cast_fp16")];
            tensor<int32, [4]> var_8537_begin_0 = const()[name = tensor<string, []>("op_8537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8537_end_0 = const()[name = tensor<string, []>("op_8537_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8537_end_mask_0 = const()[name = tensor<string, []>("op_8537_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8537_cast_fp16 = slice_by_index(begin = var_8537_begin_0, end = var_8537_end_0, end_mask = var_8537_end_mask_0, x = q_705_cast_fp16)[name = tensor<string, []>("op_8537_cast_fp16")];
            tensor<int32, [4]> var_8541_begin_0 = const()[name = tensor<string, []>("op_8541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8541_end_0 = const()[name = tensor<string, []>("op_8541_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8541_end_mask_0 = const()[name = tensor<string, []>("op_8541_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8541_cast_fp16 = slice_by_index(begin = var_8541_begin_0, end = var_8541_end_0, end_mask = var_8541_end_mask_0, x = k_423_cast_fp16)[name = tensor<string, []>("op_8541_cast_fp16")];
            tensor<int32, [4]> var_8545_begin_0 = const()[name = tensor<string, []>("op_8545_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8545_end_0 = const()[name = tensor<string, []>("op_8545_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8545_end_mask_0 = const()[name = tensor<string, []>("op_8545_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8545_cast_fp16 = slice_by_index(begin = var_8545_begin_0, end = var_8545_end_0, end_mask = var_8545_end_mask_0, x = q_705_cast_fp16)[name = tensor<string, []>("op_8545_cast_fp16")];
            tensor<int32, [4]> var_8549_begin_0 = const()[name = tensor<string, []>("op_8549_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8549_end_0 = const()[name = tensor<string, []>("op_8549_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8549_end_mask_0 = const()[name = tensor<string, []>("op_8549_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, x = k_423_cast_fp16)[name = tensor<string, []>("op_8549_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8551_cast_fp16 = mul(x = var_8545_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8551_cast_fp16")];
            tensor<int32, [5]> var_8559 = const()[name = tensor<string, []>("op_8559"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_951_cast_fp16 = reshape(shape = var_8559, x = var_8545_cast_fp16)[name = tensor<string, []>("x_951_cast_fp16")];
            tensor<int32, [2]> var_8561_split_sizes_0 = const()[name = tensor<string, []>("op_8561_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8561_axis_0 = const()[name = tensor<string, []>("op_8561_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8561_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8561_cast_fp16_1 = split(axis = var_8561_axis_0, split_sizes = var_8561_split_sizes_0, x = x_951_cast_fp16)[name = tensor<string, []>("op_8561_cast_fp16")];
            tensor<int32, [1]> squeeze_280_axes_0 = const()[name = tensor<string, []>("squeeze_280_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_280_cast_fp16 = squeeze(axes = squeeze_280_axes_0, x = var_8561_cast_fp16_0)[name = tensor<string, []>("squeeze_280_cast_fp16")];
            tensor<int32, [1]> squeeze_281_axes_0 = const()[name = tensor<string, []>("squeeze_281_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_281_cast_fp16 = squeeze(axes = squeeze_281_axes_0, x = var_8561_cast_fp16_1)[name = tensor<string, []>("squeeze_281_cast_fp16")];
            tensor<fp16, []> const_1147_promoted_to_fp16 = const()[name = tensor<string, []>("const_1147_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8564_cast_fp16 = mul(x = squeeze_281_cast_fp16, y = const_1147_promoted_to_fp16)[name = tensor<string, []>("op_8564_cast_fp16")];
            tensor<int32, []> x_953_axis_0 = const()[name = tensor<string, []>("x_953_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_953_cast_fp16 = stack(axis = x_953_axis_0, values = (var_8564_cast_fp16, squeeze_280_cast_fp16))[name = tensor<string, []>("x_953_cast_fp16")];
            tensor<int32, [4]> var_8570 = const()[name = tensor<string, []>("op_8570"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8571_cast_fp16 = reshape(shape = var_8570, x = x_953_cast_fp16)[name = tensor<string, []>("op_8571_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8572_cast_fp16 = mul(x = var_8571_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8572_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_141_cast_fp16 = add(x = var_8551_cast_fp16, y = var_8572_cast_fp16)[name = tensor<string, []>("q_patches_141_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8574_cast_fp16 = mul(x = var_8549_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8574_cast_fp16")];
            tensor<int32, [5]> var_8582 = const()[name = tensor<string, []>("op_8582"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_957_cast_fp16 = reshape(shape = var_8582, x = var_8549_cast_fp16)[name = tensor<string, []>("x_957_cast_fp16")];
            tensor<int32, [2]> var_8584_split_sizes_0 = const()[name = tensor<string, []>("op_8584_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8584_axis_0 = const()[name = tensor<string, []>("op_8584_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8584_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8584_cast_fp16_1 = split(axis = var_8584_axis_0, split_sizes = var_8584_split_sizes_0, x = x_957_cast_fp16)[name = tensor<string, []>("op_8584_cast_fp16")];
            tensor<int32, [1]> squeeze_282_axes_0 = const()[name = tensor<string, []>("squeeze_282_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_282_cast_fp16 = squeeze(axes = squeeze_282_axes_0, x = var_8584_cast_fp16_0)[name = tensor<string, []>("squeeze_282_cast_fp16")];
            tensor<int32, [1]> squeeze_283_axes_0 = const()[name = tensor<string, []>("squeeze_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_283_cast_fp16 = squeeze(axes = squeeze_283_axes_0, x = var_8584_cast_fp16_1)[name = tensor<string, []>("squeeze_283_cast_fp16")];
            tensor<fp16, []> const_1155_promoted_to_fp16 = const()[name = tensor<string, []>("const_1155_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8587_cast_fp16 = mul(x = squeeze_283_cast_fp16, y = const_1155_promoted_to_fp16)[name = tensor<string, []>("op_8587_cast_fp16")];
            tensor<int32, []> x_959_axis_0 = const()[name = tensor<string, []>("x_959_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_959_cast_fp16 = stack(axis = x_959_axis_0, values = (var_8587_cast_fp16, squeeze_282_cast_fp16))[name = tensor<string, []>("x_959_cast_fp16")];
            tensor<int32, [4]> var_8593 = const()[name = tensor<string, []>("op_8593"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8594_cast_fp16 = reshape(shape = var_8593, x = x_959_cast_fp16)[name = tensor<string, []>("op_8594_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8595_cast_fp16 = mul(x = var_8594_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8595_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_141_cast_fp16 = add(x = var_8574_cast_fp16, y = var_8595_cast_fp16)[name = tensor<string, []>("k_patches_141_cast_fp16")];
            tensor<bool, []> var_8598_interleave_0 = const()[name = tensor<string, []>("op_8598_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8598_cast_fp16 = concat(axis = var_20, interleave = var_8598_interleave_0, values = (var_8537_cast_fp16, q_patches_141_cast_fp16))[name = tensor<string, []>("op_8598_cast_fp16")];
            tensor<bool, []> var_8601_interleave_0 = const()[name = tensor<string, []>("op_8601_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8601_cast_fp16 = concat(axis = var_20, interleave = var_8601_interleave_0, values = (var_8541_cast_fp16, k_patches_141_cast_fp16))[name = tensor<string, []>("op_8601_cast_fp16")];
            tensor<fp16, []> var_8603_to_fp16 = const()[name = tensor<string, []>("op_8603_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_709_cast_fp16 = mul(x = var_8598_cast_fp16, y = var_8603_to_fp16)[name = tensor<string, []>("q_709_cast_fp16")];
            tensor<bool, []> attn_281_transpose_x_1 = const()[name = tensor<string, []>("attn_281_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_281_transpose_y_1 = const()[name = tensor<string, []>("attn_281_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_281_cast_fp16 = matmul(transpose_x = attn_281_transpose_x_1, transpose_y = attn_281_transpose_y_1, x = q_709_cast_fp16, y = var_8601_cast_fp16)[name = tensor<string, []>("attn_281_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_283_cast_fp16 = softmax(axis = var_21, x = attn_281_cast_fp16)[name = tensor<string, []>("attn_283_cast_fp16")];
            tensor<bool, []> var_8608_transpose_x_0 = const()[name = tensor<string, []>("op_8608_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8608_transpose_y_0 = const()[name = tensor<string, []>("op_8608_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8608_cast_fp16 = matmul(transpose_x = var_8608_transpose_x_0, transpose_y = var_8608_transpose_y_0, x = attn_283_cast_fp16, y = v_425_cast_fp16)[name = tensor<string, []>("op_8608_cast_fp16")];
            tensor<int32, [1]> var_8609_axes_0 = const()[name = tensor<string, []>("op_8609_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_8609_cast_fp16 = squeeze(axes = var_8609_axes_0, x = var_8608_cast_fp16)[name = tensor<string, []>("op_8609_cast_fp16")];
            tensor<string, []> var_8615_pad_type_0 = const()[name = tensor<string, []>("op_8615_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8615_strides_0 = const()[name = tensor<string, []>("op_8615_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8615_pad_0 = const()[name = tensor<string, []>("op_8615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8615_dilations_0 = const()[name = tensor<string, []>("op_8615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8615_groups_0 = const()[name = tensor<string, []>("op_8615_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75701120)))];
            tensor<fp16, [1, 64, 1, 197]> var_8615_cast_fp16 = conv(dilations = var_8615_dilations_0, groups = var_8615_groups_0, pad = var_8615_pad_0, pad_type = var_8615_pad_type_0, strides = var_8615_strides_0, weight = model_blocks_5_attn_q_projs_11_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8615_cast_fp16")];
            tensor<int32, [1]> var_8616_axes_0 = const()[name = tensor<string, []>("op_8616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8616_cast_fp16 = squeeze(axes = var_8616_axes_0, x = var_8615_cast_fp16)[name = tensor<string, []>("op_8616_cast_fp16")];
            tensor<int32, [3]> q_711_perm_0 = const()[name = tensor<string, []>("q_711_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8623_pad_type_0 = const()[name = tensor<string, []>("op_8623_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8623_strides_0 = const()[name = tensor<string, []>("op_8623_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8623_pad_0 = const()[name = tensor<string, []>("op_8623_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8623_dilations_0 = const()[name = tensor<string, []>("op_8623_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8623_groups_0 = const()[name = tensor<string, []>("op_8623_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75799488)))];
            tensor<fp16, [1, 64, 1, 197]> var_8623_cast_fp16 = conv(dilations = var_8623_dilations_0, groups = var_8623_groups_0, pad = var_8623_pad_0, pad_type = var_8623_pad_type_0, strides = var_8623_strides_0, weight = model_blocks_5_attn_k_projs_11_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8623_cast_fp16")];
            tensor<int32, [1]> var_8624_axes_0 = const()[name = tensor<string, []>("op_8624_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8624_cast_fp16 = squeeze(axes = var_8624_axes_0, x = var_8623_cast_fp16)[name = tensor<string, []>("op_8624_cast_fp16")];
            tensor<int32, [3]> k_427_perm_0 = const()[name = tensor<string, []>("k_427_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8631_pad_type_0 = const()[name = tensor<string, []>("op_8631_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8631_strides_0 = const()[name = tensor<string, []>("op_8631_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8631_pad_0 = const()[name = tensor<string, []>("op_8631_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8631_dilations_0 = const()[name = tensor<string, []>("op_8631_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8631_groups_0 = const()[name = tensor<string, []>("op_8631_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_5_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75897856)))];
            tensor<fp16, [1, 64, 1, 197]> var_8631_cast_fp16 = conv(dilations = var_8631_dilations_0, groups = var_8631_groups_0, pad = var_8631_pad_0, pad_type = var_8631_pad_type_0, strides = var_8631_strides_0, weight = model_blocks_5_attn_v_projs_11_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_8631_cast_fp16")];
            tensor<int32, [1]> var_8632_axes_0 = const()[name = tensor<string, []>("op_8632_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8632_cast_fp16 = squeeze(axes = var_8632_axes_0, x = var_8631_cast_fp16)[name = tensor<string, []>("op_8632_cast_fp16")];
            tensor<int32, [3]> v_427_perm_0 = const()[name = tensor<string, []>("v_427_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_5_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75996224)))];
            tensor<fp16, [1, 197, 64]> q_711_cast_fp16 = transpose(perm = q_711_perm_0, x = var_8616_cast_fp16)[name = tensor<string, []>("transpose_266")];
            tensor<fp16, [1, 197, 64]> q_713_cast_fp16 = add(x = q_711_cast_fp16, y = model_blocks_5_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_713_cast_fp16")];
            tensor<fp16, [64]> model_blocks_5_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75996416)))];
            tensor<fp16, [1, 197, 64]> v_427_cast_fp16 = transpose(perm = v_427_perm_0, x = var_8632_cast_fp16)[name = tensor<string, []>("transpose_264")];
            tensor<fp16, [1, 197, 64]> v_429_cast_fp16 = add(x = v_427_cast_fp16, y = model_blocks_5_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_429_cast_fp16")];
            tensor<int32, [1]> q_715_axes_0 = const()[name = tensor<string, []>("q_715_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_715_cast_fp16 = expand_dims(axes = q_715_axes_0, x = q_713_cast_fp16)[name = tensor<string, []>("q_715_cast_fp16")];
            tensor<int32, [1]> k_429_axes_0 = const()[name = tensor<string, []>("k_429_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_427_cast_fp16 = transpose(perm = k_427_perm_0, x = var_8624_cast_fp16)[name = tensor<string, []>("transpose_265")];
            tensor<fp16, [1, 1, 197, 64]> k_429_cast_fp16 = expand_dims(axes = k_429_axes_0, x = k_427_cast_fp16)[name = tensor<string, []>("k_429_cast_fp16")];
            tensor<int32, [1]> v_431_axes_0 = const()[name = tensor<string, []>("v_431_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_431_cast_fp16 = expand_dims(axes = v_431_axes_0, x = v_429_cast_fp16)[name = tensor<string, []>("v_431_cast_fp16")];
            tensor<int32, [4]> var_8641_begin_0 = const()[name = tensor<string, []>("op_8641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8641_end_0 = const()[name = tensor<string, []>("op_8641_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8641_end_mask_0 = const()[name = tensor<string, []>("op_8641_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8641_cast_fp16 = slice_by_index(begin = var_8641_begin_0, end = var_8641_end_0, end_mask = var_8641_end_mask_0, x = q_715_cast_fp16)[name = tensor<string, []>("op_8641_cast_fp16")];
            tensor<int32, [4]> var_8645_begin_0 = const()[name = tensor<string, []>("op_8645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8645_end_0 = const()[name = tensor<string, []>("op_8645_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8645_end_mask_0 = const()[name = tensor<string, []>("op_8645_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8645_cast_fp16 = slice_by_index(begin = var_8645_begin_0, end = var_8645_end_0, end_mask = var_8645_end_mask_0, x = k_429_cast_fp16)[name = tensor<string, []>("op_8645_cast_fp16")];
            tensor<int32, [4]> var_8649_begin_0 = const()[name = tensor<string, []>("op_8649_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8649_end_0 = const()[name = tensor<string, []>("op_8649_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8649_end_mask_0 = const()[name = tensor<string, []>("op_8649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8649_cast_fp16 = slice_by_index(begin = var_8649_begin_0, end = var_8649_end_0, end_mask = var_8649_end_mask_0, x = q_715_cast_fp16)[name = tensor<string, []>("op_8649_cast_fp16")];
            tensor<int32, [4]> var_8653_begin_0 = const()[name = tensor<string, []>("op_8653_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8653_end_0 = const()[name = tensor<string, []>("op_8653_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8653_end_mask_0 = const()[name = tensor<string, []>("op_8653_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8653_cast_fp16 = slice_by_index(begin = var_8653_begin_0, end = var_8653_end_0, end_mask = var_8653_end_mask_0, x = k_429_cast_fp16)[name = tensor<string, []>("op_8653_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8655_cast_fp16 = mul(x = var_8649_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8655_cast_fp16")];
            tensor<int32, [5]> var_8663 = const()[name = tensor<string, []>("op_8663"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_963_cast_fp16 = reshape(shape = var_8663, x = var_8649_cast_fp16)[name = tensor<string, []>("x_963_cast_fp16")];
            tensor<int32, [2]> var_8665_split_sizes_0 = const()[name = tensor<string, []>("op_8665_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8665_axis_0 = const()[name = tensor<string, []>("op_8665_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8665_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8665_cast_fp16_1 = split(axis = var_8665_axis_0, split_sizes = var_8665_split_sizes_0, x = x_963_cast_fp16)[name = tensor<string, []>("op_8665_cast_fp16")];
            tensor<int32, [1]> squeeze_284_axes_0 = const()[name = tensor<string, []>("squeeze_284_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_284_cast_fp16 = squeeze(axes = squeeze_284_axes_0, x = var_8665_cast_fp16_0)[name = tensor<string, []>("squeeze_284_cast_fp16")];
            tensor<int32, [1]> squeeze_285_axes_0 = const()[name = tensor<string, []>("squeeze_285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_285_cast_fp16 = squeeze(axes = squeeze_285_axes_0, x = var_8665_cast_fp16_1)[name = tensor<string, []>("squeeze_285_cast_fp16")];
            tensor<fp16, []> const_1163_promoted_to_fp16 = const()[name = tensor<string, []>("const_1163_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8668_cast_fp16 = mul(x = squeeze_285_cast_fp16, y = const_1163_promoted_to_fp16)[name = tensor<string, []>("op_8668_cast_fp16")];
            tensor<int32, []> x_965_axis_0 = const()[name = tensor<string, []>("x_965_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_965_cast_fp16 = stack(axis = x_965_axis_0, values = (var_8668_cast_fp16, squeeze_284_cast_fp16))[name = tensor<string, []>("x_965_cast_fp16")];
            tensor<int32, [4]> var_8674 = const()[name = tensor<string, []>("op_8674"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8675_cast_fp16 = reshape(shape = var_8674, x = x_965_cast_fp16)[name = tensor<string, []>("op_8675_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8676_cast_fp16 = mul(x = var_8675_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8676_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_143_cast_fp16 = add(x = var_8655_cast_fp16, y = var_8676_cast_fp16)[name = tensor<string, []>("q_patches_143_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8678_cast_fp16 = mul(x = var_8653_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8678_cast_fp16")];
            tensor<int32, [5]> var_8686 = const()[name = tensor<string, []>("op_8686"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_969_cast_fp16 = reshape(shape = var_8686, x = var_8653_cast_fp16)[name = tensor<string, []>("x_969_cast_fp16")];
            tensor<int32, [2]> var_8688_split_sizes_0 = const()[name = tensor<string, []>("op_8688_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8688_axis_0 = const()[name = tensor<string, []>("op_8688_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8688_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8688_cast_fp16_1 = split(axis = var_8688_axis_0, split_sizes = var_8688_split_sizes_0, x = x_969_cast_fp16)[name = tensor<string, []>("op_8688_cast_fp16")];
            tensor<int32, [1]> squeeze_286_axes_0 = const()[name = tensor<string, []>("squeeze_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_286_cast_fp16 = squeeze(axes = squeeze_286_axes_0, x = var_8688_cast_fp16_0)[name = tensor<string, []>("squeeze_286_cast_fp16")];
            tensor<int32, [1]> squeeze_287_axes_0 = const()[name = tensor<string, []>("squeeze_287_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_287_cast_fp16 = squeeze(axes = squeeze_287_axes_0, x = var_8688_cast_fp16_1)[name = tensor<string, []>("squeeze_287_cast_fp16")];
            tensor<fp16, []> const_1171_promoted_to_fp16 = const()[name = tensor<string, []>("const_1171_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8691_cast_fp16 = mul(x = squeeze_287_cast_fp16, y = const_1171_promoted_to_fp16)[name = tensor<string, []>("op_8691_cast_fp16")];
            tensor<int32, []> x_971_axis_0 = const()[name = tensor<string, []>("x_971_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_971_cast_fp16 = stack(axis = x_971_axis_0, values = (var_8691_cast_fp16, squeeze_286_cast_fp16))[name = tensor<string, []>("x_971_cast_fp16")];
            tensor<int32, [4]> var_8697 = const()[name = tensor<string, []>("op_8697"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8698_cast_fp16 = reshape(shape = var_8697, x = x_971_cast_fp16)[name = tensor<string, []>("op_8698_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8699_cast_fp16 = mul(x = var_8698_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8699_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_143_cast_fp16 = add(x = var_8678_cast_fp16, y = var_8699_cast_fp16)[name = tensor<string, []>("k_patches_143_cast_fp16")];
            tensor<bool, []> var_8702_interleave_0 = const()[name = tensor<string, []>("op_8702_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8702_cast_fp16 = concat(axis = var_20, interleave = var_8702_interleave_0, values = (var_8641_cast_fp16, q_patches_143_cast_fp16))[name = tensor<string, []>("op_8702_cast_fp16")];
            tensor<bool, []> var_8705_interleave_0 = const()[name = tensor<string, []>("op_8705_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8705_cast_fp16 = concat(axis = var_20, interleave = var_8705_interleave_0, values = (var_8645_cast_fp16, k_patches_143_cast_fp16))[name = tensor<string, []>("op_8705_cast_fp16")];
            tensor<fp16, []> var_8707_to_fp16 = const()[name = tensor<string, []>("op_8707_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_719_cast_fp16 = mul(x = var_8702_cast_fp16, y = var_8707_to_fp16)[name = tensor<string, []>("q_719_cast_fp16")];
            tensor<bool, []> attn_285_transpose_x_1 = const()[name = tensor<string, []>("attn_285_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_285_transpose_y_1 = const()[name = tensor<string, []>("attn_285_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_285_cast_fp16 = matmul(transpose_x = attn_285_transpose_x_1, transpose_y = attn_285_transpose_y_1, x = q_719_cast_fp16, y = var_8705_cast_fp16)[name = tensor<string, []>("attn_285_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_287_cast_fp16 = softmax(axis = var_21, x = attn_285_cast_fp16)[name = tensor<string, []>("attn_287_cast_fp16")];
            tensor<bool, []> var_8712_transpose_x_0 = const()[name = tensor<string, []>("op_8712_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_8712_transpose_y_0 = const()[name = tensor<string, []>("op_8712_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_8712_cast_fp16 = matmul(transpose_x = var_8712_transpose_x_0, transpose_y = var_8712_transpose_y_0, x = attn_287_cast_fp16, y = v_431_cast_fp16)[name = tensor<string, []>("op_8712_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_11_cast_fp16 = squeeze(axes = out_11_axes_0, x = var_8712_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<bool, []> input_95_interleave_0 = const()[name = tensor<string, []>("input_95_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_95_cast_fp16 = concat(axis = var_21, interleave = input_95_interleave_0, values = (var_7569_cast_fp16, var_7673_cast_fp16, var_7777_cast_fp16, var_7881_cast_fp16, var_7985_cast_fp16, var_8089_cast_fp16, var_8193_cast_fp16, var_8297_cast_fp16, var_8401_cast_fp16, var_8505_cast_fp16, var_8609_cast_fp16, out_11_cast_fp16))[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<int32, [1]> x_973_axes_0 = const()[name = tensor<string, []>("x_973_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_5_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75996608)))];
            tensor<fp16, [768]> model_blocks_5_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75998208)))];
            tensor<fp16, [1, 197, 768]> x_973_cast_fp16 = layer_norm(axes = x_973_axes_0, beta = model_blocks_5_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_5_attn_inner_attn_ln_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("x_973_cast_fp16")];
            tensor<int32, [4]> var_8720 = const()[name = tensor<string, []>("op_8720"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_975_cast_fp16 = reshape(shape = var_8720, x = x_973_cast_fp16)[name = tensor<string, []>("x_975_cast_fp16")];
            tensor<int32, [4]> input_97_perm_0 = const()[name = tensor<string, []>("input_97_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_977_pad_type_0 = const()[name = tensor<string, []>("x_977_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_977_strides_0 = const()[name = tensor<string, []>("x_977_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_977_pad_0 = const()[name = tensor<string, []>("x_977_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_977_dilations_0 = const()[name = tensor<string, []>("x_977_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_977_groups_0 = const()[name = tensor<string, []>("x_977_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_5_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75999808)))];
            tensor<fp16, [768]> model_blocks_5_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77179520)))];
            tensor<fp16, [1, 768, 1, 197]> input_97_cast_fp16 = transpose(perm = input_97_perm_0, x = x_975_cast_fp16)[name = tensor<string, []>("transpose_263")];
            tensor<fp16, [1, 768, 1, 197]> x_977_cast_fp16 = conv(bias = model_blocks_5_attn_proj_bias_to_fp16, dilations = x_977_dilations_0, groups = x_977_groups_0, pad = x_977_pad_0, pad_type = x_977_pad_type_0, strides = x_977_strides_0, weight = model_blocks_5_attn_proj_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("x_977_cast_fp16")];
            tensor<int32, [4]> x_979_perm_0 = const()[name = tensor<string, []>("x_979_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_8731 = const()[name = tensor<string, []>("op_8731"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_979_cast_fp16 = transpose(perm = x_979_perm_0, x = x_977_cast_fp16)[name = tensor<string, []>("transpose_262")];
            tensor<fp16, [1, 197, 768]> var_8732_cast_fp16 = reshape(shape = var_8731, x = x_979_cast_fp16)[name = tensor<string, []>("op_8732_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_99_cast_fp16 = add(x = input_91_cast_fp16, y = var_8732_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<int32, [1]> x_981_axes_0 = const()[name = tensor<string, []>("x_981_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77181120)))];
            tensor<fp16, [768]> model_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77182720)))];
            tensor<fp16, [1, 197, 768]> x_981_cast_fp16 = layer_norm(axes = x_981_axes_0, beta = model_blocks_5_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_5_norm2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_981_cast_fp16")];
            tensor<int32, [4]> var_8744 = const()[name = tensor<string, []>("op_8744"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_983_cast_fp16 = reshape(shape = var_8744, x = x_981_cast_fp16)[name = tensor<string, []>("x_983_cast_fp16")];
            tensor<int32, [4]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_5_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77184320)))];
            tensor<fp16, [2048]> model_blocks_5_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80330112)))];
            tensor<fp16, [1, 768, 1, 197]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_983_cast_fp16)[name = tensor<string, []>("transpose_261")];
            tensor<fp16, [1, 2048, 1, 197]> input_103_cast_fp16 = conv(bias = model_blocks_5_mlp_w1_bias_to_fp16, dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = model_blocks_5_mlp_w1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<string, []> x2_299_pad_type_0 = const()[name = tensor<string, []>("x2_299_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_299_strides_0 = const()[name = tensor<string, []>("x2_299_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_299_pad_0 = const()[name = tensor<string, []>("x2_299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_299_dilations_0 = const()[name = tensor<string, []>("x2_299_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_299_groups_0 = const()[name = tensor<string, []>("x2_299_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_5_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80334272)))];
            tensor<fp16, [2048]> model_blocks_5_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83480064)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_299_cast_fp16 = conv(bias = model_blocks_5_mlp_w2_bias_to_fp16, dilations = x2_299_dilations_0, groups = x2_299_groups_0, pad = x2_299_pad_0, pad_type = x2_299_pad_type_0, strides = x2_299_strides_0, weight = model_blocks_5_mlp_w2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("x2_299_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_8761_cast_fp16 = silu(x = input_103_cast_fp16)[name = tensor<string, []>("op_8761_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_41_cast_fp16 = mul(x = var_8761_cast_fp16, y = x2_299_cast_fp16)[name = tensor<string, []>("hidden_41_cast_fp16")];
            tensor<int32, [4]> hidden_43_perm_0 = const()[name = tensor<string, []>("hidden_43_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_8764 = const()[name = tensor<string, []>("op_8764"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_43_cast_fp16 = transpose(perm = hidden_43_perm_0, x = hidden_41_cast_fp16)[name = tensor<string, []>("transpose_260")];
            tensor<fp16, [1, 197, 2048]> input_105_cast_fp16 = reshape(shape = var_8764, x = hidden_43_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<int32, [1]> hidden_45_axes_0 = const()[name = tensor<string, []>("hidden_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_5_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83484224)))];
            tensor<fp16, [2048]> model_blocks_5_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83488384)))];
            tensor<fp16, [1, 197, 2048]> hidden_45_cast_fp16 = layer_norm(axes = hidden_45_axes_0, beta = model_blocks_5_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_5_mlp_ffn_ln_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("hidden_45_cast_fp16")];
            tensor<int32, [4]> var_8770 = const()[name = tensor<string, []>("op_8770"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_47_cast_fp16 = reshape(shape = var_8770, x = hidden_45_cast_fp16)[name = tensor<string, []>("hidden_47_cast_fp16")];
            tensor<int32, [4]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_985_pad_type_0 = const()[name = tensor<string, []>("x_985_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_985_strides_0 = const()[name = tensor<string, []>("x_985_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_985_pad_0 = const()[name = tensor<string, []>("x_985_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_985_dilations_0 = const()[name = tensor<string, []>("x_985_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_985_groups_0 = const()[name = tensor<string, []>("x_985_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_5_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83492544)))];
            tensor<fp16, [768]> model_blocks_5_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_5_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86638336)))];
            tensor<fp16, [1, 2048, 1, 197]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = hidden_47_cast_fp16)[name = tensor<string, []>("transpose_259")];
            tensor<fp16, [1, 768, 1, 197]> x_985_cast_fp16 = conv(bias = model_blocks_5_mlp_w3_bias_to_fp16, dilations = x_985_dilations_0, groups = x_985_groups_0, pad = x_985_pad_0, pad_type = x_985_pad_type_0, strides = x_985_strides_0, weight = model_blocks_5_mlp_w3_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_985_cast_fp16")];
            tensor<int32, [4]> x_987_perm_0 = const()[name = tensor<string, []>("x_987_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_8781 = const()[name = tensor<string, []>("op_8781"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_987_cast_fp16 = transpose(perm = x_987_perm_0, x = x_985_cast_fp16)[name = tensor<string, []>("transpose_258")];
            tensor<fp16, [1, 197, 768]> var_8782_cast_fp16 = reshape(shape = var_8781, x = x_987_cast_fp16)[name = tensor<string, []>("op_8782_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_109_cast_fp16 = add(x = input_99_cast_fp16, y = var_8782_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<int32, [1]> x_989_axes_0 = const()[name = tensor<string, []>("x_989_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86639936)))];
            tensor<fp16, [768]> model_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86641536)))];
            tensor<fp16, [1, 197, 768]> x_989_cast_fp16 = layer_norm(axes = x_989_axes_0, beta = model_blocks_6_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_6_norm1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("x_989_cast_fp16")];
            tensor<int32, [4]> var_8916 = const()[name = tensor<string, []>("op_8916"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_991_cast_fp16 = reshape(shape = var_8916, x = x_989_cast_fp16)[name = tensor<string, []>("x_991_cast_fp16")];
            tensor<int32, [4]> input_111_perm_0 = const()[name = tensor<string, []>("input_111_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_8924_pad_type_0 = const()[name = tensor<string, []>("op_8924_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8924_strides_0 = const()[name = tensor<string, []>("op_8924_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8924_pad_0 = const()[name = tensor<string, []>("op_8924_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8924_dilations_0 = const()[name = tensor<string, []>("op_8924_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8924_groups_0 = const()[name = tensor<string, []>("op_8924_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86643136)))];
            tensor<fp16, [1, 768, 1, 197]> input_111_cast_fp16 = transpose(perm = input_111_perm_0, x = x_991_cast_fp16)[name = tensor<string, []>("transpose_257")];
            tensor<fp16, [1, 64, 1, 197]> var_8924_cast_fp16 = conv(dilations = var_8924_dilations_0, groups = var_8924_groups_0, pad = var_8924_pad_0, pad_type = var_8924_pad_type_0, strides = var_8924_strides_0, weight = model_blocks_6_attn_q_projs_0_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_8924_cast_fp16")];
            tensor<int32, [1]> var_8925_axes_0 = const()[name = tensor<string, []>("op_8925_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8925_cast_fp16 = squeeze(axes = var_8925_axes_0, x = var_8924_cast_fp16)[name = tensor<string, []>("op_8925_cast_fp16")];
            tensor<int32, [3]> q_721_perm_0 = const()[name = tensor<string, []>("q_721_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8932_pad_type_0 = const()[name = tensor<string, []>("op_8932_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8932_strides_0 = const()[name = tensor<string, []>("op_8932_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8932_pad_0 = const()[name = tensor<string, []>("op_8932_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8932_dilations_0 = const()[name = tensor<string, []>("op_8932_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8932_groups_0 = const()[name = tensor<string, []>("op_8932_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86741504)))];
            tensor<fp16, [1, 64, 1, 197]> var_8932_cast_fp16 = conv(dilations = var_8932_dilations_0, groups = var_8932_groups_0, pad = var_8932_pad_0, pad_type = var_8932_pad_type_0, strides = var_8932_strides_0, weight = model_blocks_6_attn_k_projs_0_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_8932_cast_fp16")];
            tensor<int32, [1]> var_8933_axes_0 = const()[name = tensor<string, []>("op_8933_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8933_cast_fp16 = squeeze(axes = var_8933_axes_0, x = var_8932_cast_fp16)[name = tensor<string, []>("op_8933_cast_fp16")];
            tensor<int32, [3]> k_433_perm_0 = const()[name = tensor<string, []>("k_433_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_8940_pad_type_0 = const()[name = tensor<string, []>("op_8940_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_8940_strides_0 = const()[name = tensor<string, []>("op_8940_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8940_pad_0 = const()[name = tensor<string, []>("op_8940_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8940_dilations_0 = const()[name = tensor<string, []>("op_8940_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8940_groups_0 = const()[name = tensor<string, []>("op_8940_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86839872)))];
            tensor<fp16, [1, 64, 1, 197]> var_8940_cast_fp16 = conv(dilations = var_8940_dilations_0, groups = var_8940_groups_0, pad = var_8940_pad_0, pad_type = var_8940_pad_type_0, strides = var_8940_strides_0, weight = model_blocks_6_attn_v_projs_0_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_8940_cast_fp16")];
            tensor<int32, [1]> var_8941_axes_0 = const()[name = tensor<string, []>("op_8941_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_8941_cast_fp16 = squeeze(axes = var_8941_axes_0, x = var_8940_cast_fp16)[name = tensor<string, []>("op_8941_cast_fp16")];
            tensor<int32, [3]> v_433_perm_0 = const()[name = tensor<string, []>("v_433_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86938240)))];
            tensor<fp16, [1, 197, 64]> q_721_cast_fp16 = transpose(perm = q_721_perm_0, x = var_8925_cast_fp16)[name = tensor<string, []>("transpose_256")];
            tensor<fp16, [1, 197, 64]> q_723_cast_fp16 = add(x = q_721_cast_fp16, y = model_blocks_6_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_723_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86938432)))];
            tensor<fp16, [1, 197, 64]> v_433_cast_fp16 = transpose(perm = v_433_perm_0, x = var_8941_cast_fp16)[name = tensor<string, []>("transpose_254")];
            tensor<fp16, [1, 197, 64]> v_435_cast_fp16 = add(x = v_433_cast_fp16, y = model_blocks_6_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_435_cast_fp16")];
            tensor<int32, [1]> q_725_axes_0 = const()[name = tensor<string, []>("q_725_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_725_cast_fp16 = expand_dims(axes = q_725_axes_0, x = q_723_cast_fp16)[name = tensor<string, []>("q_725_cast_fp16")];
            tensor<int32, [1]> k_435_axes_0 = const()[name = tensor<string, []>("k_435_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_433_cast_fp16 = transpose(perm = k_433_perm_0, x = var_8933_cast_fp16)[name = tensor<string, []>("transpose_255")];
            tensor<fp16, [1, 1, 197, 64]> k_435_cast_fp16 = expand_dims(axes = k_435_axes_0, x = k_433_cast_fp16)[name = tensor<string, []>("k_435_cast_fp16")];
            tensor<int32, [1]> v_437_axes_0 = const()[name = tensor<string, []>("v_437_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_437_cast_fp16 = expand_dims(axes = v_437_axes_0, x = v_435_cast_fp16)[name = tensor<string, []>("v_437_cast_fp16")];
            tensor<int32, [4]> var_8950_begin_0 = const()[name = tensor<string, []>("op_8950_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8950_end_0 = const()[name = tensor<string, []>("op_8950_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8950_end_mask_0 = const()[name = tensor<string, []>("op_8950_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8950_cast_fp16 = slice_by_index(begin = var_8950_begin_0, end = var_8950_end_0, end_mask = var_8950_end_mask_0, x = q_725_cast_fp16)[name = tensor<string, []>("op_8950_cast_fp16")];
            tensor<int32, [4]> var_8954_begin_0 = const()[name = tensor<string, []>("op_8954_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8954_end_0 = const()[name = tensor<string, []>("op_8954_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_8954_end_mask_0 = const()[name = tensor<string, []>("op_8954_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_8954_cast_fp16 = slice_by_index(begin = var_8954_begin_0, end = var_8954_end_0, end_mask = var_8954_end_mask_0, x = k_435_cast_fp16)[name = tensor<string, []>("op_8954_cast_fp16")];
            tensor<int32, [4]> var_8958_begin_0 = const()[name = tensor<string, []>("op_8958_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8958_end_0 = const()[name = tensor<string, []>("op_8958_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8958_end_mask_0 = const()[name = tensor<string, []>("op_8958_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8958_cast_fp16 = slice_by_index(begin = var_8958_begin_0, end = var_8958_end_0, end_mask = var_8958_end_mask_0, x = q_725_cast_fp16)[name = tensor<string, []>("op_8958_cast_fp16")];
            tensor<int32, [4]> var_8962_begin_0 = const()[name = tensor<string, []>("op_8962_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8962_end_0 = const()[name = tensor<string, []>("op_8962_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_8962_end_mask_0 = const()[name = tensor<string, []>("op_8962_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_8962_cast_fp16 = slice_by_index(begin = var_8962_begin_0, end = var_8962_end_0, end_mask = var_8962_end_mask_0, x = k_435_cast_fp16)[name = tensor<string, []>("op_8962_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8964_cast_fp16 = mul(x = var_8958_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8964_cast_fp16")];
            tensor<int32, [5]> var_8972 = const()[name = tensor<string, []>("op_8972"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_995_cast_fp16 = reshape(shape = var_8972, x = var_8958_cast_fp16)[name = tensor<string, []>("x_995_cast_fp16")];
            tensor<int32, [2]> var_8974_split_sizes_0 = const()[name = tensor<string, []>("op_8974_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8974_axis_0 = const()[name = tensor<string, []>("op_8974_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8974_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8974_cast_fp16_1 = split(axis = var_8974_axis_0, split_sizes = var_8974_split_sizes_0, x = x_995_cast_fp16)[name = tensor<string, []>("op_8974_cast_fp16")];
            tensor<int32, [1]> squeeze_288_axes_0 = const()[name = tensor<string, []>("squeeze_288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_288_cast_fp16 = squeeze(axes = squeeze_288_axes_0, x = var_8974_cast_fp16_0)[name = tensor<string, []>("squeeze_288_cast_fp16")];
            tensor<int32, [1]> squeeze_289_axes_0 = const()[name = tensor<string, []>("squeeze_289_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_289_cast_fp16 = squeeze(axes = squeeze_289_axes_0, x = var_8974_cast_fp16_1)[name = tensor<string, []>("squeeze_289_cast_fp16")];
            tensor<fp16, []> const_1183_promoted_to_fp16 = const()[name = tensor<string, []>("const_1183_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_8977_cast_fp16 = mul(x = squeeze_289_cast_fp16, y = const_1183_promoted_to_fp16)[name = tensor<string, []>("op_8977_cast_fp16")];
            tensor<int32, []> x_997_axis_0 = const()[name = tensor<string, []>("x_997_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_997_cast_fp16 = stack(axis = x_997_axis_0, values = (var_8977_cast_fp16, squeeze_288_cast_fp16))[name = tensor<string, []>("x_997_cast_fp16")];
            tensor<int32, [4]> var_8983 = const()[name = tensor<string, []>("op_8983"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_8984_cast_fp16 = reshape(shape = var_8983, x = x_997_cast_fp16)[name = tensor<string, []>("op_8984_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8985_cast_fp16 = mul(x = var_8984_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_8985_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_145_cast_fp16 = add(x = var_8964_cast_fp16, y = var_8985_cast_fp16)[name = tensor<string, []>("q_patches_145_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_8987_cast_fp16 = mul(x = var_8962_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_8987_cast_fp16")];
            tensor<int32, [5]> var_8995 = const()[name = tensor<string, []>("op_8995"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1001_cast_fp16 = reshape(shape = var_8995, x = var_8962_cast_fp16)[name = tensor<string, []>("x_1001_cast_fp16")];
            tensor<int32, [2]> var_8997_split_sizes_0 = const()[name = tensor<string, []>("op_8997_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_8997_axis_0 = const()[name = tensor<string, []>("op_8997_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_8997_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_8997_cast_fp16_1 = split(axis = var_8997_axis_0, split_sizes = var_8997_split_sizes_0, x = x_1001_cast_fp16)[name = tensor<string, []>("op_8997_cast_fp16")];
            tensor<int32, [1]> squeeze_290_axes_0 = const()[name = tensor<string, []>("squeeze_290_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_290_cast_fp16 = squeeze(axes = squeeze_290_axes_0, x = var_8997_cast_fp16_0)[name = tensor<string, []>("squeeze_290_cast_fp16")];
            tensor<int32, [1]> squeeze_291_axes_0 = const()[name = tensor<string, []>("squeeze_291_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_291_cast_fp16 = squeeze(axes = squeeze_291_axes_0, x = var_8997_cast_fp16_1)[name = tensor<string, []>("squeeze_291_cast_fp16")];
            tensor<fp16, []> const_1191_promoted_to_fp16 = const()[name = tensor<string, []>("const_1191_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9000_cast_fp16 = mul(x = squeeze_291_cast_fp16, y = const_1191_promoted_to_fp16)[name = tensor<string, []>("op_9000_cast_fp16")];
            tensor<int32, []> x_1003_axis_0 = const()[name = tensor<string, []>("x_1003_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1003_cast_fp16 = stack(axis = x_1003_axis_0, values = (var_9000_cast_fp16, squeeze_290_cast_fp16))[name = tensor<string, []>("x_1003_cast_fp16")];
            tensor<int32, [4]> var_9006 = const()[name = tensor<string, []>("op_9006"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9007_cast_fp16 = reshape(shape = var_9006, x = x_1003_cast_fp16)[name = tensor<string, []>("op_9007_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9008_cast_fp16 = mul(x = var_9007_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9008_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_145_cast_fp16 = add(x = var_8987_cast_fp16, y = var_9008_cast_fp16)[name = tensor<string, []>("k_patches_145_cast_fp16")];
            tensor<bool, []> var_9011_interleave_0 = const()[name = tensor<string, []>("op_9011_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9011_cast_fp16 = concat(axis = var_20, interleave = var_9011_interleave_0, values = (var_8950_cast_fp16, q_patches_145_cast_fp16))[name = tensor<string, []>("op_9011_cast_fp16")];
            tensor<bool, []> var_9014_interleave_0 = const()[name = tensor<string, []>("op_9014_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9014_cast_fp16 = concat(axis = var_20, interleave = var_9014_interleave_0, values = (var_8954_cast_fp16, k_patches_145_cast_fp16))[name = tensor<string, []>("op_9014_cast_fp16")];
            tensor<fp16, []> var_9016_to_fp16 = const()[name = tensor<string, []>("op_9016_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_729_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9016_to_fp16)[name = tensor<string, []>("q_729_cast_fp16")];
            tensor<bool, []> attn_289_transpose_x_1 = const()[name = tensor<string, []>("attn_289_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_289_transpose_y_1 = const()[name = tensor<string, []>("attn_289_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_289_cast_fp16 = matmul(transpose_x = attn_289_transpose_x_1, transpose_y = attn_289_transpose_y_1, x = q_729_cast_fp16, y = var_9014_cast_fp16)[name = tensor<string, []>("attn_289_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_291_cast_fp16 = softmax(axis = var_21, x = attn_289_cast_fp16)[name = tensor<string, []>("attn_291_cast_fp16")];
            tensor<bool, []> var_9021_transpose_x_0 = const()[name = tensor<string, []>("op_9021_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9021_transpose_y_0 = const()[name = tensor<string, []>("op_9021_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9021_cast_fp16 = matmul(transpose_x = var_9021_transpose_x_0, transpose_y = var_9021_transpose_y_0, x = attn_291_cast_fp16, y = v_437_cast_fp16)[name = tensor<string, []>("op_9021_cast_fp16")];
            tensor<int32, [1]> var_9022_axes_0 = const()[name = tensor<string, []>("op_9022_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9022_cast_fp16 = squeeze(axes = var_9022_axes_0, x = var_9021_cast_fp16)[name = tensor<string, []>("op_9022_cast_fp16")];
            tensor<string, []> var_9028_pad_type_0 = const()[name = tensor<string, []>("op_9028_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9028_strides_0 = const()[name = tensor<string, []>("op_9028_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9028_pad_0 = const()[name = tensor<string, []>("op_9028_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9028_dilations_0 = const()[name = tensor<string, []>("op_9028_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9028_groups_0 = const()[name = tensor<string, []>("op_9028_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86938624)))];
            tensor<fp16, [1, 64, 1, 197]> var_9028_cast_fp16 = conv(dilations = var_9028_dilations_0, groups = var_9028_groups_0, pad = var_9028_pad_0, pad_type = var_9028_pad_type_0, strides = var_9028_strides_0, weight = model_blocks_6_attn_q_projs_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9028_cast_fp16")];
            tensor<int32, [1]> var_9029_axes_0 = const()[name = tensor<string, []>("op_9029_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9029_cast_fp16 = squeeze(axes = var_9029_axes_0, x = var_9028_cast_fp16)[name = tensor<string, []>("op_9029_cast_fp16")];
            tensor<int32, [3]> q_731_perm_0 = const()[name = tensor<string, []>("q_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9036_pad_type_0 = const()[name = tensor<string, []>("op_9036_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9036_strides_0 = const()[name = tensor<string, []>("op_9036_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9036_pad_0 = const()[name = tensor<string, []>("op_9036_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9036_dilations_0 = const()[name = tensor<string, []>("op_9036_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9036_groups_0 = const()[name = tensor<string, []>("op_9036_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87036992)))];
            tensor<fp16, [1, 64, 1, 197]> var_9036_cast_fp16 = conv(dilations = var_9036_dilations_0, groups = var_9036_groups_0, pad = var_9036_pad_0, pad_type = var_9036_pad_type_0, strides = var_9036_strides_0, weight = model_blocks_6_attn_k_projs_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9036_cast_fp16")];
            tensor<int32, [1]> var_9037_axes_0 = const()[name = tensor<string, []>("op_9037_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9037_cast_fp16 = squeeze(axes = var_9037_axes_0, x = var_9036_cast_fp16)[name = tensor<string, []>("op_9037_cast_fp16")];
            tensor<int32, [3]> k_439_perm_0 = const()[name = tensor<string, []>("k_439_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9044_pad_type_0 = const()[name = tensor<string, []>("op_9044_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9044_strides_0 = const()[name = tensor<string, []>("op_9044_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9044_pad_0 = const()[name = tensor<string, []>("op_9044_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9044_dilations_0 = const()[name = tensor<string, []>("op_9044_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9044_groups_0 = const()[name = tensor<string, []>("op_9044_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87135360)))];
            tensor<fp16, [1, 64, 1, 197]> var_9044_cast_fp16 = conv(dilations = var_9044_dilations_0, groups = var_9044_groups_0, pad = var_9044_pad_0, pad_type = var_9044_pad_type_0, strides = var_9044_strides_0, weight = model_blocks_6_attn_v_projs_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9044_cast_fp16")];
            tensor<int32, [1]> var_9045_axes_0 = const()[name = tensor<string, []>("op_9045_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9045_cast_fp16 = squeeze(axes = var_9045_axes_0, x = var_9044_cast_fp16)[name = tensor<string, []>("op_9045_cast_fp16")];
            tensor<int32, [3]> v_439_perm_0 = const()[name = tensor<string, []>("v_439_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87233728)))];
            tensor<fp16, [1, 197, 64]> q_731_cast_fp16 = transpose(perm = q_731_perm_0, x = var_9029_cast_fp16)[name = tensor<string, []>("transpose_253")];
            tensor<fp16, [1, 197, 64]> q_733_cast_fp16 = add(x = q_731_cast_fp16, y = model_blocks_6_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_733_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87233920)))];
            tensor<fp16, [1, 197, 64]> v_439_cast_fp16 = transpose(perm = v_439_perm_0, x = var_9045_cast_fp16)[name = tensor<string, []>("transpose_251")];
            tensor<fp16, [1, 197, 64]> v_441_cast_fp16 = add(x = v_439_cast_fp16, y = model_blocks_6_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_441_cast_fp16")];
            tensor<int32, [1]> q_735_axes_0 = const()[name = tensor<string, []>("q_735_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_735_cast_fp16 = expand_dims(axes = q_735_axes_0, x = q_733_cast_fp16)[name = tensor<string, []>("q_735_cast_fp16")];
            tensor<int32, [1]> k_441_axes_0 = const()[name = tensor<string, []>("k_441_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_439_cast_fp16 = transpose(perm = k_439_perm_0, x = var_9037_cast_fp16)[name = tensor<string, []>("transpose_252")];
            tensor<fp16, [1, 1, 197, 64]> k_441_cast_fp16 = expand_dims(axes = k_441_axes_0, x = k_439_cast_fp16)[name = tensor<string, []>("k_441_cast_fp16")];
            tensor<int32, [1]> v_443_axes_0 = const()[name = tensor<string, []>("v_443_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_443_cast_fp16 = expand_dims(axes = v_443_axes_0, x = v_441_cast_fp16)[name = tensor<string, []>("v_443_cast_fp16")];
            tensor<int32, [4]> var_9054_begin_0 = const()[name = tensor<string, []>("op_9054_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9054_end_0 = const()[name = tensor<string, []>("op_9054_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9054_end_mask_0 = const()[name = tensor<string, []>("op_9054_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9054_cast_fp16 = slice_by_index(begin = var_9054_begin_0, end = var_9054_end_0, end_mask = var_9054_end_mask_0, x = q_735_cast_fp16)[name = tensor<string, []>("op_9054_cast_fp16")];
            tensor<int32, [4]> var_9058_begin_0 = const()[name = tensor<string, []>("op_9058_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9058_end_0 = const()[name = tensor<string, []>("op_9058_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9058_end_mask_0 = const()[name = tensor<string, []>("op_9058_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9058_cast_fp16 = slice_by_index(begin = var_9058_begin_0, end = var_9058_end_0, end_mask = var_9058_end_mask_0, x = k_441_cast_fp16)[name = tensor<string, []>("op_9058_cast_fp16")];
            tensor<int32, [4]> var_9062_begin_0 = const()[name = tensor<string, []>("op_9062_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9062_end_0 = const()[name = tensor<string, []>("op_9062_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9062_end_mask_0 = const()[name = tensor<string, []>("op_9062_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9062_cast_fp16 = slice_by_index(begin = var_9062_begin_0, end = var_9062_end_0, end_mask = var_9062_end_mask_0, x = q_735_cast_fp16)[name = tensor<string, []>("op_9062_cast_fp16")];
            tensor<int32, [4]> var_9066_begin_0 = const()[name = tensor<string, []>("op_9066_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9066_end_0 = const()[name = tensor<string, []>("op_9066_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9066_end_mask_0 = const()[name = tensor<string, []>("op_9066_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9066_cast_fp16 = slice_by_index(begin = var_9066_begin_0, end = var_9066_end_0, end_mask = var_9066_end_mask_0, x = k_441_cast_fp16)[name = tensor<string, []>("op_9066_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9068_cast_fp16 = mul(x = var_9062_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9068_cast_fp16")];
            tensor<int32, [5]> var_9076 = const()[name = tensor<string, []>("op_9076"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1007_cast_fp16 = reshape(shape = var_9076, x = var_9062_cast_fp16)[name = tensor<string, []>("x_1007_cast_fp16")];
            tensor<int32, [2]> var_9078_split_sizes_0 = const()[name = tensor<string, []>("op_9078_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9078_axis_0 = const()[name = tensor<string, []>("op_9078_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9078_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9078_cast_fp16_1 = split(axis = var_9078_axis_0, split_sizes = var_9078_split_sizes_0, x = x_1007_cast_fp16)[name = tensor<string, []>("op_9078_cast_fp16")];
            tensor<int32, [1]> squeeze_292_axes_0 = const()[name = tensor<string, []>("squeeze_292_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_292_cast_fp16 = squeeze(axes = squeeze_292_axes_0, x = var_9078_cast_fp16_0)[name = tensor<string, []>("squeeze_292_cast_fp16")];
            tensor<int32, [1]> squeeze_293_axes_0 = const()[name = tensor<string, []>("squeeze_293_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_293_cast_fp16 = squeeze(axes = squeeze_293_axes_0, x = var_9078_cast_fp16_1)[name = tensor<string, []>("squeeze_293_cast_fp16")];
            tensor<fp16, []> const_1199_promoted_to_fp16 = const()[name = tensor<string, []>("const_1199_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9081_cast_fp16 = mul(x = squeeze_293_cast_fp16, y = const_1199_promoted_to_fp16)[name = tensor<string, []>("op_9081_cast_fp16")];
            tensor<int32, []> x_1009_axis_0 = const()[name = tensor<string, []>("x_1009_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1009_cast_fp16 = stack(axis = x_1009_axis_0, values = (var_9081_cast_fp16, squeeze_292_cast_fp16))[name = tensor<string, []>("x_1009_cast_fp16")];
            tensor<int32, [4]> var_9087 = const()[name = tensor<string, []>("op_9087"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9088_cast_fp16 = reshape(shape = var_9087, x = x_1009_cast_fp16)[name = tensor<string, []>("op_9088_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9089_cast_fp16 = mul(x = var_9088_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9089_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_147_cast_fp16 = add(x = var_9068_cast_fp16, y = var_9089_cast_fp16)[name = tensor<string, []>("q_patches_147_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9091_cast_fp16 = mul(x = var_9066_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9091_cast_fp16")];
            tensor<int32, [5]> var_9099 = const()[name = tensor<string, []>("op_9099"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1013_cast_fp16 = reshape(shape = var_9099, x = var_9066_cast_fp16)[name = tensor<string, []>("x_1013_cast_fp16")];
            tensor<int32, [2]> var_9101_split_sizes_0 = const()[name = tensor<string, []>("op_9101_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9101_axis_0 = const()[name = tensor<string, []>("op_9101_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9101_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9101_cast_fp16_1 = split(axis = var_9101_axis_0, split_sizes = var_9101_split_sizes_0, x = x_1013_cast_fp16)[name = tensor<string, []>("op_9101_cast_fp16")];
            tensor<int32, [1]> squeeze_294_axes_0 = const()[name = tensor<string, []>("squeeze_294_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_294_cast_fp16 = squeeze(axes = squeeze_294_axes_0, x = var_9101_cast_fp16_0)[name = tensor<string, []>("squeeze_294_cast_fp16")];
            tensor<int32, [1]> squeeze_295_axes_0 = const()[name = tensor<string, []>("squeeze_295_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_295_cast_fp16 = squeeze(axes = squeeze_295_axes_0, x = var_9101_cast_fp16_1)[name = tensor<string, []>("squeeze_295_cast_fp16")];
            tensor<fp16, []> const_1207_promoted_to_fp16 = const()[name = tensor<string, []>("const_1207_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9104_cast_fp16 = mul(x = squeeze_295_cast_fp16, y = const_1207_promoted_to_fp16)[name = tensor<string, []>("op_9104_cast_fp16")];
            tensor<int32, []> x_1015_axis_0 = const()[name = tensor<string, []>("x_1015_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1015_cast_fp16 = stack(axis = x_1015_axis_0, values = (var_9104_cast_fp16, squeeze_294_cast_fp16))[name = tensor<string, []>("x_1015_cast_fp16")];
            tensor<int32, [4]> var_9110 = const()[name = tensor<string, []>("op_9110"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9111_cast_fp16 = reshape(shape = var_9110, x = x_1015_cast_fp16)[name = tensor<string, []>("op_9111_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9112_cast_fp16 = mul(x = var_9111_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9112_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_147_cast_fp16 = add(x = var_9091_cast_fp16, y = var_9112_cast_fp16)[name = tensor<string, []>("k_patches_147_cast_fp16")];
            tensor<bool, []> var_9115_interleave_0 = const()[name = tensor<string, []>("op_9115_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9115_cast_fp16 = concat(axis = var_20, interleave = var_9115_interleave_0, values = (var_9054_cast_fp16, q_patches_147_cast_fp16))[name = tensor<string, []>("op_9115_cast_fp16")];
            tensor<bool, []> var_9118_interleave_0 = const()[name = tensor<string, []>("op_9118_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9118_cast_fp16 = concat(axis = var_20, interleave = var_9118_interleave_0, values = (var_9058_cast_fp16, k_patches_147_cast_fp16))[name = tensor<string, []>("op_9118_cast_fp16")];
            tensor<fp16, []> var_9120_to_fp16 = const()[name = tensor<string, []>("op_9120_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_739_cast_fp16 = mul(x = var_9115_cast_fp16, y = var_9120_to_fp16)[name = tensor<string, []>("q_739_cast_fp16")];
            tensor<bool, []> attn_293_transpose_x_1 = const()[name = tensor<string, []>("attn_293_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_293_transpose_y_1 = const()[name = tensor<string, []>("attn_293_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_293_cast_fp16 = matmul(transpose_x = attn_293_transpose_x_1, transpose_y = attn_293_transpose_y_1, x = q_739_cast_fp16, y = var_9118_cast_fp16)[name = tensor<string, []>("attn_293_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_295_cast_fp16 = softmax(axis = var_21, x = attn_293_cast_fp16)[name = tensor<string, []>("attn_295_cast_fp16")];
            tensor<bool, []> var_9125_transpose_x_0 = const()[name = tensor<string, []>("op_9125_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9125_transpose_y_0 = const()[name = tensor<string, []>("op_9125_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9125_cast_fp16 = matmul(transpose_x = var_9125_transpose_x_0, transpose_y = var_9125_transpose_y_0, x = attn_295_cast_fp16, y = v_443_cast_fp16)[name = tensor<string, []>("op_9125_cast_fp16")];
            tensor<int32, [1]> var_9126_axes_0 = const()[name = tensor<string, []>("op_9126_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9126_cast_fp16 = squeeze(axes = var_9126_axes_0, x = var_9125_cast_fp16)[name = tensor<string, []>("op_9126_cast_fp16")];
            tensor<string, []> var_9132_pad_type_0 = const()[name = tensor<string, []>("op_9132_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9132_strides_0 = const()[name = tensor<string, []>("op_9132_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9132_pad_0 = const()[name = tensor<string, []>("op_9132_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9132_dilations_0 = const()[name = tensor<string, []>("op_9132_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9132_groups_0 = const()[name = tensor<string, []>("op_9132_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87234112)))];
            tensor<fp16, [1, 64, 1, 197]> var_9132_cast_fp16 = conv(dilations = var_9132_dilations_0, groups = var_9132_groups_0, pad = var_9132_pad_0, pad_type = var_9132_pad_type_0, strides = var_9132_strides_0, weight = model_blocks_6_attn_q_projs_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9132_cast_fp16")];
            tensor<int32, [1]> var_9133_axes_0 = const()[name = tensor<string, []>("op_9133_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9133_cast_fp16 = squeeze(axes = var_9133_axes_0, x = var_9132_cast_fp16)[name = tensor<string, []>("op_9133_cast_fp16")];
            tensor<int32, [3]> q_741_perm_0 = const()[name = tensor<string, []>("q_741_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9140_pad_type_0 = const()[name = tensor<string, []>("op_9140_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9140_strides_0 = const()[name = tensor<string, []>("op_9140_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9140_pad_0 = const()[name = tensor<string, []>("op_9140_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9140_dilations_0 = const()[name = tensor<string, []>("op_9140_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9140_groups_0 = const()[name = tensor<string, []>("op_9140_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87332480)))];
            tensor<fp16, [1, 64, 1, 197]> var_9140_cast_fp16 = conv(dilations = var_9140_dilations_0, groups = var_9140_groups_0, pad = var_9140_pad_0, pad_type = var_9140_pad_type_0, strides = var_9140_strides_0, weight = model_blocks_6_attn_k_projs_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9140_cast_fp16")];
            tensor<int32, [1]> var_9141_axes_0 = const()[name = tensor<string, []>("op_9141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9141_cast_fp16 = squeeze(axes = var_9141_axes_0, x = var_9140_cast_fp16)[name = tensor<string, []>("op_9141_cast_fp16")];
            tensor<int32, [3]> k_445_perm_0 = const()[name = tensor<string, []>("k_445_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9148_pad_type_0 = const()[name = tensor<string, []>("op_9148_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9148_strides_0 = const()[name = tensor<string, []>("op_9148_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9148_pad_0 = const()[name = tensor<string, []>("op_9148_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9148_dilations_0 = const()[name = tensor<string, []>("op_9148_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9148_groups_0 = const()[name = tensor<string, []>("op_9148_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87430848)))];
            tensor<fp16, [1, 64, 1, 197]> var_9148_cast_fp16 = conv(dilations = var_9148_dilations_0, groups = var_9148_groups_0, pad = var_9148_pad_0, pad_type = var_9148_pad_type_0, strides = var_9148_strides_0, weight = model_blocks_6_attn_v_projs_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9148_cast_fp16")];
            tensor<int32, [1]> var_9149_axes_0 = const()[name = tensor<string, []>("op_9149_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9149_cast_fp16 = squeeze(axes = var_9149_axes_0, x = var_9148_cast_fp16)[name = tensor<string, []>("op_9149_cast_fp16")];
            tensor<int32, [3]> v_445_perm_0 = const()[name = tensor<string, []>("v_445_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87529216)))];
            tensor<fp16, [1, 197, 64]> q_741_cast_fp16 = transpose(perm = q_741_perm_0, x = var_9133_cast_fp16)[name = tensor<string, []>("transpose_250")];
            tensor<fp16, [1, 197, 64]> q_743_cast_fp16 = add(x = q_741_cast_fp16, y = model_blocks_6_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_743_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87529408)))];
            tensor<fp16, [1, 197, 64]> v_445_cast_fp16 = transpose(perm = v_445_perm_0, x = var_9149_cast_fp16)[name = tensor<string, []>("transpose_248")];
            tensor<fp16, [1, 197, 64]> v_447_cast_fp16 = add(x = v_445_cast_fp16, y = model_blocks_6_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_447_cast_fp16")];
            tensor<int32, [1]> q_745_axes_0 = const()[name = tensor<string, []>("q_745_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_745_cast_fp16 = expand_dims(axes = q_745_axes_0, x = q_743_cast_fp16)[name = tensor<string, []>("q_745_cast_fp16")];
            tensor<int32, [1]> k_447_axes_0 = const()[name = tensor<string, []>("k_447_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_445_cast_fp16 = transpose(perm = k_445_perm_0, x = var_9141_cast_fp16)[name = tensor<string, []>("transpose_249")];
            tensor<fp16, [1, 1, 197, 64]> k_447_cast_fp16 = expand_dims(axes = k_447_axes_0, x = k_445_cast_fp16)[name = tensor<string, []>("k_447_cast_fp16")];
            tensor<int32, [1]> v_449_axes_0 = const()[name = tensor<string, []>("v_449_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_449_cast_fp16 = expand_dims(axes = v_449_axes_0, x = v_447_cast_fp16)[name = tensor<string, []>("v_449_cast_fp16")];
            tensor<int32, [4]> var_9158_begin_0 = const()[name = tensor<string, []>("op_9158_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9158_end_0 = const()[name = tensor<string, []>("op_9158_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9158_end_mask_0 = const()[name = tensor<string, []>("op_9158_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9158_cast_fp16 = slice_by_index(begin = var_9158_begin_0, end = var_9158_end_0, end_mask = var_9158_end_mask_0, x = q_745_cast_fp16)[name = tensor<string, []>("op_9158_cast_fp16")];
            tensor<int32, [4]> var_9162_begin_0 = const()[name = tensor<string, []>("op_9162_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9162_end_0 = const()[name = tensor<string, []>("op_9162_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9162_end_mask_0 = const()[name = tensor<string, []>("op_9162_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9162_cast_fp16 = slice_by_index(begin = var_9162_begin_0, end = var_9162_end_0, end_mask = var_9162_end_mask_0, x = k_447_cast_fp16)[name = tensor<string, []>("op_9162_cast_fp16")];
            tensor<int32, [4]> var_9166_begin_0 = const()[name = tensor<string, []>("op_9166_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9166_end_0 = const()[name = tensor<string, []>("op_9166_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9166_end_mask_0 = const()[name = tensor<string, []>("op_9166_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9166_cast_fp16 = slice_by_index(begin = var_9166_begin_0, end = var_9166_end_0, end_mask = var_9166_end_mask_0, x = q_745_cast_fp16)[name = tensor<string, []>("op_9166_cast_fp16")];
            tensor<int32, [4]> var_9170_begin_0 = const()[name = tensor<string, []>("op_9170_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9170_end_0 = const()[name = tensor<string, []>("op_9170_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9170_end_mask_0 = const()[name = tensor<string, []>("op_9170_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9170_cast_fp16 = slice_by_index(begin = var_9170_begin_0, end = var_9170_end_0, end_mask = var_9170_end_mask_0, x = k_447_cast_fp16)[name = tensor<string, []>("op_9170_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9172_cast_fp16 = mul(x = var_9166_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9172_cast_fp16")];
            tensor<int32, [5]> var_9180 = const()[name = tensor<string, []>("op_9180"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1019_cast_fp16 = reshape(shape = var_9180, x = var_9166_cast_fp16)[name = tensor<string, []>("x_1019_cast_fp16")];
            tensor<int32, [2]> var_9182_split_sizes_0 = const()[name = tensor<string, []>("op_9182_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9182_axis_0 = const()[name = tensor<string, []>("op_9182_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9182_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9182_cast_fp16_1 = split(axis = var_9182_axis_0, split_sizes = var_9182_split_sizes_0, x = x_1019_cast_fp16)[name = tensor<string, []>("op_9182_cast_fp16")];
            tensor<int32, [1]> squeeze_296_axes_0 = const()[name = tensor<string, []>("squeeze_296_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_296_cast_fp16 = squeeze(axes = squeeze_296_axes_0, x = var_9182_cast_fp16_0)[name = tensor<string, []>("squeeze_296_cast_fp16")];
            tensor<int32, [1]> squeeze_297_axes_0 = const()[name = tensor<string, []>("squeeze_297_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_297_cast_fp16 = squeeze(axes = squeeze_297_axes_0, x = var_9182_cast_fp16_1)[name = tensor<string, []>("squeeze_297_cast_fp16")];
            tensor<fp16, []> const_1215_promoted_to_fp16 = const()[name = tensor<string, []>("const_1215_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9185_cast_fp16 = mul(x = squeeze_297_cast_fp16, y = const_1215_promoted_to_fp16)[name = tensor<string, []>("op_9185_cast_fp16")];
            tensor<int32, []> x_1021_axis_0 = const()[name = tensor<string, []>("x_1021_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1021_cast_fp16 = stack(axis = x_1021_axis_0, values = (var_9185_cast_fp16, squeeze_296_cast_fp16))[name = tensor<string, []>("x_1021_cast_fp16")];
            tensor<int32, [4]> var_9191 = const()[name = tensor<string, []>("op_9191"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9192_cast_fp16 = reshape(shape = var_9191, x = x_1021_cast_fp16)[name = tensor<string, []>("op_9192_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9193_cast_fp16 = mul(x = var_9192_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9193_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_149_cast_fp16 = add(x = var_9172_cast_fp16, y = var_9193_cast_fp16)[name = tensor<string, []>("q_patches_149_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9195_cast_fp16 = mul(x = var_9170_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9195_cast_fp16")];
            tensor<int32, [5]> var_9203 = const()[name = tensor<string, []>("op_9203"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1025_cast_fp16 = reshape(shape = var_9203, x = var_9170_cast_fp16)[name = tensor<string, []>("x_1025_cast_fp16")];
            tensor<int32, [2]> var_9205_split_sizes_0 = const()[name = tensor<string, []>("op_9205_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9205_axis_0 = const()[name = tensor<string, []>("op_9205_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9205_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9205_cast_fp16_1 = split(axis = var_9205_axis_0, split_sizes = var_9205_split_sizes_0, x = x_1025_cast_fp16)[name = tensor<string, []>("op_9205_cast_fp16")];
            tensor<int32, [1]> squeeze_298_axes_0 = const()[name = tensor<string, []>("squeeze_298_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_298_cast_fp16 = squeeze(axes = squeeze_298_axes_0, x = var_9205_cast_fp16_0)[name = tensor<string, []>("squeeze_298_cast_fp16")];
            tensor<int32, [1]> squeeze_299_axes_0 = const()[name = tensor<string, []>("squeeze_299_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_299_cast_fp16 = squeeze(axes = squeeze_299_axes_0, x = var_9205_cast_fp16_1)[name = tensor<string, []>("squeeze_299_cast_fp16")];
            tensor<fp16, []> const_1223_promoted_to_fp16 = const()[name = tensor<string, []>("const_1223_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9208_cast_fp16 = mul(x = squeeze_299_cast_fp16, y = const_1223_promoted_to_fp16)[name = tensor<string, []>("op_9208_cast_fp16")];
            tensor<int32, []> x_1027_axis_0 = const()[name = tensor<string, []>("x_1027_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1027_cast_fp16 = stack(axis = x_1027_axis_0, values = (var_9208_cast_fp16, squeeze_298_cast_fp16))[name = tensor<string, []>("x_1027_cast_fp16")];
            tensor<int32, [4]> var_9214 = const()[name = tensor<string, []>("op_9214"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9215_cast_fp16 = reshape(shape = var_9214, x = x_1027_cast_fp16)[name = tensor<string, []>("op_9215_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9216_cast_fp16 = mul(x = var_9215_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9216_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_149_cast_fp16 = add(x = var_9195_cast_fp16, y = var_9216_cast_fp16)[name = tensor<string, []>("k_patches_149_cast_fp16")];
            tensor<bool, []> var_9219_interleave_0 = const()[name = tensor<string, []>("op_9219_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9219_cast_fp16 = concat(axis = var_20, interleave = var_9219_interleave_0, values = (var_9158_cast_fp16, q_patches_149_cast_fp16))[name = tensor<string, []>("op_9219_cast_fp16")];
            tensor<bool, []> var_9222_interleave_0 = const()[name = tensor<string, []>("op_9222_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9222_cast_fp16 = concat(axis = var_20, interleave = var_9222_interleave_0, values = (var_9162_cast_fp16, k_patches_149_cast_fp16))[name = tensor<string, []>("op_9222_cast_fp16")];
            tensor<fp16, []> var_9224_to_fp16 = const()[name = tensor<string, []>("op_9224_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_749_cast_fp16 = mul(x = var_9219_cast_fp16, y = var_9224_to_fp16)[name = tensor<string, []>("q_749_cast_fp16")];
            tensor<bool, []> attn_297_transpose_x_1 = const()[name = tensor<string, []>("attn_297_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_297_transpose_y_1 = const()[name = tensor<string, []>("attn_297_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_297_cast_fp16 = matmul(transpose_x = attn_297_transpose_x_1, transpose_y = attn_297_transpose_y_1, x = q_749_cast_fp16, y = var_9222_cast_fp16)[name = tensor<string, []>("attn_297_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_299_cast_fp16 = softmax(axis = var_21, x = attn_297_cast_fp16)[name = tensor<string, []>("attn_299_cast_fp16")];
            tensor<bool, []> var_9229_transpose_x_0 = const()[name = tensor<string, []>("op_9229_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9229_transpose_y_0 = const()[name = tensor<string, []>("op_9229_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9229_cast_fp16 = matmul(transpose_x = var_9229_transpose_x_0, transpose_y = var_9229_transpose_y_0, x = attn_299_cast_fp16, y = v_449_cast_fp16)[name = tensor<string, []>("op_9229_cast_fp16")];
            tensor<int32, [1]> var_9230_axes_0 = const()[name = tensor<string, []>("op_9230_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9230_cast_fp16 = squeeze(axes = var_9230_axes_0, x = var_9229_cast_fp16)[name = tensor<string, []>("op_9230_cast_fp16")];
            tensor<string, []> var_9236_pad_type_0 = const()[name = tensor<string, []>("op_9236_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9236_strides_0 = const()[name = tensor<string, []>("op_9236_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9236_pad_0 = const()[name = tensor<string, []>("op_9236_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9236_dilations_0 = const()[name = tensor<string, []>("op_9236_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9236_groups_0 = const()[name = tensor<string, []>("op_9236_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87529600)))];
            tensor<fp16, [1, 64, 1, 197]> var_9236_cast_fp16 = conv(dilations = var_9236_dilations_0, groups = var_9236_groups_0, pad = var_9236_pad_0, pad_type = var_9236_pad_type_0, strides = var_9236_strides_0, weight = model_blocks_6_attn_q_projs_3_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9236_cast_fp16")];
            tensor<int32, [1]> var_9237_axes_0 = const()[name = tensor<string, []>("op_9237_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9237_cast_fp16 = squeeze(axes = var_9237_axes_0, x = var_9236_cast_fp16)[name = tensor<string, []>("op_9237_cast_fp16")];
            tensor<int32, [3]> q_751_perm_0 = const()[name = tensor<string, []>("q_751_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9244_pad_type_0 = const()[name = tensor<string, []>("op_9244_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9244_strides_0 = const()[name = tensor<string, []>("op_9244_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9244_pad_0 = const()[name = tensor<string, []>("op_9244_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9244_dilations_0 = const()[name = tensor<string, []>("op_9244_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9244_groups_0 = const()[name = tensor<string, []>("op_9244_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87627968)))];
            tensor<fp16, [1, 64, 1, 197]> var_9244_cast_fp16 = conv(dilations = var_9244_dilations_0, groups = var_9244_groups_0, pad = var_9244_pad_0, pad_type = var_9244_pad_type_0, strides = var_9244_strides_0, weight = model_blocks_6_attn_k_projs_3_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9244_cast_fp16")];
            tensor<int32, [1]> var_9245_axes_0 = const()[name = tensor<string, []>("op_9245_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9245_cast_fp16 = squeeze(axes = var_9245_axes_0, x = var_9244_cast_fp16)[name = tensor<string, []>("op_9245_cast_fp16")];
            tensor<int32, [3]> k_451_perm_0 = const()[name = tensor<string, []>("k_451_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9252_pad_type_0 = const()[name = tensor<string, []>("op_9252_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9252_strides_0 = const()[name = tensor<string, []>("op_9252_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9252_pad_0 = const()[name = tensor<string, []>("op_9252_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9252_dilations_0 = const()[name = tensor<string, []>("op_9252_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9252_groups_0 = const()[name = tensor<string, []>("op_9252_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87726336)))];
            tensor<fp16, [1, 64, 1, 197]> var_9252_cast_fp16 = conv(dilations = var_9252_dilations_0, groups = var_9252_groups_0, pad = var_9252_pad_0, pad_type = var_9252_pad_type_0, strides = var_9252_strides_0, weight = model_blocks_6_attn_v_projs_3_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9252_cast_fp16")];
            tensor<int32, [1]> var_9253_axes_0 = const()[name = tensor<string, []>("op_9253_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9253_cast_fp16 = squeeze(axes = var_9253_axes_0, x = var_9252_cast_fp16)[name = tensor<string, []>("op_9253_cast_fp16")];
            tensor<int32, [3]> v_451_perm_0 = const()[name = tensor<string, []>("v_451_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87824704)))];
            tensor<fp16, [1, 197, 64]> q_751_cast_fp16 = transpose(perm = q_751_perm_0, x = var_9237_cast_fp16)[name = tensor<string, []>("transpose_247")];
            tensor<fp16, [1, 197, 64]> q_753_cast_fp16 = add(x = q_751_cast_fp16, y = model_blocks_6_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_753_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87824896)))];
            tensor<fp16, [1, 197, 64]> v_451_cast_fp16 = transpose(perm = v_451_perm_0, x = var_9253_cast_fp16)[name = tensor<string, []>("transpose_245")];
            tensor<fp16, [1, 197, 64]> v_453_cast_fp16 = add(x = v_451_cast_fp16, y = model_blocks_6_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_453_cast_fp16")];
            tensor<int32, [1]> q_755_axes_0 = const()[name = tensor<string, []>("q_755_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_755_cast_fp16 = expand_dims(axes = q_755_axes_0, x = q_753_cast_fp16)[name = tensor<string, []>("q_755_cast_fp16")];
            tensor<int32, [1]> k_453_axes_0 = const()[name = tensor<string, []>("k_453_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_451_cast_fp16 = transpose(perm = k_451_perm_0, x = var_9245_cast_fp16)[name = tensor<string, []>("transpose_246")];
            tensor<fp16, [1, 1, 197, 64]> k_453_cast_fp16 = expand_dims(axes = k_453_axes_0, x = k_451_cast_fp16)[name = tensor<string, []>("k_453_cast_fp16")];
            tensor<int32, [1]> v_455_axes_0 = const()[name = tensor<string, []>("v_455_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_455_cast_fp16 = expand_dims(axes = v_455_axes_0, x = v_453_cast_fp16)[name = tensor<string, []>("v_455_cast_fp16")];
            tensor<int32, [4]> var_9262_begin_0 = const()[name = tensor<string, []>("op_9262_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9262_end_0 = const()[name = tensor<string, []>("op_9262_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9262_end_mask_0 = const()[name = tensor<string, []>("op_9262_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9262_cast_fp16 = slice_by_index(begin = var_9262_begin_0, end = var_9262_end_0, end_mask = var_9262_end_mask_0, x = q_755_cast_fp16)[name = tensor<string, []>("op_9262_cast_fp16")];
            tensor<int32, [4]> var_9266_begin_0 = const()[name = tensor<string, []>("op_9266_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9266_end_0 = const()[name = tensor<string, []>("op_9266_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9266_end_mask_0 = const()[name = tensor<string, []>("op_9266_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9266_cast_fp16 = slice_by_index(begin = var_9266_begin_0, end = var_9266_end_0, end_mask = var_9266_end_mask_0, x = k_453_cast_fp16)[name = tensor<string, []>("op_9266_cast_fp16")];
            tensor<int32, [4]> var_9270_begin_0 = const()[name = tensor<string, []>("op_9270_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9270_end_0 = const()[name = tensor<string, []>("op_9270_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9270_end_mask_0 = const()[name = tensor<string, []>("op_9270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9270_cast_fp16 = slice_by_index(begin = var_9270_begin_0, end = var_9270_end_0, end_mask = var_9270_end_mask_0, x = q_755_cast_fp16)[name = tensor<string, []>("op_9270_cast_fp16")];
            tensor<int32, [4]> var_9274_begin_0 = const()[name = tensor<string, []>("op_9274_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9274_end_0 = const()[name = tensor<string, []>("op_9274_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9274_end_mask_0 = const()[name = tensor<string, []>("op_9274_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9274_cast_fp16 = slice_by_index(begin = var_9274_begin_0, end = var_9274_end_0, end_mask = var_9274_end_mask_0, x = k_453_cast_fp16)[name = tensor<string, []>("op_9274_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9276_cast_fp16 = mul(x = var_9270_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9276_cast_fp16")];
            tensor<int32, [5]> var_9284 = const()[name = tensor<string, []>("op_9284"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1031_cast_fp16 = reshape(shape = var_9284, x = var_9270_cast_fp16)[name = tensor<string, []>("x_1031_cast_fp16")];
            tensor<int32, [2]> var_9286_split_sizes_0 = const()[name = tensor<string, []>("op_9286_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9286_axis_0 = const()[name = tensor<string, []>("op_9286_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9286_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9286_cast_fp16_1 = split(axis = var_9286_axis_0, split_sizes = var_9286_split_sizes_0, x = x_1031_cast_fp16)[name = tensor<string, []>("op_9286_cast_fp16")];
            tensor<int32, [1]> squeeze_300_axes_0 = const()[name = tensor<string, []>("squeeze_300_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_300_cast_fp16 = squeeze(axes = squeeze_300_axes_0, x = var_9286_cast_fp16_0)[name = tensor<string, []>("squeeze_300_cast_fp16")];
            tensor<int32, [1]> squeeze_301_axes_0 = const()[name = tensor<string, []>("squeeze_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_301_cast_fp16 = squeeze(axes = squeeze_301_axes_0, x = var_9286_cast_fp16_1)[name = tensor<string, []>("squeeze_301_cast_fp16")];
            tensor<fp16, []> const_1231_promoted_to_fp16 = const()[name = tensor<string, []>("const_1231_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9289_cast_fp16 = mul(x = squeeze_301_cast_fp16, y = const_1231_promoted_to_fp16)[name = tensor<string, []>("op_9289_cast_fp16")];
            tensor<int32, []> x_1033_axis_0 = const()[name = tensor<string, []>("x_1033_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1033_cast_fp16 = stack(axis = x_1033_axis_0, values = (var_9289_cast_fp16, squeeze_300_cast_fp16))[name = tensor<string, []>("x_1033_cast_fp16")];
            tensor<int32, [4]> var_9295 = const()[name = tensor<string, []>("op_9295"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9296_cast_fp16 = reshape(shape = var_9295, x = x_1033_cast_fp16)[name = tensor<string, []>("op_9296_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9297_cast_fp16 = mul(x = var_9296_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9297_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_151_cast_fp16 = add(x = var_9276_cast_fp16, y = var_9297_cast_fp16)[name = tensor<string, []>("q_patches_151_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9299_cast_fp16 = mul(x = var_9274_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9299_cast_fp16")];
            tensor<int32, [5]> var_9307 = const()[name = tensor<string, []>("op_9307"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1037_cast_fp16 = reshape(shape = var_9307, x = var_9274_cast_fp16)[name = tensor<string, []>("x_1037_cast_fp16")];
            tensor<int32, [2]> var_9309_split_sizes_0 = const()[name = tensor<string, []>("op_9309_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9309_axis_0 = const()[name = tensor<string, []>("op_9309_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9309_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9309_cast_fp16_1 = split(axis = var_9309_axis_0, split_sizes = var_9309_split_sizes_0, x = x_1037_cast_fp16)[name = tensor<string, []>("op_9309_cast_fp16")];
            tensor<int32, [1]> squeeze_302_axes_0 = const()[name = tensor<string, []>("squeeze_302_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_302_cast_fp16 = squeeze(axes = squeeze_302_axes_0, x = var_9309_cast_fp16_0)[name = tensor<string, []>("squeeze_302_cast_fp16")];
            tensor<int32, [1]> squeeze_303_axes_0 = const()[name = tensor<string, []>("squeeze_303_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_303_cast_fp16 = squeeze(axes = squeeze_303_axes_0, x = var_9309_cast_fp16_1)[name = tensor<string, []>("squeeze_303_cast_fp16")];
            tensor<fp16, []> const_1239_promoted_to_fp16 = const()[name = tensor<string, []>("const_1239_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9312_cast_fp16 = mul(x = squeeze_303_cast_fp16, y = const_1239_promoted_to_fp16)[name = tensor<string, []>("op_9312_cast_fp16")];
            tensor<int32, []> x_1039_axis_0 = const()[name = tensor<string, []>("x_1039_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1039_cast_fp16 = stack(axis = x_1039_axis_0, values = (var_9312_cast_fp16, squeeze_302_cast_fp16))[name = tensor<string, []>("x_1039_cast_fp16")];
            tensor<int32, [4]> var_9318 = const()[name = tensor<string, []>("op_9318"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9319_cast_fp16 = reshape(shape = var_9318, x = x_1039_cast_fp16)[name = tensor<string, []>("op_9319_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9320_cast_fp16 = mul(x = var_9319_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9320_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_151_cast_fp16 = add(x = var_9299_cast_fp16, y = var_9320_cast_fp16)[name = tensor<string, []>("k_patches_151_cast_fp16")];
            tensor<bool, []> var_9323_interleave_0 = const()[name = tensor<string, []>("op_9323_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9323_cast_fp16 = concat(axis = var_20, interleave = var_9323_interleave_0, values = (var_9262_cast_fp16, q_patches_151_cast_fp16))[name = tensor<string, []>("op_9323_cast_fp16")];
            tensor<bool, []> var_9326_interleave_0 = const()[name = tensor<string, []>("op_9326_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9326_cast_fp16 = concat(axis = var_20, interleave = var_9326_interleave_0, values = (var_9266_cast_fp16, k_patches_151_cast_fp16))[name = tensor<string, []>("op_9326_cast_fp16")];
            tensor<fp16, []> var_9328_to_fp16 = const()[name = tensor<string, []>("op_9328_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_759_cast_fp16 = mul(x = var_9323_cast_fp16, y = var_9328_to_fp16)[name = tensor<string, []>("q_759_cast_fp16")];
            tensor<bool, []> attn_301_transpose_x_1 = const()[name = tensor<string, []>("attn_301_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_301_transpose_y_1 = const()[name = tensor<string, []>("attn_301_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_301_cast_fp16 = matmul(transpose_x = attn_301_transpose_x_1, transpose_y = attn_301_transpose_y_1, x = q_759_cast_fp16, y = var_9326_cast_fp16)[name = tensor<string, []>("attn_301_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_303_cast_fp16 = softmax(axis = var_21, x = attn_301_cast_fp16)[name = tensor<string, []>("attn_303_cast_fp16")];
            tensor<bool, []> var_9333_transpose_x_0 = const()[name = tensor<string, []>("op_9333_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9333_transpose_y_0 = const()[name = tensor<string, []>("op_9333_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9333_cast_fp16 = matmul(transpose_x = var_9333_transpose_x_0, transpose_y = var_9333_transpose_y_0, x = attn_303_cast_fp16, y = v_455_cast_fp16)[name = tensor<string, []>("op_9333_cast_fp16")];
            tensor<int32, [1]> var_9334_axes_0 = const()[name = tensor<string, []>("op_9334_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9334_cast_fp16 = squeeze(axes = var_9334_axes_0, x = var_9333_cast_fp16)[name = tensor<string, []>("op_9334_cast_fp16")];
            tensor<string, []> var_9340_pad_type_0 = const()[name = tensor<string, []>("op_9340_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9340_strides_0 = const()[name = tensor<string, []>("op_9340_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9340_pad_0 = const()[name = tensor<string, []>("op_9340_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9340_dilations_0 = const()[name = tensor<string, []>("op_9340_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9340_groups_0 = const()[name = tensor<string, []>("op_9340_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87825088)))];
            tensor<fp16, [1, 64, 1, 197]> var_9340_cast_fp16 = conv(dilations = var_9340_dilations_0, groups = var_9340_groups_0, pad = var_9340_pad_0, pad_type = var_9340_pad_type_0, strides = var_9340_strides_0, weight = model_blocks_6_attn_q_projs_4_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9340_cast_fp16")];
            tensor<int32, [1]> var_9341_axes_0 = const()[name = tensor<string, []>("op_9341_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9341_cast_fp16 = squeeze(axes = var_9341_axes_0, x = var_9340_cast_fp16)[name = tensor<string, []>("op_9341_cast_fp16")];
            tensor<int32, [3]> q_761_perm_0 = const()[name = tensor<string, []>("q_761_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9348_pad_type_0 = const()[name = tensor<string, []>("op_9348_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9348_strides_0 = const()[name = tensor<string, []>("op_9348_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9348_pad_0 = const()[name = tensor<string, []>("op_9348_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9348_dilations_0 = const()[name = tensor<string, []>("op_9348_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9348_groups_0 = const()[name = tensor<string, []>("op_9348_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87923456)))];
            tensor<fp16, [1, 64, 1, 197]> var_9348_cast_fp16 = conv(dilations = var_9348_dilations_0, groups = var_9348_groups_0, pad = var_9348_pad_0, pad_type = var_9348_pad_type_0, strides = var_9348_strides_0, weight = model_blocks_6_attn_k_projs_4_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9348_cast_fp16")];
            tensor<int32, [1]> var_9349_axes_0 = const()[name = tensor<string, []>("op_9349_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9349_cast_fp16 = squeeze(axes = var_9349_axes_0, x = var_9348_cast_fp16)[name = tensor<string, []>("op_9349_cast_fp16")];
            tensor<int32, [3]> k_457_perm_0 = const()[name = tensor<string, []>("k_457_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9356_pad_type_0 = const()[name = tensor<string, []>("op_9356_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9356_strides_0 = const()[name = tensor<string, []>("op_9356_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9356_pad_0 = const()[name = tensor<string, []>("op_9356_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9356_dilations_0 = const()[name = tensor<string, []>("op_9356_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9356_groups_0 = const()[name = tensor<string, []>("op_9356_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88021824)))];
            tensor<fp16, [1, 64, 1, 197]> var_9356_cast_fp16 = conv(dilations = var_9356_dilations_0, groups = var_9356_groups_0, pad = var_9356_pad_0, pad_type = var_9356_pad_type_0, strides = var_9356_strides_0, weight = model_blocks_6_attn_v_projs_4_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9356_cast_fp16")];
            tensor<int32, [1]> var_9357_axes_0 = const()[name = tensor<string, []>("op_9357_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9357_cast_fp16 = squeeze(axes = var_9357_axes_0, x = var_9356_cast_fp16)[name = tensor<string, []>("op_9357_cast_fp16")];
            tensor<int32, [3]> v_457_perm_0 = const()[name = tensor<string, []>("v_457_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88120192)))];
            tensor<fp16, [1, 197, 64]> q_761_cast_fp16 = transpose(perm = q_761_perm_0, x = var_9341_cast_fp16)[name = tensor<string, []>("transpose_244")];
            tensor<fp16, [1, 197, 64]> q_763_cast_fp16 = add(x = q_761_cast_fp16, y = model_blocks_6_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_763_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88120384)))];
            tensor<fp16, [1, 197, 64]> v_457_cast_fp16 = transpose(perm = v_457_perm_0, x = var_9357_cast_fp16)[name = tensor<string, []>("transpose_242")];
            tensor<fp16, [1, 197, 64]> v_459_cast_fp16 = add(x = v_457_cast_fp16, y = model_blocks_6_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_459_cast_fp16")];
            tensor<int32, [1]> q_765_axes_0 = const()[name = tensor<string, []>("q_765_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_765_cast_fp16 = expand_dims(axes = q_765_axes_0, x = q_763_cast_fp16)[name = tensor<string, []>("q_765_cast_fp16")];
            tensor<int32, [1]> k_459_axes_0 = const()[name = tensor<string, []>("k_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_457_cast_fp16 = transpose(perm = k_457_perm_0, x = var_9349_cast_fp16)[name = tensor<string, []>("transpose_243")];
            tensor<fp16, [1, 1, 197, 64]> k_459_cast_fp16 = expand_dims(axes = k_459_axes_0, x = k_457_cast_fp16)[name = tensor<string, []>("k_459_cast_fp16")];
            tensor<int32, [1]> v_461_axes_0 = const()[name = tensor<string, []>("v_461_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_461_cast_fp16 = expand_dims(axes = v_461_axes_0, x = v_459_cast_fp16)[name = tensor<string, []>("v_461_cast_fp16")];
            tensor<int32, [4]> var_9366_begin_0 = const()[name = tensor<string, []>("op_9366_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9366_end_0 = const()[name = tensor<string, []>("op_9366_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9366_end_mask_0 = const()[name = tensor<string, []>("op_9366_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9366_cast_fp16 = slice_by_index(begin = var_9366_begin_0, end = var_9366_end_0, end_mask = var_9366_end_mask_0, x = q_765_cast_fp16)[name = tensor<string, []>("op_9366_cast_fp16")];
            tensor<int32, [4]> var_9370_begin_0 = const()[name = tensor<string, []>("op_9370_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9370_end_0 = const()[name = tensor<string, []>("op_9370_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9370_end_mask_0 = const()[name = tensor<string, []>("op_9370_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9370_cast_fp16 = slice_by_index(begin = var_9370_begin_0, end = var_9370_end_0, end_mask = var_9370_end_mask_0, x = k_459_cast_fp16)[name = tensor<string, []>("op_9370_cast_fp16")];
            tensor<int32, [4]> var_9374_begin_0 = const()[name = tensor<string, []>("op_9374_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9374_end_0 = const()[name = tensor<string, []>("op_9374_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9374_end_mask_0 = const()[name = tensor<string, []>("op_9374_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9374_cast_fp16 = slice_by_index(begin = var_9374_begin_0, end = var_9374_end_0, end_mask = var_9374_end_mask_0, x = q_765_cast_fp16)[name = tensor<string, []>("op_9374_cast_fp16")];
            tensor<int32, [4]> var_9378_begin_0 = const()[name = tensor<string, []>("op_9378_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9378_end_0 = const()[name = tensor<string, []>("op_9378_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9378_end_mask_0 = const()[name = tensor<string, []>("op_9378_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9378_cast_fp16 = slice_by_index(begin = var_9378_begin_0, end = var_9378_end_0, end_mask = var_9378_end_mask_0, x = k_459_cast_fp16)[name = tensor<string, []>("op_9378_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9380_cast_fp16 = mul(x = var_9374_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9380_cast_fp16")];
            tensor<int32, [5]> var_9388 = const()[name = tensor<string, []>("op_9388"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1043_cast_fp16 = reshape(shape = var_9388, x = var_9374_cast_fp16)[name = tensor<string, []>("x_1043_cast_fp16")];
            tensor<int32, [2]> var_9390_split_sizes_0 = const()[name = tensor<string, []>("op_9390_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9390_axis_0 = const()[name = tensor<string, []>("op_9390_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9390_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9390_cast_fp16_1 = split(axis = var_9390_axis_0, split_sizes = var_9390_split_sizes_0, x = x_1043_cast_fp16)[name = tensor<string, []>("op_9390_cast_fp16")];
            tensor<int32, [1]> squeeze_304_axes_0 = const()[name = tensor<string, []>("squeeze_304_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_304_cast_fp16 = squeeze(axes = squeeze_304_axes_0, x = var_9390_cast_fp16_0)[name = tensor<string, []>("squeeze_304_cast_fp16")];
            tensor<int32, [1]> squeeze_305_axes_0 = const()[name = tensor<string, []>("squeeze_305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_305_cast_fp16 = squeeze(axes = squeeze_305_axes_0, x = var_9390_cast_fp16_1)[name = tensor<string, []>("squeeze_305_cast_fp16")];
            tensor<fp16, []> const_1247_promoted_to_fp16 = const()[name = tensor<string, []>("const_1247_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9393_cast_fp16 = mul(x = squeeze_305_cast_fp16, y = const_1247_promoted_to_fp16)[name = tensor<string, []>("op_9393_cast_fp16")];
            tensor<int32, []> x_1045_axis_0 = const()[name = tensor<string, []>("x_1045_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1045_cast_fp16 = stack(axis = x_1045_axis_0, values = (var_9393_cast_fp16, squeeze_304_cast_fp16))[name = tensor<string, []>("x_1045_cast_fp16")];
            tensor<int32, [4]> var_9399 = const()[name = tensor<string, []>("op_9399"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9400_cast_fp16 = reshape(shape = var_9399, x = x_1045_cast_fp16)[name = tensor<string, []>("op_9400_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9401_cast_fp16 = mul(x = var_9400_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9401_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_153_cast_fp16 = add(x = var_9380_cast_fp16, y = var_9401_cast_fp16)[name = tensor<string, []>("q_patches_153_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9403_cast_fp16 = mul(x = var_9378_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9403_cast_fp16")];
            tensor<int32, [5]> var_9411 = const()[name = tensor<string, []>("op_9411"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1049_cast_fp16 = reshape(shape = var_9411, x = var_9378_cast_fp16)[name = tensor<string, []>("x_1049_cast_fp16")];
            tensor<int32, [2]> var_9413_split_sizes_0 = const()[name = tensor<string, []>("op_9413_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9413_axis_0 = const()[name = tensor<string, []>("op_9413_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9413_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9413_cast_fp16_1 = split(axis = var_9413_axis_0, split_sizes = var_9413_split_sizes_0, x = x_1049_cast_fp16)[name = tensor<string, []>("op_9413_cast_fp16")];
            tensor<int32, [1]> squeeze_306_axes_0 = const()[name = tensor<string, []>("squeeze_306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_306_cast_fp16 = squeeze(axes = squeeze_306_axes_0, x = var_9413_cast_fp16_0)[name = tensor<string, []>("squeeze_306_cast_fp16")];
            tensor<int32, [1]> squeeze_307_axes_0 = const()[name = tensor<string, []>("squeeze_307_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_307_cast_fp16 = squeeze(axes = squeeze_307_axes_0, x = var_9413_cast_fp16_1)[name = tensor<string, []>("squeeze_307_cast_fp16")];
            tensor<fp16, []> const_1255_promoted_to_fp16 = const()[name = tensor<string, []>("const_1255_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9416_cast_fp16 = mul(x = squeeze_307_cast_fp16, y = const_1255_promoted_to_fp16)[name = tensor<string, []>("op_9416_cast_fp16")];
            tensor<int32, []> x_1051_axis_0 = const()[name = tensor<string, []>("x_1051_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1051_cast_fp16 = stack(axis = x_1051_axis_0, values = (var_9416_cast_fp16, squeeze_306_cast_fp16))[name = tensor<string, []>("x_1051_cast_fp16")];
            tensor<int32, [4]> var_9422 = const()[name = tensor<string, []>("op_9422"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9423_cast_fp16 = reshape(shape = var_9422, x = x_1051_cast_fp16)[name = tensor<string, []>("op_9423_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9424_cast_fp16 = mul(x = var_9423_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9424_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_153_cast_fp16 = add(x = var_9403_cast_fp16, y = var_9424_cast_fp16)[name = tensor<string, []>("k_patches_153_cast_fp16")];
            tensor<bool, []> var_9427_interleave_0 = const()[name = tensor<string, []>("op_9427_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9427_cast_fp16 = concat(axis = var_20, interleave = var_9427_interleave_0, values = (var_9366_cast_fp16, q_patches_153_cast_fp16))[name = tensor<string, []>("op_9427_cast_fp16")];
            tensor<bool, []> var_9430_interleave_0 = const()[name = tensor<string, []>("op_9430_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9430_cast_fp16 = concat(axis = var_20, interleave = var_9430_interleave_0, values = (var_9370_cast_fp16, k_patches_153_cast_fp16))[name = tensor<string, []>("op_9430_cast_fp16")];
            tensor<fp16, []> var_9432_to_fp16 = const()[name = tensor<string, []>("op_9432_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_769_cast_fp16 = mul(x = var_9427_cast_fp16, y = var_9432_to_fp16)[name = tensor<string, []>("q_769_cast_fp16")];
            tensor<bool, []> attn_305_transpose_x_1 = const()[name = tensor<string, []>("attn_305_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_305_transpose_y_1 = const()[name = tensor<string, []>("attn_305_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_305_cast_fp16 = matmul(transpose_x = attn_305_transpose_x_1, transpose_y = attn_305_transpose_y_1, x = q_769_cast_fp16, y = var_9430_cast_fp16)[name = tensor<string, []>("attn_305_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_307_cast_fp16 = softmax(axis = var_21, x = attn_305_cast_fp16)[name = tensor<string, []>("attn_307_cast_fp16")];
            tensor<bool, []> var_9437_transpose_x_0 = const()[name = tensor<string, []>("op_9437_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9437_transpose_y_0 = const()[name = tensor<string, []>("op_9437_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9437_cast_fp16 = matmul(transpose_x = var_9437_transpose_x_0, transpose_y = var_9437_transpose_y_0, x = attn_307_cast_fp16, y = v_461_cast_fp16)[name = tensor<string, []>("op_9437_cast_fp16")];
            tensor<int32, [1]> var_9438_axes_0 = const()[name = tensor<string, []>("op_9438_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9438_cast_fp16 = squeeze(axes = var_9438_axes_0, x = var_9437_cast_fp16)[name = tensor<string, []>("op_9438_cast_fp16")];
            tensor<string, []> var_9444_pad_type_0 = const()[name = tensor<string, []>("op_9444_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9444_strides_0 = const()[name = tensor<string, []>("op_9444_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9444_pad_0 = const()[name = tensor<string, []>("op_9444_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9444_dilations_0 = const()[name = tensor<string, []>("op_9444_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9444_groups_0 = const()[name = tensor<string, []>("op_9444_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88120576)))];
            tensor<fp16, [1, 64, 1, 197]> var_9444_cast_fp16 = conv(dilations = var_9444_dilations_0, groups = var_9444_groups_0, pad = var_9444_pad_0, pad_type = var_9444_pad_type_0, strides = var_9444_strides_0, weight = model_blocks_6_attn_q_projs_5_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9444_cast_fp16")];
            tensor<int32, [1]> var_9445_axes_0 = const()[name = tensor<string, []>("op_9445_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9445_cast_fp16 = squeeze(axes = var_9445_axes_0, x = var_9444_cast_fp16)[name = tensor<string, []>("op_9445_cast_fp16")];
            tensor<int32, [3]> q_771_perm_0 = const()[name = tensor<string, []>("q_771_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9452_pad_type_0 = const()[name = tensor<string, []>("op_9452_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9452_strides_0 = const()[name = tensor<string, []>("op_9452_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9452_pad_0 = const()[name = tensor<string, []>("op_9452_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9452_dilations_0 = const()[name = tensor<string, []>("op_9452_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9452_groups_0 = const()[name = tensor<string, []>("op_9452_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88218944)))];
            tensor<fp16, [1, 64, 1, 197]> var_9452_cast_fp16 = conv(dilations = var_9452_dilations_0, groups = var_9452_groups_0, pad = var_9452_pad_0, pad_type = var_9452_pad_type_0, strides = var_9452_strides_0, weight = model_blocks_6_attn_k_projs_5_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9452_cast_fp16")];
            tensor<int32, [1]> var_9453_axes_0 = const()[name = tensor<string, []>("op_9453_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9453_cast_fp16 = squeeze(axes = var_9453_axes_0, x = var_9452_cast_fp16)[name = tensor<string, []>("op_9453_cast_fp16")];
            tensor<int32, [3]> k_463_perm_0 = const()[name = tensor<string, []>("k_463_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9460_pad_type_0 = const()[name = tensor<string, []>("op_9460_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9460_strides_0 = const()[name = tensor<string, []>("op_9460_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9460_pad_0 = const()[name = tensor<string, []>("op_9460_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9460_dilations_0 = const()[name = tensor<string, []>("op_9460_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9460_groups_0 = const()[name = tensor<string, []>("op_9460_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88317312)))];
            tensor<fp16, [1, 64, 1, 197]> var_9460_cast_fp16 = conv(dilations = var_9460_dilations_0, groups = var_9460_groups_0, pad = var_9460_pad_0, pad_type = var_9460_pad_type_0, strides = var_9460_strides_0, weight = model_blocks_6_attn_v_projs_5_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9460_cast_fp16")];
            tensor<int32, [1]> var_9461_axes_0 = const()[name = tensor<string, []>("op_9461_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9461_cast_fp16 = squeeze(axes = var_9461_axes_0, x = var_9460_cast_fp16)[name = tensor<string, []>("op_9461_cast_fp16")];
            tensor<int32, [3]> v_463_perm_0 = const()[name = tensor<string, []>("v_463_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88415680)))];
            tensor<fp16, [1, 197, 64]> q_771_cast_fp16 = transpose(perm = q_771_perm_0, x = var_9445_cast_fp16)[name = tensor<string, []>("transpose_241")];
            tensor<fp16, [1, 197, 64]> q_773_cast_fp16 = add(x = q_771_cast_fp16, y = model_blocks_6_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_773_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88415872)))];
            tensor<fp16, [1, 197, 64]> v_463_cast_fp16 = transpose(perm = v_463_perm_0, x = var_9461_cast_fp16)[name = tensor<string, []>("transpose_239")];
            tensor<fp16, [1, 197, 64]> v_465_cast_fp16 = add(x = v_463_cast_fp16, y = model_blocks_6_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_465_cast_fp16")];
            tensor<int32, [1]> q_775_axes_0 = const()[name = tensor<string, []>("q_775_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_775_cast_fp16 = expand_dims(axes = q_775_axes_0, x = q_773_cast_fp16)[name = tensor<string, []>("q_775_cast_fp16")];
            tensor<int32, [1]> k_465_axes_0 = const()[name = tensor<string, []>("k_465_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_463_cast_fp16 = transpose(perm = k_463_perm_0, x = var_9453_cast_fp16)[name = tensor<string, []>("transpose_240")];
            tensor<fp16, [1, 1, 197, 64]> k_465_cast_fp16 = expand_dims(axes = k_465_axes_0, x = k_463_cast_fp16)[name = tensor<string, []>("k_465_cast_fp16")];
            tensor<int32, [1]> v_467_axes_0 = const()[name = tensor<string, []>("v_467_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_467_cast_fp16 = expand_dims(axes = v_467_axes_0, x = v_465_cast_fp16)[name = tensor<string, []>("v_467_cast_fp16")];
            tensor<int32, [4]> var_9470_begin_0 = const()[name = tensor<string, []>("op_9470_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9470_end_0 = const()[name = tensor<string, []>("op_9470_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9470_end_mask_0 = const()[name = tensor<string, []>("op_9470_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9470_cast_fp16 = slice_by_index(begin = var_9470_begin_0, end = var_9470_end_0, end_mask = var_9470_end_mask_0, x = q_775_cast_fp16)[name = tensor<string, []>("op_9470_cast_fp16")];
            tensor<int32, [4]> var_9474_begin_0 = const()[name = tensor<string, []>("op_9474_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9474_end_0 = const()[name = tensor<string, []>("op_9474_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9474_end_mask_0 = const()[name = tensor<string, []>("op_9474_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9474_cast_fp16 = slice_by_index(begin = var_9474_begin_0, end = var_9474_end_0, end_mask = var_9474_end_mask_0, x = k_465_cast_fp16)[name = tensor<string, []>("op_9474_cast_fp16")];
            tensor<int32, [4]> var_9478_begin_0 = const()[name = tensor<string, []>("op_9478_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9478_end_0 = const()[name = tensor<string, []>("op_9478_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9478_end_mask_0 = const()[name = tensor<string, []>("op_9478_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9478_cast_fp16 = slice_by_index(begin = var_9478_begin_0, end = var_9478_end_0, end_mask = var_9478_end_mask_0, x = q_775_cast_fp16)[name = tensor<string, []>("op_9478_cast_fp16")];
            tensor<int32, [4]> var_9482_begin_0 = const()[name = tensor<string, []>("op_9482_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9482_end_0 = const()[name = tensor<string, []>("op_9482_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9482_end_mask_0 = const()[name = tensor<string, []>("op_9482_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, x = k_465_cast_fp16)[name = tensor<string, []>("op_9482_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9484_cast_fp16 = mul(x = var_9478_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9484_cast_fp16")];
            tensor<int32, [5]> var_9492 = const()[name = tensor<string, []>("op_9492"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1055_cast_fp16 = reshape(shape = var_9492, x = var_9478_cast_fp16)[name = tensor<string, []>("x_1055_cast_fp16")];
            tensor<int32, [2]> var_9494_split_sizes_0 = const()[name = tensor<string, []>("op_9494_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9494_axis_0 = const()[name = tensor<string, []>("op_9494_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9494_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9494_cast_fp16_1 = split(axis = var_9494_axis_0, split_sizes = var_9494_split_sizes_0, x = x_1055_cast_fp16)[name = tensor<string, []>("op_9494_cast_fp16")];
            tensor<int32, [1]> squeeze_308_axes_0 = const()[name = tensor<string, []>("squeeze_308_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_308_cast_fp16 = squeeze(axes = squeeze_308_axes_0, x = var_9494_cast_fp16_0)[name = tensor<string, []>("squeeze_308_cast_fp16")];
            tensor<int32, [1]> squeeze_309_axes_0 = const()[name = tensor<string, []>("squeeze_309_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_309_cast_fp16 = squeeze(axes = squeeze_309_axes_0, x = var_9494_cast_fp16_1)[name = tensor<string, []>("squeeze_309_cast_fp16")];
            tensor<fp16, []> const_1263_promoted_to_fp16 = const()[name = tensor<string, []>("const_1263_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9497_cast_fp16 = mul(x = squeeze_309_cast_fp16, y = const_1263_promoted_to_fp16)[name = tensor<string, []>("op_9497_cast_fp16")];
            tensor<int32, []> x_1057_axis_0 = const()[name = tensor<string, []>("x_1057_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1057_cast_fp16 = stack(axis = x_1057_axis_0, values = (var_9497_cast_fp16, squeeze_308_cast_fp16))[name = tensor<string, []>("x_1057_cast_fp16")];
            tensor<int32, [4]> var_9503 = const()[name = tensor<string, []>("op_9503"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9504_cast_fp16 = reshape(shape = var_9503, x = x_1057_cast_fp16)[name = tensor<string, []>("op_9504_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9505_cast_fp16 = mul(x = var_9504_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9505_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_155_cast_fp16 = add(x = var_9484_cast_fp16, y = var_9505_cast_fp16)[name = tensor<string, []>("q_patches_155_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9507_cast_fp16 = mul(x = var_9482_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9507_cast_fp16")];
            tensor<int32, [5]> var_9515 = const()[name = tensor<string, []>("op_9515"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1061_cast_fp16 = reshape(shape = var_9515, x = var_9482_cast_fp16)[name = tensor<string, []>("x_1061_cast_fp16")];
            tensor<int32, [2]> var_9517_split_sizes_0 = const()[name = tensor<string, []>("op_9517_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9517_axis_0 = const()[name = tensor<string, []>("op_9517_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9517_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9517_cast_fp16_1 = split(axis = var_9517_axis_0, split_sizes = var_9517_split_sizes_0, x = x_1061_cast_fp16)[name = tensor<string, []>("op_9517_cast_fp16")];
            tensor<int32, [1]> squeeze_310_axes_0 = const()[name = tensor<string, []>("squeeze_310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_310_cast_fp16 = squeeze(axes = squeeze_310_axes_0, x = var_9517_cast_fp16_0)[name = tensor<string, []>("squeeze_310_cast_fp16")];
            tensor<int32, [1]> squeeze_311_axes_0 = const()[name = tensor<string, []>("squeeze_311_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_311_cast_fp16 = squeeze(axes = squeeze_311_axes_0, x = var_9517_cast_fp16_1)[name = tensor<string, []>("squeeze_311_cast_fp16")];
            tensor<fp16, []> const_1271_promoted_to_fp16 = const()[name = tensor<string, []>("const_1271_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9520_cast_fp16 = mul(x = squeeze_311_cast_fp16, y = const_1271_promoted_to_fp16)[name = tensor<string, []>("op_9520_cast_fp16")];
            tensor<int32, []> x_1063_axis_0 = const()[name = tensor<string, []>("x_1063_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1063_cast_fp16 = stack(axis = x_1063_axis_0, values = (var_9520_cast_fp16, squeeze_310_cast_fp16))[name = tensor<string, []>("x_1063_cast_fp16")];
            tensor<int32, [4]> var_9526 = const()[name = tensor<string, []>("op_9526"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9527_cast_fp16 = reshape(shape = var_9526, x = x_1063_cast_fp16)[name = tensor<string, []>("op_9527_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9528_cast_fp16 = mul(x = var_9527_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9528_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_155_cast_fp16 = add(x = var_9507_cast_fp16, y = var_9528_cast_fp16)[name = tensor<string, []>("k_patches_155_cast_fp16")];
            tensor<bool, []> var_9531_interleave_0 = const()[name = tensor<string, []>("op_9531_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9531_cast_fp16 = concat(axis = var_20, interleave = var_9531_interleave_0, values = (var_9470_cast_fp16, q_patches_155_cast_fp16))[name = tensor<string, []>("op_9531_cast_fp16")];
            tensor<bool, []> var_9534_interleave_0 = const()[name = tensor<string, []>("op_9534_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9534_cast_fp16 = concat(axis = var_20, interleave = var_9534_interleave_0, values = (var_9474_cast_fp16, k_patches_155_cast_fp16))[name = tensor<string, []>("op_9534_cast_fp16")];
            tensor<fp16, []> var_9536_to_fp16 = const()[name = tensor<string, []>("op_9536_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_779_cast_fp16 = mul(x = var_9531_cast_fp16, y = var_9536_to_fp16)[name = tensor<string, []>("q_779_cast_fp16")];
            tensor<bool, []> attn_309_transpose_x_1 = const()[name = tensor<string, []>("attn_309_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_309_transpose_y_1 = const()[name = tensor<string, []>("attn_309_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_309_cast_fp16 = matmul(transpose_x = attn_309_transpose_x_1, transpose_y = attn_309_transpose_y_1, x = q_779_cast_fp16, y = var_9534_cast_fp16)[name = tensor<string, []>("attn_309_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_311_cast_fp16 = softmax(axis = var_21, x = attn_309_cast_fp16)[name = tensor<string, []>("attn_311_cast_fp16")];
            tensor<bool, []> var_9541_transpose_x_0 = const()[name = tensor<string, []>("op_9541_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9541_transpose_y_0 = const()[name = tensor<string, []>("op_9541_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9541_cast_fp16 = matmul(transpose_x = var_9541_transpose_x_0, transpose_y = var_9541_transpose_y_0, x = attn_311_cast_fp16, y = v_467_cast_fp16)[name = tensor<string, []>("op_9541_cast_fp16")];
            tensor<int32, [1]> var_9542_axes_0 = const()[name = tensor<string, []>("op_9542_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9542_cast_fp16 = squeeze(axes = var_9542_axes_0, x = var_9541_cast_fp16)[name = tensor<string, []>("op_9542_cast_fp16")];
            tensor<string, []> var_9548_pad_type_0 = const()[name = tensor<string, []>("op_9548_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9548_strides_0 = const()[name = tensor<string, []>("op_9548_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9548_pad_0 = const()[name = tensor<string, []>("op_9548_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9548_dilations_0 = const()[name = tensor<string, []>("op_9548_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9548_groups_0 = const()[name = tensor<string, []>("op_9548_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88416064)))];
            tensor<fp16, [1, 64, 1, 197]> var_9548_cast_fp16 = conv(dilations = var_9548_dilations_0, groups = var_9548_groups_0, pad = var_9548_pad_0, pad_type = var_9548_pad_type_0, strides = var_9548_strides_0, weight = model_blocks_6_attn_q_projs_6_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9548_cast_fp16")];
            tensor<int32, [1]> var_9549_axes_0 = const()[name = tensor<string, []>("op_9549_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9549_cast_fp16 = squeeze(axes = var_9549_axes_0, x = var_9548_cast_fp16)[name = tensor<string, []>("op_9549_cast_fp16")];
            tensor<int32, [3]> q_781_perm_0 = const()[name = tensor<string, []>("q_781_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9556_pad_type_0 = const()[name = tensor<string, []>("op_9556_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9556_strides_0 = const()[name = tensor<string, []>("op_9556_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9556_pad_0 = const()[name = tensor<string, []>("op_9556_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9556_dilations_0 = const()[name = tensor<string, []>("op_9556_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9556_groups_0 = const()[name = tensor<string, []>("op_9556_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88514432)))];
            tensor<fp16, [1, 64, 1, 197]> var_9556_cast_fp16 = conv(dilations = var_9556_dilations_0, groups = var_9556_groups_0, pad = var_9556_pad_0, pad_type = var_9556_pad_type_0, strides = var_9556_strides_0, weight = model_blocks_6_attn_k_projs_6_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9556_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = tensor<string, []>("op_9557_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9557_cast_fp16 = squeeze(axes = var_9557_axes_0, x = var_9556_cast_fp16)[name = tensor<string, []>("op_9557_cast_fp16")];
            tensor<int32, [3]> k_469_perm_0 = const()[name = tensor<string, []>("k_469_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9564_pad_type_0 = const()[name = tensor<string, []>("op_9564_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9564_strides_0 = const()[name = tensor<string, []>("op_9564_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9564_pad_0 = const()[name = tensor<string, []>("op_9564_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9564_dilations_0 = const()[name = tensor<string, []>("op_9564_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9564_groups_0 = const()[name = tensor<string, []>("op_9564_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88612800)))];
            tensor<fp16, [1, 64, 1, 197]> var_9564_cast_fp16 = conv(dilations = var_9564_dilations_0, groups = var_9564_groups_0, pad = var_9564_pad_0, pad_type = var_9564_pad_type_0, strides = var_9564_strides_0, weight = model_blocks_6_attn_v_projs_6_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9564_cast_fp16")];
            tensor<int32, [1]> var_9565_axes_0 = const()[name = tensor<string, []>("op_9565_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9565_cast_fp16 = squeeze(axes = var_9565_axes_0, x = var_9564_cast_fp16)[name = tensor<string, []>("op_9565_cast_fp16")];
            tensor<int32, [3]> v_469_perm_0 = const()[name = tensor<string, []>("v_469_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88711168)))];
            tensor<fp16, [1, 197, 64]> q_781_cast_fp16 = transpose(perm = q_781_perm_0, x = var_9549_cast_fp16)[name = tensor<string, []>("transpose_238")];
            tensor<fp16, [1, 197, 64]> q_783_cast_fp16 = add(x = q_781_cast_fp16, y = model_blocks_6_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_783_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88711360)))];
            tensor<fp16, [1, 197, 64]> v_469_cast_fp16 = transpose(perm = v_469_perm_0, x = var_9565_cast_fp16)[name = tensor<string, []>("transpose_236")];
            tensor<fp16, [1, 197, 64]> v_471_cast_fp16 = add(x = v_469_cast_fp16, y = model_blocks_6_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_471_cast_fp16")];
            tensor<int32, [1]> q_785_axes_0 = const()[name = tensor<string, []>("q_785_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_785_cast_fp16 = expand_dims(axes = q_785_axes_0, x = q_783_cast_fp16)[name = tensor<string, []>("q_785_cast_fp16")];
            tensor<int32, [1]> k_471_axes_0 = const()[name = tensor<string, []>("k_471_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_469_cast_fp16 = transpose(perm = k_469_perm_0, x = var_9557_cast_fp16)[name = tensor<string, []>("transpose_237")];
            tensor<fp16, [1, 1, 197, 64]> k_471_cast_fp16 = expand_dims(axes = k_471_axes_0, x = k_469_cast_fp16)[name = tensor<string, []>("k_471_cast_fp16")];
            tensor<int32, [1]> v_473_axes_0 = const()[name = tensor<string, []>("v_473_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_473_cast_fp16 = expand_dims(axes = v_473_axes_0, x = v_471_cast_fp16)[name = tensor<string, []>("v_473_cast_fp16")];
            tensor<int32, [4]> var_9574_begin_0 = const()[name = tensor<string, []>("op_9574_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9574_end_0 = const()[name = tensor<string, []>("op_9574_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9574_end_mask_0 = const()[name = tensor<string, []>("op_9574_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, x = q_785_cast_fp16)[name = tensor<string, []>("op_9574_cast_fp16")];
            tensor<int32, [4]> var_9578_begin_0 = const()[name = tensor<string, []>("op_9578_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9578_end_0 = const()[name = tensor<string, []>("op_9578_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9578_end_mask_0 = const()[name = tensor<string, []>("op_9578_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9578_cast_fp16 = slice_by_index(begin = var_9578_begin_0, end = var_9578_end_0, end_mask = var_9578_end_mask_0, x = k_471_cast_fp16)[name = tensor<string, []>("op_9578_cast_fp16")];
            tensor<int32, [4]> var_9582_begin_0 = const()[name = tensor<string, []>("op_9582_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9582_end_0 = const()[name = tensor<string, []>("op_9582_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9582_end_mask_0 = const()[name = tensor<string, []>("op_9582_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, x = q_785_cast_fp16)[name = tensor<string, []>("op_9582_cast_fp16")];
            tensor<int32, [4]> var_9586_begin_0 = const()[name = tensor<string, []>("op_9586_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9586_end_0 = const()[name = tensor<string, []>("op_9586_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9586_end_mask_0 = const()[name = tensor<string, []>("op_9586_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9586_cast_fp16 = slice_by_index(begin = var_9586_begin_0, end = var_9586_end_0, end_mask = var_9586_end_mask_0, x = k_471_cast_fp16)[name = tensor<string, []>("op_9586_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9588_cast_fp16 = mul(x = var_9582_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9588_cast_fp16")];
            tensor<int32, [5]> var_9596 = const()[name = tensor<string, []>("op_9596"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1067_cast_fp16 = reshape(shape = var_9596, x = var_9582_cast_fp16)[name = tensor<string, []>("x_1067_cast_fp16")];
            tensor<int32, [2]> var_9598_split_sizes_0 = const()[name = tensor<string, []>("op_9598_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9598_axis_0 = const()[name = tensor<string, []>("op_9598_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9598_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9598_cast_fp16_1 = split(axis = var_9598_axis_0, split_sizes = var_9598_split_sizes_0, x = x_1067_cast_fp16)[name = tensor<string, []>("op_9598_cast_fp16")];
            tensor<int32, [1]> squeeze_312_axes_0 = const()[name = tensor<string, []>("squeeze_312_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_312_cast_fp16 = squeeze(axes = squeeze_312_axes_0, x = var_9598_cast_fp16_0)[name = tensor<string, []>("squeeze_312_cast_fp16")];
            tensor<int32, [1]> squeeze_313_axes_0 = const()[name = tensor<string, []>("squeeze_313_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_313_cast_fp16 = squeeze(axes = squeeze_313_axes_0, x = var_9598_cast_fp16_1)[name = tensor<string, []>("squeeze_313_cast_fp16")];
            tensor<fp16, []> const_1279_promoted_to_fp16 = const()[name = tensor<string, []>("const_1279_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9601_cast_fp16 = mul(x = squeeze_313_cast_fp16, y = const_1279_promoted_to_fp16)[name = tensor<string, []>("op_9601_cast_fp16")];
            tensor<int32, []> x_1069_axis_0 = const()[name = tensor<string, []>("x_1069_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1069_cast_fp16 = stack(axis = x_1069_axis_0, values = (var_9601_cast_fp16, squeeze_312_cast_fp16))[name = tensor<string, []>("x_1069_cast_fp16")];
            tensor<int32, [4]> var_9607 = const()[name = tensor<string, []>("op_9607"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9608_cast_fp16 = reshape(shape = var_9607, x = x_1069_cast_fp16)[name = tensor<string, []>("op_9608_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9609_cast_fp16 = mul(x = var_9608_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9609_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_157_cast_fp16 = add(x = var_9588_cast_fp16, y = var_9609_cast_fp16)[name = tensor<string, []>("q_patches_157_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9611_cast_fp16 = mul(x = var_9586_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9611_cast_fp16")];
            tensor<int32, [5]> var_9619 = const()[name = tensor<string, []>("op_9619"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1073_cast_fp16 = reshape(shape = var_9619, x = var_9586_cast_fp16)[name = tensor<string, []>("x_1073_cast_fp16")];
            tensor<int32, [2]> var_9621_split_sizes_0 = const()[name = tensor<string, []>("op_9621_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9621_axis_0 = const()[name = tensor<string, []>("op_9621_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9621_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9621_cast_fp16_1 = split(axis = var_9621_axis_0, split_sizes = var_9621_split_sizes_0, x = x_1073_cast_fp16)[name = tensor<string, []>("op_9621_cast_fp16")];
            tensor<int32, [1]> squeeze_314_axes_0 = const()[name = tensor<string, []>("squeeze_314_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_314_cast_fp16 = squeeze(axes = squeeze_314_axes_0, x = var_9621_cast_fp16_0)[name = tensor<string, []>("squeeze_314_cast_fp16")];
            tensor<int32, [1]> squeeze_315_axes_0 = const()[name = tensor<string, []>("squeeze_315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_315_cast_fp16 = squeeze(axes = squeeze_315_axes_0, x = var_9621_cast_fp16_1)[name = tensor<string, []>("squeeze_315_cast_fp16")];
            tensor<fp16, []> const_1287_promoted_to_fp16 = const()[name = tensor<string, []>("const_1287_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9624_cast_fp16 = mul(x = squeeze_315_cast_fp16, y = const_1287_promoted_to_fp16)[name = tensor<string, []>("op_9624_cast_fp16")];
            tensor<int32, []> x_1075_axis_0 = const()[name = tensor<string, []>("x_1075_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1075_cast_fp16 = stack(axis = x_1075_axis_0, values = (var_9624_cast_fp16, squeeze_314_cast_fp16))[name = tensor<string, []>("x_1075_cast_fp16")];
            tensor<int32, [4]> var_9630 = const()[name = tensor<string, []>("op_9630"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9631_cast_fp16 = reshape(shape = var_9630, x = x_1075_cast_fp16)[name = tensor<string, []>("op_9631_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9632_cast_fp16 = mul(x = var_9631_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9632_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_157_cast_fp16 = add(x = var_9611_cast_fp16, y = var_9632_cast_fp16)[name = tensor<string, []>("k_patches_157_cast_fp16")];
            tensor<bool, []> var_9635_interleave_0 = const()[name = tensor<string, []>("op_9635_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9635_cast_fp16 = concat(axis = var_20, interleave = var_9635_interleave_0, values = (var_9574_cast_fp16, q_patches_157_cast_fp16))[name = tensor<string, []>("op_9635_cast_fp16")];
            tensor<bool, []> var_9638_interleave_0 = const()[name = tensor<string, []>("op_9638_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9638_cast_fp16 = concat(axis = var_20, interleave = var_9638_interleave_0, values = (var_9578_cast_fp16, k_patches_157_cast_fp16))[name = tensor<string, []>("op_9638_cast_fp16")];
            tensor<fp16, []> var_9640_to_fp16 = const()[name = tensor<string, []>("op_9640_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_789_cast_fp16 = mul(x = var_9635_cast_fp16, y = var_9640_to_fp16)[name = tensor<string, []>("q_789_cast_fp16")];
            tensor<bool, []> attn_313_transpose_x_1 = const()[name = tensor<string, []>("attn_313_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_313_transpose_y_1 = const()[name = tensor<string, []>("attn_313_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_313_cast_fp16 = matmul(transpose_x = attn_313_transpose_x_1, transpose_y = attn_313_transpose_y_1, x = q_789_cast_fp16, y = var_9638_cast_fp16)[name = tensor<string, []>("attn_313_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_315_cast_fp16 = softmax(axis = var_21, x = attn_313_cast_fp16)[name = tensor<string, []>("attn_315_cast_fp16")];
            tensor<bool, []> var_9645_transpose_x_0 = const()[name = tensor<string, []>("op_9645_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9645_transpose_y_0 = const()[name = tensor<string, []>("op_9645_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9645_cast_fp16 = matmul(transpose_x = var_9645_transpose_x_0, transpose_y = var_9645_transpose_y_0, x = attn_315_cast_fp16, y = v_473_cast_fp16)[name = tensor<string, []>("op_9645_cast_fp16")];
            tensor<int32, [1]> var_9646_axes_0 = const()[name = tensor<string, []>("op_9646_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9646_cast_fp16 = squeeze(axes = var_9646_axes_0, x = var_9645_cast_fp16)[name = tensor<string, []>("op_9646_cast_fp16")];
            tensor<string, []> var_9652_pad_type_0 = const()[name = tensor<string, []>("op_9652_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9652_strides_0 = const()[name = tensor<string, []>("op_9652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9652_pad_0 = const()[name = tensor<string, []>("op_9652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9652_dilations_0 = const()[name = tensor<string, []>("op_9652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9652_groups_0 = const()[name = tensor<string, []>("op_9652_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88711552)))];
            tensor<fp16, [1, 64, 1, 197]> var_9652_cast_fp16 = conv(dilations = var_9652_dilations_0, groups = var_9652_groups_0, pad = var_9652_pad_0, pad_type = var_9652_pad_type_0, strides = var_9652_strides_0, weight = model_blocks_6_attn_q_projs_7_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9652_cast_fp16")];
            tensor<int32, [1]> var_9653_axes_0 = const()[name = tensor<string, []>("op_9653_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9653_cast_fp16 = squeeze(axes = var_9653_axes_0, x = var_9652_cast_fp16)[name = tensor<string, []>("op_9653_cast_fp16")];
            tensor<int32, [3]> q_791_perm_0 = const()[name = tensor<string, []>("q_791_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9660_pad_type_0 = const()[name = tensor<string, []>("op_9660_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9660_strides_0 = const()[name = tensor<string, []>("op_9660_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9660_pad_0 = const()[name = tensor<string, []>("op_9660_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9660_dilations_0 = const()[name = tensor<string, []>("op_9660_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9660_groups_0 = const()[name = tensor<string, []>("op_9660_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88809920)))];
            tensor<fp16, [1, 64, 1, 197]> var_9660_cast_fp16 = conv(dilations = var_9660_dilations_0, groups = var_9660_groups_0, pad = var_9660_pad_0, pad_type = var_9660_pad_type_0, strides = var_9660_strides_0, weight = model_blocks_6_attn_k_projs_7_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9660_cast_fp16")];
            tensor<int32, [1]> var_9661_axes_0 = const()[name = tensor<string, []>("op_9661_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9661_cast_fp16 = squeeze(axes = var_9661_axes_0, x = var_9660_cast_fp16)[name = tensor<string, []>("op_9661_cast_fp16")];
            tensor<int32, [3]> k_475_perm_0 = const()[name = tensor<string, []>("k_475_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9668_pad_type_0 = const()[name = tensor<string, []>("op_9668_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9668_strides_0 = const()[name = tensor<string, []>("op_9668_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9668_pad_0 = const()[name = tensor<string, []>("op_9668_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9668_dilations_0 = const()[name = tensor<string, []>("op_9668_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9668_groups_0 = const()[name = tensor<string, []>("op_9668_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88908288)))];
            tensor<fp16, [1, 64, 1, 197]> var_9668_cast_fp16 = conv(dilations = var_9668_dilations_0, groups = var_9668_groups_0, pad = var_9668_pad_0, pad_type = var_9668_pad_type_0, strides = var_9668_strides_0, weight = model_blocks_6_attn_v_projs_7_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9668_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = tensor<string, []>("op_9669_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9669_cast_fp16 = squeeze(axes = var_9669_axes_0, x = var_9668_cast_fp16)[name = tensor<string, []>("op_9669_cast_fp16")];
            tensor<int32, [3]> v_475_perm_0 = const()[name = tensor<string, []>("v_475_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89006656)))];
            tensor<fp16, [1, 197, 64]> q_791_cast_fp16 = transpose(perm = q_791_perm_0, x = var_9653_cast_fp16)[name = tensor<string, []>("transpose_235")];
            tensor<fp16, [1, 197, 64]> q_793_cast_fp16 = add(x = q_791_cast_fp16, y = model_blocks_6_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_793_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89006848)))];
            tensor<fp16, [1, 197, 64]> v_475_cast_fp16 = transpose(perm = v_475_perm_0, x = var_9669_cast_fp16)[name = tensor<string, []>("transpose_233")];
            tensor<fp16, [1, 197, 64]> v_477_cast_fp16 = add(x = v_475_cast_fp16, y = model_blocks_6_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_477_cast_fp16")];
            tensor<int32, [1]> q_795_axes_0 = const()[name = tensor<string, []>("q_795_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_795_cast_fp16 = expand_dims(axes = q_795_axes_0, x = q_793_cast_fp16)[name = tensor<string, []>("q_795_cast_fp16")];
            tensor<int32, [1]> k_477_axes_0 = const()[name = tensor<string, []>("k_477_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_475_cast_fp16 = transpose(perm = k_475_perm_0, x = var_9661_cast_fp16)[name = tensor<string, []>("transpose_234")];
            tensor<fp16, [1, 1, 197, 64]> k_477_cast_fp16 = expand_dims(axes = k_477_axes_0, x = k_475_cast_fp16)[name = tensor<string, []>("k_477_cast_fp16")];
            tensor<int32, [1]> v_479_axes_0 = const()[name = tensor<string, []>("v_479_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_479_cast_fp16 = expand_dims(axes = v_479_axes_0, x = v_477_cast_fp16)[name = tensor<string, []>("v_479_cast_fp16")];
            tensor<int32, [4]> var_9678_begin_0 = const()[name = tensor<string, []>("op_9678_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9678_end_0 = const()[name = tensor<string, []>("op_9678_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9678_end_mask_0 = const()[name = tensor<string, []>("op_9678_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9678_cast_fp16 = slice_by_index(begin = var_9678_begin_0, end = var_9678_end_0, end_mask = var_9678_end_mask_0, x = q_795_cast_fp16)[name = tensor<string, []>("op_9678_cast_fp16")];
            tensor<int32, [4]> var_9682_begin_0 = const()[name = tensor<string, []>("op_9682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9682_end_0 = const()[name = tensor<string, []>("op_9682_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9682_end_mask_0 = const()[name = tensor<string, []>("op_9682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9682_cast_fp16 = slice_by_index(begin = var_9682_begin_0, end = var_9682_end_0, end_mask = var_9682_end_mask_0, x = k_477_cast_fp16)[name = tensor<string, []>("op_9682_cast_fp16")];
            tensor<int32, [4]> var_9686_begin_0 = const()[name = tensor<string, []>("op_9686_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9686_end_0 = const()[name = tensor<string, []>("op_9686_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9686_end_mask_0 = const()[name = tensor<string, []>("op_9686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9686_cast_fp16 = slice_by_index(begin = var_9686_begin_0, end = var_9686_end_0, end_mask = var_9686_end_mask_0, x = q_795_cast_fp16)[name = tensor<string, []>("op_9686_cast_fp16")];
            tensor<int32, [4]> var_9690_begin_0 = const()[name = tensor<string, []>("op_9690_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9690_end_0 = const()[name = tensor<string, []>("op_9690_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9690_end_mask_0 = const()[name = tensor<string, []>("op_9690_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9690_cast_fp16 = slice_by_index(begin = var_9690_begin_0, end = var_9690_end_0, end_mask = var_9690_end_mask_0, x = k_477_cast_fp16)[name = tensor<string, []>("op_9690_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9692_cast_fp16 = mul(x = var_9686_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9692_cast_fp16")];
            tensor<int32, [5]> var_9700 = const()[name = tensor<string, []>("op_9700"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1079_cast_fp16 = reshape(shape = var_9700, x = var_9686_cast_fp16)[name = tensor<string, []>("x_1079_cast_fp16")];
            tensor<int32, [2]> var_9702_split_sizes_0 = const()[name = tensor<string, []>("op_9702_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9702_axis_0 = const()[name = tensor<string, []>("op_9702_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9702_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9702_cast_fp16_1 = split(axis = var_9702_axis_0, split_sizes = var_9702_split_sizes_0, x = x_1079_cast_fp16)[name = tensor<string, []>("op_9702_cast_fp16")];
            tensor<int32, [1]> squeeze_316_axes_0 = const()[name = tensor<string, []>("squeeze_316_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_316_cast_fp16 = squeeze(axes = squeeze_316_axes_0, x = var_9702_cast_fp16_0)[name = tensor<string, []>("squeeze_316_cast_fp16")];
            tensor<int32, [1]> squeeze_317_axes_0 = const()[name = tensor<string, []>("squeeze_317_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_317_cast_fp16 = squeeze(axes = squeeze_317_axes_0, x = var_9702_cast_fp16_1)[name = tensor<string, []>("squeeze_317_cast_fp16")];
            tensor<fp16, []> const_1295_promoted_to_fp16 = const()[name = tensor<string, []>("const_1295_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9705_cast_fp16 = mul(x = squeeze_317_cast_fp16, y = const_1295_promoted_to_fp16)[name = tensor<string, []>("op_9705_cast_fp16")];
            tensor<int32, []> x_1081_axis_0 = const()[name = tensor<string, []>("x_1081_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1081_cast_fp16 = stack(axis = x_1081_axis_0, values = (var_9705_cast_fp16, squeeze_316_cast_fp16))[name = tensor<string, []>("x_1081_cast_fp16")];
            tensor<int32, [4]> var_9711 = const()[name = tensor<string, []>("op_9711"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9712_cast_fp16 = reshape(shape = var_9711, x = x_1081_cast_fp16)[name = tensor<string, []>("op_9712_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9713_cast_fp16 = mul(x = var_9712_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9713_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_159_cast_fp16 = add(x = var_9692_cast_fp16, y = var_9713_cast_fp16)[name = tensor<string, []>("q_patches_159_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9715_cast_fp16 = mul(x = var_9690_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9715_cast_fp16")];
            tensor<int32, [5]> var_9723 = const()[name = tensor<string, []>("op_9723"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1085_cast_fp16 = reshape(shape = var_9723, x = var_9690_cast_fp16)[name = tensor<string, []>("x_1085_cast_fp16")];
            tensor<int32, [2]> var_9725_split_sizes_0 = const()[name = tensor<string, []>("op_9725_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9725_axis_0 = const()[name = tensor<string, []>("op_9725_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9725_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9725_cast_fp16_1 = split(axis = var_9725_axis_0, split_sizes = var_9725_split_sizes_0, x = x_1085_cast_fp16)[name = tensor<string, []>("op_9725_cast_fp16")];
            tensor<int32, [1]> squeeze_318_axes_0 = const()[name = tensor<string, []>("squeeze_318_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_318_cast_fp16 = squeeze(axes = squeeze_318_axes_0, x = var_9725_cast_fp16_0)[name = tensor<string, []>("squeeze_318_cast_fp16")];
            tensor<int32, [1]> squeeze_319_axes_0 = const()[name = tensor<string, []>("squeeze_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_319_cast_fp16 = squeeze(axes = squeeze_319_axes_0, x = var_9725_cast_fp16_1)[name = tensor<string, []>("squeeze_319_cast_fp16")];
            tensor<fp16, []> const_1303_promoted_to_fp16 = const()[name = tensor<string, []>("const_1303_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9728_cast_fp16 = mul(x = squeeze_319_cast_fp16, y = const_1303_promoted_to_fp16)[name = tensor<string, []>("op_9728_cast_fp16")];
            tensor<int32, []> x_1087_axis_0 = const()[name = tensor<string, []>("x_1087_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1087_cast_fp16 = stack(axis = x_1087_axis_0, values = (var_9728_cast_fp16, squeeze_318_cast_fp16))[name = tensor<string, []>("x_1087_cast_fp16")];
            tensor<int32, [4]> var_9734 = const()[name = tensor<string, []>("op_9734"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9735_cast_fp16 = reshape(shape = var_9734, x = x_1087_cast_fp16)[name = tensor<string, []>("op_9735_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9736_cast_fp16 = mul(x = var_9735_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9736_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_159_cast_fp16 = add(x = var_9715_cast_fp16, y = var_9736_cast_fp16)[name = tensor<string, []>("k_patches_159_cast_fp16")];
            tensor<bool, []> var_9739_interleave_0 = const()[name = tensor<string, []>("op_9739_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9739_cast_fp16 = concat(axis = var_20, interleave = var_9739_interleave_0, values = (var_9678_cast_fp16, q_patches_159_cast_fp16))[name = tensor<string, []>("op_9739_cast_fp16")];
            tensor<bool, []> var_9742_interleave_0 = const()[name = tensor<string, []>("op_9742_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9742_cast_fp16 = concat(axis = var_20, interleave = var_9742_interleave_0, values = (var_9682_cast_fp16, k_patches_159_cast_fp16))[name = tensor<string, []>("op_9742_cast_fp16")];
            tensor<fp16, []> var_9744_to_fp16 = const()[name = tensor<string, []>("op_9744_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_799_cast_fp16 = mul(x = var_9739_cast_fp16, y = var_9744_to_fp16)[name = tensor<string, []>("q_799_cast_fp16")];
            tensor<bool, []> attn_317_transpose_x_1 = const()[name = tensor<string, []>("attn_317_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_317_transpose_y_1 = const()[name = tensor<string, []>("attn_317_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_317_cast_fp16 = matmul(transpose_x = attn_317_transpose_x_1, transpose_y = attn_317_transpose_y_1, x = q_799_cast_fp16, y = var_9742_cast_fp16)[name = tensor<string, []>("attn_317_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_319_cast_fp16 = softmax(axis = var_21, x = attn_317_cast_fp16)[name = tensor<string, []>("attn_319_cast_fp16")];
            tensor<bool, []> var_9749_transpose_x_0 = const()[name = tensor<string, []>("op_9749_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9749_transpose_y_0 = const()[name = tensor<string, []>("op_9749_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9749_cast_fp16 = matmul(transpose_x = var_9749_transpose_x_0, transpose_y = var_9749_transpose_y_0, x = attn_319_cast_fp16, y = v_479_cast_fp16)[name = tensor<string, []>("op_9749_cast_fp16")];
            tensor<int32, [1]> var_9750_axes_0 = const()[name = tensor<string, []>("op_9750_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9750_cast_fp16 = squeeze(axes = var_9750_axes_0, x = var_9749_cast_fp16)[name = tensor<string, []>("op_9750_cast_fp16")];
            tensor<string, []> var_9756_pad_type_0 = const()[name = tensor<string, []>("op_9756_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9756_strides_0 = const()[name = tensor<string, []>("op_9756_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9756_pad_0 = const()[name = tensor<string, []>("op_9756_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9756_dilations_0 = const()[name = tensor<string, []>("op_9756_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9756_groups_0 = const()[name = tensor<string, []>("op_9756_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89007040)))];
            tensor<fp16, [1, 64, 1, 197]> var_9756_cast_fp16 = conv(dilations = var_9756_dilations_0, groups = var_9756_groups_0, pad = var_9756_pad_0, pad_type = var_9756_pad_type_0, strides = var_9756_strides_0, weight = model_blocks_6_attn_q_projs_8_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9756_cast_fp16")];
            tensor<int32, [1]> var_9757_axes_0 = const()[name = tensor<string, []>("op_9757_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9757_cast_fp16 = squeeze(axes = var_9757_axes_0, x = var_9756_cast_fp16)[name = tensor<string, []>("op_9757_cast_fp16")];
            tensor<int32, [3]> q_801_perm_0 = const()[name = tensor<string, []>("q_801_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9764_pad_type_0 = const()[name = tensor<string, []>("op_9764_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9764_strides_0 = const()[name = tensor<string, []>("op_9764_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9764_pad_0 = const()[name = tensor<string, []>("op_9764_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9764_dilations_0 = const()[name = tensor<string, []>("op_9764_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9764_groups_0 = const()[name = tensor<string, []>("op_9764_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89105408)))];
            tensor<fp16, [1, 64, 1, 197]> var_9764_cast_fp16 = conv(dilations = var_9764_dilations_0, groups = var_9764_groups_0, pad = var_9764_pad_0, pad_type = var_9764_pad_type_0, strides = var_9764_strides_0, weight = model_blocks_6_attn_k_projs_8_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9764_cast_fp16")];
            tensor<int32, [1]> var_9765_axes_0 = const()[name = tensor<string, []>("op_9765_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9765_cast_fp16 = squeeze(axes = var_9765_axes_0, x = var_9764_cast_fp16)[name = tensor<string, []>("op_9765_cast_fp16")];
            tensor<int32, [3]> k_481_perm_0 = const()[name = tensor<string, []>("k_481_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9772_pad_type_0 = const()[name = tensor<string, []>("op_9772_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9772_strides_0 = const()[name = tensor<string, []>("op_9772_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9772_pad_0 = const()[name = tensor<string, []>("op_9772_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9772_dilations_0 = const()[name = tensor<string, []>("op_9772_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9772_groups_0 = const()[name = tensor<string, []>("op_9772_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89203776)))];
            tensor<fp16, [1, 64, 1, 197]> var_9772_cast_fp16 = conv(dilations = var_9772_dilations_0, groups = var_9772_groups_0, pad = var_9772_pad_0, pad_type = var_9772_pad_type_0, strides = var_9772_strides_0, weight = model_blocks_6_attn_v_projs_8_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9772_cast_fp16")];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = tensor<string, []>("op_9773_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9773_cast_fp16 = squeeze(axes = var_9773_axes_0, x = var_9772_cast_fp16)[name = tensor<string, []>("op_9773_cast_fp16")];
            tensor<int32, [3]> v_481_perm_0 = const()[name = tensor<string, []>("v_481_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89302144)))];
            tensor<fp16, [1, 197, 64]> q_801_cast_fp16 = transpose(perm = q_801_perm_0, x = var_9757_cast_fp16)[name = tensor<string, []>("transpose_232")];
            tensor<fp16, [1, 197, 64]> q_803_cast_fp16 = add(x = q_801_cast_fp16, y = model_blocks_6_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_803_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89302336)))];
            tensor<fp16, [1, 197, 64]> v_481_cast_fp16 = transpose(perm = v_481_perm_0, x = var_9773_cast_fp16)[name = tensor<string, []>("transpose_230")];
            tensor<fp16, [1, 197, 64]> v_483_cast_fp16 = add(x = v_481_cast_fp16, y = model_blocks_6_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_483_cast_fp16")];
            tensor<int32, [1]> q_805_axes_0 = const()[name = tensor<string, []>("q_805_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_805_cast_fp16 = expand_dims(axes = q_805_axes_0, x = q_803_cast_fp16)[name = tensor<string, []>("q_805_cast_fp16")];
            tensor<int32, [1]> k_483_axes_0 = const()[name = tensor<string, []>("k_483_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_481_cast_fp16 = transpose(perm = k_481_perm_0, x = var_9765_cast_fp16)[name = tensor<string, []>("transpose_231")];
            tensor<fp16, [1, 1, 197, 64]> k_483_cast_fp16 = expand_dims(axes = k_483_axes_0, x = k_481_cast_fp16)[name = tensor<string, []>("k_483_cast_fp16")];
            tensor<int32, [1]> v_485_axes_0 = const()[name = tensor<string, []>("v_485_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_485_cast_fp16 = expand_dims(axes = v_485_axes_0, x = v_483_cast_fp16)[name = tensor<string, []>("v_485_cast_fp16")];
            tensor<int32, [4]> var_9782_begin_0 = const()[name = tensor<string, []>("op_9782_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9782_end_0 = const()[name = tensor<string, []>("op_9782_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9782_end_mask_0 = const()[name = tensor<string, []>("op_9782_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9782_cast_fp16 = slice_by_index(begin = var_9782_begin_0, end = var_9782_end_0, end_mask = var_9782_end_mask_0, x = q_805_cast_fp16)[name = tensor<string, []>("op_9782_cast_fp16")];
            tensor<int32, [4]> var_9786_begin_0 = const()[name = tensor<string, []>("op_9786_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9786_end_0 = const()[name = tensor<string, []>("op_9786_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9786_end_mask_0 = const()[name = tensor<string, []>("op_9786_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9786_cast_fp16 = slice_by_index(begin = var_9786_begin_0, end = var_9786_end_0, end_mask = var_9786_end_mask_0, x = k_483_cast_fp16)[name = tensor<string, []>("op_9786_cast_fp16")];
            tensor<int32, [4]> var_9790_begin_0 = const()[name = tensor<string, []>("op_9790_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9790_end_0 = const()[name = tensor<string, []>("op_9790_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9790_end_mask_0 = const()[name = tensor<string, []>("op_9790_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9790_cast_fp16 = slice_by_index(begin = var_9790_begin_0, end = var_9790_end_0, end_mask = var_9790_end_mask_0, x = q_805_cast_fp16)[name = tensor<string, []>("op_9790_cast_fp16")];
            tensor<int32, [4]> var_9794_begin_0 = const()[name = tensor<string, []>("op_9794_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9794_end_0 = const()[name = tensor<string, []>("op_9794_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9794_end_mask_0 = const()[name = tensor<string, []>("op_9794_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9794_cast_fp16 = slice_by_index(begin = var_9794_begin_0, end = var_9794_end_0, end_mask = var_9794_end_mask_0, x = k_483_cast_fp16)[name = tensor<string, []>("op_9794_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9796_cast_fp16 = mul(x = var_9790_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9796_cast_fp16")];
            tensor<int32, [5]> var_9804 = const()[name = tensor<string, []>("op_9804"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1091_cast_fp16 = reshape(shape = var_9804, x = var_9790_cast_fp16)[name = tensor<string, []>("x_1091_cast_fp16")];
            tensor<int32, [2]> var_9806_split_sizes_0 = const()[name = tensor<string, []>("op_9806_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9806_axis_0 = const()[name = tensor<string, []>("op_9806_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9806_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9806_cast_fp16_1 = split(axis = var_9806_axis_0, split_sizes = var_9806_split_sizes_0, x = x_1091_cast_fp16)[name = tensor<string, []>("op_9806_cast_fp16")];
            tensor<int32, [1]> squeeze_320_axes_0 = const()[name = tensor<string, []>("squeeze_320_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_320_cast_fp16 = squeeze(axes = squeeze_320_axes_0, x = var_9806_cast_fp16_0)[name = tensor<string, []>("squeeze_320_cast_fp16")];
            tensor<int32, [1]> squeeze_321_axes_0 = const()[name = tensor<string, []>("squeeze_321_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_321_cast_fp16 = squeeze(axes = squeeze_321_axes_0, x = var_9806_cast_fp16_1)[name = tensor<string, []>("squeeze_321_cast_fp16")];
            tensor<fp16, []> const_1311_promoted_to_fp16 = const()[name = tensor<string, []>("const_1311_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9809_cast_fp16 = mul(x = squeeze_321_cast_fp16, y = const_1311_promoted_to_fp16)[name = tensor<string, []>("op_9809_cast_fp16")];
            tensor<int32, []> x_1093_axis_0 = const()[name = tensor<string, []>("x_1093_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1093_cast_fp16 = stack(axis = x_1093_axis_0, values = (var_9809_cast_fp16, squeeze_320_cast_fp16))[name = tensor<string, []>("x_1093_cast_fp16")];
            tensor<int32, [4]> var_9815 = const()[name = tensor<string, []>("op_9815"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9816_cast_fp16 = reshape(shape = var_9815, x = x_1093_cast_fp16)[name = tensor<string, []>("op_9816_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9817_cast_fp16 = mul(x = var_9816_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9817_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_161_cast_fp16 = add(x = var_9796_cast_fp16, y = var_9817_cast_fp16)[name = tensor<string, []>("q_patches_161_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9819_cast_fp16 = mul(x = var_9794_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9819_cast_fp16")];
            tensor<int32, [5]> var_9827 = const()[name = tensor<string, []>("op_9827"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1097_cast_fp16 = reshape(shape = var_9827, x = var_9794_cast_fp16)[name = tensor<string, []>("x_1097_cast_fp16")];
            tensor<int32, [2]> var_9829_split_sizes_0 = const()[name = tensor<string, []>("op_9829_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9829_axis_0 = const()[name = tensor<string, []>("op_9829_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9829_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9829_cast_fp16_1 = split(axis = var_9829_axis_0, split_sizes = var_9829_split_sizes_0, x = x_1097_cast_fp16)[name = tensor<string, []>("op_9829_cast_fp16")];
            tensor<int32, [1]> squeeze_322_axes_0 = const()[name = tensor<string, []>("squeeze_322_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_322_cast_fp16 = squeeze(axes = squeeze_322_axes_0, x = var_9829_cast_fp16_0)[name = tensor<string, []>("squeeze_322_cast_fp16")];
            tensor<int32, [1]> squeeze_323_axes_0 = const()[name = tensor<string, []>("squeeze_323_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_323_cast_fp16 = squeeze(axes = squeeze_323_axes_0, x = var_9829_cast_fp16_1)[name = tensor<string, []>("squeeze_323_cast_fp16")];
            tensor<fp16, []> const_1319_promoted_to_fp16 = const()[name = tensor<string, []>("const_1319_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9832_cast_fp16 = mul(x = squeeze_323_cast_fp16, y = const_1319_promoted_to_fp16)[name = tensor<string, []>("op_9832_cast_fp16")];
            tensor<int32, []> x_1099_axis_0 = const()[name = tensor<string, []>("x_1099_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1099_cast_fp16 = stack(axis = x_1099_axis_0, values = (var_9832_cast_fp16, squeeze_322_cast_fp16))[name = tensor<string, []>("x_1099_cast_fp16")];
            tensor<int32, [4]> var_9838 = const()[name = tensor<string, []>("op_9838"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9839_cast_fp16 = reshape(shape = var_9838, x = x_1099_cast_fp16)[name = tensor<string, []>("op_9839_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9840_cast_fp16 = mul(x = var_9839_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9840_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_161_cast_fp16 = add(x = var_9819_cast_fp16, y = var_9840_cast_fp16)[name = tensor<string, []>("k_patches_161_cast_fp16")];
            tensor<bool, []> var_9843_interleave_0 = const()[name = tensor<string, []>("op_9843_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9843_cast_fp16 = concat(axis = var_20, interleave = var_9843_interleave_0, values = (var_9782_cast_fp16, q_patches_161_cast_fp16))[name = tensor<string, []>("op_9843_cast_fp16")];
            tensor<bool, []> var_9846_interleave_0 = const()[name = tensor<string, []>("op_9846_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9846_cast_fp16 = concat(axis = var_20, interleave = var_9846_interleave_0, values = (var_9786_cast_fp16, k_patches_161_cast_fp16))[name = tensor<string, []>("op_9846_cast_fp16")];
            tensor<fp16, []> var_9848_to_fp16 = const()[name = tensor<string, []>("op_9848_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_809_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9848_to_fp16)[name = tensor<string, []>("q_809_cast_fp16")];
            tensor<bool, []> attn_321_transpose_x_1 = const()[name = tensor<string, []>("attn_321_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_321_transpose_y_1 = const()[name = tensor<string, []>("attn_321_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_321_cast_fp16 = matmul(transpose_x = attn_321_transpose_x_1, transpose_y = attn_321_transpose_y_1, x = q_809_cast_fp16, y = var_9846_cast_fp16)[name = tensor<string, []>("attn_321_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_323_cast_fp16 = softmax(axis = var_21, x = attn_321_cast_fp16)[name = tensor<string, []>("attn_323_cast_fp16")];
            tensor<bool, []> var_9853_transpose_x_0 = const()[name = tensor<string, []>("op_9853_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9853_transpose_y_0 = const()[name = tensor<string, []>("op_9853_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9853_cast_fp16 = matmul(transpose_x = var_9853_transpose_x_0, transpose_y = var_9853_transpose_y_0, x = attn_323_cast_fp16, y = v_485_cast_fp16)[name = tensor<string, []>("op_9853_cast_fp16")];
            tensor<int32, [1]> var_9854_axes_0 = const()[name = tensor<string, []>("op_9854_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9854_cast_fp16 = squeeze(axes = var_9854_axes_0, x = var_9853_cast_fp16)[name = tensor<string, []>("op_9854_cast_fp16")];
            tensor<string, []> var_9860_pad_type_0 = const()[name = tensor<string, []>("op_9860_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9860_strides_0 = const()[name = tensor<string, []>("op_9860_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9860_pad_0 = const()[name = tensor<string, []>("op_9860_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9860_dilations_0 = const()[name = tensor<string, []>("op_9860_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9860_groups_0 = const()[name = tensor<string, []>("op_9860_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89302528)))];
            tensor<fp16, [1, 64, 1, 197]> var_9860_cast_fp16 = conv(dilations = var_9860_dilations_0, groups = var_9860_groups_0, pad = var_9860_pad_0, pad_type = var_9860_pad_type_0, strides = var_9860_strides_0, weight = model_blocks_6_attn_q_projs_9_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9860_cast_fp16")];
            tensor<int32, [1]> var_9861_axes_0 = const()[name = tensor<string, []>("op_9861_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9861_cast_fp16 = squeeze(axes = var_9861_axes_0, x = var_9860_cast_fp16)[name = tensor<string, []>("op_9861_cast_fp16")];
            tensor<int32, [3]> q_811_perm_0 = const()[name = tensor<string, []>("q_811_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9868_pad_type_0 = const()[name = tensor<string, []>("op_9868_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9868_strides_0 = const()[name = tensor<string, []>("op_9868_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9868_pad_0 = const()[name = tensor<string, []>("op_9868_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9868_dilations_0 = const()[name = tensor<string, []>("op_9868_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9868_groups_0 = const()[name = tensor<string, []>("op_9868_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89400896)))];
            tensor<fp16, [1, 64, 1, 197]> var_9868_cast_fp16 = conv(dilations = var_9868_dilations_0, groups = var_9868_groups_0, pad = var_9868_pad_0, pad_type = var_9868_pad_type_0, strides = var_9868_strides_0, weight = model_blocks_6_attn_k_projs_9_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9868_cast_fp16")];
            tensor<int32, [1]> var_9869_axes_0 = const()[name = tensor<string, []>("op_9869_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9869_cast_fp16 = squeeze(axes = var_9869_axes_0, x = var_9868_cast_fp16)[name = tensor<string, []>("op_9869_cast_fp16")];
            tensor<int32, [3]> k_487_perm_0 = const()[name = tensor<string, []>("k_487_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9876_pad_type_0 = const()[name = tensor<string, []>("op_9876_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9876_strides_0 = const()[name = tensor<string, []>("op_9876_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9876_pad_0 = const()[name = tensor<string, []>("op_9876_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9876_dilations_0 = const()[name = tensor<string, []>("op_9876_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9876_groups_0 = const()[name = tensor<string, []>("op_9876_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89499264)))];
            tensor<fp16, [1, 64, 1, 197]> var_9876_cast_fp16 = conv(dilations = var_9876_dilations_0, groups = var_9876_groups_0, pad = var_9876_pad_0, pad_type = var_9876_pad_type_0, strides = var_9876_strides_0, weight = model_blocks_6_attn_v_projs_9_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9876_cast_fp16")];
            tensor<int32, [1]> var_9877_axes_0 = const()[name = tensor<string, []>("op_9877_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9877_cast_fp16 = squeeze(axes = var_9877_axes_0, x = var_9876_cast_fp16)[name = tensor<string, []>("op_9877_cast_fp16")];
            tensor<int32, [3]> v_487_perm_0 = const()[name = tensor<string, []>("v_487_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89597632)))];
            tensor<fp16, [1, 197, 64]> q_811_cast_fp16 = transpose(perm = q_811_perm_0, x = var_9861_cast_fp16)[name = tensor<string, []>("transpose_229")];
            tensor<fp16, [1, 197, 64]> q_813_cast_fp16 = add(x = q_811_cast_fp16, y = model_blocks_6_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_813_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89597824)))];
            tensor<fp16, [1, 197, 64]> v_487_cast_fp16 = transpose(perm = v_487_perm_0, x = var_9877_cast_fp16)[name = tensor<string, []>("transpose_227")];
            tensor<fp16, [1, 197, 64]> v_489_cast_fp16 = add(x = v_487_cast_fp16, y = model_blocks_6_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_489_cast_fp16")];
            tensor<int32, [1]> q_815_axes_0 = const()[name = tensor<string, []>("q_815_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_815_cast_fp16 = expand_dims(axes = q_815_axes_0, x = q_813_cast_fp16)[name = tensor<string, []>("q_815_cast_fp16")];
            tensor<int32, [1]> k_489_axes_0 = const()[name = tensor<string, []>("k_489_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_487_cast_fp16 = transpose(perm = k_487_perm_0, x = var_9869_cast_fp16)[name = tensor<string, []>("transpose_228")];
            tensor<fp16, [1, 1, 197, 64]> k_489_cast_fp16 = expand_dims(axes = k_489_axes_0, x = k_487_cast_fp16)[name = tensor<string, []>("k_489_cast_fp16")];
            tensor<int32, [1]> v_491_axes_0 = const()[name = tensor<string, []>("v_491_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_491_cast_fp16 = expand_dims(axes = v_491_axes_0, x = v_489_cast_fp16)[name = tensor<string, []>("v_491_cast_fp16")];
            tensor<int32, [4]> var_9886_begin_0 = const()[name = tensor<string, []>("op_9886_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9886_end_0 = const()[name = tensor<string, []>("op_9886_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9886_end_mask_0 = const()[name = tensor<string, []>("op_9886_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9886_cast_fp16 = slice_by_index(begin = var_9886_begin_0, end = var_9886_end_0, end_mask = var_9886_end_mask_0, x = q_815_cast_fp16)[name = tensor<string, []>("op_9886_cast_fp16")];
            tensor<int32, [4]> var_9890_begin_0 = const()[name = tensor<string, []>("op_9890_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9890_end_0 = const()[name = tensor<string, []>("op_9890_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9890_end_mask_0 = const()[name = tensor<string, []>("op_9890_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9890_cast_fp16 = slice_by_index(begin = var_9890_begin_0, end = var_9890_end_0, end_mask = var_9890_end_mask_0, x = k_489_cast_fp16)[name = tensor<string, []>("op_9890_cast_fp16")];
            tensor<int32, [4]> var_9894_begin_0 = const()[name = tensor<string, []>("op_9894_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9894_end_0 = const()[name = tensor<string, []>("op_9894_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9894_end_mask_0 = const()[name = tensor<string, []>("op_9894_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9894_cast_fp16 = slice_by_index(begin = var_9894_begin_0, end = var_9894_end_0, end_mask = var_9894_end_mask_0, x = q_815_cast_fp16)[name = tensor<string, []>("op_9894_cast_fp16")];
            tensor<int32, [4]> var_9898_begin_0 = const()[name = tensor<string, []>("op_9898_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9898_end_0 = const()[name = tensor<string, []>("op_9898_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9898_end_mask_0 = const()[name = tensor<string, []>("op_9898_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9898_cast_fp16 = slice_by_index(begin = var_9898_begin_0, end = var_9898_end_0, end_mask = var_9898_end_mask_0, x = k_489_cast_fp16)[name = tensor<string, []>("op_9898_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9900_cast_fp16 = mul(x = var_9894_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9900_cast_fp16")];
            tensor<int32, [5]> var_9908 = const()[name = tensor<string, []>("op_9908"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1103_cast_fp16 = reshape(shape = var_9908, x = var_9894_cast_fp16)[name = tensor<string, []>("x_1103_cast_fp16")];
            tensor<int32, [2]> var_9910_split_sizes_0 = const()[name = tensor<string, []>("op_9910_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9910_axis_0 = const()[name = tensor<string, []>("op_9910_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9910_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9910_cast_fp16_1 = split(axis = var_9910_axis_0, split_sizes = var_9910_split_sizes_0, x = x_1103_cast_fp16)[name = tensor<string, []>("op_9910_cast_fp16")];
            tensor<int32, [1]> squeeze_324_axes_0 = const()[name = tensor<string, []>("squeeze_324_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_324_cast_fp16 = squeeze(axes = squeeze_324_axes_0, x = var_9910_cast_fp16_0)[name = tensor<string, []>("squeeze_324_cast_fp16")];
            tensor<int32, [1]> squeeze_325_axes_0 = const()[name = tensor<string, []>("squeeze_325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_325_cast_fp16 = squeeze(axes = squeeze_325_axes_0, x = var_9910_cast_fp16_1)[name = tensor<string, []>("squeeze_325_cast_fp16")];
            tensor<fp16, []> const_1327_promoted_to_fp16 = const()[name = tensor<string, []>("const_1327_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9913_cast_fp16 = mul(x = squeeze_325_cast_fp16, y = const_1327_promoted_to_fp16)[name = tensor<string, []>("op_9913_cast_fp16")];
            tensor<int32, []> x_1105_axis_0 = const()[name = tensor<string, []>("x_1105_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1105_cast_fp16 = stack(axis = x_1105_axis_0, values = (var_9913_cast_fp16, squeeze_324_cast_fp16))[name = tensor<string, []>("x_1105_cast_fp16")];
            tensor<int32, [4]> var_9919 = const()[name = tensor<string, []>("op_9919"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9920_cast_fp16 = reshape(shape = var_9919, x = x_1105_cast_fp16)[name = tensor<string, []>("op_9920_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9921_cast_fp16 = mul(x = var_9920_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9921_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_163_cast_fp16 = add(x = var_9900_cast_fp16, y = var_9921_cast_fp16)[name = tensor<string, []>("q_patches_163_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9923_cast_fp16 = mul(x = var_9898_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_9923_cast_fp16")];
            tensor<int32, [5]> var_9931 = const()[name = tensor<string, []>("op_9931"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1109_cast_fp16 = reshape(shape = var_9931, x = var_9898_cast_fp16)[name = tensor<string, []>("x_1109_cast_fp16")];
            tensor<int32, [2]> var_9933_split_sizes_0 = const()[name = tensor<string, []>("op_9933_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9933_axis_0 = const()[name = tensor<string, []>("op_9933_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_9933_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_9933_cast_fp16_1 = split(axis = var_9933_axis_0, split_sizes = var_9933_split_sizes_0, x = x_1109_cast_fp16)[name = tensor<string, []>("op_9933_cast_fp16")];
            tensor<int32, [1]> squeeze_326_axes_0 = const()[name = tensor<string, []>("squeeze_326_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_326_cast_fp16 = squeeze(axes = squeeze_326_axes_0, x = var_9933_cast_fp16_0)[name = tensor<string, []>("squeeze_326_cast_fp16")];
            tensor<int32, [1]> squeeze_327_axes_0 = const()[name = tensor<string, []>("squeeze_327_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_327_cast_fp16 = squeeze(axes = squeeze_327_axes_0, x = var_9933_cast_fp16_1)[name = tensor<string, []>("squeeze_327_cast_fp16")];
            tensor<fp16, []> const_1335_promoted_to_fp16 = const()[name = tensor<string, []>("const_1335_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_9936_cast_fp16 = mul(x = squeeze_327_cast_fp16, y = const_1335_promoted_to_fp16)[name = tensor<string, []>("op_9936_cast_fp16")];
            tensor<int32, []> x_1111_axis_0 = const()[name = tensor<string, []>("x_1111_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1111_cast_fp16 = stack(axis = x_1111_axis_0, values = (var_9936_cast_fp16, squeeze_326_cast_fp16))[name = tensor<string, []>("x_1111_cast_fp16")];
            tensor<int32, [4]> var_9942 = const()[name = tensor<string, []>("op_9942"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_9943_cast_fp16 = reshape(shape = var_9942, x = x_1111_cast_fp16)[name = tensor<string, []>("op_9943_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_9944_cast_fp16 = mul(x = var_9943_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_9944_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_163_cast_fp16 = add(x = var_9923_cast_fp16, y = var_9944_cast_fp16)[name = tensor<string, []>("k_patches_163_cast_fp16")];
            tensor<bool, []> var_9947_interleave_0 = const()[name = tensor<string, []>("op_9947_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9947_cast_fp16 = concat(axis = var_20, interleave = var_9947_interleave_0, values = (var_9886_cast_fp16, q_patches_163_cast_fp16))[name = tensor<string, []>("op_9947_cast_fp16")];
            tensor<bool, []> var_9950_interleave_0 = const()[name = tensor<string, []>("op_9950_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9950_cast_fp16 = concat(axis = var_20, interleave = var_9950_interleave_0, values = (var_9890_cast_fp16, k_patches_163_cast_fp16))[name = tensor<string, []>("op_9950_cast_fp16")];
            tensor<fp16, []> var_9952_to_fp16 = const()[name = tensor<string, []>("op_9952_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_819_cast_fp16 = mul(x = var_9947_cast_fp16, y = var_9952_to_fp16)[name = tensor<string, []>("q_819_cast_fp16")];
            tensor<bool, []> attn_325_transpose_x_1 = const()[name = tensor<string, []>("attn_325_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_325_transpose_y_1 = const()[name = tensor<string, []>("attn_325_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_325_cast_fp16 = matmul(transpose_x = attn_325_transpose_x_1, transpose_y = attn_325_transpose_y_1, x = q_819_cast_fp16, y = var_9950_cast_fp16)[name = tensor<string, []>("attn_325_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_327_cast_fp16 = softmax(axis = var_21, x = attn_325_cast_fp16)[name = tensor<string, []>("attn_327_cast_fp16")];
            tensor<bool, []> var_9957_transpose_x_0 = const()[name = tensor<string, []>("op_9957_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_9957_transpose_y_0 = const()[name = tensor<string, []>("op_9957_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_9957_cast_fp16 = matmul(transpose_x = var_9957_transpose_x_0, transpose_y = var_9957_transpose_y_0, x = attn_327_cast_fp16, y = v_491_cast_fp16)[name = tensor<string, []>("op_9957_cast_fp16")];
            tensor<int32, [1]> var_9958_axes_0 = const()[name = tensor<string, []>("op_9958_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_9958_cast_fp16 = squeeze(axes = var_9958_axes_0, x = var_9957_cast_fp16)[name = tensor<string, []>("op_9958_cast_fp16")];
            tensor<string, []> var_9964_pad_type_0 = const()[name = tensor<string, []>("op_9964_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9964_strides_0 = const()[name = tensor<string, []>("op_9964_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9964_pad_0 = const()[name = tensor<string, []>("op_9964_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9964_dilations_0 = const()[name = tensor<string, []>("op_9964_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9964_groups_0 = const()[name = tensor<string, []>("op_9964_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89598016)))];
            tensor<fp16, [1, 64, 1, 197]> var_9964_cast_fp16 = conv(dilations = var_9964_dilations_0, groups = var_9964_groups_0, pad = var_9964_pad_0, pad_type = var_9964_pad_type_0, strides = var_9964_strides_0, weight = model_blocks_6_attn_q_projs_10_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9964_cast_fp16")];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = tensor<string, []>("op_9965_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9965_cast_fp16 = squeeze(axes = var_9965_axes_0, x = var_9964_cast_fp16)[name = tensor<string, []>("op_9965_cast_fp16")];
            tensor<int32, [3]> q_821_perm_0 = const()[name = tensor<string, []>("q_821_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9972_pad_type_0 = const()[name = tensor<string, []>("op_9972_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9972_strides_0 = const()[name = tensor<string, []>("op_9972_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9972_pad_0 = const()[name = tensor<string, []>("op_9972_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9972_dilations_0 = const()[name = tensor<string, []>("op_9972_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9972_groups_0 = const()[name = tensor<string, []>("op_9972_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89696384)))];
            tensor<fp16, [1, 64, 1, 197]> var_9972_cast_fp16 = conv(dilations = var_9972_dilations_0, groups = var_9972_groups_0, pad = var_9972_pad_0, pad_type = var_9972_pad_type_0, strides = var_9972_strides_0, weight = model_blocks_6_attn_k_projs_10_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9972_cast_fp16")];
            tensor<int32, [1]> var_9973_axes_0 = const()[name = tensor<string, []>("op_9973_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9973_cast_fp16 = squeeze(axes = var_9973_axes_0, x = var_9972_cast_fp16)[name = tensor<string, []>("op_9973_cast_fp16")];
            tensor<int32, [3]> k_493_perm_0 = const()[name = tensor<string, []>("k_493_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_9980_pad_type_0 = const()[name = tensor<string, []>("op_9980_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_9980_strides_0 = const()[name = tensor<string, []>("op_9980_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9980_pad_0 = const()[name = tensor<string, []>("op_9980_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9980_dilations_0 = const()[name = tensor<string, []>("op_9980_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_9980_groups_0 = const()[name = tensor<string, []>("op_9980_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89794752)))];
            tensor<fp16, [1, 64, 1, 197]> var_9980_cast_fp16 = conv(dilations = var_9980_dilations_0, groups = var_9980_groups_0, pad = var_9980_pad_0, pad_type = var_9980_pad_type_0, strides = var_9980_strides_0, weight = model_blocks_6_attn_v_projs_10_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_9980_cast_fp16")];
            tensor<int32, [1]> var_9981_axes_0 = const()[name = tensor<string, []>("op_9981_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_9981_cast_fp16 = squeeze(axes = var_9981_axes_0, x = var_9980_cast_fp16)[name = tensor<string, []>("op_9981_cast_fp16")];
            tensor<int32, [3]> v_493_perm_0 = const()[name = tensor<string, []>("v_493_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89893120)))];
            tensor<fp16, [1, 197, 64]> q_821_cast_fp16 = transpose(perm = q_821_perm_0, x = var_9965_cast_fp16)[name = tensor<string, []>("transpose_226")];
            tensor<fp16, [1, 197, 64]> q_823_cast_fp16 = add(x = q_821_cast_fp16, y = model_blocks_6_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_823_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89893312)))];
            tensor<fp16, [1, 197, 64]> v_493_cast_fp16 = transpose(perm = v_493_perm_0, x = var_9981_cast_fp16)[name = tensor<string, []>("transpose_224")];
            tensor<fp16, [1, 197, 64]> v_495_cast_fp16 = add(x = v_493_cast_fp16, y = model_blocks_6_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_495_cast_fp16")];
            tensor<int32, [1]> q_825_axes_0 = const()[name = tensor<string, []>("q_825_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_825_cast_fp16 = expand_dims(axes = q_825_axes_0, x = q_823_cast_fp16)[name = tensor<string, []>("q_825_cast_fp16")];
            tensor<int32, [1]> k_495_axes_0 = const()[name = tensor<string, []>("k_495_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_493_cast_fp16 = transpose(perm = k_493_perm_0, x = var_9973_cast_fp16)[name = tensor<string, []>("transpose_225")];
            tensor<fp16, [1, 1, 197, 64]> k_495_cast_fp16 = expand_dims(axes = k_495_axes_0, x = k_493_cast_fp16)[name = tensor<string, []>("k_495_cast_fp16")];
            tensor<int32, [1]> v_497_axes_0 = const()[name = tensor<string, []>("v_497_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_497_cast_fp16 = expand_dims(axes = v_497_axes_0, x = v_495_cast_fp16)[name = tensor<string, []>("v_497_cast_fp16")];
            tensor<int32, [4]> var_9990_begin_0 = const()[name = tensor<string, []>("op_9990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9990_end_0 = const()[name = tensor<string, []>("op_9990_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9990_end_mask_0 = const()[name = tensor<string, []>("op_9990_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9990_cast_fp16 = slice_by_index(begin = var_9990_begin_0, end = var_9990_end_0, end_mask = var_9990_end_mask_0, x = q_825_cast_fp16)[name = tensor<string, []>("op_9990_cast_fp16")];
            tensor<int32, [4]> var_9994_begin_0 = const()[name = tensor<string, []>("op_9994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9994_end_0 = const()[name = tensor<string, []>("op_9994_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_9994_end_mask_0 = const()[name = tensor<string, []>("op_9994_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_9994_cast_fp16 = slice_by_index(begin = var_9994_begin_0, end = var_9994_end_0, end_mask = var_9994_end_mask_0, x = k_495_cast_fp16)[name = tensor<string, []>("op_9994_cast_fp16")];
            tensor<int32, [4]> var_9998_begin_0 = const()[name = tensor<string, []>("op_9998_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9998_end_0 = const()[name = tensor<string, []>("op_9998_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_9998_end_mask_0 = const()[name = tensor<string, []>("op_9998_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_9998_cast_fp16 = slice_by_index(begin = var_9998_begin_0, end = var_9998_end_0, end_mask = var_9998_end_mask_0, x = q_825_cast_fp16)[name = tensor<string, []>("op_9998_cast_fp16")];
            tensor<int32, [4]> var_10002_begin_0 = const()[name = tensor<string, []>("op_10002_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10002_end_0 = const()[name = tensor<string, []>("op_10002_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10002_end_mask_0 = const()[name = tensor<string, []>("op_10002_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10002_cast_fp16 = slice_by_index(begin = var_10002_begin_0, end = var_10002_end_0, end_mask = var_10002_end_mask_0, x = k_495_cast_fp16)[name = tensor<string, []>("op_10002_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10004_cast_fp16 = mul(x = var_9998_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10004_cast_fp16")];
            tensor<int32, [5]> var_10012 = const()[name = tensor<string, []>("op_10012"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1115_cast_fp16 = reshape(shape = var_10012, x = var_9998_cast_fp16)[name = tensor<string, []>("x_1115_cast_fp16")];
            tensor<int32, [2]> var_10014_split_sizes_0 = const()[name = tensor<string, []>("op_10014_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10014_axis_0 = const()[name = tensor<string, []>("op_10014_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10014_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10014_cast_fp16_1 = split(axis = var_10014_axis_0, split_sizes = var_10014_split_sizes_0, x = x_1115_cast_fp16)[name = tensor<string, []>("op_10014_cast_fp16")];
            tensor<int32, [1]> squeeze_328_axes_0 = const()[name = tensor<string, []>("squeeze_328_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_328_cast_fp16 = squeeze(axes = squeeze_328_axes_0, x = var_10014_cast_fp16_0)[name = tensor<string, []>("squeeze_328_cast_fp16")];
            tensor<int32, [1]> squeeze_329_axes_0 = const()[name = tensor<string, []>("squeeze_329_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_329_cast_fp16 = squeeze(axes = squeeze_329_axes_0, x = var_10014_cast_fp16_1)[name = tensor<string, []>("squeeze_329_cast_fp16")];
            tensor<fp16, []> const_1343_promoted_to_fp16 = const()[name = tensor<string, []>("const_1343_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10017_cast_fp16 = mul(x = squeeze_329_cast_fp16, y = const_1343_promoted_to_fp16)[name = tensor<string, []>("op_10017_cast_fp16")];
            tensor<int32, []> x_1117_axis_0 = const()[name = tensor<string, []>("x_1117_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1117_cast_fp16 = stack(axis = x_1117_axis_0, values = (var_10017_cast_fp16, squeeze_328_cast_fp16))[name = tensor<string, []>("x_1117_cast_fp16")];
            tensor<int32, [4]> var_10023 = const()[name = tensor<string, []>("op_10023"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10024_cast_fp16 = reshape(shape = var_10023, x = x_1117_cast_fp16)[name = tensor<string, []>("op_10024_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10025_cast_fp16 = mul(x = var_10024_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10025_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_165_cast_fp16 = add(x = var_10004_cast_fp16, y = var_10025_cast_fp16)[name = tensor<string, []>("q_patches_165_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10027_cast_fp16 = mul(x = var_10002_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10027_cast_fp16")];
            tensor<int32, [5]> var_10035 = const()[name = tensor<string, []>("op_10035"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1121_cast_fp16 = reshape(shape = var_10035, x = var_10002_cast_fp16)[name = tensor<string, []>("x_1121_cast_fp16")];
            tensor<int32, [2]> var_10037_split_sizes_0 = const()[name = tensor<string, []>("op_10037_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10037_axis_0 = const()[name = tensor<string, []>("op_10037_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10037_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10037_cast_fp16_1 = split(axis = var_10037_axis_0, split_sizes = var_10037_split_sizes_0, x = x_1121_cast_fp16)[name = tensor<string, []>("op_10037_cast_fp16")];
            tensor<int32, [1]> squeeze_330_axes_0 = const()[name = tensor<string, []>("squeeze_330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_330_cast_fp16 = squeeze(axes = squeeze_330_axes_0, x = var_10037_cast_fp16_0)[name = tensor<string, []>("squeeze_330_cast_fp16")];
            tensor<int32, [1]> squeeze_331_axes_0 = const()[name = tensor<string, []>("squeeze_331_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_331_cast_fp16 = squeeze(axes = squeeze_331_axes_0, x = var_10037_cast_fp16_1)[name = tensor<string, []>("squeeze_331_cast_fp16")];
            tensor<fp16, []> const_1351_promoted_to_fp16 = const()[name = tensor<string, []>("const_1351_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10040_cast_fp16 = mul(x = squeeze_331_cast_fp16, y = const_1351_promoted_to_fp16)[name = tensor<string, []>("op_10040_cast_fp16")];
            tensor<int32, []> x_1123_axis_0 = const()[name = tensor<string, []>("x_1123_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1123_cast_fp16 = stack(axis = x_1123_axis_0, values = (var_10040_cast_fp16, squeeze_330_cast_fp16))[name = tensor<string, []>("x_1123_cast_fp16")];
            tensor<int32, [4]> var_10046 = const()[name = tensor<string, []>("op_10046"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10047_cast_fp16 = reshape(shape = var_10046, x = x_1123_cast_fp16)[name = tensor<string, []>("op_10047_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10048_cast_fp16 = mul(x = var_10047_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10048_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_165_cast_fp16 = add(x = var_10027_cast_fp16, y = var_10048_cast_fp16)[name = tensor<string, []>("k_patches_165_cast_fp16")];
            tensor<bool, []> var_10051_interleave_0 = const()[name = tensor<string, []>("op_10051_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10051_cast_fp16 = concat(axis = var_20, interleave = var_10051_interleave_0, values = (var_9990_cast_fp16, q_patches_165_cast_fp16))[name = tensor<string, []>("op_10051_cast_fp16")];
            tensor<bool, []> var_10054_interleave_0 = const()[name = tensor<string, []>("op_10054_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10054_cast_fp16 = concat(axis = var_20, interleave = var_10054_interleave_0, values = (var_9994_cast_fp16, k_patches_165_cast_fp16))[name = tensor<string, []>("op_10054_cast_fp16")];
            tensor<fp16, []> var_10056_to_fp16 = const()[name = tensor<string, []>("op_10056_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_829_cast_fp16 = mul(x = var_10051_cast_fp16, y = var_10056_to_fp16)[name = tensor<string, []>("q_829_cast_fp16")];
            tensor<bool, []> attn_329_transpose_x_1 = const()[name = tensor<string, []>("attn_329_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_329_transpose_y_1 = const()[name = tensor<string, []>("attn_329_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_329_cast_fp16 = matmul(transpose_x = attn_329_transpose_x_1, transpose_y = attn_329_transpose_y_1, x = q_829_cast_fp16, y = var_10054_cast_fp16)[name = tensor<string, []>("attn_329_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_331_cast_fp16 = softmax(axis = var_21, x = attn_329_cast_fp16)[name = tensor<string, []>("attn_331_cast_fp16")];
            tensor<bool, []> var_10061_transpose_x_0 = const()[name = tensor<string, []>("op_10061_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10061_transpose_y_0 = const()[name = tensor<string, []>("op_10061_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10061_cast_fp16 = matmul(transpose_x = var_10061_transpose_x_0, transpose_y = var_10061_transpose_y_0, x = attn_331_cast_fp16, y = v_497_cast_fp16)[name = tensor<string, []>("op_10061_cast_fp16")];
            tensor<int32, [1]> var_10062_axes_0 = const()[name = tensor<string, []>("op_10062_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10062_cast_fp16 = squeeze(axes = var_10062_axes_0, x = var_10061_cast_fp16)[name = tensor<string, []>("op_10062_cast_fp16")];
            tensor<string, []> var_10068_pad_type_0 = const()[name = tensor<string, []>("op_10068_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10068_strides_0 = const()[name = tensor<string, []>("op_10068_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10068_pad_0 = const()[name = tensor<string, []>("op_10068_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10068_dilations_0 = const()[name = tensor<string, []>("op_10068_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10068_groups_0 = const()[name = tensor<string, []>("op_10068_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89893504)))];
            tensor<fp16, [1, 64, 1, 197]> var_10068_cast_fp16 = conv(dilations = var_10068_dilations_0, groups = var_10068_groups_0, pad = var_10068_pad_0, pad_type = var_10068_pad_type_0, strides = var_10068_strides_0, weight = model_blocks_6_attn_q_projs_11_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_10068_cast_fp16")];
            tensor<int32, [1]> var_10069_axes_0 = const()[name = tensor<string, []>("op_10069_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10069_cast_fp16 = squeeze(axes = var_10069_axes_0, x = var_10068_cast_fp16)[name = tensor<string, []>("op_10069_cast_fp16")];
            tensor<int32, [3]> q_831_perm_0 = const()[name = tensor<string, []>("q_831_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10076_pad_type_0 = const()[name = tensor<string, []>("op_10076_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10076_strides_0 = const()[name = tensor<string, []>("op_10076_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10076_pad_0 = const()[name = tensor<string, []>("op_10076_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10076_dilations_0 = const()[name = tensor<string, []>("op_10076_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10076_groups_0 = const()[name = tensor<string, []>("op_10076_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89991872)))];
            tensor<fp16, [1, 64, 1, 197]> var_10076_cast_fp16 = conv(dilations = var_10076_dilations_0, groups = var_10076_groups_0, pad = var_10076_pad_0, pad_type = var_10076_pad_type_0, strides = var_10076_strides_0, weight = model_blocks_6_attn_k_projs_11_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_10076_cast_fp16")];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = tensor<string, []>("op_10077_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10077_cast_fp16 = squeeze(axes = var_10077_axes_0, x = var_10076_cast_fp16)[name = tensor<string, []>("op_10077_cast_fp16")];
            tensor<int32, [3]> k_499_perm_0 = const()[name = tensor<string, []>("k_499_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10084_pad_type_0 = const()[name = tensor<string, []>("op_10084_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10084_strides_0 = const()[name = tensor<string, []>("op_10084_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10084_pad_0 = const()[name = tensor<string, []>("op_10084_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10084_dilations_0 = const()[name = tensor<string, []>("op_10084_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10084_groups_0 = const()[name = tensor<string, []>("op_10084_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_6_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90090240)))];
            tensor<fp16, [1, 64, 1, 197]> var_10084_cast_fp16 = conv(dilations = var_10084_dilations_0, groups = var_10084_groups_0, pad = var_10084_pad_0, pad_type = var_10084_pad_type_0, strides = var_10084_strides_0, weight = model_blocks_6_attn_v_projs_11_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("op_10084_cast_fp16")];
            tensor<int32, [1]> var_10085_axes_0 = const()[name = tensor<string, []>("op_10085_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10085_cast_fp16 = squeeze(axes = var_10085_axes_0, x = var_10084_cast_fp16)[name = tensor<string, []>("op_10085_cast_fp16")];
            tensor<int32, [3]> v_499_perm_0 = const()[name = tensor<string, []>("v_499_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_6_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90188608)))];
            tensor<fp16, [1, 197, 64]> q_831_cast_fp16 = transpose(perm = q_831_perm_0, x = var_10069_cast_fp16)[name = tensor<string, []>("transpose_223")];
            tensor<fp16, [1, 197, 64]> q_833_cast_fp16 = add(x = q_831_cast_fp16, y = model_blocks_6_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_833_cast_fp16")];
            tensor<fp16, [64]> model_blocks_6_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90188800)))];
            tensor<fp16, [1, 197, 64]> v_499_cast_fp16 = transpose(perm = v_499_perm_0, x = var_10085_cast_fp16)[name = tensor<string, []>("transpose_221")];
            tensor<fp16, [1, 197, 64]> v_501_cast_fp16 = add(x = v_499_cast_fp16, y = model_blocks_6_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_501_cast_fp16")];
            tensor<int32, [1]> q_835_axes_0 = const()[name = tensor<string, []>("q_835_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_835_cast_fp16 = expand_dims(axes = q_835_axes_0, x = q_833_cast_fp16)[name = tensor<string, []>("q_835_cast_fp16")];
            tensor<int32, [1]> k_501_axes_0 = const()[name = tensor<string, []>("k_501_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_499_cast_fp16 = transpose(perm = k_499_perm_0, x = var_10077_cast_fp16)[name = tensor<string, []>("transpose_222")];
            tensor<fp16, [1, 1, 197, 64]> k_501_cast_fp16 = expand_dims(axes = k_501_axes_0, x = k_499_cast_fp16)[name = tensor<string, []>("k_501_cast_fp16")];
            tensor<int32, [1]> v_503_axes_0 = const()[name = tensor<string, []>("v_503_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_503_cast_fp16 = expand_dims(axes = v_503_axes_0, x = v_501_cast_fp16)[name = tensor<string, []>("v_503_cast_fp16")];
            tensor<int32, [4]> var_10094_begin_0 = const()[name = tensor<string, []>("op_10094_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10094_end_0 = const()[name = tensor<string, []>("op_10094_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10094_end_mask_0 = const()[name = tensor<string, []>("op_10094_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10094_cast_fp16 = slice_by_index(begin = var_10094_begin_0, end = var_10094_end_0, end_mask = var_10094_end_mask_0, x = q_835_cast_fp16)[name = tensor<string, []>("op_10094_cast_fp16")];
            tensor<int32, [4]> var_10098_begin_0 = const()[name = tensor<string, []>("op_10098_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10098_end_0 = const()[name = tensor<string, []>("op_10098_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10098_end_mask_0 = const()[name = tensor<string, []>("op_10098_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10098_cast_fp16 = slice_by_index(begin = var_10098_begin_0, end = var_10098_end_0, end_mask = var_10098_end_mask_0, x = k_501_cast_fp16)[name = tensor<string, []>("op_10098_cast_fp16")];
            tensor<int32, [4]> var_10102_begin_0 = const()[name = tensor<string, []>("op_10102_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10102_end_0 = const()[name = tensor<string, []>("op_10102_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10102_end_mask_0 = const()[name = tensor<string, []>("op_10102_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10102_cast_fp16 = slice_by_index(begin = var_10102_begin_0, end = var_10102_end_0, end_mask = var_10102_end_mask_0, x = q_835_cast_fp16)[name = tensor<string, []>("op_10102_cast_fp16")];
            tensor<int32, [4]> var_10106_begin_0 = const()[name = tensor<string, []>("op_10106_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10106_end_0 = const()[name = tensor<string, []>("op_10106_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10106_end_mask_0 = const()[name = tensor<string, []>("op_10106_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10106_cast_fp16 = slice_by_index(begin = var_10106_begin_0, end = var_10106_end_0, end_mask = var_10106_end_mask_0, x = k_501_cast_fp16)[name = tensor<string, []>("op_10106_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10108_cast_fp16 = mul(x = var_10102_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10108_cast_fp16")];
            tensor<int32, [5]> var_10116 = const()[name = tensor<string, []>("op_10116"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1127_cast_fp16 = reshape(shape = var_10116, x = var_10102_cast_fp16)[name = tensor<string, []>("x_1127_cast_fp16")];
            tensor<int32, [2]> var_10118_split_sizes_0 = const()[name = tensor<string, []>("op_10118_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10118_axis_0 = const()[name = tensor<string, []>("op_10118_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10118_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10118_cast_fp16_1 = split(axis = var_10118_axis_0, split_sizes = var_10118_split_sizes_0, x = x_1127_cast_fp16)[name = tensor<string, []>("op_10118_cast_fp16")];
            tensor<int32, [1]> squeeze_332_axes_0 = const()[name = tensor<string, []>("squeeze_332_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_332_cast_fp16 = squeeze(axes = squeeze_332_axes_0, x = var_10118_cast_fp16_0)[name = tensor<string, []>("squeeze_332_cast_fp16")];
            tensor<int32, [1]> squeeze_333_axes_0 = const()[name = tensor<string, []>("squeeze_333_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_333_cast_fp16 = squeeze(axes = squeeze_333_axes_0, x = var_10118_cast_fp16_1)[name = tensor<string, []>("squeeze_333_cast_fp16")];
            tensor<fp16, []> const_1359_promoted_to_fp16 = const()[name = tensor<string, []>("const_1359_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10121_cast_fp16 = mul(x = squeeze_333_cast_fp16, y = const_1359_promoted_to_fp16)[name = tensor<string, []>("op_10121_cast_fp16")];
            tensor<int32, []> x_1129_axis_0 = const()[name = tensor<string, []>("x_1129_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1129_cast_fp16 = stack(axis = x_1129_axis_0, values = (var_10121_cast_fp16, squeeze_332_cast_fp16))[name = tensor<string, []>("x_1129_cast_fp16")];
            tensor<int32, [4]> var_10127 = const()[name = tensor<string, []>("op_10127"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10128_cast_fp16 = reshape(shape = var_10127, x = x_1129_cast_fp16)[name = tensor<string, []>("op_10128_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10129_cast_fp16 = mul(x = var_10128_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10129_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_167_cast_fp16 = add(x = var_10108_cast_fp16, y = var_10129_cast_fp16)[name = tensor<string, []>("q_patches_167_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10131_cast_fp16 = mul(x = var_10106_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10131_cast_fp16")];
            tensor<int32, [5]> var_10139 = const()[name = tensor<string, []>("op_10139"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1133_cast_fp16 = reshape(shape = var_10139, x = var_10106_cast_fp16)[name = tensor<string, []>("x_1133_cast_fp16")];
            tensor<int32, [2]> var_10141_split_sizes_0 = const()[name = tensor<string, []>("op_10141_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10141_axis_0 = const()[name = tensor<string, []>("op_10141_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10141_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10141_cast_fp16_1 = split(axis = var_10141_axis_0, split_sizes = var_10141_split_sizes_0, x = x_1133_cast_fp16)[name = tensor<string, []>("op_10141_cast_fp16")];
            tensor<int32, [1]> squeeze_334_axes_0 = const()[name = tensor<string, []>("squeeze_334_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_334_cast_fp16 = squeeze(axes = squeeze_334_axes_0, x = var_10141_cast_fp16_0)[name = tensor<string, []>("squeeze_334_cast_fp16")];
            tensor<int32, [1]> squeeze_335_axes_0 = const()[name = tensor<string, []>("squeeze_335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_335_cast_fp16 = squeeze(axes = squeeze_335_axes_0, x = var_10141_cast_fp16_1)[name = tensor<string, []>("squeeze_335_cast_fp16")];
            tensor<fp16, []> const_1367_promoted_to_fp16 = const()[name = tensor<string, []>("const_1367_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10144_cast_fp16 = mul(x = squeeze_335_cast_fp16, y = const_1367_promoted_to_fp16)[name = tensor<string, []>("op_10144_cast_fp16")];
            tensor<int32, []> x_1135_axis_0 = const()[name = tensor<string, []>("x_1135_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1135_cast_fp16 = stack(axis = x_1135_axis_0, values = (var_10144_cast_fp16, squeeze_334_cast_fp16))[name = tensor<string, []>("x_1135_cast_fp16")];
            tensor<int32, [4]> var_10150 = const()[name = tensor<string, []>("op_10150"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10151_cast_fp16 = reshape(shape = var_10150, x = x_1135_cast_fp16)[name = tensor<string, []>("op_10151_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10152_cast_fp16 = mul(x = var_10151_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10152_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_167_cast_fp16 = add(x = var_10131_cast_fp16, y = var_10152_cast_fp16)[name = tensor<string, []>("k_patches_167_cast_fp16")];
            tensor<bool, []> var_10155_interleave_0 = const()[name = tensor<string, []>("op_10155_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10155_cast_fp16 = concat(axis = var_20, interleave = var_10155_interleave_0, values = (var_10094_cast_fp16, q_patches_167_cast_fp16))[name = tensor<string, []>("op_10155_cast_fp16")];
            tensor<bool, []> var_10158_interleave_0 = const()[name = tensor<string, []>("op_10158_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10158_cast_fp16 = concat(axis = var_20, interleave = var_10158_interleave_0, values = (var_10098_cast_fp16, k_patches_167_cast_fp16))[name = tensor<string, []>("op_10158_cast_fp16")];
            tensor<fp16, []> var_10160_to_fp16 = const()[name = tensor<string, []>("op_10160_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_839_cast_fp16 = mul(x = var_10155_cast_fp16, y = var_10160_to_fp16)[name = tensor<string, []>("q_839_cast_fp16")];
            tensor<bool, []> attn_333_transpose_x_1 = const()[name = tensor<string, []>("attn_333_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_333_transpose_y_1 = const()[name = tensor<string, []>("attn_333_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_333_cast_fp16 = matmul(transpose_x = attn_333_transpose_x_1, transpose_y = attn_333_transpose_y_1, x = q_839_cast_fp16, y = var_10158_cast_fp16)[name = tensor<string, []>("attn_333_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_335_cast_fp16 = softmax(axis = var_21, x = attn_333_cast_fp16)[name = tensor<string, []>("attn_335_cast_fp16")];
            tensor<bool, []> var_10165_transpose_x_0 = const()[name = tensor<string, []>("op_10165_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10165_transpose_y_0 = const()[name = tensor<string, []>("op_10165_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10165_cast_fp16 = matmul(transpose_x = var_10165_transpose_x_0, transpose_y = var_10165_transpose_y_0, x = attn_335_cast_fp16, y = v_503_cast_fp16)[name = tensor<string, []>("op_10165_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_13_cast_fp16 = squeeze(axes = out_13_axes_0, x = var_10165_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<bool, []> input_113_interleave_0 = const()[name = tensor<string, []>("input_113_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_113_cast_fp16 = concat(axis = var_21, interleave = input_113_interleave_0, values = (var_9022_cast_fp16, var_9126_cast_fp16, var_9230_cast_fp16, var_9334_cast_fp16, var_9438_cast_fp16, var_9542_cast_fp16, var_9646_cast_fp16, var_9750_cast_fp16, var_9854_cast_fp16, var_9958_cast_fp16, var_10062_cast_fp16, out_13_cast_fp16))[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<int32, [1]> x_1137_axes_0 = const()[name = tensor<string, []>("x_1137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_6_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90188992)))];
            tensor<fp16, [768]> model_blocks_6_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90190592)))];
            tensor<fp16, [1, 197, 768]> x_1137_cast_fp16 = layer_norm(axes = x_1137_axes_0, beta = model_blocks_6_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_6_attn_inner_attn_ln_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("x_1137_cast_fp16")];
            tensor<int32, [4]> var_10173 = const()[name = tensor<string, []>("op_10173"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1139_cast_fp16 = reshape(shape = var_10173, x = x_1137_cast_fp16)[name = tensor<string, []>("x_1139_cast_fp16")];
            tensor<int32, [4]> input_115_perm_0 = const()[name = tensor<string, []>("input_115_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1141_pad_type_0 = const()[name = tensor<string, []>("x_1141_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1141_strides_0 = const()[name = tensor<string, []>("x_1141_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1141_pad_0 = const()[name = tensor<string, []>("x_1141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1141_dilations_0 = const()[name = tensor<string, []>("x_1141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1141_groups_0 = const()[name = tensor<string, []>("x_1141_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_6_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90192192)))];
            tensor<fp16, [768]> model_blocks_6_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91371904)))];
            tensor<fp16, [1, 768, 1, 197]> input_115_cast_fp16 = transpose(perm = input_115_perm_0, x = x_1139_cast_fp16)[name = tensor<string, []>("transpose_220")];
            tensor<fp16, [1, 768, 1, 197]> x_1141_cast_fp16 = conv(bias = model_blocks_6_attn_proj_bias_to_fp16, dilations = x_1141_dilations_0, groups = x_1141_groups_0, pad = x_1141_pad_0, pad_type = x_1141_pad_type_0, strides = x_1141_strides_0, weight = model_blocks_6_attn_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("x_1141_cast_fp16")];
            tensor<int32, [4]> x_1143_perm_0 = const()[name = tensor<string, []>("x_1143_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_10184 = const()[name = tensor<string, []>("op_10184"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1143_cast_fp16 = transpose(perm = x_1143_perm_0, x = x_1141_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 197, 768]> var_10185_cast_fp16 = reshape(shape = var_10184, x = x_1143_cast_fp16)[name = tensor<string, []>("op_10185_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_117_cast_fp16 = add(x = input_109_cast_fp16, y = var_10185_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<int32, [1]> x_1145_axes_0 = const()[name = tensor<string, []>("x_1145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91373504)))];
            tensor<fp16, [768]> model_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91375104)))];
            tensor<fp16, [1, 197, 768]> x_1145_cast_fp16 = layer_norm(axes = x_1145_axes_0, beta = model_blocks_6_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_6_norm2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("x_1145_cast_fp16")];
            tensor<int32, [4]> var_10197 = const()[name = tensor<string, []>("op_10197"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1147_cast_fp16 = reshape(shape = var_10197, x = x_1145_cast_fp16)[name = tensor<string, []>("x_1147_cast_fp16")];
            tensor<int32, [4]> input_119_perm_0 = const()[name = tensor<string, []>("input_119_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_121_pad_type_0 = const()[name = tensor<string, []>("input_121_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_121_strides_0 = const()[name = tensor<string, []>("input_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_121_pad_0 = const()[name = tensor<string, []>("input_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_121_dilations_0 = const()[name = tensor<string, []>("input_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_121_groups_0 = const()[name = tensor<string, []>("input_121_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_6_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91376704)))];
            tensor<fp16, [2048]> model_blocks_6_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94522496)))];
            tensor<fp16, [1, 768, 1, 197]> input_119_cast_fp16 = transpose(perm = input_119_perm_0, x = x_1147_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 2048, 1, 197]> input_121_cast_fp16 = conv(bias = model_blocks_6_mlp_w1_bias_to_fp16, dilations = input_121_dilations_0, groups = input_121_groups_0, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = input_121_strides_0, weight = model_blocks_6_mlp_w1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<string, []> x2_349_pad_type_0 = const()[name = tensor<string, []>("x2_349_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_349_strides_0 = const()[name = tensor<string, []>("x2_349_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_349_pad_0 = const()[name = tensor<string, []>("x2_349_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_349_dilations_0 = const()[name = tensor<string, []>("x2_349_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_349_groups_0 = const()[name = tensor<string, []>("x2_349_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_6_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94526656)))];
            tensor<fp16, [2048]> model_blocks_6_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97672448)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_349_cast_fp16 = conv(bias = model_blocks_6_mlp_w2_bias_to_fp16, dilations = x2_349_dilations_0, groups = x2_349_groups_0, pad = x2_349_pad_0, pad_type = x2_349_pad_type_0, strides = x2_349_strides_0, weight = model_blocks_6_mlp_w2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("x2_349_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_10214_cast_fp16 = silu(x = input_121_cast_fp16)[name = tensor<string, []>("op_10214_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_49_cast_fp16 = mul(x = var_10214_cast_fp16, y = x2_349_cast_fp16)[name = tensor<string, []>("hidden_49_cast_fp16")];
            tensor<int32, [4]> hidden_51_perm_0 = const()[name = tensor<string, []>("hidden_51_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_10217 = const()[name = tensor<string, []>("op_10217"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_51_cast_fp16 = transpose(perm = hidden_51_perm_0, x = hidden_49_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [1, 197, 2048]> input_123_cast_fp16 = reshape(shape = var_10217, x = hidden_51_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<int32, [1]> hidden_53_axes_0 = const()[name = tensor<string, []>("hidden_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_6_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97676608)))];
            tensor<fp16, [2048]> model_blocks_6_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97680768)))];
            tensor<fp16, [1, 197, 2048]> hidden_53_cast_fp16 = layer_norm(axes = hidden_53_axes_0, beta = model_blocks_6_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_6_mlp_ffn_ln_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("hidden_53_cast_fp16")];
            tensor<int32, [4]> var_10223 = const()[name = tensor<string, []>("op_10223"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_55_cast_fp16 = reshape(shape = var_10223, x = hidden_53_cast_fp16)[name = tensor<string, []>("hidden_55_cast_fp16")];
            tensor<int32, [4]> input_125_perm_0 = const()[name = tensor<string, []>("input_125_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1149_pad_type_0 = const()[name = tensor<string, []>("x_1149_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1149_strides_0 = const()[name = tensor<string, []>("x_1149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1149_pad_0 = const()[name = tensor<string, []>("x_1149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1149_dilations_0 = const()[name = tensor<string, []>("x_1149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1149_groups_0 = const()[name = tensor<string, []>("x_1149_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_6_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97684928)))];
            tensor<fp16, [768]> model_blocks_6_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_6_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100830720)))];
            tensor<fp16, [1, 2048, 1, 197]> input_125_cast_fp16 = transpose(perm = input_125_perm_0, x = hidden_55_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [1, 768, 1, 197]> x_1149_cast_fp16 = conv(bias = model_blocks_6_mlp_w3_bias_to_fp16, dilations = x_1149_dilations_0, groups = x_1149_groups_0, pad = x_1149_pad_0, pad_type = x_1149_pad_type_0, strides = x_1149_strides_0, weight = model_blocks_6_mlp_w3_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("x_1149_cast_fp16")];
            tensor<int32, [4]> x_1151_perm_0 = const()[name = tensor<string, []>("x_1151_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_10234 = const()[name = tensor<string, []>("op_10234"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1151_cast_fp16 = transpose(perm = x_1151_perm_0, x = x_1149_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 197, 768]> var_10235_cast_fp16 = reshape(shape = var_10234, x = x_1151_cast_fp16)[name = tensor<string, []>("op_10235_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_127_cast_fp16 = add(x = input_117_cast_fp16, y = var_10235_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<int32, [1]> x_1153_axes_0 = const()[name = tensor<string, []>("x_1153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100832320)))];
            tensor<fp16, [768]> model_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100833920)))];
            tensor<fp16, [1, 197, 768]> x_1153_cast_fp16 = layer_norm(axes = x_1153_axes_0, beta = model_blocks_7_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_7_norm1_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("x_1153_cast_fp16")];
            tensor<int32, [4]> var_10369 = const()[name = tensor<string, []>("op_10369"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1155_cast_fp16 = reshape(shape = var_10369, x = x_1153_cast_fp16)[name = tensor<string, []>("x_1155_cast_fp16")];
            tensor<int32, [4]> input_129_perm_0 = const()[name = tensor<string, []>("input_129_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_10377_pad_type_0 = const()[name = tensor<string, []>("op_10377_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10377_strides_0 = const()[name = tensor<string, []>("op_10377_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10377_pad_0 = const()[name = tensor<string, []>("op_10377_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10377_dilations_0 = const()[name = tensor<string, []>("op_10377_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10377_groups_0 = const()[name = tensor<string, []>("op_10377_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100835520)))];
            tensor<fp16, [1, 768, 1, 197]> input_129_cast_fp16 = transpose(perm = input_129_perm_0, x = x_1155_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [1, 64, 1, 197]> var_10377_cast_fp16 = conv(dilations = var_10377_dilations_0, groups = var_10377_groups_0, pad = var_10377_pad_0, pad_type = var_10377_pad_type_0, strides = var_10377_strides_0, weight = model_blocks_7_attn_q_projs_0_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10377_cast_fp16")];
            tensor<int32, [1]> var_10378_axes_0 = const()[name = tensor<string, []>("op_10378_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10378_cast_fp16 = squeeze(axes = var_10378_axes_0, x = var_10377_cast_fp16)[name = tensor<string, []>("op_10378_cast_fp16")];
            tensor<int32, [3]> q_841_perm_0 = const()[name = tensor<string, []>("q_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10385_pad_type_0 = const()[name = tensor<string, []>("op_10385_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10385_strides_0 = const()[name = tensor<string, []>("op_10385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10385_pad_0 = const()[name = tensor<string, []>("op_10385_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10385_dilations_0 = const()[name = tensor<string, []>("op_10385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10385_groups_0 = const()[name = tensor<string, []>("op_10385_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100933888)))];
            tensor<fp16, [1, 64, 1, 197]> var_10385_cast_fp16 = conv(dilations = var_10385_dilations_0, groups = var_10385_groups_0, pad = var_10385_pad_0, pad_type = var_10385_pad_type_0, strides = var_10385_strides_0, weight = model_blocks_7_attn_k_projs_0_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10385_cast_fp16")];
            tensor<int32, [1]> var_10386_axes_0 = const()[name = tensor<string, []>("op_10386_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10386_cast_fp16 = squeeze(axes = var_10386_axes_0, x = var_10385_cast_fp16)[name = tensor<string, []>("op_10386_cast_fp16")];
            tensor<int32, [3]> k_505_perm_0 = const()[name = tensor<string, []>("k_505_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10393_pad_type_0 = const()[name = tensor<string, []>("op_10393_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10393_strides_0 = const()[name = tensor<string, []>("op_10393_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10393_pad_0 = const()[name = tensor<string, []>("op_10393_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10393_dilations_0 = const()[name = tensor<string, []>("op_10393_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10393_groups_0 = const()[name = tensor<string, []>("op_10393_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101032256)))];
            tensor<fp16, [1, 64, 1, 197]> var_10393_cast_fp16 = conv(dilations = var_10393_dilations_0, groups = var_10393_groups_0, pad = var_10393_pad_0, pad_type = var_10393_pad_type_0, strides = var_10393_strides_0, weight = model_blocks_7_attn_v_projs_0_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10393_cast_fp16")];
            tensor<int32, [1]> var_10394_axes_0 = const()[name = tensor<string, []>("op_10394_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10394_cast_fp16 = squeeze(axes = var_10394_axes_0, x = var_10393_cast_fp16)[name = tensor<string, []>("op_10394_cast_fp16")];
            tensor<int32, [3]> v_505_perm_0 = const()[name = tensor<string, []>("v_505_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101130624)))];
            tensor<fp16, [1, 197, 64]> q_841_cast_fp16 = transpose(perm = q_841_perm_0, x = var_10378_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 197, 64]> q_843_cast_fp16 = add(x = q_841_cast_fp16, y = model_blocks_7_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_843_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101130816)))];
            tensor<fp16, [1, 197, 64]> v_505_cast_fp16 = transpose(perm = v_505_perm_0, x = var_10394_cast_fp16)[name = tensor<string, []>("transpose_211")];
            tensor<fp16, [1, 197, 64]> v_507_cast_fp16 = add(x = v_505_cast_fp16, y = model_blocks_7_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_507_cast_fp16")];
            tensor<int32, [1]> q_845_axes_0 = const()[name = tensor<string, []>("q_845_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_845_cast_fp16 = expand_dims(axes = q_845_axes_0, x = q_843_cast_fp16)[name = tensor<string, []>("q_845_cast_fp16")];
            tensor<int32, [1]> k_507_axes_0 = const()[name = tensor<string, []>("k_507_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_505_cast_fp16 = transpose(perm = k_505_perm_0, x = var_10386_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [1, 1, 197, 64]> k_507_cast_fp16 = expand_dims(axes = k_507_axes_0, x = k_505_cast_fp16)[name = tensor<string, []>("k_507_cast_fp16")];
            tensor<int32, [1]> v_509_axes_0 = const()[name = tensor<string, []>("v_509_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_509_cast_fp16 = expand_dims(axes = v_509_axes_0, x = v_507_cast_fp16)[name = tensor<string, []>("v_509_cast_fp16")];
            tensor<int32, [4]> var_10403_begin_0 = const()[name = tensor<string, []>("op_10403_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10403_end_0 = const()[name = tensor<string, []>("op_10403_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10403_end_mask_0 = const()[name = tensor<string, []>("op_10403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10403_cast_fp16 = slice_by_index(begin = var_10403_begin_0, end = var_10403_end_0, end_mask = var_10403_end_mask_0, x = q_845_cast_fp16)[name = tensor<string, []>("op_10403_cast_fp16")];
            tensor<int32, [4]> var_10407_begin_0 = const()[name = tensor<string, []>("op_10407_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10407_end_0 = const()[name = tensor<string, []>("op_10407_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10407_end_mask_0 = const()[name = tensor<string, []>("op_10407_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10407_cast_fp16 = slice_by_index(begin = var_10407_begin_0, end = var_10407_end_0, end_mask = var_10407_end_mask_0, x = k_507_cast_fp16)[name = tensor<string, []>("op_10407_cast_fp16")];
            tensor<int32, [4]> var_10411_begin_0 = const()[name = tensor<string, []>("op_10411_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10411_end_0 = const()[name = tensor<string, []>("op_10411_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10411_end_mask_0 = const()[name = tensor<string, []>("op_10411_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10411_cast_fp16 = slice_by_index(begin = var_10411_begin_0, end = var_10411_end_0, end_mask = var_10411_end_mask_0, x = q_845_cast_fp16)[name = tensor<string, []>("op_10411_cast_fp16")];
            tensor<int32, [4]> var_10415_begin_0 = const()[name = tensor<string, []>("op_10415_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10415_end_0 = const()[name = tensor<string, []>("op_10415_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10415_end_mask_0 = const()[name = tensor<string, []>("op_10415_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10415_cast_fp16 = slice_by_index(begin = var_10415_begin_0, end = var_10415_end_0, end_mask = var_10415_end_mask_0, x = k_507_cast_fp16)[name = tensor<string, []>("op_10415_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10417_cast_fp16 = mul(x = var_10411_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10417_cast_fp16")];
            tensor<int32, [5]> var_10425 = const()[name = tensor<string, []>("op_10425"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1159_cast_fp16 = reshape(shape = var_10425, x = var_10411_cast_fp16)[name = tensor<string, []>("x_1159_cast_fp16")];
            tensor<int32, [2]> var_10427_split_sizes_0 = const()[name = tensor<string, []>("op_10427_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10427_axis_0 = const()[name = tensor<string, []>("op_10427_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10427_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10427_cast_fp16_1 = split(axis = var_10427_axis_0, split_sizes = var_10427_split_sizes_0, x = x_1159_cast_fp16)[name = tensor<string, []>("op_10427_cast_fp16")];
            tensor<int32, [1]> squeeze_336_axes_0 = const()[name = tensor<string, []>("squeeze_336_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_336_cast_fp16 = squeeze(axes = squeeze_336_axes_0, x = var_10427_cast_fp16_0)[name = tensor<string, []>("squeeze_336_cast_fp16")];
            tensor<int32, [1]> squeeze_337_axes_0 = const()[name = tensor<string, []>("squeeze_337_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_337_cast_fp16 = squeeze(axes = squeeze_337_axes_0, x = var_10427_cast_fp16_1)[name = tensor<string, []>("squeeze_337_cast_fp16")];
            tensor<fp16, []> const_1379_promoted_to_fp16 = const()[name = tensor<string, []>("const_1379_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10430_cast_fp16 = mul(x = squeeze_337_cast_fp16, y = const_1379_promoted_to_fp16)[name = tensor<string, []>("op_10430_cast_fp16")];
            tensor<int32, []> x_1161_axis_0 = const()[name = tensor<string, []>("x_1161_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1161_cast_fp16 = stack(axis = x_1161_axis_0, values = (var_10430_cast_fp16, squeeze_336_cast_fp16))[name = tensor<string, []>("x_1161_cast_fp16")];
            tensor<int32, [4]> var_10436 = const()[name = tensor<string, []>("op_10436"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10437_cast_fp16 = reshape(shape = var_10436, x = x_1161_cast_fp16)[name = tensor<string, []>("op_10437_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10438_cast_fp16 = mul(x = var_10437_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10438_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_169_cast_fp16 = add(x = var_10417_cast_fp16, y = var_10438_cast_fp16)[name = tensor<string, []>("q_patches_169_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10440_cast_fp16 = mul(x = var_10415_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10440_cast_fp16")];
            tensor<int32, [5]> var_10448 = const()[name = tensor<string, []>("op_10448"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1165_cast_fp16 = reshape(shape = var_10448, x = var_10415_cast_fp16)[name = tensor<string, []>("x_1165_cast_fp16")];
            tensor<int32, [2]> var_10450_split_sizes_0 = const()[name = tensor<string, []>("op_10450_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10450_axis_0 = const()[name = tensor<string, []>("op_10450_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10450_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10450_cast_fp16_1 = split(axis = var_10450_axis_0, split_sizes = var_10450_split_sizes_0, x = x_1165_cast_fp16)[name = tensor<string, []>("op_10450_cast_fp16")];
            tensor<int32, [1]> squeeze_338_axes_0 = const()[name = tensor<string, []>("squeeze_338_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_338_cast_fp16 = squeeze(axes = squeeze_338_axes_0, x = var_10450_cast_fp16_0)[name = tensor<string, []>("squeeze_338_cast_fp16")];
            tensor<int32, [1]> squeeze_339_axes_0 = const()[name = tensor<string, []>("squeeze_339_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_339_cast_fp16 = squeeze(axes = squeeze_339_axes_0, x = var_10450_cast_fp16_1)[name = tensor<string, []>("squeeze_339_cast_fp16")];
            tensor<fp16, []> const_1387_promoted_to_fp16 = const()[name = tensor<string, []>("const_1387_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10453_cast_fp16 = mul(x = squeeze_339_cast_fp16, y = const_1387_promoted_to_fp16)[name = tensor<string, []>("op_10453_cast_fp16")];
            tensor<int32, []> x_1167_axis_0 = const()[name = tensor<string, []>("x_1167_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1167_cast_fp16 = stack(axis = x_1167_axis_0, values = (var_10453_cast_fp16, squeeze_338_cast_fp16))[name = tensor<string, []>("x_1167_cast_fp16")];
            tensor<int32, [4]> var_10459 = const()[name = tensor<string, []>("op_10459"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10460_cast_fp16 = reshape(shape = var_10459, x = x_1167_cast_fp16)[name = tensor<string, []>("op_10460_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10461_cast_fp16 = mul(x = var_10460_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10461_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_169_cast_fp16 = add(x = var_10440_cast_fp16, y = var_10461_cast_fp16)[name = tensor<string, []>("k_patches_169_cast_fp16")];
            tensor<bool, []> var_10464_interleave_0 = const()[name = tensor<string, []>("op_10464_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10464_cast_fp16 = concat(axis = var_20, interleave = var_10464_interleave_0, values = (var_10403_cast_fp16, q_patches_169_cast_fp16))[name = tensor<string, []>("op_10464_cast_fp16")];
            tensor<bool, []> var_10467_interleave_0 = const()[name = tensor<string, []>("op_10467_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10467_cast_fp16 = concat(axis = var_20, interleave = var_10467_interleave_0, values = (var_10407_cast_fp16, k_patches_169_cast_fp16))[name = tensor<string, []>("op_10467_cast_fp16")];
            tensor<fp16, []> var_10469_to_fp16 = const()[name = tensor<string, []>("op_10469_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_849_cast_fp16 = mul(x = var_10464_cast_fp16, y = var_10469_to_fp16)[name = tensor<string, []>("q_849_cast_fp16")];
            tensor<bool, []> attn_337_transpose_x_1 = const()[name = tensor<string, []>("attn_337_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_337_transpose_y_1 = const()[name = tensor<string, []>("attn_337_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_337_cast_fp16 = matmul(transpose_x = attn_337_transpose_x_1, transpose_y = attn_337_transpose_y_1, x = q_849_cast_fp16, y = var_10467_cast_fp16)[name = tensor<string, []>("attn_337_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_339_cast_fp16 = softmax(axis = var_21, x = attn_337_cast_fp16)[name = tensor<string, []>("attn_339_cast_fp16")];
            tensor<bool, []> var_10474_transpose_x_0 = const()[name = tensor<string, []>("op_10474_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10474_transpose_y_0 = const()[name = tensor<string, []>("op_10474_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10474_cast_fp16 = matmul(transpose_x = var_10474_transpose_x_0, transpose_y = var_10474_transpose_y_0, x = attn_339_cast_fp16, y = v_509_cast_fp16)[name = tensor<string, []>("op_10474_cast_fp16")];
            tensor<int32, [1]> var_10475_axes_0 = const()[name = tensor<string, []>("op_10475_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10475_cast_fp16 = squeeze(axes = var_10475_axes_0, x = var_10474_cast_fp16)[name = tensor<string, []>("op_10475_cast_fp16")];
            tensor<string, []> var_10481_pad_type_0 = const()[name = tensor<string, []>("op_10481_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10481_strides_0 = const()[name = tensor<string, []>("op_10481_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10481_pad_0 = const()[name = tensor<string, []>("op_10481_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10481_dilations_0 = const()[name = tensor<string, []>("op_10481_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10481_groups_0 = const()[name = tensor<string, []>("op_10481_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101131008)))];
            tensor<fp16, [1, 64, 1, 197]> var_10481_cast_fp16 = conv(dilations = var_10481_dilations_0, groups = var_10481_groups_0, pad = var_10481_pad_0, pad_type = var_10481_pad_type_0, strides = var_10481_strides_0, weight = model_blocks_7_attn_q_projs_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10481_cast_fp16")];
            tensor<int32, [1]> var_10482_axes_0 = const()[name = tensor<string, []>("op_10482_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10482_cast_fp16 = squeeze(axes = var_10482_axes_0, x = var_10481_cast_fp16)[name = tensor<string, []>("op_10482_cast_fp16")];
            tensor<int32, [3]> q_851_perm_0 = const()[name = tensor<string, []>("q_851_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10489_pad_type_0 = const()[name = tensor<string, []>("op_10489_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10489_strides_0 = const()[name = tensor<string, []>("op_10489_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10489_pad_0 = const()[name = tensor<string, []>("op_10489_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10489_dilations_0 = const()[name = tensor<string, []>("op_10489_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10489_groups_0 = const()[name = tensor<string, []>("op_10489_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101229376)))];
            tensor<fp16, [1, 64, 1, 197]> var_10489_cast_fp16 = conv(dilations = var_10489_dilations_0, groups = var_10489_groups_0, pad = var_10489_pad_0, pad_type = var_10489_pad_type_0, strides = var_10489_strides_0, weight = model_blocks_7_attn_k_projs_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10489_cast_fp16")];
            tensor<int32, [1]> var_10490_axes_0 = const()[name = tensor<string, []>("op_10490_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10490_cast_fp16 = squeeze(axes = var_10490_axes_0, x = var_10489_cast_fp16)[name = tensor<string, []>("op_10490_cast_fp16")];
            tensor<int32, [3]> k_511_perm_0 = const()[name = tensor<string, []>("k_511_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10497_pad_type_0 = const()[name = tensor<string, []>("op_10497_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10497_strides_0 = const()[name = tensor<string, []>("op_10497_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10497_pad_0 = const()[name = tensor<string, []>("op_10497_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10497_dilations_0 = const()[name = tensor<string, []>("op_10497_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10497_groups_0 = const()[name = tensor<string, []>("op_10497_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101327744)))];
            tensor<fp16, [1, 64, 1, 197]> var_10497_cast_fp16 = conv(dilations = var_10497_dilations_0, groups = var_10497_groups_0, pad = var_10497_pad_0, pad_type = var_10497_pad_type_0, strides = var_10497_strides_0, weight = model_blocks_7_attn_v_projs_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10497_cast_fp16")];
            tensor<int32, [1]> var_10498_axes_0 = const()[name = tensor<string, []>("op_10498_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10498_cast_fp16 = squeeze(axes = var_10498_axes_0, x = var_10497_cast_fp16)[name = tensor<string, []>("op_10498_cast_fp16")];
            tensor<int32, [3]> v_511_perm_0 = const()[name = tensor<string, []>("v_511_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101426112)))];
            tensor<fp16, [1, 197, 64]> q_851_cast_fp16 = transpose(perm = q_851_perm_0, x = var_10482_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [1, 197, 64]> q_853_cast_fp16 = add(x = q_851_cast_fp16, y = model_blocks_7_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_853_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101426304)))];
            tensor<fp16, [1, 197, 64]> v_511_cast_fp16 = transpose(perm = v_511_perm_0, x = var_10498_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 197, 64]> v_513_cast_fp16 = add(x = v_511_cast_fp16, y = model_blocks_7_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_513_cast_fp16")];
            tensor<int32, [1]> q_855_axes_0 = const()[name = tensor<string, []>("q_855_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_855_cast_fp16 = expand_dims(axes = q_855_axes_0, x = q_853_cast_fp16)[name = tensor<string, []>("q_855_cast_fp16")];
            tensor<int32, [1]> k_513_axes_0 = const()[name = tensor<string, []>("k_513_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_511_cast_fp16 = transpose(perm = k_511_perm_0, x = var_10490_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [1, 1, 197, 64]> k_513_cast_fp16 = expand_dims(axes = k_513_axes_0, x = k_511_cast_fp16)[name = tensor<string, []>("k_513_cast_fp16")];
            tensor<int32, [1]> v_515_axes_0 = const()[name = tensor<string, []>("v_515_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_515_cast_fp16 = expand_dims(axes = v_515_axes_0, x = v_513_cast_fp16)[name = tensor<string, []>("v_515_cast_fp16")];
            tensor<int32, [4]> var_10507_begin_0 = const()[name = tensor<string, []>("op_10507_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10507_end_0 = const()[name = tensor<string, []>("op_10507_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10507_end_mask_0 = const()[name = tensor<string, []>("op_10507_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10507_cast_fp16 = slice_by_index(begin = var_10507_begin_0, end = var_10507_end_0, end_mask = var_10507_end_mask_0, x = q_855_cast_fp16)[name = tensor<string, []>("op_10507_cast_fp16")];
            tensor<int32, [4]> var_10511_begin_0 = const()[name = tensor<string, []>("op_10511_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10511_end_0 = const()[name = tensor<string, []>("op_10511_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10511_end_mask_0 = const()[name = tensor<string, []>("op_10511_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10511_cast_fp16 = slice_by_index(begin = var_10511_begin_0, end = var_10511_end_0, end_mask = var_10511_end_mask_0, x = k_513_cast_fp16)[name = tensor<string, []>("op_10511_cast_fp16")];
            tensor<int32, [4]> var_10515_begin_0 = const()[name = tensor<string, []>("op_10515_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10515_end_0 = const()[name = tensor<string, []>("op_10515_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10515_end_mask_0 = const()[name = tensor<string, []>("op_10515_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10515_cast_fp16 = slice_by_index(begin = var_10515_begin_0, end = var_10515_end_0, end_mask = var_10515_end_mask_0, x = q_855_cast_fp16)[name = tensor<string, []>("op_10515_cast_fp16")];
            tensor<int32, [4]> var_10519_begin_0 = const()[name = tensor<string, []>("op_10519_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10519_end_0 = const()[name = tensor<string, []>("op_10519_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10519_end_mask_0 = const()[name = tensor<string, []>("op_10519_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10519_cast_fp16 = slice_by_index(begin = var_10519_begin_0, end = var_10519_end_0, end_mask = var_10519_end_mask_0, x = k_513_cast_fp16)[name = tensor<string, []>("op_10519_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10521_cast_fp16 = mul(x = var_10515_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10521_cast_fp16")];
            tensor<int32, [5]> var_10529 = const()[name = tensor<string, []>("op_10529"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1171_cast_fp16 = reshape(shape = var_10529, x = var_10515_cast_fp16)[name = tensor<string, []>("x_1171_cast_fp16")];
            tensor<int32, [2]> var_10531_split_sizes_0 = const()[name = tensor<string, []>("op_10531_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10531_axis_0 = const()[name = tensor<string, []>("op_10531_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10531_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10531_cast_fp16_1 = split(axis = var_10531_axis_0, split_sizes = var_10531_split_sizes_0, x = x_1171_cast_fp16)[name = tensor<string, []>("op_10531_cast_fp16")];
            tensor<int32, [1]> squeeze_340_axes_0 = const()[name = tensor<string, []>("squeeze_340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_340_cast_fp16 = squeeze(axes = squeeze_340_axes_0, x = var_10531_cast_fp16_0)[name = tensor<string, []>("squeeze_340_cast_fp16")];
            tensor<int32, [1]> squeeze_341_axes_0 = const()[name = tensor<string, []>("squeeze_341_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_341_cast_fp16 = squeeze(axes = squeeze_341_axes_0, x = var_10531_cast_fp16_1)[name = tensor<string, []>("squeeze_341_cast_fp16")];
            tensor<fp16, []> const_1395_promoted_to_fp16 = const()[name = tensor<string, []>("const_1395_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10534_cast_fp16 = mul(x = squeeze_341_cast_fp16, y = const_1395_promoted_to_fp16)[name = tensor<string, []>("op_10534_cast_fp16")];
            tensor<int32, []> x_1173_axis_0 = const()[name = tensor<string, []>("x_1173_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1173_cast_fp16 = stack(axis = x_1173_axis_0, values = (var_10534_cast_fp16, squeeze_340_cast_fp16))[name = tensor<string, []>("x_1173_cast_fp16")];
            tensor<int32, [4]> var_10540 = const()[name = tensor<string, []>("op_10540"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10541_cast_fp16 = reshape(shape = var_10540, x = x_1173_cast_fp16)[name = tensor<string, []>("op_10541_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10542_cast_fp16 = mul(x = var_10541_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10542_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_171_cast_fp16 = add(x = var_10521_cast_fp16, y = var_10542_cast_fp16)[name = tensor<string, []>("q_patches_171_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10544_cast_fp16 = mul(x = var_10519_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10544_cast_fp16")];
            tensor<int32, [5]> var_10552 = const()[name = tensor<string, []>("op_10552"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1177_cast_fp16 = reshape(shape = var_10552, x = var_10519_cast_fp16)[name = tensor<string, []>("x_1177_cast_fp16")];
            tensor<int32, [2]> var_10554_split_sizes_0 = const()[name = tensor<string, []>("op_10554_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10554_axis_0 = const()[name = tensor<string, []>("op_10554_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10554_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10554_cast_fp16_1 = split(axis = var_10554_axis_0, split_sizes = var_10554_split_sizes_0, x = x_1177_cast_fp16)[name = tensor<string, []>("op_10554_cast_fp16")];
            tensor<int32, [1]> squeeze_342_axes_0 = const()[name = tensor<string, []>("squeeze_342_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_342_cast_fp16 = squeeze(axes = squeeze_342_axes_0, x = var_10554_cast_fp16_0)[name = tensor<string, []>("squeeze_342_cast_fp16")];
            tensor<int32, [1]> squeeze_343_axes_0 = const()[name = tensor<string, []>("squeeze_343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_343_cast_fp16 = squeeze(axes = squeeze_343_axes_0, x = var_10554_cast_fp16_1)[name = tensor<string, []>("squeeze_343_cast_fp16")];
            tensor<fp16, []> const_1403_promoted_to_fp16 = const()[name = tensor<string, []>("const_1403_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10557_cast_fp16 = mul(x = squeeze_343_cast_fp16, y = const_1403_promoted_to_fp16)[name = tensor<string, []>("op_10557_cast_fp16")];
            tensor<int32, []> x_1179_axis_0 = const()[name = tensor<string, []>("x_1179_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1179_cast_fp16 = stack(axis = x_1179_axis_0, values = (var_10557_cast_fp16, squeeze_342_cast_fp16))[name = tensor<string, []>("x_1179_cast_fp16")];
            tensor<int32, [4]> var_10563 = const()[name = tensor<string, []>("op_10563"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10564_cast_fp16 = reshape(shape = var_10563, x = x_1179_cast_fp16)[name = tensor<string, []>("op_10564_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10565_cast_fp16 = mul(x = var_10564_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10565_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_171_cast_fp16 = add(x = var_10544_cast_fp16, y = var_10565_cast_fp16)[name = tensor<string, []>("k_patches_171_cast_fp16")];
            tensor<bool, []> var_10568_interleave_0 = const()[name = tensor<string, []>("op_10568_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10568_cast_fp16 = concat(axis = var_20, interleave = var_10568_interleave_0, values = (var_10507_cast_fp16, q_patches_171_cast_fp16))[name = tensor<string, []>("op_10568_cast_fp16")];
            tensor<bool, []> var_10571_interleave_0 = const()[name = tensor<string, []>("op_10571_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10571_cast_fp16 = concat(axis = var_20, interleave = var_10571_interleave_0, values = (var_10511_cast_fp16, k_patches_171_cast_fp16))[name = tensor<string, []>("op_10571_cast_fp16")];
            tensor<fp16, []> var_10573_to_fp16 = const()[name = tensor<string, []>("op_10573_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_859_cast_fp16 = mul(x = var_10568_cast_fp16, y = var_10573_to_fp16)[name = tensor<string, []>("q_859_cast_fp16")];
            tensor<bool, []> attn_341_transpose_x_1 = const()[name = tensor<string, []>("attn_341_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_341_transpose_y_1 = const()[name = tensor<string, []>("attn_341_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_341_cast_fp16 = matmul(transpose_x = attn_341_transpose_x_1, transpose_y = attn_341_transpose_y_1, x = q_859_cast_fp16, y = var_10571_cast_fp16)[name = tensor<string, []>("attn_341_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_343_cast_fp16 = softmax(axis = var_21, x = attn_341_cast_fp16)[name = tensor<string, []>("attn_343_cast_fp16")];
            tensor<bool, []> var_10578_transpose_x_0 = const()[name = tensor<string, []>("op_10578_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10578_transpose_y_0 = const()[name = tensor<string, []>("op_10578_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10578_cast_fp16 = matmul(transpose_x = var_10578_transpose_x_0, transpose_y = var_10578_transpose_y_0, x = attn_343_cast_fp16, y = v_515_cast_fp16)[name = tensor<string, []>("op_10578_cast_fp16")];
            tensor<int32, [1]> var_10579_axes_0 = const()[name = tensor<string, []>("op_10579_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10579_cast_fp16 = squeeze(axes = var_10579_axes_0, x = var_10578_cast_fp16)[name = tensor<string, []>("op_10579_cast_fp16")];
            tensor<string, []> var_10585_pad_type_0 = const()[name = tensor<string, []>("op_10585_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10585_strides_0 = const()[name = tensor<string, []>("op_10585_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10585_pad_0 = const()[name = tensor<string, []>("op_10585_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10585_dilations_0 = const()[name = tensor<string, []>("op_10585_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10585_groups_0 = const()[name = tensor<string, []>("op_10585_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101426496)))];
            tensor<fp16, [1, 64, 1, 197]> var_10585_cast_fp16 = conv(dilations = var_10585_dilations_0, groups = var_10585_groups_0, pad = var_10585_pad_0, pad_type = var_10585_pad_type_0, strides = var_10585_strides_0, weight = model_blocks_7_attn_q_projs_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10585_cast_fp16")];
            tensor<int32, [1]> var_10586_axes_0 = const()[name = tensor<string, []>("op_10586_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10586_cast_fp16 = squeeze(axes = var_10586_axes_0, x = var_10585_cast_fp16)[name = tensor<string, []>("op_10586_cast_fp16")];
            tensor<int32, [3]> q_861_perm_0 = const()[name = tensor<string, []>("q_861_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10593_pad_type_0 = const()[name = tensor<string, []>("op_10593_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10593_strides_0 = const()[name = tensor<string, []>("op_10593_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10593_pad_0 = const()[name = tensor<string, []>("op_10593_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10593_dilations_0 = const()[name = tensor<string, []>("op_10593_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10593_groups_0 = const()[name = tensor<string, []>("op_10593_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101524864)))];
            tensor<fp16, [1, 64, 1, 197]> var_10593_cast_fp16 = conv(dilations = var_10593_dilations_0, groups = var_10593_groups_0, pad = var_10593_pad_0, pad_type = var_10593_pad_type_0, strides = var_10593_strides_0, weight = model_blocks_7_attn_k_projs_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10593_cast_fp16")];
            tensor<int32, [1]> var_10594_axes_0 = const()[name = tensor<string, []>("op_10594_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10594_cast_fp16 = squeeze(axes = var_10594_axes_0, x = var_10593_cast_fp16)[name = tensor<string, []>("op_10594_cast_fp16")];
            tensor<int32, [3]> k_517_perm_0 = const()[name = tensor<string, []>("k_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10601_pad_type_0 = const()[name = tensor<string, []>("op_10601_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10601_strides_0 = const()[name = tensor<string, []>("op_10601_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10601_pad_0 = const()[name = tensor<string, []>("op_10601_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10601_dilations_0 = const()[name = tensor<string, []>("op_10601_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10601_groups_0 = const()[name = tensor<string, []>("op_10601_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101623232)))];
            tensor<fp16, [1, 64, 1, 197]> var_10601_cast_fp16 = conv(dilations = var_10601_dilations_0, groups = var_10601_groups_0, pad = var_10601_pad_0, pad_type = var_10601_pad_type_0, strides = var_10601_strides_0, weight = model_blocks_7_attn_v_projs_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10601_cast_fp16")];
            tensor<int32, [1]> var_10602_axes_0 = const()[name = tensor<string, []>("op_10602_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10602_cast_fp16 = squeeze(axes = var_10602_axes_0, x = var_10601_cast_fp16)[name = tensor<string, []>("op_10602_cast_fp16")];
            tensor<int32, [3]> v_517_perm_0 = const()[name = tensor<string, []>("v_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101721600)))];
            tensor<fp16, [1, 197, 64]> q_861_cast_fp16 = transpose(perm = q_861_perm_0, x = var_10586_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 197, 64]> q_863_cast_fp16 = add(x = q_861_cast_fp16, y = model_blocks_7_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_863_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101721792)))];
            tensor<fp16, [1, 197, 64]> v_517_cast_fp16 = transpose(perm = v_517_perm_0, x = var_10602_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [1, 197, 64]> v_519_cast_fp16 = add(x = v_517_cast_fp16, y = model_blocks_7_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_519_cast_fp16")];
            tensor<int32, [1]> q_865_axes_0 = const()[name = tensor<string, []>("q_865_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_865_cast_fp16 = expand_dims(axes = q_865_axes_0, x = q_863_cast_fp16)[name = tensor<string, []>("q_865_cast_fp16")];
            tensor<int32, [1]> k_519_axes_0 = const()[name = tensor<string, []>("k_519_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_517_cast_fp16 = transpose(perm = k_517_perm_0, x = var_10594_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 1, 197, 64]> k_519_cast_fp16 = expand_dims(axes = k_519_axes_0, x = k_517_cast_fp16)[name = tensor<string, []>("k_519_cast_fp16")];
            tensor<int32, [1]> v_521_axes_0 = const()[name = tensor<string, []>("v_521_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_521_cast_fp16 = expand_dims(axes = v_521_axes_0, x = v_519_cast_fp16)[name = tensor<string, []>("v_521_cast_fp16")];
            tensor<int32, [4]> var_10611_begin_0 = const()[name = tensor<string, []>("op_10611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10611_end_0 = const()[name = tensor<string, []>("op_10611_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10611_end_mask_0 = const()[name = tensor<string, []>("op_10611_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10611_cast_fp16 = slice_by_index(begin = var_10611_begin_0, end = var_10611_end_0, end_mask = var_10611_end_mask_0, x = q_865_cast_fp16)[name = tensor<string, []>("op_10611_cast_fp16")];
            tensor<int32, [4]> var_10615_begin_0 = const()[name = tensor<string, []>("op_10615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10615_end_0 = const()[name = tensor<string, []>("op_10615_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10615_end_mask_0 = const()[name = tensor<string, []>("op_10615_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10615_cast_fp16 = slice_by_index(begin = var_10615_begin_0, end = var_10615_end_0, end_mask = var_10615_end_mask_0, x = k_519_cast_fp16)[name = tensor<string, []>("op_10615_cast_fp16")];
            tensor<int32, [4]> var_10619_begin_0 = const()[name = tensor<string, []>("op_10619_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10619_end_0 = const()[name = tensor<string, []>("op_10619_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10619_end_mask_0 = const()[name = tensor<string, []>("op_10619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10619_cast_fp16 = slice_by_index(begin = var_10619_begin_0, end = var_10619_end_0, end_mask = var_10619_end_mask_0, x = q_865_cast_fp16)[name = tensor<string, []>("op_10619_cast_fp16")];
            tensor<int32, [4]> var_10623_begin_0 = const()[name = tensor<string, []>("op_10623_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10623_end_0 = const()[name = tensor<string, []>("op_10623_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10623_end_mask_0 = const()[name = tensor<string, []>("op_10623_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10623_cast_fp16 = slice_by_index(begin = var_10623_begin_0, end = var_10623_end_0, end_mask = var_10623_end_mask_0, x = k_519_cast_fp16)[name = tensor<string, []>("op_10623_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10625_cast_fp16 = mul(x = var_10619_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10625_cast_fp16")];
            tensor<int32, [5]> var_10633 = const()[name = tensor<string, []>("op_10633"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1183_cast_fp16 = reshape(shape = var_10633, x = var_10619_cast_fp16)[name = tensor<string, []>("x_1183_cast_fp16")];
            tensor<int32, [2]> var_10635_split_sizes_0 = const()[name = tensor<string, []>("op_10635_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10635_axis_0 = const()[name = tensor<string, []>("op_10635_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10635_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10635_cast_fp16_1 = split(axis = var_10635_axis_0, split_sizes = var_10635_split_sizes_0, x = x_1183_cast_fp16)[name = tensor<string, []>("op_10635_cast_fp16")];
            tensor<int32, [1]> squeeze_344_axes_0 = const()[name = tensor<string, []>("squeeze_344_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_344_cast_fp16 = squeeze(axes = squeeze_344_axes_0, x = var_10635_cast_fp16_0)[name = tensor<string, []>("squeeze_344_cast_fp16")];
            tensor<int32, [1]> squeeze_345_axes_0 = const()[name = tensor<string, []>("squeeze_345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_345_cast_fp16 = squeeze(axes = squeeze_345_axes_0, x = var_10635_cast_fp16_1)[name = tensor<string, []>("squeeze_345_cast_fp16")];
            tensor<fp16, []> const_1411_promoted_to_fp16 = const()[name = tensor<string, []>("const_1411_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10638_cast_fp16 = mul(x = squeeze_345_cast_fp16, y = const_1411_promoted_to_fp16)[name = tensor<string, []>("op_10638_cast_fp16")];
            tensor<int32, []> x_1185_axis_0 = const()[name = tensor<string, []>("x_1185_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1185_cast_fp16 = stack(axis = x_1185_axis_0, values = (var_10638_cast_fp16, squeeze_344_cast_fp16))[name = tensor<string, []>("x_1185_cast_fp16")];
            tensor<int32, [4]> var_10644 = const()[name = tensor<string, []>("op_10644"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10645_cast_fp16 = reshape(shape = var_10644, x = x_1185_cast_fp16)[name = tensor<string, []>("op_10645_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10646_cast_fp16 = mul(x = var_10645_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10646_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_173_cast_fp16 = add(x = var_10625_cast_fp16, y = var_10646_cast_fp16)[name = tensor<string, []>("q_patches_173_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10648_cast_fp16 = mul(x = var_10623_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10648_cast_fp16")];
            tensor<int32, [5]> var_10656 = const()[name = tensor<string, []>("op_10656"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1189_cast_fp16 = reshape(shape = var_10656, x = var_10623_cast_fp16)[name = tensor<string, []>("x_1189_cast_fp16")];
            tensor<int32, [2]> var_10658_split_sizes_0 = const()[name = tensor<string, []>("op_10658_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10658_axis_0 = const()[name = tensor<string, []>("op_10658_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10658_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10658_cast_fp16_1 = split(axis = var_10658_axis_0, split_sizes = var_10658_split_sizes_0, x = x_1189_cast_fp16)[name = tensor<string, []>("op_10658_cast_fp16")];
            tensor<int32, [1]> squeeze_346_axes_0 = const()[name = tensor<string, []>("squeeze_346_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_346_cast_fp16 = squeeze(axes = squeeze_346_axes_0, x = var_10658_cast_fp16_0)[name = tensor<string, []>("squeeze_346_cast_fp16")];
            tensor<int32, [1]> squeeze_347_axes_0 = const()[name = tensor<string, []>("squeeze_347_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_347_cast_fp16 = squeeze(axes = squeeze_347_axes_0, x = var_10658_cast_fp16_1)[name = tensor<string, []>("squeeze_347_cast_fp16")];
            tensor<fp16, []> const_1419_promoted_to_fp16 = const()[name = tensor<string, []>("const_1419_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10661_cast_fp16 = mul(x = squeeze_347_cast_fp16, y = const_1419_promoted_to_fp16)[name = tensor<string, []>("op_10661_cast_fp16")];
            tensor<int32, []> x_1191_axis_0 = const()[name = tensor<string, []>("x_1191_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1191_cast_fp16 = stack(axis = x_1191_axis_0, values = (var_10661_cast_fp16, squeeze_346_cast_fp16))[name = tensor<string, []>("x_1191_cast_fp16")];
            tensor<int32, [4]> var_10667 = const()[name = tensor<string, []>("op_10667"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10668_cast_fp16 = reshape(shape = var_10667, x = x_1191_cast_fp16)[name = tensor<string, []>("op_10668_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10669_cast_fp16 = mul(x = var_10668_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10669_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_173_cast_fp16 = add(x = var_10648_cast_fp16, y = var_10669_cast_fp16)[name = tensor<string, []>("k_patches_173_cast_fp16")];
            tensor<bool, []> var_10672_interleave_0 = const()[name = tensor<string, []>("op_10672_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10672_cast_fp16 = concat(axis = var_20, interleave = var_10672_interleave_0, values = (var_10611_cast_fp16, q_patches_173_cast_fp16))[name = tensor<string, []>("op_10672_cast_fp16")];
            tensor<bool, []> var_10675_interleave_0 = const()[name = tensor<string, []>("op_10675_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10675_cast_fp16 = concat(axis = var_20, interleave = var_10675_interleave_0, values = (var_10615_cast_fp16, k_patches_173_cast_fp16))[name = tensor<string, []>("op_10675_cast_fp16")];
            tensor<fp16, []> var_10677_to_fp16 = const()[name = tensor<string, []>("op_10677_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_869_cast_fp16 = mul(x = var_10672_cast_fp16, y = var_10677_to_fp16)[name = tensor<string, []>("q_869_cast_fp16")];
            tensor<bool, []> attn_345_transpose_x_1 = const()[name = tensor<string, []>("attn_345_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_345_transpose_y_1 = const()[name = tensor<string, []>("attn_345_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_345_cast_fp16 = matmul(transpose_x = attn_345_transpose_x_1, transpose_y = attn_345_transpose_y_1, x = q_869_cast_fp16, y = var_10675_cast_fp16)[name = tensor<string, []>("attn_345_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_347_cast_fp16 = softmax(axis = var_21, x = attn_345_cast_fp16)[name = tensor<string, []>("attn_347_cast_fp16")];
            tensor<bool, []> var_10682_transpose_x_0 = const()[name = tensor<string, []>("op_10682_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10682_transpose_y_0 = const()[name = tensor<string, []>("op_10682_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10682_cast_fp16 = matmul(transpose_x = var_10682_transpose_x_0, transpose_y = var_10682_transpose_y_0, x = attn_347_cast_fp16, y = v_521_cast_fp16)[name = tensor<string, []>("op_10682_cast_fp16")];
            tensor<int32, [1]> var_10683_axes_0 = const()[name = tensor<string, []>("op_10683_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10683_cast_fp16 = squeeze(axes = var_10683_axes_0, x = var_10682_cast_fp16)[name = tensor<string, []>("op_10683_cast_fp16")];
            tensor<string, []> var_10689_pad_type_0 = const()[name = tensor<string, []>("op_10689_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10689_strides_0 = const()[name = tensor<string, []>("op_10689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10689_pad_0 = const()[name = tensor<string, []>("op_10689_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10689_dilations_0 = const()[name = tensor<string, []>("op_10689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10689_groups_0 = const()[name = tensor<string, []>("op_10689_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101721984)))];
            tensor<fp16, [1, 64, 1, 197]> var_10689_cast_fp16 = conv(dilations = var_10689_dilations_0, groups = var_10689_groups_0, pad = var_10689_pad_0, pad_type = var_10689_pad_type_0, strides = var_10689_strides_0, weight = model_blocks_7_attn_q_projs_3_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10689_cast_fp16")];
            tensor<int32, [1]> var_10690_axes_0 = const()[name = tensor<string, []>("op_10690_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10690_cast_fp16 = squeeze(axes = var_10690_axes_0, x = var_10689_cast_fp16)[name = tensor<string, []>("op_10690_cast_fp16")];
            tensor<int32, [3]> q_871_perm_0 = const()[name = tensor<string, []>("q_871_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10697_pad_type_0 = const()[name = tensor<string, []>("op_10697_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10697_strides_0 = const()[name = tensor<string, []>("op_10697_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10697_pad_0 = const()[name = tensor<string, []>("op_10697_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10697_dilations_0 = const()[name = tensor<string, []>("op_10697_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10697_groups_0 = const()[name = tensor<string, []>("op_10697_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101820352)))];
            tensor<fp16, [1, 64, 1, 197]> var_10697_cast_fp16 = conv(dilations = var_10697_dilations_0, groups = var_10697_groups_0, pad = var_10697_pad_0, pad_type = var_10697_pad_type_0, strides = var_10697_strides_0, weight = model_blocks_7_attn_k_projs_3_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10697_cast_fp16")];
            tensor<int32, [1]> var_10698_axes_0 = const()[name = tensor<string, []>("op_10698_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10698_cast_fp16 = squeeze(axes = var_10698_axes_0, x = var_10697_cast_fp16)[name = tensor<string, []>("op_10698_cast_fp16")];
            tensor<int32, [3]> k_523_perm_0 = const()[name = tensor<string, []>("k_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10705_pad_type_0 = const()[name = tensor<string, []>("op_10705_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10705_strides_0 = const()[name = tensor<string, []>("op_10705_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10705_pad_0 = const()[name = tensor<string, []>("op_10705_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10705_dilations_0 = const()[name = tensor<string, []>("op_10705_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10705_groups_0 = const()[name = tensor<string, []>("op_10705_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101918720)))];
            tensor<fp16, [1, 64, 1, 197]> var_10705_cast_fp16 = conv(dilations = var_10705_dilations_0, groups = var_10705_groups_0, pad = var_10705_pad_0, pad_type = var_10705_pad_type_0, strides = var_10705_strides_0, weight = model_blocks_7_attn_v_projs_3_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10705_cast_fp16")];
            tensor<int32, [1]> var_10706_axes_0 = const()[name = tensor<string, []>("op_10706_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10706_cast_fp16 = squeeze(axes = var_10706_axes_0, x = var_10705_cast_fp16)[name = tensor<string, []>("op_10706_cast_fp16")];
            tensor<int32, [3]> v_523_perm_0 = const()[name = tensor<string, []>("v_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102017088)))];
            tensor<fp16, [1, 197, 64]> q_871_cast_fp16 = transpose(perm = q_871_perm_0, x = var_10690_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [1, 197, 64]> q_873_cast_fp16 = add(x = q_871_cast_fp16, y = model_blocks_7_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_873_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102017280)))];
            tensor<fp16, [1, 197, 64]> v_523_cast_fp16 = transpose(perm = v_523_perm_0, x = var_10706_cast_fp16)[name = tensor<string, []>("transpose_202")];
            tensor<fp16, [1, 197, 64]> v_525_cast_fp16 = add(x = v_523_cast_fp16, y = model_blocks_7_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_525_cast_fp16")];
            tensor<int32, [1]> q_875_axes_0 = const()[name = tensor<string, []>("q_875_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_875_cast_fp16 = expand_dims(axes = q_875_axes_0, x = q_873_cast_fp16)[name = tensor<string, []>("q_875_cast_fp16")];
            tensor<int32, [1]> k_525_axes_0 = const()[name = tensor<string, []>("k_525_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_523_cast_fp16 = transpose(perm = k_523_perm_0, x = var_10698_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 1, 197, 64]> k_525_cast_fp16 = expand_dims(axes = k_525_axes_0, x = k_523_cast_fp16)[name = tensor<string, []>("k_525_cast_fp16")];
            tensor<int32, [1]> v_527_axes_0 = const()[name = tensor<string, []>("v_527_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_527_cast_fp16 = expand_dims(axes = v_527_axes_0, x = v_525_cast_fp16)[name = tensor<string, []>("v_527_cast_fp16")];
            tensor<int32, [4]> var_10715_begin_0 = const()[name = tensor<string, []>("op_10715_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10715_end_0 = const()[name = tensor<string, []>("op_10715_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10715_end_mask_0 = const()[name = tensor<string, []>("op_10715_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10715_cast_fp16 = slice_by_index(begin = var_10715_begin_0, end = var_10715_end_0, end_mask = var_10715_end_mask_0, x = q_875_cast_fp16)[name = tensor<string, []>("op_10715_cast_fp16")];
            tensor<int32, [4]> var_10719_begin_0 = const()[name = tensor<string, []>("op_10719_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10719_end_0 = const()[name = tensor<string, []>("op_10719_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10719_end_mask_0 = const()[name = tensor<string, []>("op_10719_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10719_cast_fp16 = slice_by_index(begin = var_10719_begin_0, end = var_10719_end_0, end_mask = var_10719_end_mask_0, x = k_525_cast_fp16)[name = tensor<string, []>("op_10719_cast_fp16")];
            tensor<int32, [4]> var_10723_begin_0 = const()[name = tensor<string, []>("op_10723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10723_end_0 = const()[name = tensor<string, []>("op_10723_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10723_end_mask_0 = const()[name = tensor<string, []>("op_10723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10723_cast_fp16 = slice_by_index(begin = var_10723_begin_0, end = var_10723_end_0, end_mask = var_10723_end_mask_0, x = q_875_cast_fp16)[name = tensor<string, []>("op_10723_cast_fp16")];
            tensor<int32, [4]> var_10727_begin_0 = const()[name = tensor<string, []>("op_10727_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10727_end_0 = const()[name = tensor<string, []>("op_10727_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10727_end_mask_0 = const()[name = tensor<string, []>("op_10727_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10727_cast_fp16 = slice_by_index(begin = var_10727_begin_0, end = var_10727_end_0, end_mask = var_10727_end_mask_0, x = k_525_cast_fp16)[name = tensor<string, []>("op_10727_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10729_cast_fp16 = mul(x = var_10723_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10729_cast_fp16")];
            tensor<int32, [5]> var_10737 = const()[name = tensor<string, []>("op_10737"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1195_cast_fp16 = reshape(shape = var_10737, x = var_10723_cast_fp16)[name = tensor<string, []>("x_1195_cast_fp16")];
            tensor<int32, [2]> var_10739_split_sizes_0 = const()[name = tensor<string, []>("op_10739_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10739_axis_0 = const()[name = tensor<string, []>("op_10739_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10739_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10739_cast_fp16_1 = split(axis = var_10739_axis_0, split_sizes = var_10739_split_sizes_0, x = x_1195_cast_fp16)[name = tensor<string, []>("op_10739_cast_fp16")];
            tensor<int32, [1]> squeeze_348_axes_0 = const()[name = tensor<string, []>("squeeze_348_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_348_cast_fp16 = squeeze(axes = squeeze_348_axes_0, x = var_10739_cast_fp16_0)[name = tensor<string, []>("squeeze_348_cast_fp16")];
            tensor<int32, [1]> squeeze_349_axes_0 = const()[name = tensor<string, []>("squeeze_349_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_349_cast_fp16 = squeeze(axes = squeeze_349_axes_0, x = var_10739_cast_fp16_1)[name = tensor<string, []>("squeeze_349_cast_fp16")];
            tensor<fp16, []> const_1427_promoted_to_fp16 = const()[name = tensor<string, []>("const_1427_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10742_cast_fp16 = mul(x = squeeze_349_cast_fp16, y = const_1427_promoted_to_fp16)[name = tensor<string, []>("op_10742_cast_fp16")];
            tensor<int32, []> x_1197_axis_0 = const()[name = tensor<string, []>("x_1197_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1197_cast_fp16 = stack(axis = x_1197_axis_0, values = (var_10742_cast_fp16, squeeze_348_cast_fp16))[name = tensor<string, []>("x_1197_cast_fp16")];
            tensor<int32, [4]> var_10748 = const()[name = tensor<string, []>("op_10748"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10749_cast_fp16 = reshape(shape = var_10748, x = x_1197_cast_fp16)[name = tensor<string, []>("op_10749_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10750_cast_fp16 = mul(x = var_10749_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10750_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_175_cast_fp16 = add(x = var_10729_cast_fp16, y = var_10750_cast_fp16)[name = tensor<string, []>("q_patches_175_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10752_cast_fp16 = mul(x = var_10727_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10752_cast_fp16")];
            tensor<int32, [5]> var_10760 = const()[name = tensor<string, []>("op_10760"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1201_cast_fp16 = reshape(shape = var_10760, x = var_10727_cast_fp16)[name = tensor<string, []>("x_1201_cast_fp16")];
            tensor<int32, [2]> var_10762_split_sizes_0 = const()[name = tensor<string, []>("op_10762_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10762_axis_0 = const()[name = tensor<string, []>("op_10762_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10762_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10762_cast_fp16_1 = split(axis = var_10762_axis_0, split_sizes = var_10762_split_sizes_0, x = x_1201_cast_fp16)[name = tensor<string, []>("op_10762_cast_fp16")];
            tensor<int32, [1]> squeeze_350_axes_0 = const()[name = tensor<string, []>("squeeze_350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_350_cast_fp16 = squeeze(axes = squeeze_350_axes_0, x = var_10762_cast_fp16_0)[name = tensor<string, []>("squeeze_350_cast_fp16")];
            tensor<int32, [1]> squeeze_351_axes_0 = const()[name = tensor<string, []>("squeeze_351_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_351_cast_fp16 = squeeze(axes = squeeze_351_axes_0, x = var_10762_cast_fp16_1)[name = tensor<string, []>("squeeze_351_cast_fp16")];
            tensor<fp16, []> const_1435_promoted_to_fp16 = const()[name = tensor<string, []>("const_1435_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10765_cast_fp16 = mul(x = squeeze_351_cast_fp16, y = const_1435_promoted_to_fp16)[name = tensor<string, []>("op_10765_cast_fp16")];
            tensor<int32, []> x_1203_axis_0 = const()[name = tensor<string, []>("x_1203_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1203_cast_fp16 = stack(axis = x_1203_axis_0, values = (var_10765_cast_fp16, squeeze_350_cast_fp16))[name = tensor<string, []>("x_1203_cast_fp16")];
            tensor<int32, [4]> var_10771 = const()[name = tensor<string, []>("op_10771"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10772_cast_fp16 = reshape(shape = var_10771, x = x_1203_cast_fp16)[name = tensor<string, []>("op_10772_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10773_cast_fp16 = mul(x = var_10772_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10773_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_175_cast_fp16 = add(x = var_10752_cast_fp16, y = var_10773_cast_fp16)[name = tensor<string, []>("k_patches_175_cast_fp16")];
            tensor<bool, []> var_10776_interleave_0 = const()[name = tensor<string, []>("op_10776_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10776_cast_fp16 = concat(axis = var_20, interleave = var_10776_interleave_0, values = (var_10715_cast_fp16, q_patches_175_cast_fp16))[name = tensor<string, []>("op_10776_cast_fp16")];
            tensor<bool, []> var_10779_interleave_0 = const()[name = tensor<string, []>("op_10779_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10779_cast_fp16 = concat(axis = var_20, interleave = var_10779_interleave_0, values = (var_10719_cast_fp16, k_patches_175_cast_fp16))[name = tensor<string, []>("op_10779_cast_fp16")];
            tensor<fp16, []> var_10781_to_fp16 = const()[name = tensor<string, []>("op_10781_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_879_cast_fp16 = mul(x = var_10776_cast_fp16, y = var_10781_to_fp16)[name = tensor<string, []>("q_879_cast_fp16")];
            tensor<bool, []> attn_349_transpose_x_1 = const()[name = tensor<string, []>("attn_349_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_349_transpose_y_1 = const()[name = tensor<string, []>("attn_349_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_349_cast_fp16 = matmul(transpose_x = attn_349_transpose_x_1, transpose_y = attn_349_transpose_y_1, x = q_879_cast_fp16, y = var_10779_cast_fp16)[name = tensor<string, []>("attn_349_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_351_cast_fp16 = softmax(axis = var_21, x = attn_349_cast_fp16)[name = tensor<string, []>("attn_351_cast_fp16")];
            tensor<bool, []> var_10786_transpose_x_0 = const()[name = tensor<string, []>("op_10786_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10786_transpose_y_0 = const()[name = tensor<string, []>("op_10786_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10786_cast_fp16 = matmul(transpose_x = var_10786_transpose_x_0, transpose_y = var_10786_transpose_y_0, x = attn_351_cast_fp16, y = v_527_cast_fp16)[name = tensor<string, []>("op_10786_cast_fp16")];
            tensor<int32, [1]> var_10787_axes_0 = const()[name = tensor<string, []>("op_10787_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10787_cast_fp16 = squeeze(axes = var_10787_axes_0, x = var_10786_cast_fp16)[name = tensor<string, []>("op_10787_cast_fp16")];
            tensor<string, []> var_10793_pad_type_0 = const()[name = tensor<string, []>("op_10793_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10793_strides_0 = const()[name = tensor<string, []>("op_10793_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10793_pad_0 = const()[name = tensor<string, []>("op_10793_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10793_dilations_0 = const()[name = tensor<string, []>("op_10793_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10793_groups_0 = const()[name = tensor<string, []>("op_10793_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102017472)))];
            tensor<fp16, [1, 64, 1, 197]> var_10793_cast_fp16 = conv(dilations = var_10793_dilations_0, groups = var_10793_groups_0, pad = var_10793_pad_0, pad_type = var_10793_pad_type_0, strides = var_10793_strides_0, weight = model_blocks_7_attn_q_projs_4_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10793_cast_fp16")];
            tensor<int32, [1]> var_10794_axes_0 = const()[name = tensor<string, []>("op_10794_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10794_cast_fp16 = squeeze(axes = var_10794_axes_0, x = var_10793_cast_fp16)[name = tensor<string, []>("op_10794_cast_fp16")];
            tensor<int32, [3]> q_881_perm_0 = const()[name = tensor<string, []>("q_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10801_pad_type_0 = const()[name = tensor<string, []>("op_10801_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10801_strides_0 = const()[name = tensor<string, []>("op_10801_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10801_pad_0 = const()[name = tensor<string, []>("op_10801_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10801_dilations_0 = const()[name = tensor<string, []>("op_10801_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10801_groups_0 = const()[name = tensor<string, []>("op_10801_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102115840)))];
            tensor<fp16, [1, 64, 1, 197]> var_10801_cast_fp16 = conv(dilations = var_10801_dilations_0, groups = var_10801_groups_0, pad = var_10801_pad_0, pad_type = var_10801_pad_type_0, strides = var_10801_strides_0, weight = model_blocks_7_attn_k_projs_4_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10801_cast_fp16")];
            tensor<int32, [1]> var_10802_axes_0 = const()[name = tensor<string, []>("op_10802_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10802_cast_fp16 = squeeze(axes = var_10802_axes_0, x = var_10801_cast_fp16)[name = tensor<string, []>("op_10802_cast_fp16")];
            tensor<int32, [3]> k_529_perm_0 = const()[name = tensor<string, []>("k_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10809_pad_type_0 = const()[name = tensor<string, []>("op_10809_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10809_strides_0 = const()[name = tensor<string, []>("op_10809_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10809_pad_0 = const()[name = tensor<string, []>("op_10809_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10809_dilations_0 = const()[name = tensor<string, []>("op_10809_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10809_groups_0 = const()[name = tensor<string, []>("op_10809_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102214208)))];
            tensor<fp16, [1, 64, 1, 197]> var_10809_cast_fp16 = conv(dilations = var_10809_dilations_0, groups = var_10809_groups_0, pad = var_10809_pad_0, pad_type = var_10809_pad_type_0, strides = var_10809_strides_0, weight = model_blocks_7_attn_v_projs_4_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10809_cast_fp16")];
            tensor<int32, [1]> var_10810_axes_0 = const()[name = tensor<string, []>("op_10810_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10810_cast_fp16 = squeeze(axes = var_10810_axes_0, x = var_10809_cast_fp16)[name = tensor<string, []>("op_10810_cast_fp16")];
            tensor<int32, [3]> v_529_perm_0 = const()[name = tensor<string, []>("v_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102312576)))];
            tensor<fp16, [1, 197, 64]> q_881_cast_fp16 = transpose(perm = q_881_perm_0, x = var_10794_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 197, 64]> q_883_cast_fp16 = add(x = q_881_cast_fp16, y = model_blocks_7_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_883_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102312768)))];
            tensor<fp16, [1, 197, 64]> v_529_cast_fp16 = transpose(perm = v_529_perm_0, x = var_10810_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [1, 197, 64]> v_531_cast_fp16 = add(x = v_529_cast_fp16, y = model_blocks_7_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_531_cast_fp16")];
            tensor<int32, [1]> q_885_axes_0 = const()[name = tensor<string, []>("q_885_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_885_cast_fp16 = expand_dims(axes = q_885_axes_0, x = q_883_cast_fp16)[name = tensor<string, []>("q_885_cast_fp16")];
            tensor<int32, [1]> k_531_axes_0 = const()[name = tensor<string, []>("k_531_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_529_cast_fp16 = transpose(perm = k_529_perm_0, x = var_10802_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [1, 1, 197, 64]> k_531_cast_fp16 = expand_dims(axes = k_531_axes_0, x = k_529_cast_fp16)[name = tensor<string, []>("k_531_cast_fp16")];
            tensor<int32, [1]> v_533_axes_0 = const()[name = tensor<string, []>("v_533_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_533_cast_fp16 = expand_dims(axes = v_533_axes_0, x = v_531_cast_fp16)[name = tensor<string, []>("v_533_cast_fp16")];
            tensor<int32, [4]> var_10819_begin_0 = const()[name = tensor<string, []>("op_10819_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10819_end_0 = const()[name = tensor<string, []>("op_10819_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10819_end_mask_0 = const()[name = tensor<string, []>("op_10819_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10819_cast_fp16 = slice_by_index(begin = var_10819_begin_0, end = var_10819_end_0, end_mask = var_10819_end_mask_0, x = q_885_cast_fp16)[name = tensor<string, []>("op_10819_cast_fp16")];
            tensor<int32, [4]> var_10823_begin_0 = const()[name = tensor<string, []>("op_10823_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10823_end_0 = const()[name = tensor<string, []>("op_10823_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10823_end_mask_0 = const()[name = tensor<string, []>("op_10823_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10823_cast_fp16 = slice_by_index(begin = var_10823_begin_0, end = var_10823_end_0, end_mask = var_10823_end_mask_0, x = k_531_cast_fp16)[name = tensor<string, []>("op_10823_cast_fp16")];
            tensor<int32, [4]> var_10827_begin_0 = const()[name = tensor<string, []>("op_10827_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10827_end_0 = const()[name = tensor<string, []>("op_10827_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10827_end_mask_0 = const()[name = tensor<string, []>("op_10827_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10827_cast_fp16 = slice_by_index(begin = var_10827_begin_0, end = var_10827_end_0, end_mask = var_10827_end_mask_0, x = q_885_cast_fp16)[name = tensor<string, []>("op_10827_cast_fp16")];
            tensor<int32, [4]> var_10831_begin_0 = const()[name = tensor<string, []>("op_10831_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10831_end_0 = const()[name = tensor<string, []>("op_10831_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10831_end_mask_0 = const()[name = tensor<string, []>("op_10831_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10831_cast_fp16 = slice_by_index(begin = var_10831_begin_0, end = var_10831_end_0, end_mask = var_10831_end_mask_0, x = k_531_cast_fp16)[name = tensor<string, []>("op_10831_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10833_cast_fp16 = mul(x = var_10827_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10833_cast_fp16")];
            tensor<int32, [5]> var_10841 = const()[name = tensor<string, []>("op_10841"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1207_cast_fp16 = reshape(shape = var_10841, x = var_10827_cast_fp16)[name = tensor<string, []>("x_1207_cast_fp16")];
            tensor<int32, [2]> var_10843_split_sizes_0 = const()[name = tensor<string, []>("op_10843_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10843_axis_0 = const()[name = tensor<string, []>("op_10843_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10843_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10843_cast_fp16_1 = split(axis = var_10843_axis_0, split_sizes = var_10843_split_sizes_0, x = x_1207_cast_fp16)[name = tensor<string, []>("op_10843_cast_fp16")];
            tensor<int32, [1]> squeeze_352_axes_0 = const()[name = tensor<string, []>("squeeze_352_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_352_cast_fp16 = squeeze(axes = squeeze_352_axes_0, x = var_10843_cast_fp16_0)[name = tensor<string, []>("squeeze_352_cast_fp16")];
            tensor<int32, [1]> squeeze_353_axes_0 = const()[name = tensor<string, []>("squeeze_353_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_353_cast_fp16 = squeeze(axes = squeeze_353_axes_0, x = var_10843_cast_fp16_1)[name = tensor<string, []>("squeeze_353_cast_fp16")];
            tensor<fp16, []> const_1443_promoted_to_fp16 = const()[name = tensor<string, []>("const_1443_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10846_cast_fp16 = mul(x = squeeze_353_cast_fp16, y = const_1443_promoted_to_fp16)[name = tensor<string, []>("op_10846_cast_fp16")];
            tensor<int32, []> x_1209_axis_0 = const()[name = tensor<string, []>("x_1209_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1209_cast_fp16 = stack(axis = x_1209_axis_0, values = (var_10846_cast_fp16, squeeze_352_cast_fp16))[name = tensor<string, []>("x_1209_cast_fp16")];
            tensor<int32, [4]> var_10852 = const()[name = tensor<string, []>("op_10852"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10853_cast_fp16 = reshape(shape = var_10852, x = x_1209_cast_fp16)[name = tensor<string, []>("op_10853_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10854_cast_fp16 = mul(x = var_10853_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10854_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_177_cast_fp16 = add(x = var_10833_cast_fp16, y = var_10854_cast_fp16)[name = tensor<string, []>("q_patches_177_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10856_cast_fp16 = mul(x = var_10831_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10856_cast_fp16")];
            tensor<int32, [5]> var_10864 = const()[name = tensor<string, []>("op_10864"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1213_cast_fp16 = reshape(shape = var_10864, x = var_10831_cast_fp16)[name = tensor<string, []>("x_1213_cast_fp16")];
            tensor<int32, [2]> var_10866_split_sizes_0 = const()[name = tensor<string, []>("op_10866_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10866_axis_0 = const()[name = tensor<string, []>("op_10866_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10866_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10866_cast_fp16_1 = split(axis = var_10866_axis_0, split_sizes = var_10866_split_sizes_0, x = x_1213_cast_fp16)[name = tensor<string, []>("op_10866_cast_fp16")];
            tensor<int32, [1]> squeeze_354_axes_0 = const()[name = tensor<string, []>("squeeze_354_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_354_cast_fp16 = squeeze(axes = squeeze_354_axes_0, x = var_10866_cast_fp16_0)[name = tensor<string, []>("squeeze_354_cast_fp16")];
            tensor<int32, [1]> squeeze_355_axes_0 = const()[name = tensor<string, []>("squeeze_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_355_cast_fp16 = squeeze(axes = squeeze_355_axes_0, x = var_10866_cast_fp16_1)[name = tensor<string, []>("squeeze_355_cast_fp16")];
            tensor<fp16, []> const_1451_promoted_to_fp16 = const()[name = tensor<string, []>("const_1451_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10869_cast_fp16 = mul(x = squeeze_355_cast_fp16, y = const_1451_promoted_to_fp16)[name = tensor<string, []>("op_10869_cast_fp16")];
            tensor<int32, []> x_1215_axis_0 = const()[name = tensor<string, []>("x_1215_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1215_cast_fp16 = stack(axis = x_1215_axis_0, values = (var_10869_cast_fp16, squeeze_354_cast_fp16))[name = tensor<string, []>("x_1215_cast_fp16")];
            tensor<int32, [4]> var_10875 = const()[name = tensor<string, []>("op_10875"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10876_cast_fp16 = reshape(shape = var_10875, x = x_1215_cast_fp16)[name = tensor<string, []>("op_10876_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10877_cast_fp16 = mul(x = var_10876_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10877_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_177_cast_fp16 = add(x = var_10856_cast_fp16, y = var_10877_cast_fp16)[name = tensor<string, []>("k_patches_177_cast_fp16")];
            tensor<bool, []> var_10880_interleave_0 = const()[name = tensor<string, []>("op_10880_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10880_cast_fp16 = concat(axis = var_20, interleave = var_10880_interleave_0, values = (var_10819_cast_fp16, q_patches_177_cast_fp16))[name = tensor<string, []>("op_10880_cast_fp16")];
            tensor<bool, []> var_10883_interleave_0 = const()[name = tensor<string, []>("op_10883_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10883_cast_fp16 = concat(axis = var_20, interleave = var_10883_interleave_0, values = (var_10823_cast_fp16, k_patches_177_cast_fp16))[name = tensor<string, []>("op_10883_cast_fp16")];
            tensor<fp16, []> var_10885_to_fp16 = const()[name = tensor<string, []>("op_10885_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_889_cast_fp16 = mul(x = var_10880_cast_fp16, y = var_10885_to_fp16)[name = tensor<string, []>("q_889_cast_fp16")];
            tensor<bool, []> attn_353_transpose_x_1 = const()[name = tensor<string, []>("attn_353_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_353_transpose_y_1 = const()[name = tensor<string, []>("attn_353_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_353_cast_fp16 = matmul(transpose_x = attn_353_transpose_x_1, transpose_y = attn_353_transpose_y_1, x = q_889_cast_fp16, y = var_10883_cast_fp16)[name = tensor<string, []>("attn_353_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_355_cast_fp16 = softmax(axis = var_21, x = attn_353_cast_fp16)[name = tensor<string, []>("attn_355_cast_fp16")];
            tensor<bool, []> var_10890_transpose_x_0 = const()[name = tensor<string, []>("op_10890_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10890_transpose_y_0 = const()[name = tensor<string, []>("op_10890_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10890_cast_fp16 = matmul(transpose_x = var_10890_transpose_x_0, transpose_y = var_10890_transpose_y_0, x = attn_355_cast_fp16, y = v_533_cast_fp16)[name = tensor<string, []>("op_10890_cast_fp16")];
            tensor<int32, [1]> var_10891_axes_0 = const()[name = tensor<string, []>("op_10891_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10891_cast_fp16 = squeeze(axes = var_10891_axes_0, x = var_10890_cast_fp16)[name = tensor<string, []>("op_10891_cast_fp16")];
            tensor<string, []> var_10897_pad_type_0 = const()[name = tensor<string, []>("op_10897_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10897_strides_0 = const()[name = tensor<string, []>("op_10897_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10897_pad_0 = const()[name = tensor<string, []>("op_10897_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10897_dilations_0 = const()[name = tensor<string, []>("op_10897_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10897_groups_0 = const()[name = tensor<string, []>("op_10897_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102312960)))];
            tensor<fp16, [1, 64, 1, 197]> var_10897_cast_fp16 = conv(dilations = var_10897_dilations_0, groups = var_10897_groups_0, pad = var_10897_pad_0, pad_type = var_10897_pad_type_0, strides = var_10897_strides_0, weight = model_blocks_7_attn_q_projs_5_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10897_cast_fp16")];
            tensor<int32, [1]> var_10898_axes_0 = const()[name = tensor<string, []>("op_10898_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10898_cast_fp16 = squeeze(axes = var_10898_axes_0, x = var_10897_cast_fp16)[name = tensor<string, []>("op_10898_cast_fp16")];
            tensor<int32, [3]> q_891_perm_0 = const()[name = tensor<string, []>("q_891_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10905_pad_type_0 = const()[name = tensor<string, []>("op_10905_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10905_strides_0 = const()[name = tensor<string, []>("op_10905_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10905_pad_0 = const()[name = tensor<string, []>("op_10905_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10905_dilations_0 = const()[name = tensor<string, []>("op_10905_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10905_groups_0 = const()[name = tensor<string, []>("op_10905_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102411328)))];
            tensor<fp16, [1, 64, 1, 197]> var_10905_cast_fp16 = conv(dilations = var_10905_dilations_0, groups = var_10905_groups_0, pad = var_10905_pad_0, pad_type = var_10905_pad_type_0, strides = var_10905_strides_0, weight = model_blocks_7_attn_k_projs_5_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10905_cast_fp16")];
            tensor<int32, [1]> var_10906_axes_0 = const()[name = tensor<string, []>("op_10906_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10906_cast_fp16 = squeeze(axes = var_10906_axes_0, x = var_10905_cast_fp16)[name = tensor<string, []>("op_10906_cast_fp16")];
            tensor<int32, [3]> k_535_perm_0 = const()[name = tensor<string, []>("k_535_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_10913_pad_type_0 = const()[name = tensor<string, []>("op_10913_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_10913_strides_0 = const()[name = tensor<string, []>("op_10913_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10913_pad_0 = const()[name = tensor<string, []>("op_10913_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10913_dilations_0 = const()[name = tensor<string, []>("op_10913_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10913_groups_0 = const()[name = tensor<string, []>("op_10913_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102509696)))];
            tensor<fp16, [1, 64, 1, 197]> var_10913_cast_fp16 = conv(dilations = var_10913_dilations_0, groups = var_10913_groups_0, pad = var_10913_pad_0, pad_type = var_10913_pad_type_0, strides = var_10913_strides_0, weight = model_blocks_7_attn_v_projs_5_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_10913_cast_fp16")];
            tensor<int32, [1]> var_10914_axes_0 = const()[name = tensor<string, []>("op_10914_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_10914_cast_fp16 = squeeze(axes = var_10914_axes_0, x = var_10913_cast_fp16)[name = tensor<string, []>("op_10914_cast_fp16")];
            tensor<int32, [3]> v_535_perm_0 = const()[name = tensor<string, []>("v_535_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102608064)))];
            tensor<fp16, [1, 197, 64]> q_891_cast_fp16 = transpose(perm = q_891_perm_0, x = var_10898_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [1, 197, 64]> q_893_cast_fp16 = add(x = q_891_cast_fp16, y = model_blocks_7_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_893_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102608256)))];
            tensor<fp16, [1, 197, 64]> v_535_cast_fp16 = transpose(perm = v_535_perm_0, x = var_10914_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 197, 64]> v_537_cast_fp16 = add(x = v_535_cast_fp16, y = model_blocks_7_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_537_cast_fp16")];
            tensor<int32, [1]> q_895_axes_0 = const()[name = tensor<string, []>("q_895_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_895_cast_fp16 = expand_dims(axes = q_895_axes_0, x = q_893_cast_fp16)[name = tensor<string, []>("q_895_cast_fp16")];
            tensor<int32, [1]> k_537_axes_0 = const()[name = tensor<string, []>("k_537_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_535_cast_fp16 = transpose(perm = k_535_perm_0, x = var_10906_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [1, 1, 197, 64]> k_537_cast_fp16 = expand_dims(axes = k_537_axes_0, x = k_535_cast_fp16)[name = tensor<string, []>("k_537_cast_fp16")];
            tensor<int32, [1]> v_539_axes_0 = const()[name = tensor<string, []>("v_539_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_539_cast_fp16 = expand_dims(axes = v_539_axes_0, x = v_537_cast_fp16)[name = tensor<string, []>("v_539_cast_fp16")];
            tensor<int32, [4]> var_10923_begin_0 = const()[name = tensor<string, []>("op_10923_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10923_end_0 = const()[name = tensor<string, []>("op_10923_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10923_end_mask_0 = const()[name = tensor<string, []>("op_10923_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10923_cast_fp16 = slice_by_index(begin = var_10923_begin_0, end = var_10923_end_0, end_mask = var_10923_end_mask_0, x = q_895_cast_fp16)[name = tensor<string, []>("op_10923_cast_fp16")];
            tensor<int32, [4]> var_10927_begin_0 = const()[name = tensor<string, []>("op_10927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10927_end_0 = const()[name = tensor<string, []>("op_10927_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_10927_end_mask_0 = const()[name = tensor<string, []>("op_10927_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_10927_cast_fp16 = slice_by_index(begin = var_10927_begin_0, end = var_10927_end_0, end_mask = var_10927_end_mask_0, x = k_537_cast_fp16)[name = tensor<string, []>("op_10927_cast_fp16")];
            tensor<int32, [4]> var_10931_begin_0 = const()[name = tensor<string, []>("op_10931_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10931_end_0 = const()[name = tensor<string, []>("op_10931_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10931_end_mask_0 = const()[name = tensor<string, []>("op_10931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10931_cast_fp16 = slice_by_index(begin = var_10931_begin_0, end = var_10931_end_0, end_mask = var_10931_end_mask_0, x = q_895_cast_fp16)[name = tensor<string, []>("op_10931_cast_fp16")];
            tensor<int32, [4]> var_10935_begin_0 = const()[name = tensor<string, []>("op_10935_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10935_end_0 = const()[name = tensor<string, []>("op_10935_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_10935_end_mask_0 = const()[name = tensor<string, []>("op_10935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_10935_cast_fp16 = slice_by_index(begin = var_10935_begin_0, end = var_10935_end_0, end_mask = var_10935_end_mask_0, x = k_537_cast_fp16)[name = tensor<string, []>("op_10935_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10937_cast_fp16 = mul(x = var_10931_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10937_cast_fp16")];
            tensor<int32, [5]> var_10945 = const()[name = tensor<string, []>("op_10945"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1219_cast_fp16 = reshape(shape = var_10945, x = var_10931_cast_fp16)[name = tensor<string, []>("x_1219_cast_fp16")];
            tensor<int32, [2]> var_10947_split_sizes_0 = const()[name = tensor<string, []>("op_10947_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10947_axis_0 = const()[name = tensor<string, []>("op_10947_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10947_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10947_cast_fp16_1 = split(axis = var_10947_axis_0, split_sizes = var_10947_split_sizes_0, x = x_1219_cast_fp16)[name = tensor<string, []>("op_10947_cast_fp16")];
            tensor<int32, [1]> squeeze_356_axes_0 = const()[name = tensor<string, []>("squeeze_356_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_356_cast_fp16 = squeeze(axes = squeeze_356_axes_0, x = var_10947_cast_fp16_0)[name = tensor<string, []>("squeeze_356_cast_fp16")];
            tensor<int32, [1]> squeeze_357_axes_0 = const()[name = tensor<string, []>("squeeze_357_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_357_cast_fp16 = squeeze(axes = squeeze_357_axes_0, x = var_10947_cast_fp16_1)[name = tensor<string, []>("squeeze_357_cast_fp16")];
            tensor<fp16, []> const_1459_promoted_to_fp16 = const()[name = tensor<string, []>("const_1459_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10950_cast_fp16 = mul(x = squeeze_357_cast_fp16, y = const_1459_promoted_to_fp16)[name = tensor<string, []>("op_10950_cast_fp16")];
            tensor<int32, []> x_1221_axis_0 = const()[name = tensor<string, []>("x_1221_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1221_cast_fp16 = stack(axis = x_1221_axis_0, values = (var_10950_cast_fp16, squeeze_356_cast_fp16))[name = tensor<string, []>("x_1221_cast_fp16")];
            tensor<int32, [4]> var_10956 = const()[name = tensor<string, []>("op_10956"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10957_cast_fp16 = reshape(shape = var_10956, x = x_1221_cast_fp16)[name = tensor<string, []>("op_10957_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10958_cast_fp16 = mul(x = var_10957_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10958_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_179_cast_fp16 = add(x = var_10937_cast_fp16, y = var_10958_cast_fp16)[name = tensor<string, []>("q_patches_179_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10960_cast_fp16 = mul(x = var_10935_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_10960_cast_fp16")];
            tensor<int32, [5]> var_10968 = const()[name = tensor<string, []>("op_10968"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1225_cast_fp16 = reshape(shape = var_10968, x = var_10935_cast_fp16)[name = tensor<string, []>("x_1225_cast_fp16")];
            tensor<int32, [2]> var_10970_split_sizes_0 = const()[name = tensor<string, []>("op_10970_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_10970_axis_0 = const()[name = tensor<string, []>("op_10970_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_10970_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_10970_cast_fp16_1 = split(axis = var_10970_axis_0, split_sizes = var_10970_split_sizes_0, x = x_1225_cast_fp16)[name = tensor<string, []>("op_10970_cast_fp16")];
            tensor<int32, [1]> squeeze_358_axes_0 = const()[name = tensor<string, []>("squeeze_358_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_358_cast_fp16 = squeeze(axes = squeeze_358_axes_0, x = var_10970_cast_fp16_0)[name = tensor<string, []>("squeeze_358_cast_fp16")];
            tensor<int32, [1]> squeeze_359_axes_0 = const()[name = tensor<string, []>("squeeze_359_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_359_cast_fp16 = squeeze(axes = squeeze_359_axes_0, x = var_10970_cast_fp16_1)[name = tensor<string, []>("squeeze_359_cast_fp16")];
            tensor<fp16, []> const_1467_promoted_to_fp16 = const()[name = tensor<string, []>("const_1467_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_10973_cast_fp16 = mul(x = squeeze_359_cast_fp16, y = const_1467_promoted_to_fp16)[name = tensor<string, []>("op_10973_cast_fp16")];
            tensor<int32, []> x_1227_axis_0 = const()[name = tensor<string, []>("x_1227_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1227_cast_fp16 = stack(axis = x_1227_axis_0, values = (var_10973_cast_fp16, squeeze_358_cast_fp16))[name = tensor<string, []>("x_1227_cast_fp16")];
            tensor<int32, [4]> var_10979 = const()[name = tensor<string, []>("op_10979"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_10980_cast_fp16 = reshape(shape = var_10979, x = x_1227_cast_fp16)[name = tensor<string, []>("op_10980_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_10981_cast_fp16 = mul(x = var_10980_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_10981_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_179_cast_fp16 = add(x = var_10960_cast_fp16, y = var_10981_cast_fp16)[name = tensor<string, []>("k_patches_179_cast_fp16")];
            tensor<bool, []> var_10984_interleave_0 = const()[name = tensor<string, []>("op_10984_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10984_cast_fp16 = concat(axis = var_20, interleave = var_10984_interleave_0, values = (var_10923_cast_fp16, q_patches_179_cast_fp16))[name = tensor<string, []>("op_10984_cast_fp16")];
            tensor<bool, []> var_10987_interleave_0 = const()[name = tensor<string, []>("op_10987_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10987_cast_fp16 = concat(axis = var_20, interleave = var_10987_interleave_0, values = (var_10927_cast_fp16, k_patches_179_cast_fp16))[name = tensor<string, []>("op_10987_cast_fp16")];
            tensor<fp16, []> var_10989_to_fp16 = const()[name = tensor<string, []>("op_10989_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_899_cast_fp16 = mul(x = var_10984_cast_fp16, y = var_10989_to_fp16)[name = tensor<string, []>("q_899_cast_fp16")];
            tensor<bool, []> attn_357_transpose_x_1 = const()[name = tensor<string, []>("attn_357_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_357_transpose_y_1 = const()[name = tensor<string, []>("attn_357_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_357_cast_fp16 = matmul(transpose_x = attn_357_transpose_x_1, transpose_y = attn_357_transpose_y_1, x = q_899_cast_fp16, y = var_10987_cast_fp16)[name = tensor<string, []>("attn_357_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_359_cast_fp16 = softmax(axis = var_21, x = attn_357_cast_fp16)[name = tensor<string, []>("attn_359_cast_fp16")];
            tensor<bool, []> var_10994_transpose_x_0 = const()[name = tensor<string, []>("op_10994_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_10994_transpose_y_0 = const()[name = tensor<string, []>("op_10994_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_10994_cast_fp16 = matmul(transpose_x = var_10994_transpose_x_0, transpose_y = var_10994_transpose_y_0, x = attn_359_cast_fp16, y = v_539_cast_fp16)[name = tensor<string, []>("op_10994_cast_fp16")];
            tensor<int32, [1]> var_10995_axes_0 = const()[name = tensor<string, []>("op_10995_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_10995_cast_fp16 = squeeze(axes = var_10995_axes_0, x = var_10994_cast_fp16)[name = tensor<string, []>("op_10995_cast_fp16")];
            tensor<string, []> var_11001_pad_type_0 = const()[name = tensor<string, []>("op_11001_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11001_strides_0 = const()[name = tensor<string, []>("op_11001_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11001_pad_0 = const()[name = tensor<string, []>("op_11001_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11001_dilations_0 = const()[name = tensor<string, []>("op_11001_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11001_groups_0 = const()[name = tensor<string, []>("op_11001_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102608448)))];
            tensor<fp16, [1, 64, 1, 197]> var_11001_cast_fp16 = conv(dilations = var_11001_dilations_0, groups = var_11001_groups_0, pad = var_11001_pad_0, pad_type = var_11001_pad_type_0, strides = var_11001_strides_0, weight = model_blocks_7_attn_q_projs_6_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11001_cast_fp16")];
            tensor<int32, [1]> var_11002_axes_0 = const()[name = tensor<string, []>("op_11002_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11002_cast_fp16 = squeeze(axes = var_11002_axes_0, x = var_11001_cast_fp16)[name = tensor<string, []>("op_11002_cast_fp16")];
            tensor<int32, [3]> q_901_perm_0 = const()[name = tensor<string, []>("q_901_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11009_pad_type_0 = const()[name = tensor<string, []>("op_11009_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11009_strides_0 = const()[name = tensor<string, []>("op_11009_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11009_pad_0 = const()[name = tensor<string, []>("op_11009_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11009_dilations_0 = const()[name = tensor<string, []>("op_11009_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11009_groups_0 = const()[name = tensor<string, []>("op_11009_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102706816)))];
            tensor<fp16, [1, 64, 1, 197]> var_11009_cast_fp16 = conv(dilations = var_11009_dilations_0, groups = var_11009_groups_0, pad = var_11009_pad_0, pad_type = var_11009_pad_type_0, strides = var_11009_strides_0, weight = model_blocks_7_attn_k_projs_6_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11009_cast_fp16")];
            tensor<int32, [1]> var_11010_axes_0 = const()[name = tensor<string, []>("op_11010_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11010_cast_fp16 = squeeze(axes = var_11010_axes_0, x = var_11009_cast_fp16)[name = tensor<string, []>("op_11010_cast_fp16")];
            tensor<int32, [3]> k_541_perm_0 = const()[name = tensor<string, []>("k_541_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11017_pad_type_0 = const()[name = tensor<string, []>("op_11017_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11017_strides_0 = const()[name = tensor<string, []>("op_11017_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11017_pad_0 = const()[name = tensor<string, []>("op_11017_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11017_dilations_0 = const()[name = tensor<string, []>("op_11017_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11017_groups_0 = const()[name = tensor<string, []>("op_11017_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102805184)))];
            tensor<fp16, [1, 64, 1, 197]> var_11017_cast_fp16 = conv(dilations = var_11017_dilations_0, groups = var_11017_groups_0, pad = var_11017_pad_0, pad_type = var_11017_pad_type_0, strides = var_11017_strides_0, weight = model_blocks_7_attn_v_projs_6_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11017_cast_fp16")];
            tensor<int32, [1]> var_11018_axes_0 = const()[name = tensor<string, []>("op_11018_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11018_cast_fp16 = squeeze(axes = var_11018_axes_0, x = var_11017_cast_fp16)[name = tensor<string, []>("op_11018_cast_fp16")];
            tensor<int32, [3]> v_541_perm_0 = const()[name = tensor<string, []>("v_541_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102903552)))];
            tensor<fp16, [1, 197, 64]> q_901_cast_fp16 = transpose(perm = q_901_perm_0, x = var_11002_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 197, 64]> q_903_cast_fp16 = add(x = q_901_cast_fp16, y = model_blocks_7_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_903_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102903744)))];
            tensor<fp16, [1, 197, 64]> v_541_cast_fp16 = transpose(perm = v_541_perm_0, x = var_11018_cast_fp16)[name = tensor<string, []>("transpose_193")];
            tensor<fp16, [1, 197, 64]> v_543_cast_fp16 = add(x = v_541_cast_fp16, y = model_blocks_7_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_543_cast_fp16")];
            tensor<int32, [1]> q_905_axes_0 = const()[name = tensor<string, []>("q_905_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_905_cast_fp16 = expand_dims(axes = q_905_axes_0, x = q_903_cast_fp16)[name = tensor<string, []>("q_905_cast_fp16")];
            tensor<int32, [1]> k_543_axes_0 = const()[name = tensor<string, []>("k_543_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_541_cast_fp16 = transpose(perm = k_541_perm_0, x = var_11010_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 1, 197, 64]> k_543_cast_fp16 = expand_dims(axes = k_543_axes_0, x = k_541_cast_fp16)[name = tensor<string, []>("k_543_cast_fp16")];
            tensor<int32, [1]> v_545_axes_0 = const()[name = tensor<string, []>("v_545_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_545_cast_fp16 = expand_dims(axes = v_545_axes_0, x = v_543_cast_fp16)[name = tensor<string, []>("v_545_cast_fp16")];
            tensor<int32, [4]> var_11027_begin_0 = const()[name = tensor<string, []>("op_11027_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11027_end_0 = const()[name = tensor<string, []>("op_11027_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11027_end_mask_0 = const()[name = tensor<string, []>("op_11027_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11027_cast_fp16 = slice_by_index(begin = var_11027_begin_0, end = var_11027_end_0, end_mask = var_11027_end_mask_0, x = q_905_cast_fp16)[name = tensor<string, []>("op_11027_cast_fp16")];
            tensor<int32, [4]> var_11031_begin_0 = const()[name = tensor<string, []>("op_11031_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11031_end_0 = const()[name = tensor<string, []>("op_11031_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11031_end_mask_0 = const()[name = tensor<string, []>("op_11031_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11031_cast_fp16 = slice_by_index(begin = var_11031_begin_0, end = var_11031_end_0, end_mask = var_11031_end_mask_0, x = k_543_cast_fp16)[name = tensor<string, []>("op_11031_cast_fp16")];
            tensor<int32, [4]> var_11035_begin_0 = const()[name = tensor<string, []>("op_11035_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11035_end_0 = const()[name = tensor<string, []>("op_11035_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11035_end_mask_0 = const()[name = tensor<string, []>("op_11035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11035_cast_fp16 = slice_by_index(begin = var_11035_begin_0, end = var_11035_end_0, end_mask = var_11035_end_mask_0, x = q_905_cast_fp16)[name = tensor<string, []>("op_11035_cast_fp16")];
            tensor<int32, [4]> var_11039_begin_0 = const()[name = tensor<string, []>("op_11039_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11039_end_0 = const()[name = tensor<string, []>("op_11039_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11039_end_mask_0 = const()[name = tensor<string, []>("op_11039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11039_cast_fp16 = slice_by_index(begin = var_11039_begin_0, end = var_11039_end_0, end_mask = var_11039_end_mask_0, x = k_543_cast_fp16)[name = tensor<string, []>("op_11039_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11041_cast_fp16 = mul(x = var_11035_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11041_cast_fp16")];
            tensor<int32, [5]> var_11049 = const()[name = tensor<string, []>("op_11049"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1231_cast_fp16 = reshape(shape = var_11049, x = var_11035_cast_fp16)[name = tensor<string, []>("x_1231_cast_fp16")];
            tensor<int32, [2]> var_11051_split_sizes_0 = const()[name = tensor<string, []>("op_11051_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11051_axis_0 = const()[name = tensor<string, []>("op_11051_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11051_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11051_cast_fp16_1 = split(axis = var_11051_axis_0, split_sizes = var_11051_split_sizes_0, x = x_1231_cast_fp16)[name = tensor<string, []>("op_11051_cast_fp16")];
            tensor<int32, [1]> squeeze_360_axes_0 = const()[name = tensor<string, []>("squeeze_360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_360_cast_fp16 = squeeze(axes = squeeze_360_axes_0, x = var_11051_cast_fp16_0)[name = tensor<string, []>("squeeze_360_cast_fp16")];
            tensor<int32, [1]> squeeze_361_axes_0 = const()[name = tensor<string, []>("squeeze_361_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_361_cast_fp16 = squeeze(axes = squeeze_361_axes_0, x = var_11051_cast_fp16_1)[name = tensor<string, []>("squeeze_361_cast_fp16")];
            tensor<fp16, []> const_1475_promoted_to_fp16 = const()[name = tensor<string, []>("const_1475_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11054_cast_fp16 = mul(x = squeeze_361_cast_fp16, y = const_1475_promoted_to_fp16)[name = tensor<string, []>("op_11054_cast_fp16")];
            tensor<int32, []> x_1233_axis_0 = const()[name = tensor<string, []>("x_1233_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1233_cast_fp16 = stack(axis = x_1233_axis_0, values = (var_11054_cast_fp16, squeeze_360_cast_fp16))[name = tensor<string, []>("x_1233_cast_fp16")];
            tensor<int32, [4]> var_11060 = const()[name = tensor<string, []>("op_11060"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11061_cast_fp16 = reshape(shape = var_11060, x = x_1233_cast_fp16)[name = tensor<string, []>("op_11061_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11062_cast_fp16 = mul(x = var_11061_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11062_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_181_cast_fp16 = add(x = var_11041_cast_fp16, y = var_11062_cast_fp16)[name = tensor<string, []>("q_patches_181_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11064_cast_fp16 = mul(x = var_11039_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11064_cast_fp16")];
            tensor<int32, [5]> var_11072 = const()[name = tensor<string, []>("op_11072"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1237_cast_fp16 = reshape(shape = var_11072, x = var_11039_cast_fp16)[name = tensor<string, []>("x_1237_cast_fp16")];
            tensor<int32, [2]> var_11074_split_sizes_0 = const()[name = tensor<string, []>("op_11074_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11074_axis_0 = const()[name = tensor<string, []>("op_11074_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11074_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11074_cast_fp16_1 = split(axis = var_11074_axis_0, split_sizes = var_11074_split_sizes_0, x = x_1237_cast_fp16)[name = tensor<string, []>("op_11074_cast_fp16")];
            tensor<int32, [1]> squeeze_362_axes_0 = const()[name = tensor<string, []>("squeeze_362_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_362_cast_fp16 = squeeze(axes = squeeze_362_axes_0, x = var_11074_cast_fp16_0)[name = tensor<string, []>("squeeze_362_cast_fp16")];
            tensor<int32, [1]> squeeze_363_axes_0 = const()[name = tensor<string, []>("squeeze_363_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_363_cast_fp16 = squeeze(axes = squeeze_363_axes_0, x = var_11074_cast_fp16_1)[name = tensor<string, []>("squeeze_363_cast_fp16")];
            tensor<fp16, []> const_1483_promoted_to_fp16 = const()[name = tensor<string, []>("const_1483_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11077_cast_fp16 = mul(x = squeeze_363_cast_fp16, y = const_1483_promoted_to_fp16)[name = tensor<string, []>("op_11077_cast_fp16")];
            tensor<int32, []> x_1239_axis_0 = const()[name = tensor<string, []>("x_1239_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1239_cast_fp16 = stack(axis = x_1239_axis_0, values = (var_11077_cast_fp16, squeeze_362_cast_fp16))[name = tensor<string, []>("x_1239_cast_fp16")];
            tensor<int32, [4]> var_11083 = const()[name = tensor<string, []>("op_11083"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11084_cast_fp16 = reshape(shape = var_11083, x = x_1239_cast_fp16)[name = tensor<string, []>("op_11084_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11085_cast_fp16 = mul(x = var_11084_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11085_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_181_cast_fp16 = add(x = var_11064_cast_fp16, y = var_11085_cast_fp16)[name = tensor<string, []>("k_patches_181_cast_fp16")];
            tensor<bool, []> var_11088_interleave_0 = const()[name = tensor<string, []>("op_11088_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11088_cast_fp16 = concat(axis = var_20, interleave = var_11088_interleave_0, values = (var_11027_cast_fp16, q_patches_181_cast_fp16))[name = tensor<string, []>("op_11088_cast_fp16")];
            tensor<bool, []> var_11091_interleave_0 = const()[name = tensor<string, []>("op_11091_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11091_cast_fp16 = concat(axis = var_20, interleave = var_11091_interleave_0, values = (var_11031_cast_fp16, k_patches_181_cast_fp16))[name = tensor<string, []>("op_11091_cast_fp16")];
            tensor<fp16, []> var_11093_to_fp16 = const()[name = tensor<string, []>("op_11093_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_909_cast_fp16 = mul(x = var_11088_cast_fp16, y = var_11093_to_fp16)[name = tensor<string, []>("q_909_cast_fp16")];
            tensor<bool, []> attn_361_transpose_x_1 = const()[name = tensor<string, []>("attn_361_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_361_transpose_y_1 = const()[name = tensor<string, []>("attn_361_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_361_cast_fp16 = matmul(transpose_x = attn_361_transpose_x_1, transpose_y = attn_361_transpose_y_1, x = q_909_cast_fp16, y = var_11091_cast_fp16)[name = tensor<string, []>("attn_361_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_363_cast_fp16 = softmax(axis = var_21, x = attn_361_cast_fp16)[name = tensor<string, []>("attn_363_cast_fp16")];
            tensor<bool, []> var_11098_transpose_x_0 = const()[name = tensor<string, []>("op_11098_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11098_transpose_y_0 = const()[name = tensor<string, []>("op_11098_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11098_cast_fp16 = matmul(transpose_x = var_11098_transpose_x_0, transpose_y = var_11098_transpose_y_0, x = attn_363_cast_fp16, y = v_545_cast_fp16)[name = tensor<string, []>("op_11098_cast_fp16")];
            tensor<int32, [1]> var_11099_axes_0 = const()[name = tensor<string, []>("op_11099_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11099_cast_fp16 = squeeze(axes = var_11099_axes_0, x = var_11098_cast_fp16)[name = tensor<string, []>("op_11099_cast_fp16")];
            tensor<string, []> var_11105_pad_type_0 = const()[name = tensor<string, []>("op_11105_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11105_strides_0 = const()[name = tensor<string, []>("op_11105_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11105_pad_0 = const()[name = tensor<string, []>("op_11105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11105_dilations_0 = const()[name = tensor<string, []>("op_11105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11105_groups_0 = const()[name = tensor<string, []>("op_11105_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102903936)))];
            tensor<fp16, [1, 64, 1, 197]> var_11105_cast_fp16 = conv(dilations = var_11105_dilations_0, groups = var_11105_groups_0, pad = var_11105_pad_0, pad_type = var_11105_pad_type_0, strides = var_11105_strides_0, weight = model_blocks_7_attn_q_projs_7_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11105_cast_fp16")];
            tensor<int32, [1]> var_11106_axes_0 = const()[name = tensor<string, []>("op_11106_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11106_cast_fp16 = squeeze(axes = var_11106_axes_0, x = var_11105_cast_fp16)[name = tensor<string, []>("op_11106_cast_fp16")];
            tensor<int32, [3]> q_911_perm_0 = const()[name = tensor<string, []>("q_911_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11113_pad_type_0 = const()[name = tensor<string, []>("op_11113_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11113_strides_0 = const()[name = tensor<string, []>("op_11113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11113_pad_0 = const()[name = tensor<string, []>("op_11113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11113_dilations_0 = const()[name = tensor<string, []>("op_11113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11113_groups_0 = const()[name = tensor<string, []>("op_11113_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103002304)))];
            tensor<fp16, [1, 64, 1, 197]> var_11113_cast_fp16 = conv(dilations = var_11113_dilations_0, groups = var_11113_groups_0, pad = var_11113_pad_0, pad_type = var_11113_pad_type_0, strides = var_11113_strides_0, weight = model_blocks_7_attn_k_projs_7_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11113_cast_fp16")];
            tensor<int32, [1]> var_11114_axes_0 = const()[name = tensor<string, []>("op_11114_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11114_cast_fp16 = squeeze(axes = var_11114_axes_0, x = var_11113_cast_fp16)[name = tensor<string, []>("op_11114_cast_fp16")];
            tensor<int32, [3]> k_547_perm_0 = const()[name = tensor<string, []>("k_547_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11121_pad_type_0 = const()[name = tensor<string, []>("op_11121_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11121_strides_0 = const()[name = tensor<string, []>("op_11121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11121_pad_0 = const()[name = tensor<string, []>("op_11121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11121_dilations_0 = const()[name = tensor<string, []>("op_11121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11121_groups_0 = const()[name = tensor<string, []>("op_11121_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103100672)))];
            tensor<fp16, [1, 64, 1, 197]> var_11121_cast_fp16 = conv(dilations = var_11121_dilations_0, groups = var_11121_groups_0, pad = var_11121_pad_0, pad_type = var_11121_pad_type_0, strides = var_11121_strides_0, weight = model_blocks_7_attn_v_projs_7_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11121_cast_fp16")];
            tensor<int32, [1]> var_11122_axes_0 = const()[name = tensor<string, []>("op_11122_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11122_cast_fp16 = squeeze(axes = var_11122_axes_0, x = var_11121_cast_fp16)[name = tensor<string, []>("op_11122_cast_fp16")];
            tensor<int32, [3]> v_547_perm_0 = const()[name = tensor<string, []>("v_547_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103199040)))];
            tensor<fp16, [1, 197, 64]> q_911_cast_fp16 = transpose(perm = q_911_perm_0, x = var_11106_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [1, 197, 64]> q_913_cast_fp16 = add(x = q_911_cast_fp16, y = model_blocks_7_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_913_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103199232)))];
            tensor<fp16, [1, 197, 64]> v_547_cast_fp16 = transpose(perm = v_547_perm_0, x = var_11122_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [1, 197, 64]> v_549_cast_fp16 = add(x = v_547_cast_fp16, y = model_blocks_7_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_549_cast_fp16")];
            tensor<int32, [1]> q_915_axes_0 = const()[name = tensor<string, []>("q_915_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_915_cast_fp16 = expand_dims(axes = q_915_axes_0, x = q_913_cast_fp16)[name = tensor<string, []>("q_915_cast_fp16")];
            tensor<int32, [1]> k_549_axes_0 = const()[name = tensor<string, []>("k_549_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_547_cast_fp16 = transpose(perm = k_547_perm_0, x = var_11114_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 1, 197, 64]> k_549_cast_fp16 = expand_dims(axes = k_549_axes_0, x = k_547_cast_fp16)[name = tensor<string, []>("k_549_cast_fp16")];
            tensor<int32, [1]> v_551_axes_0 = const()[name = tensor<string, []>("v_551_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_551_cast_fp16 = expand_dims(axes = v_551_axes_0, x = v_549_cast_fp16)[name = tensor<string, []>("v_551_cast_fp16")];
            tensor<int32, [4]> var_11131_begin_0 = const()[name = tensor<string, []>("op_11131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11131_end_0 = const()[name = tensor<string, []>("op_11131_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11131_end_mask_0 = const()[name = tensor<string, []>("op_11131_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11131_cast_fp16 = slice_by_index(begin = var_11131_begin_0, end = var_11131_end_0, end_mask = var_11131_end_mask_0, x = q_915_cast_fp16)[name = tensor<string, []>("op_11131_cast_fp16")];
            tensor<int32, [4]> var_11135_begin_0 = const()[name = tensor<string, []>("op_11135_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11135_end_0 = const()[name = tensor<string, []>("op_11135_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11135_end_mask_0 = const()[name = tensor<string, []>("op_11135_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11135_cast_fp16 = slice_by_index(begin = var_11135_begin_0, end = var_11135_end_0, end_mask = var_11135_end_mask_0, x = k_549_cast_fp16)[name = tensor<string, []>("op_11135_cast_fp16")];
            tensor<int32, [4]> var_11139_begin_0 = const()[name = tensor<string, []>("op_11139_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11139_end_0 = const()[name = tensor<string, []>("op_11139_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11139_end_mask_0 = const()[name = tensor<string, []>("op_11139_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11139_cast_fp16 = slice_by_index(begin = var_11139_begin_0, end = var_11139_end_0, end_mask = var_11139_end_mask_0, x = q_915_cast_fp16)[name = tensor<string, []>("op_11139_cast_fp16")];
            tensor<int32, [4]> var_11143_begin_0 = const()[name = tensor<string, []>("op_11143_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11143_end_0 = const()[name = tensor<string, []>("op_11143_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11143_end_mask_0 = const()[name = tensor<string, []>("op_11143_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11143_cast_fp16 = slice_by_index(begin = var_11143_begin_0, end = var_11143_end_0, end_mask = var_11143_end_mask_0, x = k_549_cast_fp16)[name = tensor<string, []>("op_11143_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11145_cast_fp16 = mul(x = var_11139_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11145_cast_fp16")];
            tensor<int32, [5]> var_11153 = const()[name = tensor<string, []>("op_11153"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1243_cast_fp16 = reshape(shape = var_11153, x = var_11139_cast_fp16)[name = tensor<string, []>("x_1243_cast_fp16")];
            tensor<int32, [2]> var_11155_split_sizes_0 = const()[name = tensor<string, []>("op_11155_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11155_axis_0 = const()[name = tensor<string, []>("op_11155_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11155_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11155_cast_fp16_1 = split(axis = var_11155_axis_0, split_sizes = var_11155_split_sizes_0, x = x_1243_cast_fp16)[name = tensor<string, []>("op_11155_cast_fp16")];
            tensor<int32, [1]> squeeze_364_axes_0 = const()[name = tensor<string, []>("squeeze_364_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_364_cast_fp16 = squeeze(axes = squeeze_364_axes_0, x = var_11155_cast_fp16_0)[name = tensor<string, []>("squeeze_364_cast_fp16")];
            tensor<int32, [1]> squeeze_365_axes_0 = const()[name = tensor<string, []>("squeeze_365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_365_cast_fp16 = squeeze(axes = squeeze_365_axes_0, x = var_11155_cast_fp16_1)[name = tensor<string, []>("squeeze_365_cast_fp16")];
            tensor<fp16, []> const_1491_promoted_to_fp16 = const()[name = tensor<string, []>("const_1491_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11158_cast_fp16 = mul(x = squeeze_365_cast_fp16, y = const_1491_promoted_to_fp16)[name = tensor<string, []>("op_11158_cast_fp16")];
            tensor<int32, []> x_1245_axis_0 = const()[name = tensor<string, []>("x_1245_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1245_cast_fp16 = stack(axis = x_1245_axis_0, values = (var_11158_cast_fp16, squeeze_364_cast_fp16))[name = tensor<string, []>("x_1245_cast_fp16")];
            tensor<int32, [4]> var_11164 = const()[name = tensor<string, []>("op_11164"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11165_cast_fp16 = reshape(shape = var_11164, x = x_1245_cast_fp16)[name = tensor<string, []>("op_11165_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11166_cast_fp16 = mul(x = var_11165_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11166_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_183_cast_fp16 = add(x = var_11145_cast_fp16, y = var_11166_cast_fp16)[name = tensor<string, []>("q_patches_183_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11168_cast_fp16 = mul(x = var_11143_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11168_cast_fp16")];
            tensor<int32, [5]> var_11176 = const()[name = tensor<string, []>("op_11176"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1249_cast_fp16 = reshape(shape = var_11176, x = var_11143_cast_fp16)[name = tensor<string, []>("x_1249_cast_fp16")];
            tensor<int32, [2]> var_11178_split_sizes_0 = const()[name = tensor<string, []>("op_11178_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11178_axis_0 = const()[name = tensor<string, []>("op_11178_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11178_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11178_cast_fp16_1 = split(axis = var_11178_axis_0, split_sizes = var_11178_split_sizes_0, x = x_1249_cast_fp16)[name = tensor<string, []>("op_11178_cast_fp16")];
            tensor<int32, [1]> squeeze_366_axes_0 = const()[name = tensor<string, []>("squeeze_366_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_366_cast_fp16 = squeeze(axes = squeeze_366_axes_0, x = var_11178_cast_fp16_0)[name = tensor<string, []>("squeeze_366_cast_fp16")];
            tensor<int32, [1]> squeeze_367_axes_0 = const()[name = tensor<string, []>("squeeze_367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_367_cast_fp16 = squeeze(axes = squeeze_367_axes_0, x = var_11178_cast_fp16_1)[name = tensor<string, []>("squeeze_367_cast_fp16")];
            tensor<fp16, []> const_1499_promoted_to_fp16 = const()[name = tensor<string, []>("const_1499_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11181_cast_fp16 = mul(x = squeeze_367_cast_fp16, y = const_1499_promoted_to_fp16)[name = tensor<string, []>("op_11181_cast_fp16")];
            tensor<int32, []> x_1251_axis_0 = const()[name = tensor<string, []>("x_1251_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1251_cast_fp16 = stack(axis = x_1251_axis_0, values = (var_11181_cast_fp16, squeeze_366_cast_fp16))[name = tensor<string, []>("x_1251_cast_fp16")];
            tensor<int32, [4]> var_11187 = const()[name = tensor<string, []>("op_11187"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11188_cast_fp16 = reshape(shape = var_11187, x = x_1251_cast_fp16)[name = tensor<string, []>("op_11188_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11189_cast_fp16 = mul(x = var_11188_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11189_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_183_cast_fp16 = add(x = var_11168_cast_fp16, y = var_11189_cast_fp16)[name = tensor<string, []>("k_patches_183_cast_fp16")];
            tensor<bool, []> var_11192_interleave_0 = const()[name = tensor<string, []>("op_11192_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11192_cast_fp16 = concat(axis = var_20, interleave = var_11192_interleave_0, values = (var_11131_cast_fp16, q_patches_183_cast_fp16))[name = tensor<string, []>("op_11192_cast_fp16")];
            tensor<bool, []> var_11195_interleave_0 = const()[name = tensor<string, []>("op_11195_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11195_cast_fp16 = concat(axis = var_20, interleave = var_11195_interleave_0, values = (var_11135_cast_fp16, k_patches_183_cast_fp16))[name = tensor<string, []>("op_11195_cast_fp16")];
            tensor<fp16, []> var_11197_to_fp16 = const()[name = tensor<string, []>("op_11197_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_919_cast_fp16 = mul(x = var_11192_cast_fp16, y = var_11197_to_fp16)[name = tensor<string, []>("q_919_cast_fp16")];
            tensor<bool, []> attn_365_transpose_x_1 = const()[name = tensor<string, []>("attn_365_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_365_transpose_y_1 = const()[name = tensor<string, []>("attn_365_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_365_cast_fp16 = matmul(transpose_x = attn_365_transpose_x_1, transpose_y = attn_365_transpose_y_1, x = q_919_cast_fp16, y = var_11195_cast_fp16)[name = tensor<string, []>("attn_365_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_367_cast_fp16 = softmax(axis = var_21, x = attn_365_cast_fp16)[name = tensor<string, []>("attn_367_cast_fp16")];
            tensor<bool, []> var_11202_transpose_x_0 = const()[name = tensor<string, []>("op_11202_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11202_transpose_y_0 = const()[name = tensor<string, []>("op_11202_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11202_cast_fp16 = matmul(transpose_x = var_11202_transpose_x_0, transpose_y = var_11202_transpose_y_0, x = attn_367_cast_fp16, y = v_551_cast_fp16)[name = tensor<string, []>("op_11202_cast_fp16")];
            tensor<int32, [1]> var_11203_axes_0 = const()[name = tensor<string, []>("op_11203_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11203_cast_fp16 = squeeze(axes = var_11203_axes_0, x = var_11202_cast_fp16)[name = tensor<string, []>("op_11203_cast_fp16")];
            tensor<string, []> var_11209_pad_type_0 = const()[name = tensor<string, []>("op_11209_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11209_strides_0 = const()[name = tensor<string, []>("op_11209_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11209_pad_0 = const()[name = tensor<string, []>("op_11209_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11209_dilations_0 = const()[name = tensor<string, []>("op_11209_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11209_groups_0 = const()[name = tensor<string, []>("op_11209_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103199424)))];
            tensor<fp16, [1, 64, 1, 197]> var_11209_cast_fp16 = conv(dilations = var_11209_dilations_0, groups = var_11209_groups_0, pad = var_11209_pad_0, pad_type = var_11209_pad_type_0, strides = var_11209_strides_0, weight = model_blocks_7_attn_q_projs_8_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11209_cast_fp16")];
            tensor<int32, [1]> var_11210_axes_0 = const()[name = tensor<string, []>("op_11210_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11210_cast_fp16 = squeeze(axes = var_11210_axes_0, x = var_11209_cast_fp16)[name = tensor<string, []>("op_11210_cast_fp16")];
            tensor<int32, [3]> q_921_perm_0 = const()[name = tensor<string, []>("q_921_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11217_pad_type_0 = const()[name = tensor<string, []>("op_11217_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11217_strides_0 = const()[name = tensor<string, []>("op_11217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11217_pad_0 = const()[name = tensor<string, []>("op_11217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11217_dilations_0 = const()[name = tensor<string, []>("op_11217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11217_groups_0 = const()[name = tensor<string, []>("op_11217_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103297792)))];
            tensor<fp16, [1, 64, 1, 197]> var_11217_cast_fp16 = conv(dilations = var_11217_dilations_0, groups = var_11217_groups_0, pad = var_11217_pad_0, pad_type = var_11217_pad_type_0, strides = var_11217_strides_0, weight = model_blocks_7_attn_k_projs_8_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11217_cast_fp16")];
            tensor<int32, [1]> var_11218_axes_0 = const()[name = tensor<string, []>("op_11218_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11218_cast_fp16 = squeeze(axes = var_11218_axes_0, x = var_11217_cast_fp16)[name = tensor<string, []>("op_11218_cast_fp16")];
            tensor<int32, [3]> k_553_perm_0 = const()[name = tensor<string, []>("k_553_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11225_pad_type_0 = const()[name = tensor<string, []>("op_11225_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11225_strides_0 = const()[name = tensor<string, []>("op_11225_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11225_pad_0 = const()[name = tensor<string, []>("op_11225_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11225_dilations_0 = const()[name = tensor<string, []>("op_11225_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11225_groups_0 = const()[name = tensor<string, []>("op_11225_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103396160)))];
            tensor<fp16, [1, 64, 1, 197]> var_11225_cast_fp16 = conv(dilations = var_11225_dilations_0, groups = var_11225_groups_0, pad = var_11225_pad_0, pad_type = var_11225_pad_type_0, strides = var_11225_strides_0, weight = model_blocks_7_attn_v_projs_8_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11225_cast_fp16")];
            tensor<int32, [1]> var_11226_axes_0 = const()[name = tensor<string, []>("op_11226_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11226_cast_fp16 = squeeze(axes = var_11226_axes_0, x = var_11225_cast_fp16)[name = tensor<string, []>("op_11226_cast_fp16")];
            tensor<int32, [3]> v_553_perm_0 = const()[name = tensor<string, []>("v_553_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103494528)))];
            tensor<fp16, [1, 197, 64]> q_921_cast_fp16 = transpose(perm = q_921_perm_0, x = var_11210_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 197, 64]> q_923_cast_fp16 = add(x = q_921_cast_fp16, y = model_blocks_7_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_923_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103494720)))];
            tensor<fp16, [1, 197, 64]> v_553_cast_fp16 = transpose(perm = v_553_perm_0, x = var_11226_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [1, 197, 64]> v_555_cast_fp16 = add(x = v_553_cast_fp16, y = model_blocks_7_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_555_cast_fp16")];
            tensor<int32, [1]> q_925_axes_0 = const()[name = tensor<string, []>("q_925_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_925_cast_fp16 = expand_dims(axes = q_925_axes_0, x = q_923_cast_fp16)[name = tensor<string, []>("q_925_cast_fp16")];
            tensor<int32, [1]> k_555_axes_0 = const()[name = tensor<string, []>("k_555_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_553_cast_fp16 = transpose(perm = k_553_perm_0, x = var_11218_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [1, 1, 197, 64]> k_555_cast_fp16 = expand_dims(axes = k_555_axes_0, x = k_553_cast_fp16)[name = tensor<string, []>("k_555_cast_fp16")];
            tensor<int32, [1]> v_557_axes_0 = const()[name = tensor<string, []>("v_557_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_557_cast_fp16 = expand_dims(axes = v_557_axes_0, x = v_555_cast_fp16)[name = tensor<string, []>("v_557_cast_fp16")];
            tensor<int32, [4]> var_11235_begin_0 = const()[name = tensor<string, []>("op_11235_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11235_end_0 = const()[name = tensor<string, []>("op_11235_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11235_end_mask_0 = const()[name = tensor<string, []>("op_11235_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11235_cast_fp16 = slice_by_index(begin = var_11235_begin_0, end = var_11235_end_0, end_mask = var_11235_end_mask_0, x = q_925_cast_fp16)[name = tensor<string, []>("op_11235_cast_fp16")];
            tensor<int32, [4]> var_11239_begin_0 = const()[name = tensor<string, []>("op_11239_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11239_end_0 = const()[name = tensor<string, []>("op_11239_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11239_end_mask_0 = const()[name = tensor<string, []>("op_11239_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11239_cast_fp16 = slice_by_index(begin = var_11239_begin_0, end = var_11239_end_0, end_mask = var_11239_end_mask_0, x = k_555_cast_fp16)[name = tensor<string, []>("op_11239_cast_fp16")];
            tensor<int32, [4]> var_11243_begin_0 = const()[name = tensor<string, []>("op_11243_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11243_end_0 = const()[name = tensor<string, []>("op_11243_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11243_end_mask_0 = const()[name = tensor<string, []>("op_11243_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11243_cast_fp16 = slice_by_index(begin = var_11243_begin_0, end = var_11243_end_0, end_mask = var_11243_end_mask_0, x = q_925_cast_fp16)[name = tensor<string, []>("op_11243_cast_fp16")];
            tensor<int32, [4]> var_11247_begin_0 = const()[name = tensor<string, []>("op_11247_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11247_end_0 = const()[name = tensor<string, []>("op_11247_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11247_end_mask_0 = const()[name = tensor<string, []>("op_11247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11247_cast_fp16 = slice_by_index(begin = var_11247_begin_0, end = var_11247_end_0, end_mask = var_11247_end_mask_0, x = k_555_cast_fp16)[name = tensor<string, []>("op_11247_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11249_cast_fp16 = mul(x = var_11243_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11249_cast_fp16")];
            tensor<int32, [5]> var_11257 = const()[name = tensor<string, []>("op_11257"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1255_cast_fp16 = reshape(shape = var_11257, x = var_11243_cast_fp16)[name = tensor<string, []>("x_1255_cast_fp16")];
            tensor<int32, [2]> var_11259_split_sizes_0 = const()[name = tensor<string, []>("op_11259_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11259_axis_0 = const()[name = tensor<string, []>("op_11259_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11259_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11259_cast_fp16_1 = split(axis = var_11259_axis_0, split_sizes = var_11259_split_sizes_0, x = x_1255_cast_fp16)[name = tensor<string, []>("op_11259_cast_fp16")];
            tensor<int32, [1]> squeeze_368_axes_0 = const()[name = tensor<string, []>("squeeze_368_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_368_cast_fp16 = squeeze(axes = squeeze_368_axes_0, x = var_11259_cast_fp16_0)[name = tensor<string, []>("squeeze_368_cast_fp16")];
            tensor<int32, [1]> squeeze_369_axes_0 = const()[name = tensor<string, []>("squeeze_369_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_369_cast_fp16 = squeeze(axes = squeeze_369_axes_0, x = var_11259_cast_fp16_1)[name = tensor<string, []>("squeeze_369_cast_fp16")];
            tensor<fp16, []> const_1507_promoted_to_fp16 = const()[name = tensor<string, []>("const_1507_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11262_cast_fp16 = mul(x = squeeze_369_cast_fp16, y = const_1507_promoted_to_fp16)[name = tensor<string, []>("op_11262_cast_fp16")];
            tensor<int32, []> x_1257_axis_0 = const()[name = tensor<string, []>("x_1257_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1257_cast_fp16 = stack(axis = x_1257_axis_0, values = (var_11262_cast_fp16, squeeze_368_cast_fp16))[name = tensor<string, []>("x_1257_cast_fp16")];
            tensor<int32, [4]> var_11268 = const()[name = tensor<string, []>("op_11268"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11269_cast_fp16 = reshape(shape = var_11268, x = x_1257_cast_fp16)[name = tensor<string, []>("op_11269_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11270_cast_fp16 = mul(x = var_11269_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11270_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_185_cast_fp16 = add(x = var_11249_cast_fp16, y = var_11270_cast_fp16)[name = tensor<string, []>("q_patches_185_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11272_cast_fp16 = mul(x = var_11247_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11272_cast_fp16")];
            tensor<int32, [5]> var_11280 = const()[name = tensor<string, []>("op_11280"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1261_cast_fp16 = reshape(shape = var_11280, x = var_11247_cast_fp16)[name = tensor<string, []>("x_1261_cast_fp16")];
            tensor<int32, [2]> var_11282_split_sizes_0 = const()[name = tensor<string, []>("op_11282_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11282_axis_0 = const()[name = tensor<string, []>("op_11282_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11282_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11282_cast_fp16_1 = split(axis = var_11282_axis_0, split_sizes = var_11282_split_sizes_0, x = x_1261_cast_fp16)[name = tensor<string, []>("op_11282_cast_fp16")];
            tensor<int32, [1]> squeeze_370_axes_0 = const()[name = tensor<string, []>("squeeze_370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_370_cast_fp16 = squeeze(axes = squeeze_370_axes_0, x = var_11282_cast_fp16_0)[name = tensor<string, []>("squeeze_370_cast_fp16")];
            tensor<int32, [1]> squeeze_371_axes_0 = const()[name = tensor<string, []>("squeeze_371_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_371_cast_fp16 = squeeze(axes = squeeze_371_axes_0, x = var_11282_cast_fp16_1)[name = tensor<string, []>("squeeze_371_cast_fp16")];
            tensor<fp16, []> const_1515_promoted_to_fp16 = const()[name = tensor<string, []>("const_1515_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11285_cast_fp16 = mul(x = squeeze_371_cast_fp16, y = const_1515_promoted_to_fp16)[name = tensor<string, []>("op_11285_cast_fp16")];
            tensor<int32, []> x_1263_axis_0 = const()[name = tensor<string, []>("x_1263_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1263_cast_fp16 = stack(axis = x_1263_axis_0, values = (var_11285_cast_fp16, squeeze_370_cast_fp16))[name = tensor<string, []>("x_1263_cast_fp16")];
            tensor<int32, [4]> var_11291 = const()[name = tensor<string, []>("op_11291"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11292_cast_fp16 = reshape(shape = var_11291, x = x_1263_cast_fp16)[name = tensor<string, []>("op_11292_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11293_cast_fp16 = mul(x = var_11292_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11293_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_185_cast_fp16 = add(x = var_11272_cast_fp16, y = var_11293_cast_fp16)[name = tensor<string, []>("k_patches_185_cast_fp16")];
            tensor<bool, []> var_11296_interleave_0 = const()[name = tensor<string, []>("op_11296_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11296_cast_fp16 = concat(axis = var_20, interleave = var_11296_interleave_0, values = (var_11235_cast_fp16, q_patches_185_cast_fp16))[name = tensor<string, []>("op_11296_cast_fp16")];
            tensor<bool, []> var_11299_interleave_0 = const()[name = tensor<string, []>("op_11299_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11299_cast_fp16 = concat(axis = var_20, interleave = var_11299_interleave_0, values = (var_11239_cast_fp16, k_patches_185_cast_fp16))[name = tensor<string, []>("op_11299_cast_fp16")];
            tensor<fp16, []> var_11301_to_fp16 = const()[name = tensor<string, []>("op_11301_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_929_cast_fp16 = mul(x = var_11296_cast_fp16, y = var_11301_to_fp16)[name = tensor<string, []>("q_929_cast_fp16")];
            tensor<bool, []> attn_369_transpose_x_1 = const()[name = tensor<string, []>("attn_369_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_369_transpose_y_1 = const()[name = tensor<string, []>("attn_369_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_369_cast_fp16 = matmul(transpose_x = attn_369_transpose_x_1, transpose_y = attn_369_transpose_y_1, x = q_929_cast_fp16, y = var_11299_cast_fp16)[name = tensor<string, []>("attn_369_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_371_cast_fp16 = softmax(axis = var_21, x = attn_369_cast_fp16)[name = tensor<string, []>("attn_371_cast_fp16")];
            tensor<bool, []> var_11306_transpose_x_0 = const()[name = tensor<string, []>("op_11306_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11306_transpose_y_0 = const()[name = tensor<string, []>("op_11306_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11306_cast_fp16 = matmul(transpose_x = var_11306_transpose_x_0, transpose_y = var_11306_transpose_y_0, x = attn_371_cast_fp16, y = v_557_cast_fp16)[name = tensor<string, []>("op_11306_cast_fp16")];
            tensor<int32, [1]> var_11307_axes_0 = const()[name = tensor<string, []>("op_11307_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11307_cast_fp16 = squeeze(axes = var_11307_axes_0, x = var_11306_cast_fp16)[name = tensor<string, []>("op_11307_cast_fp16")];
            tensor<string, []> var_11313_pad_type_0 = const()[name = tensor<string, []>("op_11313_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11313_strides_0 = const()[name = tensor<string, []>("op_11313_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11313_pad_0 = const()[name = tensor<string, []>("op_11313_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11313_dilations_0 = const()[name = tensor<string, []>("op_11313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11313_groups_0 = const()[name = tensor<string, []>("op_11313_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103494912)))];
            tensor<fp16, [1, 64, 1, 197]> var_11313_cast_fp16 = conv(dilations = var_11313_dilations_0, groups = var_11313_groups_0, pad = var_11313_pad_0, pad_type = var_11313_pad_type_0, strides = var_11313_strides_0, weight = model_blocks_7_attn_q_projs_9_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11313_cast_fp16")];
            tensor<int32, [1]> var_11314_axes_0 = const()[name = tensor<string, []>("op_11314_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11314_cast_fp16 = squeeze(axes = var_11314_axes_0, x = var_11313_cast_fp16)[name = tensor<string, []>("op_11314_cast_fp16")];
            tensor<int32, [3]> q_931_perm_0 = const()[name = tensor<string, []>("q_931_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11321_pad_type_0 = const()[name = tensor<string, []>("op_11321_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11321_strides_0 = const()[name = tensor<string, []>("op_11321_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11321_pad_0 = const()[name = tensor<string, []>("op_11321_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11321_dilations_0 = const()[name = tensor<string, []>("op_11321_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11321_groups_0 = const()[name = tensor<string, []>("op_11321_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103593280)))];
            tensor<fp16, [1, 64, 1, 197]> var_11321_cast_fp16 = conv(dilations = var_11321_dilations_0, groups = var_11321_groups_0, pad = var_11321_pad_0, pad_type = var_11321_pad_type_0, strides = var_11321_strides_0, weight = model_blocks_7_attn_k_projs_9_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11321_cast_fp16")];
            tensor<int32, [1]> var_11322_axes_0 = const()[name = tensor<string, []>("op_11322_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11322_cast_fp16 = squeeze(axes = var_11322_axes_0, x = var_11321_cast_fp16)[name = tensor<string, []>("op_11322_cast_fp16")];
            tensor<int32, [3]> k_559_perm_0 = const()[name = tensor<string, []>("k_559_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11329_pad_type_0 = const()[name = tensor<string, []>("op_11329_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11329_strides_0 = const()[name = tensor<string, []>("op_11329_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11329_pad_0 = const()[name = tensor<string, []>("op_11329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11329_dilations_0 = const()[name = tensor<string, []>("op_11329_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11329_groups_0 = const()[name = tensor<string, []>("op_11329_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103691648)))];
            tensor<fp16, [1, 64, 1, 197]> var_11329_cast_fp16 = conv(dilations = var_11329_dilations_0, groups = var_11329_groups_0, pad = var_11329_pad_0, pad_type = var_11329_pad_type_0, strides = var_11329_strides_0, weight = model_blocks_7_attn_v_projs_9_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11329_cast_fp16")];
            tensor<int32, [1]> var_11330_axes_0 = const()[name = tensor<string, []>("op_11330_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11330_cast_fp16 = squeeze(axes = var_11330_axes_0, x = var_11329_cast_fp16)[name = tensor<string, []>("op_11330_cast_fp16")];
            tensor<int32, [3]> v_559_perm_0 = const()[name = tensor<string, []>("v_559_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103790016)))];
            tensor<fp16, [1, 197, 64]> q_931_cast_fp16 = transpose(perm = q_931_perm_0, x = var_11314_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [1, 197, 64]> q_933_cast_fp16 = add(x = q_931_cast_fp16, y = model_blocks_7_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_933_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103790208)))];
            tensor<fp16, [1, 197, 64]> v_559_cast_fp16 = transpose(perm = v_559_perm_0, x = var_11330_cast_fp16)[name = tensor<string, []>("transpose_184")];
            tensor<fp16, [1, 197, 64]> v_561_cast_fp16 = add(x = v_559_cast_fp16, y = model_blocks_7_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_561_cast_fp16")];
            tensor<int32, [1]> q_935_axes_0 = const()[name = tensor<string, []>("q_935_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_935_cast_fp16 = expand_dims(axes = q_935_axes_0, x = q_933_cast_fp16)[name = tensor<string, []>("q_935_cast_fp16")];
            tensor<int32, [1]> k_561_axes_0 = const()[name = tensor<string, []>("k_561_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_559_cast_fp16 = transpose(perm = k_559_perm_0, x = var_11322_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [1, 1, 197, 64]> k_561_cast_fp16 = expand_dims(axes = k_561_axes_0, x = k_559_cast_fp16)[name = tensor<string, []>("k_561_cast_fp16")];
            tensor<int32, [1]> v_563_axes_0 = const()[name = tensor<string, []>("v_563_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_563_cast_fp16 = expand_dims(axes = v_563_axes_0, x = v_561_cast_fp16)[name = tensor<string, []>("v_563_cast_fp16")];
            tensor<int32, [4]> var_11339_begin_0 = const()[name = tensor<string, []>("op_11339_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11339_end_0 = const()[name = tensor<string, []>("op_11339_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11339_end_mask_0 = const()[name = tensor<string, []>("op_11339_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11339_cast_fp16 = slice_by_index(begin = var_11339_begin_0, end = var_11339_end_0, end_mask = var_11339_end_mask_0, x = q_935_cast_fp16)[name = tensor<string, []>("op_11339_cast_fp16")];
            tensor<int32, [4]> var_11343_begin_0 = const()[name = tensor<string, []>("op_11343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11343_end_0 = const()[name = tensor<string, []>("op_11343_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11343_end_mask_0 = const()[name = tensor<string, []>("op_11343_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11343_cast_fp16 = slice_by_index(begin = var_11343_begin_0, end = var_11343_end_0, end_mask = var_11343_end_mask_0, x = k_561_cast_fp16)[name = tensor<string, []>("op_11343_cast_fp16")];
            tensor<int32, [4]> var_11347_begin_0 = const()[name = tensor<string, []>("op_11347_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11347_end_0 = const()[name = tensor<string, []>("op_11347_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11347_end_mask_0 = const()[name = tensor<string, []>("op_11347_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11347_cast_fp16 = slice_by_index(begin = var_11347_begin_0, end = var_11347_end_0, end_mask = var_11347_end_mask_0, x = q_935_cast_fp16)[name = tensor<string, []>("op_11347_cast_fp16")];
            tensor<int32, [4]> var_11351_begin_0 = const()[name = tensor<string, []>("op_11351_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11351_end_0 = const()[name = tensor<string, []>("op_11351_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11351_end_mask_0 = const()[name = tensor<string, []>("op_11351_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11351_cast_fp16 = slice_by_index(begin = var_11351_begin_0, end = var_11351_end_0, end_mask = var_11351_end_mask_0, x = k_561_cast_fp16)[name = tensor<string, []>("op_11351_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11353_cast_fp16 = mul(x = var_11347_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11353_cast_fp16")];
            tensor<int32, [5]> var_11361 = const()[name = tensor<string, []>("op_11361"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1267_cast_fp16 = reshape(shape = var_11361, x = var_11347_cast_fp16)[name = tensor<string, []>("x_1267_cast_fp16")];
            tensor<int32, [2]> var_11363_split_sizes_0 = const()[name = tensor<string, []>("op_11363_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11363_axis_0 = const()[name = tensor<string, []>("op_11363_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11363_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11363_cast_fp16_1 = split(axis = var_11363_axis_0, split_sizes = var_11363_split_sizes_0, x = x_1267_cast_fp16)[name = tensor<string, []>("op_11363_cast_fp16")];
            tensor<int32, [1]> squeeze_372_axes_0 = const()[name = tensor<string, []>("squeeze_372_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_372_cast_fp16 = squeeze(axes = squeeze_372_axes_0, x = var_11363_cast_fp16_0)[name = tensor<string, []>("squeeze_372_cast_fp16")];
            tensor<int32, [1]> squeeze_373_axes_0 = const()[name = tensor<string, []>("squeeze_373_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_373_cast_fp16 = squeeze(axes = squeeze_373_axes_0, x = var_11363_cast_fp16_1)[name = tensor<string, []>("squeeze_373_cast_fp16")];
            tensor<fp16, []> const_1523_promoted_to_fp16 = const()[name = tensor<string, []>("const_1523_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11366_cast_fp16 = mul(x = squeeze_373_cast_fp16, y = const_1523_promoted_to_fp16)[name = tensor<string, []>("op_11366_cast_fp16")];
            tensor<int32, []> x_1269_axis_0 = const()[name = tensor<string, []>("x_1269_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1269_cast_fp16 = stack(axis = x_1269_axis_0, values = (var_11366_cast_fp16, squeeze_372_cast_fp16))[name = tensor<string, []>("x_1269_cast_fp16")];
            tensor<int32, [4]> var_11372 = const()[name = tensor<string, []>("op_11372"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11373_cast_fp16 = reshape(shape = var_11372, x = x_1269_cast_fp16)[name = tensor<string, []>("op_11373_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11374_cast_fp16 = mul(x = var_11373_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11374_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_187_cast_fp16 = add(x = var_11353_cast_fp16, y = var_11374_cast_fp16)[name = tensor<string, []>("q_patches_187_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11376_cast_fp16 = mul(x = var_11351_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11376_cast_fp16")];
            tensor<int32, [5]> var_11384 = const()[name = tensor<string, []>("op_11384"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1273_cast_fp16 = reshape(shape = var_11384, x = var_11351_cast_fp16)[name = tensor<string, []>("x_1273_cast_fp16")];
            tensor<int32, [2]> var_11386_split_sizes_0 = const()[name = tensor<string, []>("op_11386_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11386_axis_0 = const()[name = tensor<string, []>("op_11386_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11386_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11386_cast_fp16_1 = split(axis = var_11386_axis_0, split_sizes = var_11386_split_sizes_0, x = x_1273_cast_fp16)[name = tensor<string, []>("op_11386_cast_fp16")];
            tensor<int32, [1]> squeeze_374_axes_0 = const()[name = tensor<string, []>("squeeze_374_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_374_cast_fp16 = squeeze(axes = squeeze_374_axes_0, x = var_11386_cast_fp16_0)[name = tensor<string, []>("squeeze_374_cast_fp16")];
            tensor<int32, [1]> squeeze_375_axes_0 = const()[name = tensor<string, []>("squeeze_375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_375_cast_fp16 = squeeze(axes = squeeze_375_axes_0, x = var_11386_cast_fp16_1)[name = tensor<string, []>("squeeze_375_cast_fp16")];
            tensor<fp16, []> const_1531_promoted_to_fp16 = const()[name = tensor<string, []>("const_1531_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11389_cast_fp16 = mul(x = squeeze_375_cast_fp16, y = const_1531_promoted_to_fp16)[name = tensor<string, []>("op_11389_cast_fp16")];
            tensor<int32, []> x_1275_axis_0 = const()[name = tensor<string, []>("x_1275_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1275_cast_fp16 = stack(axis = x_1275_axis_0, values = (var_11389_cast_fp16, squeeze_374_cast_fp16))[name = tensor<string, []>("x_1275_cast_fp16")];
            tensor<int32, [4]> var_11395 = const()[name = tensor<string, []>("op_11395"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11396_cast_fp16 = reshape(shape = var_11395, x = x_1275_cast_fp16)[name = tensor<string, []>("op_11396_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11397_cast_fp16 = mul(x = var_11396_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11397_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_187_cast_fp16 = add(x = var_11376_cast_fp16, y = var_11397_cast_fp16)[name = tensor<string, []>("k_patches_187_cast_fp16")];
            tensor<bool, []> var_11400_interleave_0 = const()[name = tensor<string, []>("op_11400_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11400_cast_fp16 = concat(axis = var_20, interleave = var_11400_interleave_0, values = (var_11339_cast_fp16, q_patches_187_cast_fp16))[name = tensor<string, []>("op_11400_cast_fp16")];
            tensor<bool, []> var_11403_interleave_0 = const()[name = tensor<string, []>("op_11403_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11403_cast_fp16 = concat(axis = var_20, interleave = var_11403_interleave_0, values = (var_11343_cast_fp16, k_patches_187_cast_fp16))[name = tensor<string, []>("op_11403_cast_fp16")];
            tensor<fp16, []> var_11405_to_fp16 = const()[name = tensor<string, []>("op_11405_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_939_cast_fp16 = mul(x = var_11400_cast_fp16, y = var_11405_to_fp16)[name = tensor<string, []>("q_939_cast_fp16")];
            tensor<bool, []> attn_373_transpose_x_1 = const()[name = tensor<string, []>("attn_373_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_373_transpose_y_1 = const()[name = tensor<string, []>("attn_373_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_373_cast_fp16 = matmul(transpose_x = attn_373_transpose_x_1, transpose_y = attn_373_transpose_y_1, x = q_939_cast_fp16, y = var_11403_cast_fp16)[name = tensor<string, []>("attn_373_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_375_cast_fp16 = softmax(axis = var_21, x = attn_373_cast_fp16)[name = tensor<string, []>("attn_375_cast_fp16")];
            tensor<bool, []> var_11410_transpose_x_0 = const()[name = tensor<string, []>("op_11410_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11410_transpose_y_0 = const()[name = tensor<string, []>("op_11410_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11410_cast_fp16 = matmul(transpose_x = var_11410_transpose_x_0, transpose_y = var_11410_transpose_y_0, x = attn_375_cast_fp16, y = v_563_cast_fp16)[name = tensor<string, []>("op_11410_cast_fp16")];
            tensor<int32, [1]> var_11411_axes_0 = const()[name = tensor<string, []>("op_11411_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11411_cast_fp16 = squeeze(axes = var_11411_axes_0, x = var_11410_cast_fp16)[name = tensor<string, []>("op_11411_cast_fp16")];
            tensor<string, []> var_11417_pad_type_0 = const()[name = tensor<string, []>("op_11417_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11417_strides_0 = const()[name = tensor<string, []>("op_11417_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11417_pad_0 = const()[name = tensor<string, []>("op_11417_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11417_dilations_0 = const()[name = tensor<string, []>("op_11417_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11417_groups_0 = const()[name = tensor<string, []>("op_11417_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103790400)))];
            tensor<fp16, [1, 64, 1, 197]> var_11417_cast_fp16 = conv(dilations = var_11417_dilations_0, groups = var_11417_groups_0, pad = var_11417_pad_0, pad_type = var_11417_pad_type_0, strides = var_11417_strides_0, weight = model_blocks_7_attn_q_projs_10_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11417_cast_fp16")];
            tensor<int32, [1]> var_11418_axes_0 = const()[name = tensor<string, []>("op_11418_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11418_cast_fp16 = squeeze(axes = var_11418_axes_0, x = var_11417_cast_fp16)[name = tensor<string, []>("op_11418_cast_fp16")];
            tensor<int32, [3]> q_941_perm_0 = const()[name = tensor<string, []>("q_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11425_pad_type_0 = const()[name = tensor<string, []>("op_11425_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11425_strides_0 = const()[name = tensor<string, []>("op_11425_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11425_pad_0 = const()[name = tensor<string, []>("op_11425_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11425_dilations_0 = const()[name = tensor<string, []>("op_11425_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11425_groups_0 = const()[name = tensor<string, []>("op_11425_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103888768)))];
            tensor<fp16, [1, 64, 1, 197]> var_11425_cast_fp16 = conv(dilations = var_11425_dilations_0, groups = var_11425_groups_0, pad = var_11425_pad_0, pad_type = var_11425_pad_type_0, strides = var_11425_strides_0, weight = model_blocks_7_attn_k_projs_10_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11425_cast_fp16")];
            tensor<int32, [1]> var_11426_axes_0 = const()[name = tensor<string, []>("op_11426_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11426_cast_fp16 = squeeze(axes = var_11426_axes_0, x = var_11425_cast_fp16)[name = tensor<string, []>("op_11426_cast_fp16")];
            tensor<int32, [3]> k_565_perm_0 = const()[name = tensor<string, []>("k_565_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11433_pad_type_0 = const()[name = tensor<string, []>("op_11433_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11433_strides_0 = const()[name = tensor<string, []>("op_11433_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11433_pad_0 = const()[name = tensor<string, []>("op_11433_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11433_dilations_0 = const()[name = tensor<string, []>("op_11433_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11433_groups_0 = const()[name = tensor<string, []>("op_11433_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103987136)))];
            tensor<fp16, [1, 64, 1, 197]> var_11433_cast_fp16 = conv(dilations = var_11433_dilations_0, groups = var_11433_groups_0, pad = var_11433_pad_0, pad_type = var_11433_pad_type_0, strides = var_11433_strides_0, weight = model_blocks_7_attn_v_projs_10_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11433_cast_fp16")];
            tensor<int32, [1]> var_11434_axes_0 = const()[name = tensor<string, []>("op_11434_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11434_cast_fp16 = squeeze(axes = var_11434_axes_0, x = var_11433_cast_fp16)[name = tensor<string, []>("op_11434_cast_fp16")];
            tensor<int32, [3]> v_565_perm_0 = const()[name = tensor<string, []>("v_565_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104085504)))];
            tensor<fp16, [1, 197, 64]> q_941_cast_fp16 = transpose(perm = q_941_perm_0, x = var_11418_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 197, 64]> q_943_cast_fp16 = add(x = q_941_cast_fp16, y = model_blocks_7_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_943_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104085696)))];
            tensor<fp16, [1, 197, 64]> v_565_cast_fp16 = transpose(perm = v_565_perm_0, x = var_11434_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [1, 197, 64]> v_567_cast_fp16 = add(x = v_565_cast_fp16, y = model_blocks_7_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_567_cast_fp16")];
            tensor<int32, [1]> q_945_axes_0 = const()[name = tensor<string, []>("q_945_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_945_cast_fp16 = expand_dims(axes = q_945_axes_0, x = q_943_cast_fp16)[name = tensor<string, []>("q_945_cast_fp16")];
            tensor<int32, [1]> k_567_axes_0 = const()[name = tensor<string, []>("k_567_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_565_cast_fp16 = transpose(perm = k_565_perm_0, x = var_11426_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 1, 197, 64]> k_567_cast_fp16 = expand_dims(axes = k_567_axes_0, x = k_565_cast_fp16)[name = tensor<string, []>("k_567_cast_fp16")];
            tensor<int32, [1]> v_569_axes_0 = const()[name = tensor<string, []>("v_569_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_569_cast_fp16 = expand_dims(axes = v_569_axes_0, x = v_567_cast_fp16)[name = tensor<string, []>("v_569_cast_fp16")];
            tensor<int32, [4]> var_11443_begin_0 = const()[name = tensor<string, []>("op_11443_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11443_end_0 = const()[name = tensor<string, []>("op_11443_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11443_end_mask_0 = const()[name = tensor<string, []>("op_11443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11443_cast_fp16 = slice_by_index(begin = var_11443_begin_0, end = var_11443_end_0, end_mask = var_11443_end_mask_0, x = q_945_cast_fp16)[name = tensor<string, []>("op_11443_cast_fp16")];
            tensor<int32, [4]> var_11447_begin_0 = const()[name = tensor<string, []>("op_11447_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11447_end_0 = const()[name = tensor<string, []>("op_11447_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11447_end_mask_0 = const()[name = tensor<string, []>("op_11447_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11447_cast_fp16 = slice_by_index(begin = var_11447_begin_0, end = var_11447_end_0, end_mask = var_11447_end_mask_0, x = k_567_cast_fp16)[name = tensor<string, []>("op_11447_cast_fp16")];
            tensor<int32, [4]> var_11451_begin_0 = const()[name = tensor<string, []>("op_11451_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11451_end_0 = const()[name = tensor<string, []>("op_11451_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11451_end_mask_0 = const()[name = tensor<string, []>("op_11451_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11451_cast_fp16 = slice_by_index(begin = var_11451_begin_0, end = var_11451_end_0, end_mask = var_11451_end_mask_0, x = q_945_cast_fp16)[name = tensor<string, []>("op_11451_cast_fp16")];
            tensor<int32, [4]> var_11455_begin_0 = const()[name = tensor<string, []>("op_11455_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11455_end_0 = const()[name = tensor<string, []>("op_11455_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11455_end_mask_0 = const()[name = tensor<string, []>("op_11455_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11455_cast_fp16 = slice_by_index(begin = var_11455_begin_0, end = var_11455_end_0, end_mask = var_11455_end_mask_0, x = k_567_cast_fp16)[name = tensor<string, []>("op_11455_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11457_cast_fp16 = mul(x = var_11451_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11457_cast_fp16")];
            tensor<int32, [5]> var_11465 = const()[name = tensor<string, []>("op_11465"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1279_cast_fp16 = reshape(shape = var_11465, x = var_11451_cast_fp16)[name = tensor<string, []>("x_1279_cast_fp16")];
            tensor<int32, [2]> var_11467_split_sizes_0 = const()[name = tensor<string, []>("op_11467_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11467_axis_0 = const()[name = tensor<string, []>("op_11467_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11467_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11467_cast_fp16_1 = split(axis = var_11467_axis_0, split_sizes = var_11467_split_sizes_0, x = x_1279_cast_fp16)[name = tensor<string, []>("op_11467_cast_fp16")];
            tensor<int32, [1]> squeeze_376_axes_0 = const()[name = tensor<string, []>("squeeze_376_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_376_cast_fp16 = squeeze(axes = squeeze_376_axes_0, x = var_11467_cast_fp16_0)[name = tensor<string, []>("squeeze_376_cast_fp16")];
            tensor<int32, [1]> squeeze_377_axes_0 = const()[name = tensor<string, []>("squeeze_377_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_377_cast_fp16 = squeeze(axes = squeeze_377_axes_0, x = var_11467_cast_fp16_1)[name = tensor<string, []>("squeeze_377_cast_fp16")];
            tensor<fp16, []> const_1539_promoted_to_fp16 = const()[name = tensor<string, []>("const_1539_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11470_cast_fp16 = mul(x = squeeze_377_cast_fp16, y = const_1539_promoted_to_fp16)[name = tensor<string, []>("op_11470_cast_fp16")];
            tensor<int32, []> x_1281_axis_0 = const()[name = tensor<string, []>("x_1281_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1281_cast_fp16 = stack(axis = x_1281_axis_0, values = (var_11470_cast_fp16, squeeze_376_cast_fp16))[name = tensor<string, []>("x_1281_cast_fp16")];
            tensor<int32, [4]> var_11476 = const()[name = tensor<string, []>("op_11476"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11477_cast_fp16 = reshape(shape = var_11476, x = x_1281_cast_fp16)[name = tensor<string, []>("op_11477_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11478_cast_fp16 = mul(x = var_11477_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11478_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_189_cast_fp16 = add(x = var_11457_cast_fp16, y = var_11478_cast_fp16)[name = tensor<string, []>("q_patches_189_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11480_cast_fp16 = mul(x = var_11455_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11480_cast_fp16")];
            tensor<int32, [5]> var_11488 = const()[name = tensor<string, []>("op_11488"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1285_cast_fp16 = reshape(shape = var_11488, x = var_11455_cast_fp16)[name = tensor<string, []>("x_1285_cast_fp16")];
            tensor<int32, [2]> var_11490_split_sizes_0 = const()[name = tensor<string, []>("op_11490_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11490_axis_0 = const()[name = tensor<string, []>("op_11490_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11490_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11490_cast_fp16_1 = split(axis = var_11490_axis_0, split_sizes = var_11490_split_sizes_0, x = x_1285_cast_fp16)[name = tensor<string, []>("op_11490_cast_fp16")];
            tensor<int32, [1]> squeeze_378_axes_0 = const()[name = tensor<string, []>("squeeze_378_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_378_cast_fp16 = squeeze(axes = squeeze_378_axes_0, x = var_11490_cast_fp16_0)[name = tensor<string, []>("squeeze_378_cast_fp16")];
            tensor<int32, [1]> squeeze_379_axes_0 = const()[name = tensor<string, []>("squeeze_379_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_379_cast_fp16 = squeeze(axes = squeeze_379_axes_0, x = var_11490_cast_fp16_1)[name = tensor<string, []>("squeeze_379_cast_fp16")];
            tensor<fp16, []> const_1547_promoted_to_fp16 = const()[name = tensor<string, []>("const_1547_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11493_cast_fp16 = mul(x = squeeze_379_cast_fp16, y = const_1547_promoted_to_fp16)[name = tensor<string, []>("op_11493_cast_fp16")];
            tensor<int32, []> x_1287_axis_0 = const()[name = tensor<string, []>("x_1287_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1287_cast_fp16 = stack(axis = x_1287_axis_0, values = (var_11493_cast_fp16, squeeze_378_cast_fp16))[name = tensor<string, []>("x_1287_cast_fp16")];
            tensor<int32, [4]> var_11499 = const()[name = tensor<string, []>("op_11499"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11500_cast_fp16 = reshape(shape = var_11499, x = x_1287_cast_fp16)[name = tensor<string, []>("op_11500_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11501_cast_fp16 = mul(x = var_11500_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11501_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_189_cast_fp16 = add(x = var_11480_cast_fp16, y = var_11501_cast_fp16)[name = tensor<string, []>("k_patches_189_cast_fp16")];
            tensor<bool, []> var_11504_interleave_0 = const()[name = tensor<string, []>("op_11504_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11504_cast_fp16 = concat(axis = var_20, interleave = var_11504_interleave_0, values = (var_11443_cast_fp16, q_patches_189_cast_fp16))[name = tensor<string, []>("op_11504_cast_fp16")];
            tensor<bool, []> var_11507_interleave_0 = const()[name = tensor<string, []>("op_11507_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11507_cast_fp16 = concat(axis = var_20, interleave = var_11507_interleave_0, values = (var_11447_cast_fp16, k_patches_189_cast_fp16))[name = tensor<string, []>("op_11507_cast_fp16")];
            tensor<fp16, []> var_11509_to_fp16 = const()[name = tensor<string, []>("op_11509_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_949_cast_fp16 = mul(x = var_11504_cast_fp16, y = var_11509_to_fp16)[name = tensor<string, []>("q_949_cast_fp16")];
            tensor<bool, []> attn_377_transpose_x_1 = const()[name = tensor<string, []>("attn_377_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_377_transpose_y_1 = const()[name = tensor<string, []>("attn_377_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_377_cast_fp16 = matmul(transpose_x = attn_377_transpose_x_1, transpose_y = attn_377_transpose_y_1, x = q_949_cast_fp16, y = var_11507_cast_fp16)[name = tensor<string, []>("attn_377_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_379_cast_fp16 = softmax(axis = var_21, x = attn_377_cast_fp16)[name = tensor<string, []>("attn_379_cast_fp16")];
            tensor<bool, []> var_11514_transpose_x_0 = const()[name = tensor<string, []>("op_11514_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11514_transpose_y_0 = const()[name = tensor<string, []>("op_11514_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11514_cast_fp16 = matmul(transpose_x = var_11514_transpose_x_0, transpose_y = var_11514_transpose_y_0, x = attn_379_cast_fp16, y = v_569_cast_fp16)[name = tensor<string, []>("op_11514_cast_fp16")];
            tensor<int32, [1]> var_11515_axes_0 = const()[name = tensor<string, []>("op_11515_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11515_cast_fp16 = squeeze(axes = var_11515_axes_0, x = var_11514_cast_fp16)[name = tensor<string, []>("op_11515_cast_fp16")];
            tensor<string, []> var_11521_pad_type_0 = const()[name = tensor<string, []>("op_11521_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11521_strides_0 = const()[name = tensor<string, []>("op_11521_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11521_pad_0 = const()[name = tensor<string, []>("op_11521_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11521_dilations_0 = const()[name = tensor<string, []>("op_11521_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11521_groups_0 = const()[name = tensor<string, []>("op_11521_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104085888)))];
            tensor<fp16, [1, 64, 1, 197]> var_11521_cast_fp16 = conv(dilations = var_11521_dilations_0, groups = var_11521_groups_0, pad = var_11521_pad_0, pad_type = var_11521_pad_type_0, strides = var_11521_strides_0, weight = model_blocks_7_attn_q_projs_11_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11521_cast_fp16")];
            tensor<int32, [1]> var_11522_axes_0 = const()[name = tensor<string, []>("op_11522_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11522_cast_fp16 = squeeze(axes = var_11522_axes_0, x = var_11521_cast_fp16)[name = tensor<string, []>("op_11522_cast_fp16")];
            tensor<int32, [3]> q_951_perm_0 = const()[name = tensor<string, []>("q_951_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11529_pad_type_0 = const()[name = tensor<string, []>("op_11529_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11529_strides_0 = const()[name = tensor<string, []>("op_11529_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11529_pad_0 = const()[name = tensor<string, []>("op_11529_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11529_dilations_0 = const()[name = tensor<string, []>("op_11529_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11529_groups_0 = const()[name = tensor<string, []>("op_11529_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104184256)))];
            tensor<fp16, [1, 64, 1, 197]> var_11529_cast_fp16 = conv(dilations = var_11529_dilations_0, groups = var_11529_groups_0, pad = var_11529_pad_0, pad_type = var_11529_pad_type_0, strides = var_11529_strides_0, weight = model_blocks_7_attn_k_projs_11_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11529_cast_fp16")];
            tensor<int32, [1]> var_11530_axes_0 = const()[name = tensor<string, []>("op_11530_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11530_cast_fp16 = squeeze(axes = var_11530_axes_0, x = var_11529_cast_fp16)[name = tensor<string, []>("op_11530_cast_fp16")];
            tensor<int32, [3]> k_571_perm_0 = const()[name = tensor<string, []>("k_571_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11537_pad_type_0 = const()[name = tensor<string, []>("op_11537_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11537_strides_0 = const()[name = tensor<string, []>("op_11537_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11537_pad_0 = const()[name = tensor<string, []>("op_11537_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11537_dilations_0 = const()[name = tensor<string, []>("op_11537_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11537_groups_0 = const()[name = tensor<string, []>("op_11537_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_7_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104282624)))];
            tensor<fp16, [1, 64, 1, 197]> var_11537_cast_fp16 = conv(dilations = var_11537_dilations_0, groups = var_11537_groups_0, pad = var_11537_pad_0, pad_type = var_11537_pad_type_0, strides = var_11537_strides_0, weight = model_blocks_7_attn_v_projs_11_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_11537_cast_fp16")];
            tensor<int32, [1]> var_11538_axes_0 = const()[name = tensor<string, []>("op_11538_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11538_cast_fp16 = squeeze(axes = var_11538_axes_0, x = var_11537_cast_fp16)[name = tensor<string, []>("op_11538_cast_fp16")];
            tensor<int32, [3]> v_571_perm_0 = const()[name = tensor<string, []>("v_571_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_7_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104380992)))];
            tensor<fp16, [1, 197, 64]> q_951_cast_fp16 = transpose(perm = q_951_perm_0, x = var_11522_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [1, 197, 64]> q_953_cast_fp16 = add(x = q_951_cast_fp16, y = model_blocks_7_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_953_cast_fp16")];
            tensor<fp16, [64]> model_blocks_7_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104381184)))];
            tensor<fp16, [1, 197, 64]> v_571_cast_fp16 = transpose(perm = v_571_perm_0, x = var_11538_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [1, 197, 64]> v_573_cast_fp16 = add(x = v_571_cast_fp16, y = model_blocks_7_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_573_cast_fp16")];
            tensor<int32, [1]> q_955_axes_0 = const()[name = tensor<string, []>("q_955_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_955_cast_fp16 = expand_dims(axes = q_955_axes_0, x = q_953_cast_fp16)[name = tensor<string, []>("q_955_cast_fp16")];
            tensor<int32, [1]> k_573_axes_0 = const()[name = tensor<string, []>("k_573_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_571_cast_fp16 = transpose(perm = k_571_perm_0, x = var_11530_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 1, 197, 64]> k_573_cast_fp16 = expand_dims(axes = k_573_axes_0, x = k_571_cast_fp16)[name = tensor<string, []>("k_573_cast_fp16")];
            tensor<int32, [1]> v_575_axes_0 = const()[name = tensor<string, []>("v_575_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_575_cast_fp16 = expand_dims(axes = v_575_axes_0, x = v_573_cast_fp16)[name = tensor<string, []>("v_575_cast_fp16")];
            tensor<int32, [4]> var_11547_begin_0 = const()[name = tensor<string, []>("op_11547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11547_end_0 = const()[name = tensor<string, []>("op_11547_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11547_end_mask_0 = const()[name = tensor<string, []>("op_11547_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11547_cast_fp16 = slice_by_index(begin = var_11547_begin_0, end = var_11547_end_0, end_mask = var_11547_end_mask_0, x = q_955_cast_fp16)[name = tensor<string, []>("op_11547_cast_fp16")];
            tensor<int32, [4]> var_11551_begin_0 = const()[name = tensor<string, []>("op_11551_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11551_end_0 = const()[name = tensor<string, []>("op_11551_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11551_end_mask_0 = const()[name = tensor<string, []>("op_11551_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11551_cast_fp16 = slice_by_index(begin = var_11551_begin_0, end = var_11551_end_0, end_mask = var_11551_end_mask_0, x = k_573_cast_fp16)[name = tensor<string, []>("op_11551_cast_fp16")];
            tensor<int32, [4]> var_11555_begin_0 = const()[name = tensor<string, []>("op_11555_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11555_end_0 = const()[name = tensor<string, []>("op_11555_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11555_end_mask_0 = const()[name = tensor<string, []>("op_11555_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11555_cast_fp16 = slice_by_index(begin = var_11555_begin_0, end = var_11555_end_0, end_mask = var_11555_end_mask_0, x = q_955_cast_fp16)[name = tensor<string, []>("op_11555_cast_fp16")];
            tensor<int32, [4]> var_11559_begin_0 = const()[name = tensor<string, []>("op_11559_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11559_end_0 = const()[name = tensor<string, []>("op_11559_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11559_end_mask_0 = const()[name = tensor<string, []>("op_11559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11559_cast_fp16 = slice_by_index(begin = var_11559_begin_0, end = var_11559_end_0, end_mask = var_11559_end_mask_0, x = k_573_cast_fp16)[name = tensor<string, []>("op_11559_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11561_cast_fp16 = mul(x = var_11555_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11561_cast_fp16")];
            tensor<int32, [5]> var_11569 = const()[name = tensor<string, []>("op_11569"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1291_cast_fp16 = reshape(shape = var_11569, x = var_11555_cast_fp16)[name = tensor<string, []>("x_1291_cast_fp16")];
            tensor<int32, [2]> var_11571_split_sizes_0 = const()[name = tensor<string, []>("op_11571_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11571_axis_0 = const()[name = tensor<string, []>("op_11571_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11571_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11571_cast_fp16_1 = split(axis = var_11571_axis_0, split_sizes = var_11571_split_sizes_0, x = x_1291_cast_fp16)[name = tensor<string, []>("op_11571_cast_fp16")];
            tensor<int32, [1]> squeeze_380_axes_0 = const()[name = tensor<string, []>("squeeze_380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_380_cast_fp16 = squeeze(axes = squeeze_380_axes_0, x = var_11571_cast_fp16_0)[name = tensor<string, []>("squeeze_380_cast_fp16")];
            tensor<int32, [1]> squeeze_381_axes_0 = const()[name = tensor<string, []>("squeeze_381_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_381_cast_fp16 = squeeze(axes = squeeze_381_axes_0, x = var_11571_cast_fp16_1)[name = tensor<string, []>("squeeze_381_cast_fp16")];
            tensor<fp16, []> const_1555_promoted_to_fp16 = const()[name = tensor<string, []>("const_1555_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11574_cast_fp16 = mul(x = squeeze_381_cast_fp16, y = const_1555_promoted_to_fp16)[name = tensor<string, []>("op_11574_cast_fp16")];
            tensor<int32, []> x_1293_axis_0 = const()[name = tensor<string, []>("x_1293_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1293_cast_fp16 = stack(axis = x_1293_axis_0, values = (var_11574_cast_fp16, squeeze_380_cast_fp16))[name = tensor<string, []>("x_1293_cast_fp16")];
            tensor<int32, [4]> var_11580 = const()[name = tensor<string, []>("op_11580"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11581_cast_fp16 = reshape(shape = var_11580, x = x_1293_cast_fp16)[name = tensor<string, []>("op_11581_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11582_cast_fp16 = mul(x = var_11581_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11582_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_191_cast_fp16 = add(x = var_11561_cast_fp16, y = var_11582_cast_fp16)[name = tensor<string, []>("q_patches_191_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11584_cast_fp16 = mul(x = var_11559_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11584_cast_fp16")];
            tensor<int32, [5]> var_11592 = const()[name = tensor<string, []>("op_11592"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1297_cast_fp16 = reshape(shape = var_11592, x = var_11559_cast_fp16)[name = tensor<string, []>("x_1297_cast_fp16")];
            tensor<int32, [2]> var_11594_split_sizes_0 = const()[name = tensor<string, []>("op_11594_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11594_axis_0 = const()[name = tensor<string, []>("op_11594_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11594_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11594_cast_fp16_1 = split(axis = var_11594_axis_0, split_sizes = var_11594_split_sizes_0, x = x_1297_cast_fp16)[name = tensor<string, []>("op_11594_cast_fp16")];
            tensor<int32, [1]> squeeze_382_axes_0 = const()[name = tensor<string, []>("squeeze_382_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_382_cast_fp16 = squeeze(axes = squeeze_382_axes_0, x = var_11594_cast_fp16_0)[name = tensor<string, []>("squeeze_382_cast_fp16")];
            tensor<int32, [1]> squeeze_383_axes_0 = const()[name = tensor<string, []>("squeeze_383_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_383_cast_fp16 = squeeze(axes = squeeze_383_axes_0, x = var_11594_cast_fp16_1)[name = tensor<string, []>("squeeze_383_cast_fp16")];
            tensor<fp16, []> const_1563_promoted_to_fp16 = const()[name = tensor<string, []>("const_1563_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11597_cast_fp16 = mul(x = squeeze_383_cast_fp16, y = const_1563_promoted_to_fp16)[name = tensor<string, []>("op_11597_cast_fp16")];
            tensor<int32, []> x_1299_axis_0 = const()[name = tensor<string, []>("x_1299_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1299_cast_fp16 = stack(axis = x_1299_axis_0, values = (var_11597_cast_fp16, squeeze_382_cast_fp16))[name = tensor<string, []>("x_1299_cast_fp16")];
            tensor<int32, [4]> var_11603 = const()[name = tensor<string, []>("op_11603"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11604_cast_fp16 = reshape(shape = var_11603, x = x_1299_cast_fp16)[name = tensor<string, []>("op_11604_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11605_cast_fp16 = mul(x = var_11604_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11605_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_191_cast_fp16 = add(x = var_11584_cast_fp16, y = var_11605_cast_fp16)[name = tensor<string, []>("k_patches_191_cast_fp16")];
            tensor<bool, []> var_11608_interleave_0 = const()[name = tensor<string, []>("op_11608_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11608_cast_fp16 = concat(axis = var_20, interleave = var_11608_interleave_0, values = (var_11547_cast_fp16, q_patches_191_cast_fp16))[name = tensor<string, []>("op_11608_cast_fp16")];
            tensor<bool, []> var_11611_interleave_0 = const()[name = tensor<string, []>("op_11611_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11611_cast_fp16 = concat(axis = var_20, interleave = var_11611_interleave_0, values = (var_11551_cast_fp16, k_patches_191_cast_fp16))[name = tensor<string, []>("op_11611_cast_fp16")];
            tensor<fp16, []> var_11613_to_fp16 = const()[name = tensor<string, []>("op_11613_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_959_cast_fp16 = mul(x = var_11608_cast_fp16, y = var_11613_to_fp16)[name = tensor<string, []>("q_959_cast_fp16")];
            tensor<bool, []> attn_381_transpose_x_1 = const()[name = tensor<string, []>("attn_381_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_381_transpose_y_1 = const()[name = tensor<string, []>("attn_381_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_381_cast_fp16 = matmul(transpose_x = attn_381_transpose_x_1, transpose_y = attn_381_transpose_y_1, x = q_959_cast_fp16, y = var_11611_cast_fp16)[name = tensor<string, []>("attn_381_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_383_cast_fp16 = softmax(axis = var_21, x = attn_381_cast_fp16)[name = tensor<string, []>("attn_383_cast_fp16")];
            tensor<bool, []> var_11618_transpose_x_0 = const()[name = tensor<string, []>("op_11618_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11618_transpose_y_0 = const()[name = tensor<string, []>("op_11618_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11618_cast_fp16 = matmul(transpose_x = var_11618_transpose_x_0, transpose_y = var_11618_transpose_y_0, x = attn_383_cast_fp16, y = v_575_cast_fp16)[name = tensor<string, []>("op_11618_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_15_cast_fp16 = squeeze(axes = out_15_axes_0, x = var_11618_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<bool, []> input_131_interleave_0 = const()[name = tensor<string, []>("input_131_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_131_cast_fp16 = concat(axis = var_21, interleave = input_131_interleave_0, values = (var_10475_cast_fp16, var_10579_cast_fp16, var_10683_cast_fp16, var_10787_cast_fp16, var_10891_cast_fp16, var_10995_cast_fp16, var_11099_cast_fp16, var_11203_cast_fp16, var_11307_cast_fp16, var_11411_cast_fp16, var_11515_cast_fp16, out_15_cast_fp16))[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<int32, [1]> x_1301_axes_0 = const()[name = tensor<string, []>("x_1301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_7_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104381376)))];
            tensor<fp16, [768]> model_blocks_7_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104382976)))];
            tensor<fp16, [1, 197, 768]> x_1301_cast_fp16 = layer_norm(axes = x_1301_axes_0, beta = model_blocks_7_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_7_attn_inner_attn_ln_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("x_1301_cast_fp16")];
            tensor<int32, [4]> var_11626 = const()[name = tensor<string, []>("op_11626"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1303_cast_fp16 = reshape(shape = var_11626, x = x_1301_cast_fp16)[name = tensor<string, []>("x_1303_cast_fp16")];
            tensor<int32, [4]> input_133_perm_0 = const()[name = tensor<string, []>("input_133_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1305_pad_type_0 = const()[name = tensor<string, []>("x_1305_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1305_strides_0 = const()[name = tensor<string, []>("x_1305_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1305_pad_0 = const()[name = tensor<string, []>("x_1305_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1305_dilations_0 = const()[name = tensor<string, []>("x_1305_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1305_groups_0 = const()[name = tensor<string, []>("x_1305_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_7_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104384576)))];
            tensor<fp16, [768]> model_blocks_7_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105564288)))];
            tensor<fp16, [1, 768, 1, 197]> input_133_cast_fp16 = transpose(perm = input_133_perm_0, x = x_1303_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 768, 1, 197]> x_1305_cast_fp16 = conv(bias = model_blocks_7_attn_proj_bias_to_fp16, dilations = x_1305_dilations_0, groups = x_1305_groups_0, pad = x_1305_pad_0, pad_type = x_1305_pad_type_0, strides = x_1305_strides_0, weight = model_blocks_7_attn_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("x_1305_cast_fp16")];
            tensor<int32, [4]> x_1307_perm_0 = const()[name = tensor<string, []>("x_1307_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_11637 = const()[name = tensor<string, []>("op_11637"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1307_cast_fp16 = transpose(perm = x_1307_perm_0, x = x_1305_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [1, 197, 768]> var_11638_cast_fp16 = reshape(shape = var_11637, x = x_1307_cast_fp16)[name = tensor<string, []>("op_11638_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_135_cast_fp16 = add(x = input_127_cast_fp16, y = var_11638_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [1]> x_1309_axes_0 = const()[name = tensor<string, []>("x_1309_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105565888)))];
            tensor<fp16, [768]> model_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105567488)))];
            tensor<fp16, [1, 197, 768]> x_1309_cast_fp16 = layer_norm(axes = x_1309_axes_0, beta = model_blocks_7_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_7_norm2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("x_1309_cast_fp16")];
            tensor<int32, [4]> var_11650 = const()[name = tensor<string, []>("op_11650"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1311_cast_fp16 = reshape(shape = var_11650, x = x_1309_cast_fp16)[name = tensor<string, []>("x_1311_cast_fp16")];
            tensor<int32, [4]> input_137_perm_0 = const()[name = tensor<string, []>("input_137_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_139_pad_type_0 = const()[name = tensor<string, []>("input_139_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_139_strides_0 = const()[name = tensor<string, []>("input_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_139_pad_0 = const()[name = tensor<string, []>("input_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_139_dilations_0 = const()[name = tensor<string, []>("input_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_139_groups_0 = const()[name = tensor<string, []>("input_139_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_7_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105569088)))];
            tensor<fp16, [2048]> model_blocks_7_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108714880)))];
            tensor<fp16, [1, 768, 1, 197]> input_137_cast_fp16 = transpose(perm = input_137_perm_0, x = x_1311_cast_fp16)[name = tensor<string, []>("transpose_175")];
            tensor<fp16, [1, 2048, 1, 197]> input_139_cast_fp16 = conv(bias = model_blocks_7_mlp_w1_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = model_blocks_7_mlp_w1_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<string, []> x2_399_pad_type_0 = const()[name = tensor<string, []>("x2_399_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_399_strides_0 = const()[name = tensor<string, []>("x2_399_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_399_pad_0 = const()[name = tensor<string, []>("x2_399_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_399_dilations_0 = const()[name = tensor<string, []>("x2_399_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_399_groups_0 = const()[name = tensor<string, []>("x2_399_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_7_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108719040)))];
            tensor<fp16, [2048]> model_blocks_7_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111864832)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_399_cast_fp16 = conv(bias = model_blocks_7_mlp_w2_bias_to_fp16, dilations = x2_399_dilations_0, groups = x2_399_groups_0, pad = x2_399_pad_0, pad_type = x2_399_pad_type_0, strides = x2_399_strides_0, weight = model_blocks_7_mlp_w2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("x2_399_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_11667_cast_fp16 = silu(x = input_139_cast_fp16)[name = tensor<string, []>("op_11667_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_57_cast_fp16 = mul(x = var_11667_cast_fp16, y = x2_399_cast_fp16)[name = tensor<string, []>("hidden_57_cast_fp16")];
            tensor<int32, [4]> hidden_59_perm_0 = const()[name = tensor<string, []>("hidden_59_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_11670 = const()[name = tensor<string, []>("op_11670"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_59_cast_fp16 = transpose(perm = hidden_59_perm_0, x = hidden_57_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [1, 197, 2048]> input_141_cast_fp16 = reshape(shape = var_11670, x = hidden_59_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<int32, [1]> hidden_61_axes_0 = const()[name = tensor<string, []>("hidden_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_7_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111868992)))];
            tensor<fp16, [2048]> model_blocks_7_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111873152)))];
            tensor<fp16, [1, 197, 2048]> hidden_61_cast_fp16 = layer_norm(axes = hidden_61_axes_0, beta = model_blocks_7_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_7_mlp_ffn_ln_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("hidden_61_cast_fp16")];
            tensor<int32, [4]> var_11676 = const()[name = tensor<string, []>("op_11676"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_63_cast_fp16 = reshape(shape = var_11676, x = hidden_61_cast_fp16)[name = tensor<string, []>("hidden_63_cast_fp16")];
            tensor<int32, [4]> input_143_perm_0 = const()[name = tensor<string, []>("input_143_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1313_pad_type_0 = const()[name = tensor<string, []>("x_1313_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1313_strides_0 = const()[name = tensor<string, []>("x_1313_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1313_pad_0 = const()[name = tensor<string, []>("x_1313_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1313_dilations_0 = const()[name = tensor<string, []>("x_1313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1313_groups_0 = const()[name = tensor<string, []>("x_1313_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_7_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111877312)))];
            tensor<fp16, [768]> model_blocks_7_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_7_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115023104)))];
            tensor<fp16, [1, 2048, 1, 197]> input_143_cast_fp16 = transpose(perm = input_143_perm_0, x = hidden_63_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [1, 768, 1, 197]> x_1313_cast_fp16 = conv(bias = model_blocks_7_mlp_w3_bias_to_fp16, dilations = x_1313_dilations_0, groups = x_1313_groups_0, pad = x_1313_pad_0, pad_type = x_1313_pad_type_0, strides = x_1313_strides_0, weight = model_blocks_7_mlp_w3_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("x_1313_cast_fp16")];
            tensor<int32, [4]> x_1315_perm_0 = const()[name = tensor<string, []>("x_1315_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_11687 = const()[name = tensor<string, []>("op_11687"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1315_cast_fp16 = transpose(perm = x_1315_perm_0, x = x_1313_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 197, 768]> var_11688_cast_fp16 = reshape(shape = var_11687, x = x_1315_cast_fp16)[name = tensor<string, []>("op_11688_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_145_cast_fp16 = add(x = input_135_cast_fp16, y = var_11688_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<int32, [1]> x_1317_axes_0 = const()[name = tensor<string, []>("x_1317_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115024704)))];
            tensor<fp16, [768]> model_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115026304)))];
            tensor<fp16, [1, 197, 768]> x_1317_cast_fp16 = layer_norm(axes = x_1317_axes_0, beta = model_blocks_8_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_8_norm1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("x_1317_cast_fp16")];
            tensor<int32, [4]> var_11822 = const()[name = tensor<string, []>("op_11822"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1319_cast_fp16 = reshape(shape = var_11822, x = x_1317_cast_fp16)[name = tensor<string, []>("x_1319_cast_fp16")];
            tensor<int32, [4]> input_147_perm_0 = const()[name = tensor<string, []>("input_147_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_11830_pad_type_0 = const()[name = tensor<string, []>("op_11830_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11830_strides_0 = const()[name = tensor<string, []>("op_11830_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11830_pad_0 = const()[name = tensor<string, []>("op_11830_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11830_dilations_0 = const()[name = tensor<string, []>("op_11830_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11830_groups_0 = const()[name = tensor<string, []>("op_11830_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115027904)))];
            tensor<fp16, [1, 768, 1, 197]> input_147_cast_fp16 = transpose(perm = input_147_perm_0, x = x_1319_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 64, 1, 197]> var_11830_cast_fp16 = conv(dilations = var_11830_dilations_0, groups = var_11830_groups_0, pad = var_11830_pad_0, pad_type = var_11830_pad_type_0, strides = var_11830_strides_0, weight = model_blocks_8_attn_q_projs_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11830_cast_fp16")];
            tensor<int32, [1]> var_11831_axes_0 = const()[name = tensor<string, []>("op_11831_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11831_cast_fp16 = squeeze(axes = var_11831_axes_0, x = var_11830_cast_fp16)[name = tensor<string, []>("op_11831_cast_fp16")];
            tensor<int32, [3]> q_961_perm_0 = const()[name = tensor<string, []>("q_961_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11838_pad_type_0 = const()[name = tensor<string, []>("op_11838_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11838_strides_0 = const()[name = tensor<string, []>("op_11838_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11838_pad_0 = const()[name = tensor<string, []>("op_11838_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11838_dilations_0 = const()[name = tensor<string, []>("op_11838_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11838_groups_0 = const()[name = tensor<string, []>("op_11838_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115126272)))];
            tensor<fp16, [1, 64, 1, 197]> var_11838_cast_fp16 = conv(dilations = var_11838_dilations_0, groups = var_11838_groups_0, pad = var_11838_pad_0, pad_type = var_11838_pad_type_0, strides = var_11838_strides_0, weight = model_blocks_8_attn_k_projs_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11838_cast_fp16")];
            tensor<int32, [1]> var_11839_axes_0 = const()[name = tensor<string, []>("op_11839_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11839_cast_fp16 = squeeze(axes = var_11839_axes_0, x = var_11838_cast_fp16)[name = tensor<string, []>("op_11839_cast_fp16")];
            tensor<int32, [3]> k_577_perm_0 = const()[name = tensor<string, []>("k_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11846_pad_type_0 = const()[name = tensor<string, []>("op_11846_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11846_strides_0 = const()[name = tensor<string, []>("op_11846_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11846_pad_0 = const()[name = tensor<string, []>("op_11846_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11846_dilations_0 = const()[name = tensor<string, []>("op_11846_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11846_groups_0 = const()[name = tensor<string, []>("op_11846_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115224640)))];
            tensor<fp16, [1, 64, 1, 197]> var_11846_cast_fp16 = conv(dilations = var_11846_dilations_0, groups = var_11846_groups_0, pad = var_11846_pad_0, pad_type = var_11846_pad_type_0, strides = var_11846_strides_0, weight = model_blocks_8_attn_v_projs_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11846_cast_fp16")];
            tensor<int32, [1]> var_11847_axes_0 = const()[name = tensor<string, []>("op_11847_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11847_cast_fp16 = squeeze(axes = var_11847_axes_0, x = var_11846_cast_fp16)[name = tensor<string, []>("op_11847_cast_fp16")];
            tensor<int32, [3]> v_577_perm_0 = const()[name = tensor<string, []>("v_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115323008)))];
            tensor<fp16, [1, 197, 64]> q_961_cast_fp16 = transpose(perm = q_961_perm_0, x = var_11831_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 197, 64]> q_963_cast_fp16 = add(x = q_961_cast_fp16, y = model_blocks_8_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_963_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115323200)))];
            tensor<fp16, [1, 197, 64]> v_577_cast_fp16 = transpose(perm = v_577_perm_0, x = var_11847_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [1, 197, 64]> v_579_cast_fp16 = add(x = v_577_cast_fp16, y = model_blocks_8_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_579_cast_fp16")];
            tensor<int32, [1]> q_965_axes_0 = const()[name = tensor<string, []>("q_965_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_965_cast_fp16 = expand_dims(axes = q_965_axes_0, x = q_963_cast_fp16)[name = tensor<string, []>("q_965_cast_fp16")];
            tensor<int32, [1]> k_579_axes_0 = const()[name = tensor<string, []>("k_579_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_577_cast_fp16 = transpose(perm = k_577_perm_0, x = var_11839_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [1, 1, 197, 64]> k_579_cast_fp16 = expand_dims(axes = k_579_axes_0, x = k_577_cast_fp16)[name = tensor<string, []>("k_579_cast_fp16")];
            tensor<int32, [1]> v_581_axes_0 = const()[name = tensor<string, []>("v_581_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_581_cast_fp16 = expand_dims(axes = v_581_axes_0, x = v_579_cast_fp16)[name = tensor<string, []>("v_581_cast_fp16")];
            tensor<int32, [4]> var_11856_begin_0 = const()[name = tensor<string, []>("op_11856_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11856_end_0 = const()[name = tensor<string, []>("op_11856_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11856_end_mask_0 = const()[name = tensor<string, []>("op_11856_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11856_cast_fp16 = slice_by_index(begin = var_11856_begin_0, end = var_11856_end_0, end_mask = var_11856_end_mask_0, x = q_965_cast_fp16)[name = tensor<string, []>("op_11856_cast_fp16")];
            tensor<int32, [4]> var_11860_begin_0 = const()[name = tensor<string, []>("op_11860_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11860_end_0 = const()[name = tensor<string, []>("op_11860_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11860_end_mask_0 = const()[name = tensor<string, []>("op_11860_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11860_cast_fp16 = slice_by_index(begin = var_11860_begin_0, end = var_11860_end_0, end_mask = var_11860_end_mask_0, x = k_579_cast_fp16)[name = tensor<string, []>("op_11860_cast_fp16")];
            tensor<int32, [4]> var_11864_begin_0 = const()[name = tensor<string, []>("op_11864_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11864_end_0 = const()[name = tensor<string, []>("op_11864_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11864_end_mask_0 = const()[name = tensor<string, []>("op_11864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11864_cast_fp16 = slice_by_index(begin = var_11864_begin_0, end = var_11864_end_0, end_mask = var_11864_end_mask_0, x = q_965_cast_fp16)[name = tensor<string, []>("op_11864_cast_fp16")];
            tensor<int32, [4]> var_11868_begin_0 = const()[name = tensor<string, []>("op_11868_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11868_end_0 = const()[name = tensor<string, []>("op_11868_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11868_end_mask_0 = const()[name = tensor<string, []>("op_11868_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11868_cast_fp16 = slice_by_index(begin = var_11868_begin_0, end = var_11868_end_0, end_mask = var_11868_end_mask_0, x = k_579_cast_fp16)[name = tensor<string, []>("op_11868_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11870_cast_fp16 = mul(x = var_11864_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11870_cast_fp16")];
            tensor<int32, [5]> var_11878 = const()[name = tensor<string, []>("op_11878"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1323_cast_fp16 = reshape(shape = var_11878, x = var_11864_cast_fp16)[name = tensor<string, []>("x_1323_cast_fp16")];
            tensor<int32, [2]> var_11880_split_sizes_0 = const()[name = tensor<string, []>("op_11880_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11880_axis_0 = const()[name = tensor<string, []>("op_11880_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11880_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11880_cast_fp16_1 = split(axis = var_11880_axis_0, split_sizes = var_11880_split_sizes_0, x = x_1323_cast_fp16)[name = tensor<string, []>("op_11880_cast_fp16")];
            tensor<int32, [1]> squeeze_384_axes_0 = const()[name = tensor<string, []>("squeeze_384_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_384_cast_fp16 = squeeze(axes = squeeze_384_axes_0, x = var_11880_cast_fp16_0)[name = tensor<string, []>("squeeze_384_cast_fp16")];
            tensor<int32, [1]> squeeze_385_axes_0 = const()[name = tensor<string, []>("squeeze_385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_385_cast_fp16 = squeeze(axes = squeeze_385_axes_0, x = var_11880_cast_fp16_1)[name = tensor<string, []>("squeeze_385_cast_fp16")];
            tensor<fp16, []> const_1575_promoted_to_fp16 = const()[name = tensor<string, []>("const_1575_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11883_cast_fp16 = mul(x = squeeze_385_cast_fp16, y = const_1575_promoted_to_fp16)[name = tensor<string, []>("op_11883_cast_fp16")];
            tensor<int32, []> x_1325_axis_0 = const()[name = tensor<string, []>("x_1325_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1325_cast_fp16 = stack(axis = x_1325_axis_0, values = (var_11883_cast_fp16, squeeze_384_cast_fp16))[name = tensor<string, []>("x_1325_cast_fp16")];
            tensor<int32, [4]> var_11889 = const()[name = tensor<string, []>("op_11889"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11890_cast_fp16 = reshape(shape = var_11889, x = x_1325_cast_fp16)[name = tensor<string, []>("op_11890_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11891_cast_fp16 = mul(x = var_11890_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11891_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_193_cast_fp16 = add(x = var_11870_cast_fp16, y = var_11891_cast_fp16)[name = tensor<string, []>("q_patches_193_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11893_cast_fp16 = mul(x = var_11868_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11893_cast_fp16")];
            tensor<int32, [5]> var_11901 = const()[name = tensor<string, []>("op_11901"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1329_cast_fp16 = reshape(shape = var_11901, x = var_11868_cast_fp16)[name = tensor<string, []>("x_1329_cast_fp16")];
            tensor<int32, [2]> var_11903_split_sizes_0 = const()[name = tensor<string, []>("op_11903_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11903_axis_0 = const()[name = tensor<string, []>("op_11903_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11903_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11903_cast_fp16_1 = split(axis = var_11903_axis_0, split_sizes = var_11903_split_sizes_0, x = x_1329_cast_fp16)[name = tensor<string, []>("op_11903_cast_fp16")];
            tensor<int32, [1]> squeeze_386_axes_0 = const()[name = tensor<string, []>("squeeze_386_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_386_cast_fp16 = squeeze(axes = squeeze_386_axes_0, x = var_11903_cast_fp16_0)[name = tensor<string, []>("squeeze_386_cast_fp16")];
            tensor<int32, [1]> squeeze_387_axes_0 = const()[name = tensor<string, []>("squeeze_387_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_387_cast_fp16 = squeeze(axes = squeeze_387_axes_0, x = var_11903_cast_fp16_1)[name = tensor<string, []>("squeeze_387_cast_fp16")];
            tensor<fp16, []> const_1583_promoted_to_fp16 = const()[name = tensor<string, []>("const_1583_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11906_cast_fp16 = mul(x = squeeze_387_cast_fp16, y = const_1583_promoted_to_fp16)[name = tensor<string, []>("op_11906_cast_fp16")];
            tensor<int32, []> x_1331_axis_0 = const()[name = tensor<string, []>("x_1331_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1331_cast_fp16 = stack(axis = x_1331_axis_0, values = (var_11906_cast_fp16, squeeze_386_cast_fp16))[name = tensor<string, []>("x_1331_cast_fp16")];
            tensor<int32, [4]> var_11912 = const()[name = tensor<string, []>("op_11912"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11913_cast_fp16 = reshape(shape = var_11912, x = x_1331_cast_fp16)[name = tensor<string, []>("op_11913_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11914_cast_fp16 = mul(x = var_11913_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11914_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_193_cast_fp16 = add(x = var_11893_cast_fp16, y = var_11914_cast_fp16)[name = tensor<string, []>("k_patches_193_cast_fp16")];
            tensor<bool, []> var_11917_interleave_0 = const()[name = tensor<string, []>("op_11917_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11917_cast_fp16 = concat(axis = var_20, interleave = var_11917_interleave_0, values = (var_11856_cast_fp16, q_patches_193_cast_fp16))[name = tensor<string, []>("op_11917_cast_fp16")];
            tensor<bool, []> var_11920_interleave_0 = const()[name = tensor<string, []>("op_11920_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11920_cast_fp16 = concat(axis = var_20, interleave = var_11920_interleave_0, values = (var_11860_cast_fp16, k_patches_193_cast_fp16))[name = tensor<string, []>("op_11920_cast_fp16")];
            tensor<fp16, []> var_11922_to_fp16 = const()[name = tensor<string, []>("op_11922_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_969_cast_fp16 = mul(x = var_11917_cast_fp16, y = var_11922_to_fp16)[name = tensor<string, []>("q_969_cast_fp16")];
            tensor<bool, []> attn_385_transpose_x_1 = const()[name = tensor<string, []>("attn_385_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_385_transpose_y_1 = const()[name = tensor<string, []>("attn_385_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_385_cast_fp16 = matmul(transpose_x = attn_385_transpose_x_1, transpose_y = attn_385_transpose_y_1, x = q_969_cast_fp16, y = var_11920_cast_fp16)[name = tensor<string, []>("attn_385_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_387_cast_fp16 = softmax(axis = var_21, x = attn_385_cast_fp16)[name = tensor<string, []>("attn_387_cast_fp16")];
            tensor<bool, []> var_11927_transpose_x_0 = const()[name = tensor<string, []>("op_11927_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_11927_transpose_y_0 = const()[name = tensor<string, []>("op_11927_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_11927_cast_fp16 = matmul(transpose_x = var_11927_transpose_x_0, transpose_y = var_11927_transpose_y_0, x = attn_387_cast_fp16, y = v_581_cast_fp16)[name = tensor<string, []>("op_11927_cast_fp16")];
            tensor<int32, [1]> var_11928_axes_0 = const()[name = tensor<string, []>("op_11928_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_11928_cast_fp16 = squeeze(axes = var_11928_axes_0, x = var_11927_cast_fp16)[name = tensor<string, []>("op_11928_cast_fp16")];
            tensor<string, []> var_11934_pad_type_0 = const()[name = tensor<string, []>("op_11934_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11934_strides_0 = const()[name = tensor<string, []>("op_11934_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11934_pad_0 = const()[name = tensor<string, []>("op_11934_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11934_dilations_0 = const()[name = tensor<string, []>("op_11934_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11934_groups_0 = const()[name = tensor<string, []>("op_11934_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115323392)))];
            tensor<fp16, [1, 64, 1, 197]> var_11934_cast_fp16 = conv(dilations = var_11934_dilations_0, groups = var_11934_groups_0, pad = var_11934_pad_0, pad_type = var_11934_pad_type_0, strides = var_11934_strides_0, weight = model_blocks_8_attn_q_projs_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11934_cast_fp16")];
            tensor<int32, [1]> var_11935_axes_0 = const()[name = tensor<string, []>("op_11935_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11935_cast_fp16 = squeeze(axes = var_11935_axes_0, x = var_11934_cast_fp16)[name = tensor<string, []>("op_11935_cast_fp16")];
            tensor<int32, [3]> q_971_perm_0 = const()[name = tensor<string, []>("q_971_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11942_pad_type_0 = const()[name = tensor<string, []>("op_11942_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11942_strides_0 = const()[name = tensor<string, []>("op_11942_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11942_pad_0 = const()[name = tensor<string, []>("op_11942_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11942_dilations_0 = const()[name = tensor<string, []>("op_11942_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11942_groups_0 = const()[name = tensor<string, []>("op_11942_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115421760)))];
            tensor<fp16, [1, 64, 1, 197]> var_11942_cast_fp16 = conv(dilations = var_11942_dilations_0, groups = var_11942_groups_0, pad = var_11942_pad_0, pad_type = var_11942_pad_type_0, strides = var_11942_strides_0, weight = model_blocks_8_attn_k_projs_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11942_cast_fp16")];
            tensor<int32, [1]> var_11943_axes_0 = const()[name = tensor<string, []>("op_11943_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11943_cast_fp16 = squeeze(axes = var_11943_axes_0, x = var_11942_cast_fp16)[name = tensor<string, []>("op_11943_cast_fp16")];
            tensor<int32, [3]> k_583_perm_0 = const()[name = tensor<string, []>("k_583_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_11950_pad_type_0 = const()[name = tensor<string, []>("op_11950_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_11950_strides_0 = const()[name = tensor<string, []>("op_11950_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11950_pad_0 = const()[name = tensor<string, []>("op_11950_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11950_dilations_0 = const()[name = tensor<string, []>("op_11950_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11950_groups_0 = const()[name = tensor<string, []>("op_11950_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115520128)))];
            tensor<fp16, [1, 64, 1, 197]> var_11950_cast_fp16 = conv(dilations = var_11950_dilations_0, groups = var_11950_groups_0, pad = var_11950_pad_0, pad_type = var_11950_pad_type_0, strides = var_11950_strides_0, weight = model_blocks_8_attn_v_projs_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_11950_cast_fp16")];
            tensor<int32, [1]> var_11951_axes_0 = const()[name = tensor<string, []>("op_11951_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_11951_cast_fp16 = squeeze(axes = var_11951_axes_0, x = var_11950_cast_fp16)[name = tensor<string, []>("op_11951_cast_fp16")];
            tensor<int32, [3]> v_583_perm_0 = const()[name = tensor<string, []>("v_583_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115618496)))];
            tensor<fp16, [1, 197, 64]> q_971_cast_fp16 = transpose(perm = q_971_perm_0, x = var_11935_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 197, 64]> q_973_cast_fp16 = add(x = q_971_cast_fp16, y = model_blocks_8_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_973_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115618688)))];
            tensor<fp16, [1, 197, 64]> v_583_cast_fp16 = transpose(perm = v_583_perm_0, x = var_11951_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 197, 64]> v_585_cast_fp16 = add(x = v_583_cast_fp16, y = model_blocks_8_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_585_cast_fp16")];
            tensor<int32, [1]> q_975_axes_0 = const()[name = tensor<string, []>("q_975_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_975_cast_fp16 = expand_dims(axes = q_975_axes_0, x = q_973_cast_fp16)[name = tensor<string, []>("q_975_cast_fp16")];
            tensor<int32, [1]> k_585_axes_0 = const()[name = tensor<string, []>("k_585_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_583_cast_fp16 = transpose(perm = k_583_perm_0, x = var_11943_cast_fp16)[name = tensor<string, []>("transpose_166")];
            tensor<fp16, [1, 1, 197, 64]> k_585_cast_fp16 = expand_dims(axes = k_585_axes_0, x = k_583_cast_fp16)[name = tensor<string, []>("k_585_cast_fp16")];
            tensor<int32, [1]> v_587_axes_0 = const()[name = tensor<string, []>("v_587_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_587_cast_fp16 = expand_dims(axes = v_587_axes_0, x = v_585_cast_fp16)[name = tensor<string, []>("v_587_cast_fp16")];
            tensor<int32, [4]> var_11960_begin_0 = const()[name = tensor<string, []>("op_11960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11960_end_0 = const()[name = tensor<string, []>("op_11960_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11960_end_mask_0 = const()[name = tensor<string, []>("op_11960_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11960_cast_fp16 = slice_by_index(begin = var_11960_begin_0, end = var_11960_end_0, end_mask = var_11960_end_mask_0, x = q_975_cast_fp16)[name = tensor<string, []>("op_11960_cast_fp16")];
            tensor<int32, [4]> var_11964_begin_0 = const()[name = tensor<string, []>("op_11964_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11964_end_0 = const()[name = tensor<string, []>("op_11964_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_11964_end_mask_0 = const()[name = tensor<string, []>("op_11964_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_11964_cast_fp16 = slice_by_index(begin = var_11964_begin_0, end = var_11964_end_0, end_mask = var_11964_end_mask_0, x = k_585_cast_fp16)[name = tensor<string, []>("op_11964_cast_fp16")];
            tensor<int32, [4]> var_11968_begin_0 = const()[name = tensor<string, []>("op_11968_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11968_end_0 = const()[name = tensor<string, []>("op_11968_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11968_end_mask_0 = const()[name = tensor<string, []>("op_11968_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11968_cast_fp16 = slice_by_index(begin = var_11968_begin_0, end = var_11968_end_0, end_mask = var_11968_end_mask_0, x = q_975_cast_fp16)[name = tensor<string, []>("op_11968_cast_fp16")];
            tensor<int32, [4]> var_11972_begin_0 = const()[name = tensor<string, []>("op_11972_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11972_end_0 = const()[name = tensor<string, []>("op_11972_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_11972_end_mask_0 = const()[name = tensor<string, []>("op_11972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_11972_cast_fp16 = slice_by_index(begin = var_11972_begin_0, end = var_11972_end_0, end_mask = var_11972_end_mask_0, x = k_585_cast_fp16)[name = tensor<string, []>("op_11972_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11974_cast_fp16 = mul(x = var_11968_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11974_cast_fp16")];
            tensor<int32, [5]> var_11982 = const()[name = tensor<string, []>("op_11982"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1335_cast_fp16 = reshape(shape = var_11982, x = var_11968_cast_fp16)[name = tensor<string, []>("x_1335_cast_fp16")];
            tensor<int32, [2]> var_11984_split_sizes_0 = const()[name = tensor<string, []>("op_11984_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_11984_axis_0 = const()[name = tensor<string, []>("op_11984_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_11984_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_11984_cast_fp16_1 = split(axis = var_11984_axis_0, split_sizes = var_11984_split_sizes_0, x = x_1335_cast_fp16)[name = tensor<string, []>("op_11984_cast_fp16")];
            tensor<int32, [1]> squeeze_388_axes_0 = const()[name = tensor<string, []>("squeeze_388_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_388_cast_fp16 = squeeze(axes = squeeze_388_axes_0, x = var_11984_cast_fp16_0)[name = tensor<string, []>("squeeze_388_cast_fp16")];
            tensor<int32, [1]> squeeze_389_axes_0 = const()[name = tensor<string, []>("squeeze_389_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_389_cast_fp16 = squeeze(axes = squeeze_389_axes_0, x = var_11984_cast_fp16_1)[name = tensor<string, []>("squeeze_389_cast_fp16")];
            tensor<fp16, []> const_1591_promoted_to_fp16 = const()[name = tensor<string, []>("const_1591_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_11987_cast_fp16 = mul(x = squeeze_389_cast_fp16, y = const_1591_promoted_to_fp16)[name = tensor<string, []>("op_11987_cast_fp16")];
            tensor<int32, []> x_1337_axis_0 = const()[name = tensor<string, []>("x_1337_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1337_cast_fp16 = stack(axis = x_1337_axis_0, values = (var_11987_cast_fp16, squeeze_388_cast_fp16))[name = tensor<string, []>("x_1337_cast_fp16")];
            tensor<int32, [4]> var_11993 = const()[name = tensor<string, []>("op_11993"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_11994_cast_fp16 = reshape(shape = var_11993, x = x_1337_cast_fp16)[name = tensor<string, []>("op_11994_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11995_cast_fp16 = mul(x = var_11994_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_11995_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_195_cast_fp16 = add(x = var_11974_cast_fp16, y = var_11995_cast_fp16)[name = tensor<string, []>("q_patches_195_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_11997_cast_fp16 = mul(x = var_11972_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_11997_cast_fp16")];
            tensor<int32, [5]> var_12005 = const()[name = tensor<string, []>("op_12005"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1341_cast_fp16 = reshape(shape = var_12005, x = var_11972_cast_fp16)[name = tensor<string, []>("x_1341_cast_fp16")];
            tensor<int32, [2]> var_12007_split_sizes_0 = const()[name = tensor<string, []>("op_12007_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12007_axis_0 = const()[name = tensor<string, []>("op_12007_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12007_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12007_cast_fp16_1 = split(axis = var_12007_axis_0, split_sizes = var_12007_split_sizes_0, x = x_1341_cast_fp16)[name = tensor<string, []>("op_12007_cast_fp16")];
            tensor<int32, [1]> squeeze_390_axes_0 = const()[name = tensor<string, []>("squeeze_390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_390_cast_fp16 = squeeze(axes = squeeze_390_axes_0, x = var_12007_cast_fp16_0)[name = tensor<string, []>("squeeze_390_cast_fp16")];
            tensor<int32, [1]> squeeze_391_axes_0 = const()[name = tensor<string, []>("squeeze_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_391_cast_fp16 = squeeze(axes = squeeze_391_axes_0, x = var_12007_cast_fp16_1)[name = tensor<string, []>("squeeze_391_cast_fp16")];
            tensor<fp16, []> const_1599_promoted_to_fp16 = const()[name = tensor<string, []>("const_1599_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12010_cast_fp16 = mul(x = squeeze_391_cast_fp16, y = const_1599_promoted_to_fp16)[name = tensor<string, []>("op_12010_cast_fp16")];
            tensor<int32, []> x_1343_axis_0 = const()[name = tensor<string, []>("x_1343_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1343_cast_fp16 = stack(axis = x_1343_axis_0, values = (var_12010_cast_fp16, squeeze_390_cast_fp16))[name = tensor<string, []>("x_1343_cast_fp16")];
            tensor<int32, [4]> var_12016 = const()[name = tensor<string, []>("op_12016"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12017_cast_fp16 = reshape(shape = var_12016, x = x_1343_cast_fp16)[name = tensor<string, []>("op_12017_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12018_cast_fp16 = mul(x = var_12017_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12018_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_195_cast_fp16 = add(x = var_11997_cast_fp16, y = var_12018_cast_fp16)[name = tensor<string, []>("k_patches_195_cast_fp16")];
            tensor<bool, []> var_12021_interleave_0 = const()[name = tensor<string, []>("op_12021_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12021_cast_fp16 = concat(axis = var_20, interleave = var_12021_interleave_0, values = (var_11960_cast_fp16, q_patches_195_cast_fp16))[name = tensor<string, []>("op_12021_cast_fp16")];
            tensor<bool, []> var_12024_interleave_0 = const()[name = tensor<string, []>("op_12024_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12024_cast_fp16 = concat(axis = var_20, interleave = var_12024_interleave_0, values = (var_11964_cast_fp16, k_patches_195_cast_fp16))[name = tensor<string, []>("op_12024_cast_fp16")];
            tensor<fp16, []> var_12026_to_fp16 = const()[name = tensor<string, []>("op_12026_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_979_cast_fp16 = mul(x = var_12021_cast_fp16, y = var_12026_to_fp16)[name = tensor<string, []>("q_979_cast_fp16")];
            tensor<bool, []> attn_389_transpose_x_1 = const()[name = tensor<string, []>("attn_389_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_389_transpose_y_1 = const()[name = tensor<string, []>("attn_389_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_389_cast_fp16 = matmul(transpose_x = attn_389_transpose_x_1, transpose_y = attn_389_transpose_y_1, x = q_979_cast_fp16, y = var_12024_cast_fp16)[name = tensor<string, []>("attn_389_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_391_cast_fp16 = softmax(axis = var_21, x = attn_389_cast_fp16)[name = tensor<string, []>("attn_391_cast_fp16")];
            tensor<bool, []> var_12031_transpose_x_0 = const()[name = tensor<string, []>("op_12031_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12031_transpose_y_0 = const()[name = tensor<string, []>("op_12031_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12031_cast_fp16 = matmul(transpose_x = var_12031_transpose_x_0, transpose_y = var_12031_transpose_y_0, x = attn_391_cast_fp16, y = v_587_cast_fp16)[name = tensor<string, []>("op_12031_cast_fp16")];
            tensor<int32, [1]> var_12032_axes_0 = const()[name = tensor<string, []>("op_12032_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12032_cast_fp16 = squeeze(axes = var_12032_axes_0, x = var_12031_cast_fp16)[name = tensor<string, []>("op_12032_cast_fp16")];
            tensor<string, []> var_12038_pad_type_0 = const()[name = tensor<string, []>("op_12038_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12038_strides_0 = const()[name = tensor<string, []>("op_12038_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12038_pad_0 = const()[name = tensor<string, []>("op_12038_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12038_dilations_0 = const()[name = tensor<string, []>("op_12038_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12038_groups_0 = const()[name = tensor<string, []>("op_12038_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115618880)))];
            tensor<fp16, [1, 64, 1, 197]> var_12038_cast_fp16 = conv(dilations = var_12038_dilations_0, groups = var_12038_groups_0, pad = var_12038_pad_0, pad_type = var_12038_pad_type_0, strides = var_12038_strides_0, weight = model_blocks_8_attn_q_projs_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12038_cast_fp16")];
            tensor<int32, [1]> var_12039_axes_0 = const()[name = tensor<string, []>("op_12039_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12039_cast_fp16 = squeeze(axes = var_12039_axes_0, x = var_12038_cast_fp16)[name = tensor<string, []>("op_12039_cast_fp16")];
            tensor<int32, [3]> q_981_perm_0 = const()[name = tensor<string, []>("q_981_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12046_pad_type_0 = const()[name = tensor<string, []>("op_12046_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12046_strides_0 = const()[name = tensor<string, []>("op_12046_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12046_pad_0 = const()[name = tensor<string, []>("op_12046_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12046_dilations_0 = const()[name = tensor<string, []>("op_12046_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12046_groups_0 = const()[name = tensor<string, []>("op_12046_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115717248)))];
            tensor<fp16, [1, 64, 1, 197]> var_12046_cast_fp16 = conv(dilations = var_12046_dilations_0, groups = var_12046_groups_0, pad = var_12046_pad_0, pad_type = var_12046_pad_type_0, strides = var_12046_strides_0, weight = model_blocks_8_attn_k_projs_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12046_cast_fp16")];
            tensor<int32, [1]> var_12047_axes_0 = const()[name = tensor<string, []>("op_12047_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12047_cast_fp16 = squeeze(axes = var_12047_axes_0, x = var_12046_cast_fp16)[name = tensor<string, []>("op_12047_cast_fp16")];
            tensor<int32, [3]> k_589_perm_0 = const()[name = tensor<string, []>("k_589_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12054_pad_type_0 = const()[name = tensor<string, []>("op_12054_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12054_strides_0 = const()[name = tensor<string, []>("op_12054_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12054_pad_0 = const()[name = tensor<string, []>("op_12054_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12054_dilations_0 = const()[name = tensor<string, []>("op_12054_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12054_groups_0 = const()[name = tensor<string, []>("op_12054_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115815616)))];
            tensor<fp16, [1, 64, 1, 197]> var_12054_cast_fp16 = conv(dilations = var_12054_dilations_0, groups = var_12054_groups_0, pad = var_12054_pad_0, pad_type = var_12054_pad_type_0, strides = var_12054_strides_0, weight = model_blocks_8_attn_v_projs_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12054_cast_fp16")];
            tensor<int32, [1]> var_12055_axes_0 = const()[name = tensor<string, []>("op_12055_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12055_cast_fp16 = squeeze(axes = var_12055_axes_0, x = var_12054_cast_fp16)[name = tensor<string, []>("op_12055_cast_fp16")];
            tensor<int32, [3]> v_589_perm_0 = const()[name = tensor<string, []>("v_589_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115913984)))];
            tensor<fp16, [1, 197, 64]> q_981_cast_fp16 = transpose(perm = q_981_perm_0, x = var_12039_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [1, 197, 64]> q_983_cast_fp16 = add(x = q_981_cast_fp16, y = model_blocks_8_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_983_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115914176)))];
            tensor<fp16, [1, 197, 64]> v_589_cast_fp16 = transpose(perm = v_589_perm_0, x = var_12055_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [1, 197, 64]> v_591_cast_fp16 = add(x = v_589_cast_fp16, y = model_blocks_8_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_591_cast_fp16")];
            tensor<int32, [1]> q_985_axes_0 = const()[name = tensor<string, []>("q_985_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_985_cast_fp16 = expand_dims(axes = q_985_axes_0, x = q_983_cast_fp16)[name = tensor<string, []>("q_985_cast_fp16")];
            tensor<int32, [1]> k_591_axes_0 = const()[name = tensor<string, []>("k_591_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_589_cast_fp16 = transpose(perm = k_589_perm_0, x = var_12047_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [1, 1, 197, 64]> k_591_cast_fp16 = expand_dims(axes = k_591_axes_0, x = k_589_cast_fp16)[name = tensor<string, []>("k_591_cast_fp16")];
            tensor<int32, [1]> v_593_axes_0 = const()[name = tensor<string, []>("v_593_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_593_cast_fp16 = expand_dims(axes = v_593_axes_0, x = v_591_cast_fp16)[name = tensor<string, []>("v_593_cast_fp16")];
            tensor<int32, [4]> var_12064_begin_0 = const()[name = tensor<string, []>("op_12064_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12064_end_0 = const()[name = tensor<string, []>("op_12064_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12064_end_mask_0 = const()[name = tensor<string, []>("op_12064_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12064_cast_fp16 = slice_by_index(begin = var_12064_begin_0, end = var_12064_end_0, end_mask = var_12064_end_mask_0, x = q_985_cast_fp16)[name = tensor<string, []>("op_12064_cast_fp16")];
            tensor<int32, [4]> var_12068_begin_0 = const()[name = tensor<string, []>("op_12068_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12068_end_0 = const()[name = tensor<string, []>("op_12068_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12068_end_mask_0 = const()[name = tensor<string, []>("op_12068_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12068_cast_fp16 = slice_by_index(begin = var_12068_begin_0, end = var_12068_end_0, end_mask = var_12068_end_mask_0, x = k_591_cast_fp16)[name = tensor<string, []>("op_12068_cast_fp16")];
            tensor<int32, [4]> var_12072_begin_0 = const()[name = tensor<string, []>("op_12072_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12072_end_0 = const()[name = tensor<string, []>("op_12072_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12072_end_mask_0 = const()[name = tensor<string, []>("op_12072_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12072_cast_fp16 = slice_by_index(begin = var_12072_begin_0, end = var_12072_end_0, end_mask = var_12072_end_mask_0, x = q_985_cast_fp16)[name = tensor<string, []>("op_12072_cast_fp16")];
            tensor<int32, [4]> var_12076_begin_0 = const()[name = tensor<string, []>("op_12076_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12076_end_0 = const()[name = tensor<string, []>("op_12076_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12076_end_mask_0 = const()[name = tensor<string, []>("op_12076_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12076_cast_fp16 = slice_by_index(begin = var_12076_begin_0, end = var_12076_end_0, end_mask = var_12076_end_mask_0, x = k_591_cast_fp16)[name = tensor<string, []>("op_12076_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12078_cast_fp16 = mul(x = var_12072_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12078_cast_fp16")];
            tensor<int32, [5]> var_12086 = const()[name = tensor<string, []>("op_12086"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1347_cast_fp16 = reshape(shape = var_12086, x = var_12072_cast_fp16)[name = tensor<string, []>("x_1347_cast_fp16")];
            tensor<int32, [2]> var_12088_split_sizes_0 = const()[name = tensor<string, []>("op_12088_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12088_axis_0 = const()[name = tensor<string, []>("op_12088_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12088_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12088_cast_fp16_1 = split(axis = var_12088_axis_0, split_sizes = var_12088_split_sizes_0, x = x_1347_cast_fp16)[name = tensor<string, []>("op_12088_cast_fp16")];
            tensor<int32, [1]> squeeze_392_axes_0 = const()[name = tensor<string, []>("squeeze_392_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_392_cast_fp16 = squeeze(axes = squeeze_392_axes_0, x = var_12088_cast_fp16_0)[name = tensor<string, []>("squeeze_392_cast_fp16")];
            tensor<int32, [1]> squeeze_393_axes_0 = const()[name = tensor<string, []>("squeeze_393_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_393_cast_fp16 = squeeze(axes = squeeze_393_axes_0, x = var_12088_cast_fp16_1)[name = tensor<string, []>("squeeze_393_cast_fp16")];
            tensor<fp16, []> const_1607_promoted_to_fp16 = const()[name = tensor<string, []>("const_1607_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12091_cast_fp16 = mul(x = squeeze_393_cast_fp16, y = const_1607_promoted_to_fp16)[name = tensor<string, []>("op_12091_cast_fp16")];
            tensor<int32, []> x_1349_axis_0 = const()[name = tensor<string, []>("x_1349_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1349_cast_fp16 = stack(axis = x_1349_axis_0, values = (var_12091_cast_fp16, squeeze_392_cast_fp16))[name = tensor<string, []>("x_1349_cast_fp16")];
            tensor<int32, [4]> var_12097 = const()[name = tensor<string, []>("op_12097"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12098_cast_fp16 = reshape(shape = var_12097, x = x_1349_cast_fp16)[name = tensor<string, []>("op_12098_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12099_cast_fp16 = mul(x = var_12098_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12099_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_197_cast_fp16 = add(x = var_12078_cast_fp16, y = var_12099_cast_fp16)[name = tensor<string, []>("q_patches_197_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12101_cast_fp16 = mul(x = var_12076_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12101_cast_fp16")];
            tensor<int32, [5]> var_12109 = const()[name = tensor<string, []>("op_12109"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1353_cast_fp16 = reshape(shape = var_12109, x = var_12076_cast_fp16)[name = tensor<string, []>("x_1353_cast_fp16")];
            tensor<int32, [2]> var_12111_split_sizes_0 = const()[name = tensor<string, []>("op_12111_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12111_axis_0 = const()[name = tensor<string, []>("op_12111_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12111_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12111_cast_fp16_1 = split(axis = var_12111_axis_0, split_sizes = var_12111_split_sizes_0, x = x_1353_cast_fp16)[name = tensor<string, []>("op_12111_cast_fp16")];
            tensor<int32, [1]> squeeze_394_axes_0 = const()[name = tensor<string, []>("squeeze_394_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_394_cast_fp16 = squeeze(axes = squeeze_394_axes_0, x = var_12111_cast_fp16_0)[name = tensor<string, []>("squeeze_394_cast_fp16")];
            tensor<int32, [1]> squeeze_395_axes_0 = const()[name = tensor<string, []>("squeeze_395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_395_cast_fp16 = squeeze(axes = squeeze_395_axes_0, x = var_12111_cast_fp16_1)[name = tensor<string, []>("squeeze_395_cast_fp16")];
            tensor<fp16, []> const_1615_promoted_to_fp16 = const()[name = tensor<string, []>("const_1615_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12114_cast_fp16 = mul(x = squeeze_395_cast_fp16, y = const_1615_promoted_to_fp16)[name = tensor<string, []>("op_12114_cast_fp16")];
            tensor<int32, []> x_1355_axis_0 = const()[name = tensor<string, []>("x_1355_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1355_cast_fp16 = stack(axis = x_1355_axis_0, values = (var_12114_cast_fp16, squeeze_394_cast_fp16))[name = tensor<string, []>("x_1355_cast_fp16")];
            tensor<int32, [4]> var_12120 = const()[name = tensor<string, []>("op_12120"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12121_cast_fp16 = reshape(shape = var_12120, x = x_1355_cast_fp16)[name = tensor<string, []>("op_12121_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12122_cast_fp16 = mul(x = var_12121_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12122_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_197_cast_fp16 = add(x = var_12101_cast_fp16, y = var_12122_cast_fp16)[name = tensor<string, []>("k_patches_197_cast_fp16")];
            tensor<bool, []> var_12125_interleave_0 = const()[name = tensor<string, []>("op_12125_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12125_cast_fp16 = concat(axis = var_20, interleave = var_12125_interleave_0, values = (var_12064_cast_fp16, q_patches_197_cast_fp16))[name = tensor<string, []>("op_12125_cast_fp16")];
            tensor<bool, []> var_12128_interleave_0 = const()[name = tensor<string, []>("op_12128_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12128_cast_fp16 = concat(axis = var_20, interleave = var_12128_interleave_0, values = (var_12068_cast_fp16, k_patches_197_cast_fp16))[name = tensor<string, []>("op_12128_cast_fp16")];
            tensor<fp16, []> var_12130_to_fp16 = const()[name = tensor<string, []>("op_12130_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_989_cast_fp16 = mul(x = var_12125_cast_fp16, y = var_12130_to_fp16)[name = tensor<string, []>("q_989_cast_fp16")];
            tensor<bool, []> attn_393_transpose_x_1 = const()[name = tensor<string, []>("attn_393_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_393_transpose_y_1 = const()[name = tensor<string, []>("attn_393_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_393_cast_fp16 = matmul(transpose_x = attn_393_transpose_x_1, transpose_y = attn_393_transpose_y_1, x = q_989_cast_fp16, y = var_12128_cast_fp16)[name = tensor<string, []>("attn_393_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_395_cast_fp16 = softmax(axis = var_21, x = attn_393_cast_fp16)[name = tensor<string, []>("attn_395_cast_fp16")];
            tensor<bool, []> var_12135_transpose_x_0 = const()[name = tensor<string, []>("op_12135_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12135_transpose_y_0 = const()[name = tensor<string, []>("op_12135_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12135_cast_fp16 = matmul(transpose_x = var_12135_transpose_x_0, transpose_y = var_12135_transpose_y_0, x = attn_395_cast_fp16, y = v_593_cast_fp16)[name = tensor<string, []>("op_12135_cast_fp16")];
            tensor<int32, [1]> var_12136_axes_0 = const()[name = tensor<string, []>("op_12136_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12136_cast_fp16 = squeeze(axes = var_12136_axes_0, x = var_12135_cast_fp16)[name = tensor<string, []>("op_12136_cast_fp16")];
            tensor<string, []> var_12142_pad_type_0 = const()[name = tensor<string, []>("op_12142_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12142_strides_0 = const()[name = tensor<string, []>("op_12142_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12142_pad_0 = const()[name = tensor<string, []>("op_12142_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12142_dilations_0 = const()[name = tensor<string, []>("op_12142_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12142_groups_0 = const()[name = tensor<string, []>("op_12142_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115914368)))];
            tensor<fp16, [1, 64, 1, 197]> var_12142_cast_fp16 = conv(dilations = var_12142_dilations_0, groups = var_12142_groups_0, pad = var_12142_pad_0, pad_type = var_12142_pad_type_0, strides = var_12142_strides_0, weight = model_blocks_8_attn_q_projs_3_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12142_cast_fp16")];
            tensor<int32, [1]> var_12143_axes_0 = const()[name = tensor<string, []>("op_12143_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12143_cast_fp16 = squeeze(axes = var_12143_axes_0, x = var_12142_cast_fp16)[name = tensor<string, []>("op_12143_cast_fp16")];
            tensor<int32, [3]> q_991_perm_0 = const()[name = tensor<string, []>("q_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12150_pad_type_0 = const()[name = tensor<string, []>("op_12150_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12150_strides_0 = const()[name = tensor<string, []>("op_12150_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12150_pad_0 = const()[name = tensor<string, []>("op_12150_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12150_dilations_0 = const()[name = tensor<string, []>("op_12150_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12150_groups_0 = const()[name = tensor<string, []>("op_12150_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116012736)))];
            tensor<fp16, [1, 64, 1, 197]> var_12150_cast_fp16 = conv(dilations = var_12150_dilations_0, groups = var_12150_groups_0, pad = var_12150_pad_0, pad_type = var_12150_pad_type_0, strides = var_12150_strides_0, weight = model_blocks_8_attn_k_projs_3_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12150_cast_fp16")];
            tensor<int32, [1]> var_12151_axes_0 = const()[name = tensor<string, []>("op_12151_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12151_cast_fp16 = squeeze(axes = var_12151_axes_0, x = var_12150_cast_fp16)[name = tensor<string, []>("op_12151_cast_fp16")];
            tensor<int32, [3]> k_595_perm_0 = const()[name = tensor<string, []>("k_595_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12158_pad_type_0 = const()[name = tensor<string, []>("op_12158_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12158_strides_0 = const()[name = tensor<string, []>("op_12158_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12158_pad_0 = const()[name = tensor<string, []>("op_12158_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12158_dilations_0 = const()[name = tensor<string, []>("op_12158_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12158_groups_0 = const()[name = tensor<string, []>("op_12158_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116111104)))];
            tensor<fp16, [1, 64, 1, 197]> var_12158_cast_fp16 = conv(dilations = var_12158_dilations_0, groups = var_12158_groups_0, pad = var_12158_pad_0, pad_type = var_12158_pad_type_0, strides = var_12158_strides_0, weight = model_blocks_8_attn_v_projs_3_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12158_cast_fp16")];
            tensor<int32, [1]> var_12159_axes_0 = const()[name = tensor<string, []>("op_12159_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12159_cast_fp16 = squeeze(axes = var_12159_axes_0, x = var_12158_cast_fp16)[name = tensor<string, []>("op_12159_cast_fp16")];
            tensor<int32, [3]> v_595_perm_0 = const()[name = tensor<string, []>("v_595_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116209472)))];
            tensor<fp16, [1, 197, 64]> q_991_cast_fp16 = transpose(perm = q_991_perm_0, x = var_12143_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [1, 197, 64]> q_993_cast_fp16 = add(x = q_991_cast_fp16, y = model_blocks_8_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_993_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116209664)))];
            tensor<fp16, [1, 197, 64]> v_595_cast_fp16 = transpose(perm = v_595_perm_0, x = var_12159_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 197, 64]> v_597_cast_fp16 = add(x = v_595_cast_fp16, y = model_blocks_8_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_597_cast_fp16")];
            tensor<int32, [1]> q_995_axes_0 = const()[name = tensor<string, []>("q_995_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_995_cast_fp16 = expand_dims(axes = q_995_axes_0, x = q_993_cast_fp16)[name = tensor<string, []>("q_995_cast_fp16")];
            tensor<int32, [1]> k_597_axes_0 = const()[name = tensor<string, []>("k_597_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_595_cast_fp16 = transpose(perm = k_595_perm_0, x = var_12151_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 1, 197, 64]> k_597_cast_fp16 = expand_dims(axes = k_597_axes_0, x = k_595_cast_fp16)[name = tensor<string, []>("k_597_cast_fp16")];
            tensor<int32, [1]> v_599_axes_0 = const()[name = tensor<string, []>("v_599_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_599_cast_fp16 = expand_dims(axes = v_599_axes_0, x = v_597_cast_fp16)[name = tensor<string, []>("v_599_cast_fp16")];
            tensor<int32, [4]> var_12168_begin_0 = const()[name = tensor<string, []>("op_12168_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12168_end_0 = const()[name = tensor<string, []>("op_12168_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12168_end_mask_0 = const()[name = tensor<string, []>("op_12168_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12168_cast_fp16 = slice_by_index(begin = var_12168_begin_0, end = var_12168_end_0, end_mask = var_12168_end_mask_0, x = q_995_cast_fp16)[name = tensor<string, []>("op_12168_cast_fp16")];
            tensor<int32, [4]> var_12172_begin_0 = const()[name = tensor<string, []>("op_12172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12172_end_0 = const()[name = tensor<string, []>("op_12172_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12172_end_mask_0 = const()[name = tensor<string, []>("op_12172_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12172_cast_fp16 = slice_by_index(begin = var_12172_begin_0, end = var_12172_end_0, end_mask = var_12172_end_mask_0, x = k_597_cast_fp16)[name = tensor<string, []>("op_12172_cast_fp16")];
            tensor<int32, [4]> var_12176_begin_0 = const()[name = tensor<string, []>("op_12176_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12176_end_0 = const()[name = tensor<string, []>("op_12176_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12176_end_mask_0 = const()[name = tensor<string, []>("op_12176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12176_cast_fp16 = slice_by_index(begin = var_12176_begin_0, end = var_12176_end_0, end_mask = var_12176_end_mask_0, x = q_995_cast_fp16)[name = tensor<string, []>("op_12176_cast_fp16")];
            tensor<int32, [4]> var_12180_begin_0 = const()[name = tensor<string, []>("op_12180_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12180_end_0 = const()[name = tensor<string, []>("op_12180_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12180_end_mask_0 = const()[name = tensor<string, []>("op_12180_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12180_cast_fp16 = slice_by_index(begin = var_12180_begin_0, end = var_12180_end_0, end_mask = var_12180_end_mask_0, x = k_597_cast_fp16)[name = tensor<string, []>("op_12180_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12182_cast_fp16 = mul(x = var_12176_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12182_cast_fp16")];
            tensor<int32, [5]> var_12190 = const()[name = tensor<string, []>("op_12190"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1359_cast_fp16 = reshape(shape = var_12190, x = var_12176_cast_fp16)[name = tensor<string, []>("x_1359_cast_fp16")];
            tensor<int32, [2]> var_12192_split_sizes_0 = const()[name = tensor<string, []>("op_12192_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12192_axis_0 = const()[name = tensor<string, []>("op_12192_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12192_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12192_cast_fp16_1 = split(axis = var_12192_axis_0, split_sizes = var_12192_split_sizes_0, x = x_1359_cast_fp16)[name = tensor<string, []>("op_12192_cast_fp16")];
            tensor<int32, [1]> squeeze_396_axes_0 = const()[name = tensor<string, []>("squeeze_396_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_396_cast_fp16 = squeeze(axes = squeeze_396_axes_0, x = var_12192_cast_fp16_0)[name = tensor<string, []>("squeeze_396_cast_fp16")];
            tensor<int32, [1]> squeeze_397_axes_0 = const()[name = tensor<string, []>("squeeze_397_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_397_cast_fp16 = squeeze(axes = squeeze_397_axes_0, x = var_12192_cast_fp16_1)[name = tensor<string, []>("squeeze_397_cast_fp16")];
            tensor<fp16, []> const_1623_promoted_to_fp16 = const()[name = tensor<string, []>("const_1623_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12195_cast_fp16 = mul(x = squeeze_397_cast_fp16, y = const_1623_promoted_to_fp16)[name = tensor<string, []>("op_12195_cast_fp16")];
            tensor<int32, []> x_1361_axis_0 = const()[name = tensor<string, []>("x_1361_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1361_cast_fp16 = stack(axis = x_1361_axis_0, values = (var_12195_cast_fp16, squeeze_396_cast_fp16))[name = tensor<string, []>("x_1361_cast_fp16")];
            tensor<int32, [4]> var_12201 = const()[name = tensor<string, []>("op_12201"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12202_cast_fp16 = reshape(shape = var_12201, x = x_1361_cast_fp16)[name = tensor<string, []>("op_12202_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12203_cast_fp16 = mul(x = var_12202_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12203_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_199_cast_fp16 = add(x = var_12182_cast_fp16, y = var_12203_cast_fp16)[name = tensor<string, []>("q_patches_199_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12205_cast_fp16 = mul(x = var_12180_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12205_cast_fp16")];
            tensor<int32, [5]> var_12213 = const()[name = tensor<string, []>("op_12213"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1365_cast_fp16 = reshape(shape = var_12213, x = var_12180_cast_fp16)[name = tensor<string, []>("x_1365_cast_fp16")];
            tensor<int32, [2]> var_12215_split_sizes_0 = const()[name = tensor<string, []>("op_12215_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12215_axis_0 = const()[name = tensor<string, []>("op_12215_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12215_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12215_cast_fp16_1 = split(axis = var_12215_axis_0, split_sizes = var_12215_split_sizes_0, x = x_1365_cast_fp16)[name = tensor<string, []>("op_12215_cast_fp16")];
            tensor<int32, [1]> squeeze_398_axes_0 = const()[name = tensor<string, []>("squeeze_398_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_398_cast_fp16 = squeeze(axes = squeeze_398_axes_0, x = var_12215_cast_fp16_0)[name = tensor<string, []>("squeeze_398_cast_fp16")];
            tensor<int32, [1]> squeeze_399_axes_0 = const()[name = tensor<string, []>("squeeze_399_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_399_cast_fp16 = squeeze(axes = squeeze_399_axes_0, x = var_12215_cast_fp16_1)[name = tensor<string, []>("squeeze_399_cast_fp16")];
            tensor<fp16, []> const_1631_promoted_to_fp16 = const()[name = tensor<string, []>("const_1631_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12218_cast_fp16 = mul(x = squeeze_399_cast_fp16, y = const_1631_promoted_to_fp16)[name = tensor<string, []>("op_12218_cast_fp16")];
            tensor<int32, []> x_1367_axis_0 = const()[name = tensor<string, []>("x_1367_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1367_cast_fp16 = stack(axis = x_1367_axis_0, values = (var_12218_cast_fp16, squeeze_398_cast_fp16))[name = tensor<string, []>("x_1367_cast_fp16")];
            tensor<int32, [4]> var_12224 = const()[name = tensor<string, []>("op_12224"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12225_cast_fp16 = reshape(shape = var_12224, x = x_1367_cast_fp16)[name = tensor<string, []>("op_12225_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12226_cast_fp16 = mul(x = var_12225_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12226_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_199_cast_fp16 = add(x = var_12205_cast_fp16, y = var_12226_cast_fp16)[name = tensor<string, []>("k_patches_199_cast_fp16")];
            tensor<bool, []> var_12229_interleave_0 = const()[name = tensor<string, []>("op_12229_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12229_cast_fp16 = concat(axis = var_20, interleave = var_12229_interleave_0, values = (var_12168_cast_fp16, q_patches_199_cast_fp16))[name = tensor<string, []>("op_12229_cast_fp16")];
            tensor<bool, []> var_12232_interleave_0 = const()[name = tensor<string, []>("op_12232_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12232_cast_fp16 = concat(axis = var_20, interleave = var_12232_interleave_0, values = (var_12172_cast_fp16, k_patches_199_cast_fp16))[name = tensor<string, []>("op_12232_cast_fp16")];
            tensor<fp16, []> var_12234_to_fp16 = const()[name = tensor<string, []>("op_12234_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_999_cast_fp16 = mul(x = var_12229_cast_fp16, y = var_12234_to_fp16)[name = tensor<string, []>("q_999_cast_fp16")];
            tensor<bool, []> attn_397_transpose_x_1 = const()[name = tensor<string, []>("attn_397_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_397_transpose_y_1 = const()[name = tensor<string, []>("attn_397_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_397_cast_fp16 = matmul(transpose_x = attn_397_transpose_x_1, transpose_y = attn_397_transpose_y_1, x = q_999_cast_fp16, y = var_12232_cast_fp16)[name = tensor<string, []>("attn_397_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_399_cast_fp16 = softmax(axis = var_21, x = attn_397_cast_fp16)[name = tensor<string, []>("attn_399_cast_fp16")];
            tensor<bool, []> var_12239_transpose_x_0 = const()[name = tensor<string, []>("op_12239_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12239_transpose_y_0 = const()[name = tensor<string, []>("op_12239_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12239_cast_fp16 = matmul(transpose_x = var_12239_transpose_x_0, transpose_y = var_12239_transpose_y_0, x = attn_399_cast_fp16, y = v_599_cast_fp16)[name = tensor<string, []>("op_12239_cast_fp16")];
            tensor<int32, [1]> var_12240_axes_0 = const()[name = tensor<string, []>("op_12240_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12240_cast_fp16 = squeeze(axes = var_12240_axes_0, x = var_12239_cast_fp16)[name = tensor<string, []>("op_12240_cast_fp16")];
            tensor<string, []> var_12246_pad_type_0 = const()[name = tensor<string, []>("op_12246_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12246_strides_0 = const()[name = tensor<string, []>("op_12246_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12246_pad_0 = const()[name = tensor<string, []>("op_12246_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12246_dilations_0 = const()[name = tensor<string, []>("op_12246_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12246_groups_0 = const()[name = tensor<string, []>("op_12246_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116209856)))];
            tensor<fp16, [1, 64, 1, 197]> var_12246_cast_fp16 = conv(dilations = var_12246_dilations_0, groups = var_12246_groups_0, pad = var_12246_pad_0, pad_type = var_12246_pad_type_0, strides = var_12246_strides_0, weight = model_blocks_8_attn_q_projs_4_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12246_cast_fp16")];
            tensor<int32, [1]> var_12247_axes_0 = const()[name = tensor<string, []>("op_12247_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12247_cast_fp16 = squeeze(axes = var_12247_axes_0, x = var_12246_cast_fp16)[name = tensor<string, []>("op_12247_cast_fp16")];
            tensor<int32, [3]> q_1001_perm_0 = const()[name = tensor<string, []>("q_1001_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12254_pad_type_0 = const()[name = tensor<string, []>("op_12254_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12254_strides_0 = const()[name = tensor<string, []>("op_12254_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12254_pad_0 = const()[name = tensor<string, []>("op_12254_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12254_dilations_0 = const()[name = tensor<string, []>("op_12254_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12254_groups_0 = const()[name = tensor<string, []>("op_12254_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116308224)))];
            tensor<fp16, [1, 64, 1, 197]> var_12254_cast_fp16 = conv(dilations = var_12254_dilations_0, groups = var_12254_groups_0, pad = var_12254_pad_0, pad_type = var_12254_pad_type_0, strides = var_12254_strides_0, weight = model_blocks_8_attn_k_projs_4_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12254_cast_fp16")];
            tensor<int32, [1]> var_12255_axes_0 = const()[name = tensor<string, []>("op_12255_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12255_cast_fp16 = squeeze(axes = var_12255_axes_0, x = var_12254_cast_fp16)[name = tensor<string, []>("op_12255_cast_fp16")];
            tensor<int32, [3]> k_601_perm_0 = const()[name = tensor<string, []>("k_601_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12262_pad_type_0 = const()[name = tensor<string, []>("op_12262_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12262_strides_0 = const()[name = tensor<string, []>("op_12262_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12262_pad_0 = const()[name = tensor<string, []>("op_12262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12262_dilations_0 = const()[name = tensor<string, []>("op_12262_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12262_groups_0 = const()[name = tensor<string, []>("op_12262_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116406592)))];
            tensor<fp16, [1, 64, 1, 197]> var_12262_cast_fp16 = conv(dilations = var_12262_dilations_0, groups = var_12262_groups_0, pad = var_12262_pad_0, pad_type = var_12262_pad_type_0, strides = var_12262_strides_0, weight = model_blocks_8_attn_v_projs_4_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12262_cast_fp16")];
            tensor<int32, [1]> var_12263_axes_0 = const()[name = tensor<string, []>("op_12263_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12263_cast_fp16 = squeeze(axes = var_12263_axes_0, x = var_12262_cast_fp16)[name = tensor<string, []>("op_12263_cast_fp16")];
            tensor<int32, [3]> v_601_perm_0 = const()[name = tensor<string, []>("v_601_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116504960)))];
            tensor<fp16, [1, 197, 64]> q_1001_cast_fp16 = transpose(perm = q_1001_perm_0, x = var_12247_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 197, 64]> q_1003_cast_fp16 = add(x = q_1001_cast_fp16, y = model_blocks_8_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_1003_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116505152)))];
            tensor<fp16, [1, 197, 64]> v_601_cast_fp16 = transpose(perm = v_601_perm_0, x = var_12263_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [1, 197, 64]> v_603_cast_fp16 = add(x = v_601_cast_fp16, y = model_blocks_8_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_603_cast_fp16")];
            tensor<int32, [1]> q_1005_axes_0 = const()[name = tensor<string, []>("q_1005_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1005_cast_fp16 = expand_dims(axes = q_1005_axes_0, x = q_1003_cast_fp16)[name = tensor<string, []>("q_1005_cast_fp16")];
            tensor<int32, [1]> k_603_axes_0 = const()[name = tensor<string, []>("k_603_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_601_cast_fp16 = transpose(perm = k_601_perm_0, x = var_12255_cast_fp16)[name = tensor<string, []>("transpose_157")];
            tensor<fp16, [1, 1, 197, 64]> k_603_cast_fp16 = expand_dims(axes = k_603_axes_0, x = k_601_cast_fp16)[name = tensor<string, []>("k_603_cast_fp16")];
            tensor<int32, [1]> v_605_axes_0 = const()[name = tensor<string, []>("v_605_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_605_cast_fp16 = expand_dims(axes = v_605_axes_0, x = v_603_cast_fp16)[name = tensor<string, []>("v_605_cast_fp16")];
            tensor<int32, [4]> var_12272_begin_0 = const()[name = tensor<string, []>("op_12272_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12272_end_0 = const()[name = tensor<string, []>("op_12272_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12272_end_mask_0 = const()[name = tensor<string, []>("op_12272_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12272_cast_fp16 = slice_by_index(begin = var_12272_begin_0, end = var_12272_end_0, end_mask = var_12272_end_mask_0, x = q_1005_cast_fp16)[name = tensor<string, []>("op_12272_cast_fp16")];
            tensor<int32, [4]> var_12276_begin_0 = const()[name = tensor<string, []>("op_12276_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12276_end_0 = const()[name = tensor<string, []>("op_12276_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12276_end_mask_0 = const()[name = tensor<string, []>("op_12276_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12276_cast_fp16 = slice_by_index(begin = var_12276_begin_0, end = var_12276_end_0, end_mask = var_12276_end_mask_0, x = k_603_cast_fp16)[name = tensor<string, []>("op_12276_cast_fp16")];
            tensor<int32, [4]> var_12280_begin_0 = const()[name = tensor<string, []>("op_12280_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12280_end_0 = const()[name = tensor<string, []>("op_12280_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12280_end_mask_0 = const()[name = tensor<string, []>("op_12280_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12280_cast_fp16 = slice_by_index(begin = var_12280_begin_0, end = var_12280_end_0, end_mask = var_12280_end_mask_0, x = q_1005_cast_fp16)[name = tensor<string, []>("op_12280_cast_fp16")];
            tensor<int32, [4]> var_12284_begin_0 = const()[name = tensor<string, []>("op_12284_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12284_end_0 = const()[name = tensor<string, []>("op_12284_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12284_end_mask_0 = const()[name = tensor<string, []>("op_12284_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12284_cast_fp16 = slice_by_index(begin = var_12284_begin_0, end = var_12284_end_0, end_mask = var_12284_end_mask_0, x = k_603_cast_fp16)[name = tensor<string, []>("op_12284_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12286_cast_fp16 = mul(x = var_12280_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12286_cast_fp16")];
            tensor<int32, [5]> var_12294 = const()[name = tensor<string, []>("op_12294"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1371_cast_fp16 = reshape(shape = var_12294, x = var_12280_cast_fp16)[name = tensor<string, []>("x_1371_cast_fp16")];
            tensor<int32, [2]> var_12296_split_sizes_0 = const()[name = tensor<string, []>("op_12296_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12296_axis_0 = const()[name = tensor<string, []>("op_12296_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12296_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12296_cast_fp16_1 = split(axis = var_12296_axis_0, split_sizes = var_12296_split_sizes_0, x = x_1371_cast_fp16)[name = tensor<string, []>("op_12296_cast_fp16")];
            tensor<int32, [1]> squeeze_400_axes_0 = const()[name = tensor<string, []>("squeeze_400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_400_cast_fp16 = squeeze(axes = squeeze_400_axes_0, x = var_12296_cast_fp16_0)[name = tensor<string, []>("squeeze_400_cast_fp16")];
            tensor<int32, [1]> squeeze_401_axes_0 = const()[name = tensor<string, []>("squeeze_401_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_401_cast_fp16 = squeeze(axes = squeeze_401_axes_0, x = var_12296_cast_fp16_1)[name = tensor<string, []>("squeeze_401_cast_fp16")];
            tensor<fp16, []> const_1639_promoted_to_fp16 = const()[name = tensor<string, []>("const_1639_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12299_cast_fp16 = mul(x = squeeze_401_cast_fp16, y = const_1639_promoted_to_fp16)[name = tensor<string, []>("op_12299_cast_fp16")];
            tensor<int32, []> x_1373_axis_0 = const()[name = tensor<string, []>("x_1373_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1373_cast_fp16 = stack(axis = x_1373_axis_0, values = (var_12299_cast_fp16, squeeze_400_cast_fp16))[name = tensor<string, []>("x_1373_cast_fp16")];
            tensor<int32, [4]> var_12305 = const()[name = tensor<string, []>("op_12305"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12306_cast_fp16 = reshape(shape = var_12305, x = x_1373_cast_fp16)[name = tensor<string, []>("op_12306_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12307_cast_fp16 = mul(x = var_12306_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12307_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_201_cast_fp16 = add(x = var_12286_cast_fp16, y = var_12307_cast_fp16)[name = tensor<string, []>("q_patches_201_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12309_cast_fp16 = mul(x = var_12284_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12309_cast_fp16")];
            tensor<int32, [5]> var_12317 = const()[name = tensor<string, []>("op_12317"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1377_cast_fp16 = reshape(shape = var_12317, x = var_12284_cast_fp16)[name = tensor<string, []>("x_1377_cast_fp16")];
            tensor<int32, [2]> var_12319_split_sizes_0 = const()[name = tensor<string, []>("op_12319_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12319_axis_0 = const()[name = tensor<string, []>("op_12319_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12319_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12319_cast_fp16_1 = split(axis = var_12319_axis_0, split_sizes = var_12319_split_sizes_0, x = x_1377_cast_fp16)[name = tensor<string, []>("op_12319_cast_fp16")];
            tensor<int32, [1]> squeeze_402_axes_0 = const()[name = tensor<string, []>("squeeze_402_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_402_cast_fp16 = squeeze(axes = squeeze_402_axes_0, x = var_12319_cast_fp16_0)[name = tensor<string, []>("squeeze_402_cast_fp16")];
            tensor<int32, [1]> squeeze_403_axes_0 = const()[name = tensor<string, []>("squeeze_403_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_403_cast_fp16 = squeeze(axes = squeeze_403_axes_0, x = var_12319_cast_fp16_1)[name = tensor<string, []>("squeeze_403_cast_fp16")];
            tensor<fp16, []> const_1647_promoted_to_fp16 = const()[name = tensor<string, []>("const_1647_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12322_cast_fp16 = mul(x = squeeze_403_cast_fp16, y = const_1647_promoted_to_fp16)[name = tensor<string, []>("op_12322_cast_fp16")];
            tensor<int32, []> x_1379_axis_0 = const()[name = tensor<string, []>("x_1379_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1379_cast_fp16 = stack(axis = x_1379_axis_0, values = (var_12322_cast_fp16, squeeze_402_cast_fp16))[name = tensor<string, []>("x_1379_cast_fp16")];
            tensor<int32, [4]> var_12328 = const()[name = tensor<string, []>("op_12328"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12329_cast_fp16 = reshape(shape = var_12328, x = x_1379_cast_fp16)[name = tensor<string, []>("op_12329_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12330_cast_fp16 = mul(x = var_12329_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12330_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_201_cast_fp16 = add(x = var_12309_cast_fp16, y = var_12330_cast_fp16)[name = tensor<string, []>("k_patches_201_cast_fp16")];
            tensor<bool, []> var_12333_interleave_0 = const()[name = tensor<string, []>("op_12333_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12333_cast_fp16 = concat(axis = var_20, interleave = var_12333_interleave_0, values = (var_12272_cast_fp16, q_patches_201_cast_fp16))[name = tensor<string, []>("op_12333_cast_fp16")];
            tensor<bool, []> var_12336_interleave_0 = const()[name = tensor<string, []>("op_12336_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12336_cast_fp16 = concat(axis = var_20, interleave = var_12336_interleave_0, values = (var_12276_cast_fp16, k_patches_201_cast_fp16))[name = tensor<string, []>("op_12336_cast_fp16")];
            tensor<fp16, []> var_12338_to_fp16 = const()[name = tensor<string, []>("op_12338_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1009_cast_fp16 = mul(x = var_12333_cast_fp16, y = var_12338_to_fp16)[name = tensor<string, []>("q_1009_cast_fp16")];
            tensor<bool, []> attn_401_transpose_x_1 = const()[name = tensor<string, []>("attn_401_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_401_transpose_y_1 = const()[name = tensor<string, []>("attn_401_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_401_cast_fp16 = matmul(transpose_x = attn_401_transpose_x_1, transpose_y = attn_401_transpose_y_1, x = q_1009_cast_fp16, y = var_12336_cast_fp16)[name = tensor<string, []>("attn_401_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_403_cast_fp16 = softmax(axis = var_21, x = attn_401_cast_fp16)[name = tensor<string, []>("attn_403_cast_fp16")];
            tensor<bool, []> var_12343_transpose_x_0 = const()[name = tensor<string, []>("op_12343_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12343_transpose_y_0 = const()[name = tensor<string, []>("op_12343_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12343_cast_fp16 = matmul(transpose_x = var_12343_transpose_x_0, transpose_y = var_12343_transpose_y_0, x = attn_403_cast_fp16, y = v_605_cast_fp16)[name = tensor<string, []>("op_12343_cast_fp16")];
            tensor<int32, [1]> var_12344_axes_0 = const()[name = tensor<string, []>("op_12344_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12344_cast_fp16 = squeeze(axes = var_12344_axes_0, x = var_12343_cast_fp16)[name = tensor<string, []>("op_12344_cast_fp16")];
            tensor<string, []> var_12350_pad_type_0 = const()[name = tensor<string, []>("op_12350_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12350_strides_0 = const()[name = tensor<string, []>("op_12350_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12350_pad_0 = const()[name = tensor<string, []>("op_12350_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12350_dilations_0 = const()[name = tensor<string, []>("op_12350_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12350_groups_0 = const()[name = tensor<string, []>("op_12350_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116505344)))];
            tensor<fp16, [1, 64, 1, 197]> var_12350_cast_fp16 = conv(dilations = var_12350_dilations_0, groups = var_12350_groups_0, pad = var_12350_pad_0, pad_type = var_12350_pad_type_0, strides = var_12350_strides_0, weight = model_blocks_8_attn_q_projs_5_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12350_cast_fp16")];
            tensor<int32, [1]> var_12351_axes_0 = const()[name = tensor<string, []>("op_12351_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12351_cast_fp16 = squeeze(axes = var_12351_axes_0, x = var_12350_cast_fp16)[name = tensor<string, []>("op_12351_cast_fp16")];
            tensor<int32, [3]> q_1011_perm_0 = const()[name = tensor<string, []>("q_1011_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12358_pad_type_0 = const()[name = tensor<string, []>("op_12358_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12358_strides_0 = const()[name = tensor<string, []>("op_12358_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12358_pad_0 = const()[name = tensor<string, []>("op_12358_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12358_dilations_0 = const()[name = tensor<string, []>("op_12358_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12358_groups_0 = const()[name = tensor<string, []>("op_12358_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116603712)))];
            tensor<fp16, [1, 64, 1, 197]> var_12358_cast_fp16 = conv(dilations = var_12358_dilations_0, groups = var_12358_groups_0, pad = var_12358_pad_0, pad_type = var_12358_pad_type_0, strides = var_12358_strides_0, weight = model_blocks_8_attn_k_projs_5_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12358_cast_fp16")];
            tensor<int32, [1]> var_12359_axes_0 = const()[name = tensor<string, []>("op_12359_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12359_cast_fp16 = squeeze(axes = var_12359_axes_0, x = var_12358_cast_fp16)[name = tensor<string, []>("op_12359_cast_fp16")];
            tensor<int32, [3]> k_607_perm_0 = const()[name = tensor<string, []>("k_607_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12366_pad_type_0 = const()[name = tensor<string, []>("op_12366_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12366_strides_0 = const()[name = tensor<string, []>("op_12366_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12366_pad_0 = const()[name = tensor<string, []>("op_12366_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12366_dilations_0 = const()[name = tensor<string, []>("op_12366_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12366_groups_0 = const()[name = tensor<string, []>("op_12366_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116702080)))];
            tensor<fp16, [1, 64, 1, 197]> var_12366_cast_fp16 = conv(dilations = var_12366_dilations_0, groups = var_12366_groups_0, pad = var_12366_pad_0, pad_type = var_12366_pad_type_0, strides = var_12366_strides_0, weight = model_blocks_8_attn_v_projs_5_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12366_cast_fp16")];
            tensor<int32, [1]> var_12367_axes_0 = const()[name = tensor<string, []>("op_12367_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12367_cast_fp16 = squeeze(axes = var_12367_axes_0, x = var_12366_cast_fp16)[name = tensor<string, []>("op_12367_cast_fp16")];
            tensor<int32, [3]> v_607_perm_0 = const()[name = tensor<string, []>("v_607_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116800448)))];
            tensor<fp16, [1, 197, 64]> q_1011_cast_fp16 = transpose(perm = q_1011_perm_0, x = var_12351_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 197, 64]> q_1013_cast_fp16 = add(x = q_1011_cast_fp16, y = model_blocks_8_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_1013_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116800640)))];
            tensor<fp16, [1, 197, 64]> v_607_cast_fp16 = transpose(perm = v_607_perm_0, x = var_12367_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 197, 64]> v_609_cast_fp16 = add(x = v_607_cast_fp16, y = model_blocks_8_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_609_cast_fp16")];
            tensor<int32, [1]> q_1015_axes_0 = const()[name = tensor<string, []>("q_1015_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1015_cast_fp16 = expand_dims(axes = q_1015_axes_0, x = q_1013_cast_fp16)[name = tensor<string, []>("q_1015_cast_fp16")];
            tensor<int32, [1]> k_609_axes_0 = const()[name = tensor<string, []>("k_609_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_607_cast_fp16 = transpose(perm = k_607_perm_0, x = var_12359_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [1, 1, 197, 64]> k_609_cast_fp16 = expand_dims(axes = k_609_axes_0, x = k_607_cast_fp16)[name = tensor<string, []>("k_609_cast_fp16")];
            tensor<int32, [1]> v_611_axes_0 = const()[name = tensor<string, []>("v_611_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_611_cast_fp16 = expand_dims(axes = v_611_axes_0, x = v_609_cast_fp16)[name = tensor<string, []>("v_611_cast_fp16")];
            tensor<int32, [4]> var_12376_begin_0 = const()[name = tensor<string, []>("op_12376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12376_end_0 = const()[name = tensor<string, []>("op_12376_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12376_end_mask_0 = const()[name = tensor<string, []>("op_12376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12376_cast_fp16 = slice_by_index(begin = var_12376_begin_0, end = var_12376_end_0, end_mask = var_12376_end_mask_0, x = q_1015_cast_fp16)[name = tensor<string, []>("op_12376_cast_fp16")];
            tensor<int32, [4]> var_12380_begin_0 = const()[name = tensor<string, []>("op_12380_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12380_end_0 = const()[name = tensor<string, []>("op_12380_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12380_end_mask_0 = const()[name = tensor<string, []>("op_12380_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12380_cast_fp16 = slice_by_index(begin = var_12380_begin_0, end = var_12380_end_0, end_mask = var_12380_end_mask_0, x = k_609_cast_fp16)[name = tensor<string, []>("op_12380_cast_fp16")];
            tensor<int32, [4]> var_12384_begin_0 = const()[name = tensor<string, []>("op_12384_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12384_end_0 = const()[name = tensor<string, []>("op_12384_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12384_end_mask_0 = const()[name = tensor<string, []>("op_12384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12384_cast_fp16 = slice_by_index(begin = var_12384_begin_0, end = var_12384_end_0, end_mask = var_12384_end_mask_0, x = q_1015_cast_fp16)[name = tensor<string, []>("op_12384_cast_fp16")];
            tensor<int32, [4]> var_12388_begin_0 = const()[name = tensor<string, []>("op_12388_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12388_end_0 = const()[name = tensor<string, []>("op_12388_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12388_end_mask_0 = const()[name = tensor<string, []>("op_12388_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12388_cast_fp16 = slice_by_index(begin = var_12388_begin_0, end = var_12388_end_0, end_mask = var_12388_end_mask_0, x = k_609_cast_fp16)[name = tensor<string, []>("op_12388_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12390_cast_fp16 = mul(x = var_12384_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12390_cast_fp16")];
            tensor<int32, [5]> var_12398 = const()[name = tensor<string, []>("op_12398"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1383_cast_fp16 = reshape(shape = var_12398, x = var_12384_cast_fp16)[name = tensor<string, []>("x_1383_cast_fp16")];
            tensor<int32, [2]> var_12400_split_sizes_0 = const()[name = tensor<string, []>("op_12400_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12400_axis_0 = const()[name = tensor<string, []>("op_12400_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12400_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12400_cast_fp16_1 = split(axis = var_12400_axis_0, split_sizes = var_12400_split_sizes_0, x = x_1383_cast_fp16)[name = tensor<string, []>("op_12400_cast_fp16")];
            tensor<int32, [1]> squeeze_404_axes_0 = const()[name = tensor<string, []>("squeeze_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_404_cast_fp16 = squeeze(axes = squeeze_404_axes_0, x = var_12400_cast_fp16_0)[name = tensor<string, []>("squeeze_404_cast_fp16")];
            tensor<int32, [1]> squeeze_405_axes_0 = const()[name = tensor<string, []>("squeeze_405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_405_cast_fp16 = squeeze(axes = squeeze_405_axes_0, x = var_12400_cast_fp16_1)[name = tensor<string, []>("squeeze_405_cast_fp16")];
            tensor<fp16, []> const_1655_promoted_to_fp16 = const()[name = tensor<string, []>("const_1655_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12403_cast_fp16 = mul(x = squeeze_405_cast_fp16, y = const_1655_promoted_to_fp16)[name = tensor<string, []>("op_12403_cast_fp16")];
            tensor<int32, []> x_1385_axis_0 = const()[name = tensor<string, []>("x_1385_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1385_cast_fp16 = stack(axis = x_1385_axis_0, values = (var_12403_cast_fp16, squeeze_404_cast_fp16))[name = tensor<string, []>("x_1385_cast_fp16")];
            tensor<int32, [4]> var_12409 = const()[name = tensor<string, []>("op_12409"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12410_cast_fp16 = reshape(shape = var_12409, x = x_1385_cast_fp16)[name = tensor<string, []>("op_12410_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12411_cast_fp16 = mul(x = var_12410_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12411_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_203_cast_fp16 = add(x = var_12390_cast_fp16, y = var_12411_cast_fp16)[name = tensor<string, []>("q_patches_203_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12413_cast_fp16 = mul(x = var_12388_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12413_cast_fp16")];
            tensor<int32, [5]> var_12421 = const()[name = tensor<string, []>("op_12421"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1389_cast_fp16 = reshape(shape = var_12421, x = var_12388_cast_fp16)[name = tensor<string, []>("x_1389_cast_fp16")];
            tensor<int32, [2]> var_12423_split_sizes_0 = const()[name = tensor<string, []>("op_12423_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12423_axis_0 = const()[name = tensor<string, []>("op_12423_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12423_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12423_cast_fp16_1 = split(axis = var_12423_axis_0, split_sizes = var_12423_split_sizes_0, x = x_1389_cast_fp16)[name = tensor<string, []>("op_12423_cast_fp16")];
            tensor<int32, [1]> squeeze_406_axes_0 = const()[name = tensor<string, []>("squeeze_406_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_406_cast_fp16 = squeeze(axes = squeeze_406_axes_0, x = var_12423_cast_fp16_0)[name = tensor<string, []>("squeeze_406_cast_fp16")];
            tensor<int32, [1]> squeeze_407_axes_0 = const()[name = tensor<string, []>("squeeze_407_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_407_cast_fp16 = squeeze(axes = squeeze_407_axes_0, x = var_12423_cast_fp16_1)[name = tensor<string, []>("squeeze_407_cast_fp16")];
            tensor<fp16, []> const_1663_promoted_to_fp16 = const()[name = tensor<string, []>("const_1663_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12426_cast_fp16 = mul(x = squeeze_407_cast_fp16, y = const_1663_promoted_to_fp16)[name = tensor<string, []>("op_12426_cast_fp16")];
            tensor<int32, []> x_1391_axis_0 = const()[name = tensor<string, []>("x_1391_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1391_cast_fp16 = stack(axis = x_1391_axis_0, values = (var_12426_cast_fp16, squeeze_406_cast_fp16))[name = tensor<string, []>("x_1391_cast_fp16")];
            tensor<int32, [4]> var_12432 = const()[name = tensor<string, []>("op_12432"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12433_cast_fp16 = reshape(shape = var_12432, x = x_1391_cast_fp16)[name = tensor<string, []>("op_12433_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12434_cast_fp16 = mul(x = var_12433_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12434_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_203_cast_fp16 = add(x = var_12413_cast_fp16, y = var_12434_cast_fp16)[name = tensor<string, []>("k_patches_203_cast_fp16")];
            tensor<bool, []> var_12437_interleave_0 = const()[name = tensor<string, []>("op_12437_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12437_cast_fp16 = concat(axis = var_20, interleave = var_12437_interleave_0, values = (var_12376_cast_fp16, q_patches_203_cast_fp16))[name = tensor<string, []>("op_12437_cast_fp16")];
            tensor<bool, []> var_12440_interleave_0 = const()[name = tensor<string, []>("op_12440_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12440_cast_fp16 = concat(axis = var_20, interleave = var_12440_interleave_0, values = (var_12380_cast_fp16, k_patches_203_cast_fp16))[name = tensor<string, []>("op_12440_cast_fp16")];
            tensor<fp16, []> var_12442_to_fp16 = const()[name = tensor<string, []>("op_12442_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1019_cast_fp16 = mul(x = var_12437_cast_fp16, y = var_12442_to_fp16)[name = tensor<string, []>("q_1019_cast_fp16")];
            tensor<bool, []> attn_405_transpose_x_1 = const()[name = tensor<string, []>("attn_405_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_405_transpose_y_1 = const()[name = tensor<string, []>("attn_405_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_405_cast_fp16 = matmul(transpose_x = attn_405_transpose_x_1, transpose_y = attn_405_transpose_y_1, x = q_1019_cast_fp16, y = var_12440_cast_fp16)[name = tensor<string, []>("attn_405_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_407_cast_fp16 = softmax(axis = var_21, x = attn_405_cast_fp16)[name = tensor<string, []>("attn_407_cast_fp16")];
            tensor<bool, []> var_12447_transpose_x_0 = const()[name = tensor<string, []>("op_12447_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12447_transpose_y_0 = const()[name = tensor<string, []>("op_12447_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12447_cast_fp16 = matmul(transpose_x = var_12447_transpose_x_0, transpose_y = var_12447_transpose_y_0, x = attn_407_cast_fp16, y = v_611_cast_fp16)[name = tensor<string, []>("op_12447_cast_fp16")];
            tensor<int32, [1]> var_12448_axes_0 = const()[name = tensor<string, []>("op_12448_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12448_cast_fp16 = squeeze(axes = var_12448_axes_0, x = var_12447_cast_fp16)[name = tensor<string, []>("op_12448_cast_fp16")];
            tensor<string, []> var_12454_pad_type_0 = const()[name = tensor<string, []>("op_12454_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12454_strides_0 = const()[name = tensor<string, []>("op_12454_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12454_pad_0 = const()[name = tensor<string, []>("op_12454_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12454_dilations_0 = const()[name = tensor<string, []>("op_12454_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12454_groups_0 = const()[name = tensor<string, []>("op_12454_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116800832)))];
            tensor<fp16, [1, 64, 1, 197]> var_12454_cast_fp16 = conv(dilations = var_12454_dilations_0, groups = var_12454_groups_0, pad = var_12454_pad_0, pad_type = var_12454_pad_type_0, strides = var_12454_strides_0, weight = model_blocks_8_attn_q_projs_6_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12454_cast_fp16")];
            tensor<int32, [1]> var_12455_axes_0 = const()[name = tensor<string, []>("op_12455_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12455_cast_fp16 = squeeze(axes = var_12455_axes_0, x = var_12454_cast_fp16)[name = tensor<string, []>("op_12455_cast_fp16")];
            tensor<int32, [3]> q_1021_perm_0 = const()[name = tensor<string, []>("q_1021_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12462_pad_type_0 = const()[name = tensor<string, []>("op_12462_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12462_strides_0 = const()[name = tensor<string, []>("op_12462_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12462_pad_0 = const()[name = tensor<string, []>("op_12462_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12462_dilations_0 = const()[name = tensor<string, []>("op_12462_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12462_groups_0 = const()[name = tensor<string, []>("op_12462_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116899200)))];
            tensor<fp16, [1, 64, 1, 197]> var_12462_cast_fp16 = conv(dilations = var_12462_dilations_0, groups = var_12462_groups_0, pad = var_12462_pad_0, pad_type = var_12462_pad_type_0, strides = var_12462_strides_0, weight = model_blocks_8_attn_k_projs_6_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12462_cast_fp16")];
            tensor<int32, [1]> var_12463_axes_0 = const()[name = tensor<string, []>("op_12463_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12463_cast_fp16 = squeeze(axes = var_12463_axes_0, x = var_12462_cast_fp16)[name = tensor<string, []>("op_12463_cast_fp16")];
            tensor<int32, [3]> k_613_perm_0 = const()[name = tensor<string, []>("k_613_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12470_pad_type_0 = const()[name = tensor<string, []>("op_12470_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12470_strides_0 = const()[name = tensor<string, []>("op_12470_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12470_pad_0 = const()[name = tensor<string, []>("op_12470_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12470_dilations_0 = const()[name = tensor<string, []>("op_12470_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12470_groups_0 = const()[name = tensor<string, []>("op_12470_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116997568)))];
            tensor<fp16, [1, 64, 1, 197]> var_12470_cast_fp16 = conv(dilations = var_12470_dilations_0, groups = var_12470_groups_0, pad = var_12470_pad_0, pad_type = var_12470_pad_type_0, strides = var_12470_strides_0, weight = model_blocks_8_attn_v_projs_6_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12470_cast_fp16")];
            tensor<int32, [1]> var_12471_axes_0 = const()[name = tensor<string, []>("op_12471_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12471_cast_fp16 = squeeze(axes = var_12471_axes_0, x = var_12470_cast_fp16)[name = tensor<string, []>("op_12471_cast_fp16")];
            tensor<int32, [3]> v_613_perm_0 = const()[name = tensor<string, []>("v_613_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117095936)))];
            tensor<fp16, [1, 197, 64]> q_1021_cast_fp16 = transpose(perm = q_1021_perm_0, x = var_12455_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [1, 197, 64]> q_1023_cast_fp16 = add(x = q_1021_cast_fp16, y = model_blocks_8_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_1023_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117096128)))];
            tensor<fp16, [1, 197, 64]> v_613_cast_fp16 = transpose(perm = v_613_perm_0, x = var_12471_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [1, 197, 64]> v_615_cast_fp16 = add(x = v_613_cast_fp16, y = model_blocks_8_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_615_cast_fp16")];
            tensor<int32, [1]> q_1025_axes_0 = const()[name = tensor<string, []>("q_1025_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1025_cast_fp16 = expand_dims(axes = q_1025_axes_0, x = q_1023_cast_fp16)[name = tensor<string, []>("q_1025_cast_fp16")];
            tensor<int32, [1]> k_615_axes_0 = const()[name = tensor<string, []>("k_615_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_613_cast_fp16 = transpose(perm = k_613_perm_0, x = var_12463_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [1, 1, 197, 64]> k_615_cast_fp16 = expand_dims(axes = k_615_axes_0, x = k_613_cast_fp16)[name = tensor<string, []>("k_615_cast_fp16")];
            tensor<int32, [1]> v_617_axes_0 = const()[name = tensor<string, []>("v_617_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_617_cast_fp16 = expand_dims(axes = v_617_axes_0, x = v_615_cast_fp16)[name = tensor<string, []>("v_617_cast_fp16")];
            tensor<int32, [4]> var_12480_begin_0 = const()[name = tensor<string, []>("op_12480_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12480_end_0 = const()[name = tensor<string, []>("op_12480_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12480_end_mask_0 = const()[name = tensor<string, []>("op_12480_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12480_cast_fp16 = slice_by_index(begin = var_12480_begin_0, end = var_12480_end_0, end_mask = var_12480_end_mask_0, x = q_1025_cast_fp16)[name = tensor<string, []>("op_12480_cast_fp16")];
            tensor<int32, [4]> var_12484_begin_0 = const()[name = tensor<string, []>("op_12484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12484_end_0 = const()[name = tensor<string, []>("op_12484_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12484_end_mask_0 = const()[name = tensor<string, []>("op_12484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12484_cast_fp16 = slice_by_index(begin = var_12484_begin_0, end = var_12484_end_0, end_mask = var_12484_end_mask_0, x = k_615_cast_fp16)[name = tensor<string, []>("op_12484_cast_fp16")];
            tensor<int32, [4]> var_12488_begin_0 = const()[name = tensor<string, []>("op_12488_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12488_end_0 = const()[name = tensor<string, []>("op_12488_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12488_end_mask_0 = const()[name = tensor<string, []>("op_12488_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12488_cast_fp16 = slice_by_index(begin = var_12488_begin_0, end = var_12488_end_0, end_mask = var_12488_end_mask_0, x = q_1025_cast_fp16)[name = tensor<string, []>("op_12488_cast_fp16")];
            tensor<int32, [4]> var_12492_begin_0 = const()[name = tensor<string, []>("op_12492_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12492_end_0 = const()[name = tensor<string, []>("op_12492_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12492_end_mask_0 = const()[name = tensor<string, []>("op_12492_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12492_cast_fp16 = slice_by_index(begin = var_12492_begin_0, end = var_12492_end_0, end_mask = var_12492_end_mask_0, x = k_615_cast_fp16)[name = tensor<string, []>("op_12492_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12494_cast_fp16 = mul(x = var_12488_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12494_cast_fp16")];
            tensor<int32, [5]> var_12502 = const()[name = tensor<string, []>("op_12502"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1395_cast_fp16 = reshape(shape = var_12502, x = var_12488_cast_fp16)[name = tensor<string, []>("x_1395_cast_fp16")];
            tensor<int32, [2]> var_12504_split_sizes_0 = const()[name = tensor<string, []>("op_12504_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12504_axis_0 = const()[name = tensor<string, []>("op_12504_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12504_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12504_cast_fp16_1 = split(axis = var_12504_axis_0, split_sizes = var_12504_split_sizes_0, x = x_1395_cast_fp16)[name = tensor<string, []>("op_12504_cast_fp16")];
            tensor<int32, [1]> squeeze_408_axes_0 = const()[name = tensor<string, []>("squeeze_408_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_408_cast_fp16 = squeeze(axes = squeeze_408_axes_0, x = var_12504_cast_fp16_0)[name = tensor<string, []>("squeeze_408_cast_fp16")];
            tensor<int32, [1]> squeeze_409_axes_0 = const()[name = tensor<string, []>("squeeze_409_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_409_cast_fp16 = squeeze(axes = squeeze_409_axes_0, x = var_12504_cast_fp16_1)[name = tensor<string, []>("squeeze_409_cast_fp16")];
            tensor<fp16, []> const_1671_promoted_to_fp16 = const()[name = tensor<string, []>("const_1671_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12507_cast_fp16 = mul(x = squeeze_409_cast_fp16, y = const_1671_promoted_to_fp16)[name = tensor<string, []>("op_12507_cast_fp16")];
            tensor<int32, []> x_1397_axis_0 = const()[name = tensor<string, []>("x_1397_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1397_cast_fp16 = stack(axis = x_1397_axis_0, values = (var_12507_cast_fp16, squeeze_408_cast_fp16))[name = tensor<string, []>("x_1397_cast_fp16")];
            tensor<int32, [4]> var_12513 = const()[name = tensor<string, []>("op_12513"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12514_cast_fp16 = reshape(shape = var_12513, x = x_1397_cast_fp16)[name = tensor<string, []>("op_12514_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12515_cast_fp16 = mul(x = var_12514_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12515_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_205_cast_fp16 = add(x = var_12494_cast_fp16, y = var_12515_cast_fp16)[name = tensor<string, []>("q_patches_205_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12517_cast_fp16 = mul(x = var_12492_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12517_cast_fp16")];
            tensor<int32, [5]> var_12525 = const()[name = tensor<string, []>("op_12525"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1401_cast_fp16 = reshape(shape = var_12525, x = var_12492_cast_fp16)[name = tensor<string, []>("x_1401_cast_fp16")];
            tensor<int32, [2]> var_12527_split_sizes_0 = const()[name = tensor<string, []>("op_12527_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12527_axis_0 = const()[name = tensor<string, []>("op_12527_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12527_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12527_cast_fp16_1 = split(axis = var_12527_axis_0, split_sizes = var_12527_split_sizes_0, x = x_1401_cast_fp16)[name = tensor<string, []>("op_12527_cast_fp16")];
            tensor<int32, [1]> squeeze_410_axes_0 = const()[name = tensor<string, []>("squeeze_410_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_410_cast_fp16 = squeeze(axes = squeeze_410_axes_0, x = var_12527_cast_fp16_0)[name = tensor<string, []>("squeeze_410_cast_fp16")];
            tensor<int32, [1]> squeeze_411_axes_0 = const()[name = tensor<string, []>("squeeze_411_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_411_cast_fp16 = squeeze(axes = squeeze_411_axes_0, x = var_12527_cast_fp16_1)[name = tensor<string, []>("squeeze_411_cast_fp16")];
            tensor<fp16, []> const_1679_promoted_to_fp16 = const()[name = tensor<string, []>("const_1679_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12530_cast_fp16 = mul(x = squeeze_411_cast_fp16, y = const_1679_promoted_to_fp16)[name = tensor<string, []>("op_12530_cast_fp16")];
            tensor<int32, []> x_1403_axis_0 = const()[name = tensor<string, []>("x_1403_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1403_cast_fp16 = stack(axis = x_1403_axis_0, values = (var_12530_cast_fp16, squeeze_410_cast_fp16))[name = tensor<string, []>("x_1403_cast_fp16")];
            tensor<int32, [4]> var_12536 = const()[name = tensor<string, []>("op_12536"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12537_cast_fp16 = reshape(shape = var_12536, x = x_1403_cast_fp16)[name = tensor<string, []>("op_12537_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12538_cast_fp16 = mul(x = var_12537_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12538_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_205_cast_fp16 = add(x = var_12517_cast_fp16, y = var_12538_cast_fp16)[name = tensor<string, []>("k_patches_205_cast_fp16")];
            tensor<bool, []> var_12541_interleave_0 = const()[name = tensor<string, []>("op_12541_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12541_cast_fp16 = concat(axis = var_20, interleave = var_12541_interleave_0, values = (var_12480_cast_fp16, q_patches_205_cast_fp16))[name = tensor<string, []>("op_12541_cast_fp16")];
            tensor<bool, []> var_12544_interleave_0 = const()[name = tensor<string, []>("op_12544_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12544_cast_fp16 = concat(axis = var_20, interleave = var_12544_interleave_0, values = (var_12484_cast_fp16, k_patches_205_cast_fp16))[name = tensor<string, []>("op_12544_cast_fp16")];
            tensor<fp16, []> var_12546_to_fp16 = const()[name = tensor<string, []>("op_12546_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1029_cast_fp16 = mul(x = var_12541_cast_fp16, y = var_12546_to_fp16)[name = tensor<string, []>("q_1029_cast_fp16")];
            tensor<bool, []> attn_409_transpose_x_1 = const()[name = tensor<string, []>("attn_409_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_409_transpose_y_1 = const()[name = tensor<string, []>("attn_409_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_409_cast_fp16 = matmul(transpose_x = attn_409_transpose_x_1, transpose_y = attn_409_transpose_y_1, x = q_1029_cast_fp16, y = var_12544_cast_fp16)[name = tensor<string, []>("attn_409_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_411_cast_fp16 = softmax(axis = var_21, x = attn_409_cast_fp16)[name = tensor<string, []>("attn_411_cast_fp16")];
            tensor<bool, []> var_12551_transpose_x_0 = const()[name = tensor<string, []>("op_12551_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12551_transpose_y_0 = const()[name = tensor<string, []>("op_12551_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12551_cast_fp16 = matmul(transpose_x = var_12551_transpose_x_0, transpose_y = var_12551_transpose_y_0, x = attn_411_cast_fp16, y = v_617_cast_fp16)[name = tensor<string, []>("op_12551_cast_fp16")];
            tensor<int32, [1]> var_12552_axes_0 = const()[name = tensor<string, []>("op_12552_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12552_cast_fp16 = squeeze(axes = var_12552_axes_0, x = var_12551_cast_fp16)[name = tensor<string, []>("op_12552_cast_fp16")];
            tensor<string, []> var_12558_pad_type_0 = const()[name = tensor<string, []>("op_12558_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12558_strides_0 = const()[name = tensor<string, []>("op_12558_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12558_pad_0 = const()[name = tensor<string, []>("op_12558_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12558_dilations_0 = const()[name = tensor<string, []>("op_12558_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12558_groups_0 = const()[name = tensor<string, []>("op_12558_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117096320)))];
            tensor<fp16, [1, 64, 1, 197]> var_12558_cast_fp16 = conv(dilations = var_12558_dilations_0, groups = var_12558_groups_0, pad = var_12558_pad_0, pad_type = var_12558_pad_type_0, strides = var_12558_strides_0, weight = model_blocks_8_attn_q_projs_7_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12558_cast_fp16")];
            tensor<int32, [1]> var_12559_axes_0 = const()[name = tensor<string, []>("op_12559_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12559_cast_fp16 = squeeze(axes = var_12559_axes_0, x = var_12558_cast_fp16)[name = tensor<string, []>("op_12559_cast_fp16")];
            tensor<int32, [3]> q_1031_perm_0 = const()[name = tensor<string, []>("q_1031_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12566_pad_type_0 = const()[name = tensor<string, []>("op_12566_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12566_strides_0 = const()[name = tensor<string, []>("op_12566_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12566_pad_0 = const()[name = tensor<string, []>("op_12566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12566_dilations_0 = const()[name = tensor<string, []>("op_12566_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12566_groups_0 = const()[name = tensor<string, []>("op_12566_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117194688)))];
            tensor<fp16, [1, 64, 1, 197]> var_12566_cast_fp16 = conv(dilations = var_12566_dilations_0, groups = var_12566_groups_0, pad = var_12566_pad_0, pad_type = var_12566_pad_type_0, strides = var_12566_strides_0, weight = model_blocks_8_attn_k_projs_7_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12566_cast_fp16")];
            tensor<int32, [1]> var_12567_axes_0 = const()[name = tensor<string, []>("op_12567_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12567_cast_fp16 = squeeze(axes = var_12567_axes_0, x = var_12566_cast_fp16)[name = tensor<string, []>("op_12567_cast_fp16")];
            tensor<int32, [3]> k_619_perm_0 = const()[name = tensor<string, []>("k_619_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12574_pad_type_0 = const()[name = tensor<string, []>("op_12574_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12574_strides_0 = const()[name = tensor<string, []>("op_12574_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12574_pad_0 = const()[name = tensor<string, []>("op_12574_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12574_dilations_0 = const()[name = tensor<string, []>("op_12574_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12574_groups_0 = const()[name = tensor<string, []>("op_12574_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117293056)))];
            tensor<fp16, [1, 64, 1, 197]> var_12574_cast_fp16 = conv(dilations = var_12574_dilations_0, groups = var_12574_groups_0, pad = var_12574_pad_0, pad_type = var_12574_pad_type_0, strides = var_12574_strides_0, weight = model_blocks_8_attn_v_projs_7_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12574_cast_fp16")];
            tensor<int32, [1]> var_12575_axes_0 = const()[name = tensor<string, []>("op_12575_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12575_cast_fp16 = squeeze(axes = var_12575_axes_0, x = var_12574_cast_fp16)[name = tensor<string, []>("op_12575_cast_fp16")];
            tensor<int32, [3]> v_619_perm_0 = const()[name = tensor<string, []>("v_619_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117391424)))];
            tensor<fp16, [1, 197, 64]> q_1031_cast_fp16 = transpose(perm = q_1031_perm_0, x = var_12559_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [1, 197, 64]> q_1033_cast_fp16 = add(x = q_1031_cast_fp16, y = model_blocks_8_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_1033_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117391616)))];
            tensor<fp16, [1, 197, 64]> v_619_cast_fp16 = transpose(perm = v_619_perm_0, x = var_12575_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 197, 64]> v_621_cast_fp16 = add(x = v_619_cast_fp16, y = model_blocks_8_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_621_cast_fp16")];
            tensor<int32, [1]> q_1035_axes_0 = const()[name = tensor<string, []>("q_1035_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1035_cast_fp16 = expand_dims(axes = q_1035_axes_0, x = q_1033_cast_fp16)[name = tensor<string, []>("q_1035_cast_fp16")];
            tensor<int32, [1]> k_621_axes_0 = const()[name = tensor<string, []>("k_621_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_619_cast_fp16 = transpose(perm = k_619_perm_0, x = var_12567_cast_fp16)[name = tensor<string, []>("transpose_148")];
            tensor<fp16, [1, 1, 197, 64]> k_621_cast_fp16 = expand_dims(axes = k_621_axes_0, x = k_619_cast_fp16)[name = tensor<string, []>("k_621_cast_fp16")];
            tensor<int32, [1]> v_623_axes_0 = const()[name = tensor<string, []>("v_623_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_623_cast_fp16 = expand_dims(axes = v_623_axes_0, x = v_621_cast_fp16)[name = tensor<string, []>("v_623_cast_fp16")];
            tensor<int32, [4]> var_12584_begin_0 = const()[name = tensor<string, []>("op_12584_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12584_end_0 = const()[name = tensor<string, []>("op_12584_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12584_end_mask_0 = const()[name = tensor<string, []>("op_12584_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12584_cast_fp16 = slice_by_index(begin = var_12584_begin_0, end = var_12584_end_0, end_mask = var_12584_end_mask_0, x = q_1035_cast_fp16)[name = tensor<string, []>("op_12584_cast_fp16")];
            tensor<int32, [4]> var_12588_begin_0 = const()[name = tensor<string, []>("op_12588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12588_end_0 = const()[name = tensor<string, []>("op_12588_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12588_end_mask_0 = const()[name = tensor<string, []>("op_12588_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12588_cast_fp16 = slice_by_index(begin = var_12588_begin_0, end = var_12588_end_0, end_mask = var_12588_end_mask_0, x = k_621_cast_fp16)[name = tensor<string, []>("op_12588_cast_fp16")];
            tensor<int32, [4]> var_12592_begin_0 = const()[name = tensor<string, []>("op_12592_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12592_end_0 = const()[name = tensor<string, []>("op_12592_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12592_end_mask_0 = const()[name = tensor<string, []>("op_12592_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12592_cast_fp16 = slice_by_index(begin = var_12592_begin_0, end = var_12592_end_0, end_mask = var_12592_end_mask_0, x = q_1035_cast_fp16)[name = tensor<string, []>("op_12592_cast_fp16")];
            tensor<int32, [4]> var_12596_begin_0 = const()[name = tensor<string, []>("op_12596_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12596_end_0 = const()[name = tensor<string, []>("op_12596_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12596_end_mask_0 = const()[name = tensor<string, []>("op_12596_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12596_cast_fp16 = slice_by_index(begin = var_12596_begin_0, end = var_12596_end_0, end_mask = var_12596_end_mask_0, x = k_621_cast_fp16)[name = tensor<string, []>("op_12596_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12598_cast_fp16 = mul(x = var_12592_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12598_cast_fp16")];
            tensor<int32, [5]> var_12606 = const()[name = tensor<string, []>("op_12606"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1407_cast_fp16 = reshape(shape = var_12606, x = var_12592_cast_fp16)[name = tensor<string, []>("x_1407_cast_fp16")];
            tensor<int32, [2]> var_12608_split_sizes_0 = const()[name = tensor<string, []>("op_12608_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12608_axis_0 = const()[name = tensor<string, []>("op_12608_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12608_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12608_cast_fp16_1 = split(axis = var_12608_axis_0, split_sizes = var_12608_split_sizes_0, x = x_1407_cast_fp16)[name = tensor<string, []>("op_12608_cast_fp16")];
            tensor<int32, [1]> squeeze_412_axes_0 = const()[name = tensor<string, []>("squeeze_412_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_412_cast_fp16 = squeeze(axes = squeeze_412_axes_0, x = var_12608_cast_fp16_0)[name = tensor<string, []>("squeeze_412_cast_fp16")];
            tensor<int32, [1]> squeeze_413_axes_0 = const()[name = tensor<string, []>("squeeze_413_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_413_cast_fp16 = squeeze(axes = squeeze_413_axes_0, x = var_12608_cast_fp16_1)[name = tensor<string, []>("squeeze_413_cast_fp16")];
            tensor<fp16, []> const_1687_promoted_to_fp16 = const()[name = tensor<string, []>("const_1687_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12611_cast_fp16 = mul(x = squeeze_413_cast_fp16, y = const_1687_promoted_to_fp16)[name = tensor<string, []>("op_12611_cast_fp16")];
            tensor<int32, []> x_1409_axis_0 = const()[name = tensor<string, []>("x_1409_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1409_cast_fp16 = stack(axis = x_1409_axis_0, values = (var_12611_cast_fp16, squeeze_412_cast_fp16))[name = tensor<string, []>("x_1409_cast_fp16")];
            tensor<int32, [4]> var_12617 = const()[name = tensor<string, []>("op_12617"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12618_cast_fp16 = reshape(shape = var_12617, x = x_1409_cast_fp16)[name = tensor<string, []>("op_12618_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12619_cast_fp16 = mul(x = var_12618_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12619_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_207_cast_fp16 = add(x = var_12598_cast_fp16, y = var_12619_cast_fp16)[name = tensor<string, []>("q_patches_207_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12621_cast_fp16 = mul(x = var_12596_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12621_cast_fp16")];
            tensor<int32, [5]> var_12629 = const()[name = tensor<string, []>("op_12629"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1413_cast_fp16 = reshape(shape = var_12629, x = var_12596_cast_fp16)[name = tensor<string, []>("x_1413_cast_fp16")];
            tensor<int32, [2]> var_12631_split_sizes_0 = const()[name = tensor<string, []>("op_12631_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12631_axis_0 = const()[name = tensor<string, []>("op_12631_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12631_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12631_cast_fp16_1 = split(axis = var_12631_axis_0, split_sizes = var_12631_split_sizes_0, x = x_1413_cast_fp16)[name = tensor<string, []>("op_12631_cast_fp16")];
            tensor<int32, [1]> squeeze_414_axes_0 = const()[name = tensor<string, []>("squeeze_414_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_414_cast_fp16 = squeeze(axes = squeeze_414_axes_0, x = var_12631_cast_fp16_0)[name = tensor<string, []>("squeeze_414_cast_fp16")];
            tensor<int32, [1]> squeeze_415_axes_0 = const()[name = tensor<string, []>("squeeze_415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_415_cast_fp16 = squeeze(axes = squeeze_415_axes_0, x = var_12631_cast_fp16_1)[name = tensor<string, []>("squeeze_415_cast_fp16")];
            tensor<fp16, []> const_1695_promoted_to_fp16 = const()[name = tensor<string, []>("const_1695_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12634_cast_fp16 = mul(x = squeeze_415_cast_fp16, y = const_1695_promoted_to_fp16)[name = tensor<string, []>("op_12634_cast_fp16")];
            tensor<int32, []> x_1415_axis_0 = const()[name = tensor<string, []>("x_1415_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1415_cast_fp16 = stack(axis = x_1415_axis_0, values = (var_12634_cast_fp16, squeeze_414_cast_fp16))[name = tensor<string, []>("x_1415_cast_fp16")];
            tensor<int32, [4]> var_12640 = const()[name = tensor<string, []>("op_12640"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12641_cast_fp16 = reshape(shape = var_12640, x = x_1415_cast_fp16)[name = tensor<string, []>("op_12641_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12642_cast_fp16 = mul(x = var_12641_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12642_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_207_cast_fp16 = add(x = var_12621_cast_fp16, y = var_12642_cast_fp16)[name = tensor<string, []>("k_patches_207_cast_fp16")];
            tensor<bool, []> var_12645_interleave_0 = const()[name = tensor<string, []>("op_12645_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12645_cast_fp16 = concat(axis = var_20, interleave = var_12645_interleave_0, values = (var_12584_cast_fp16, q_patches_207_cast_fp16))[name = tensor<string, []>("op_12645_cast_fp16")];
            tensor<bool, []> var_12648_interleave_0 = const()[name = tensor<string, []>("op_12648_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12648_cast_fp16 = concat(axis = var_20, interleave = var_12648_interleave_0, values = (var_12588_cast_fp16, k_patches_207_cast_fp16))[name = tensor<string, []>("op_12648_cast_fp16")];
            tensor<fp16, []> var_12650_to_fp16 = const()[name = tensor<string, []>("op_12650_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1039_cast_fp16 = mul(x = var_12645_cast_fp16, y = var_12650_to_fp16)[name = tensor<string, []>("q_1039_cast_fp16")];
            tensor<bool, []> attn_413_transpose_x_1 = const()[name = tensor<string, []>("attn_413_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_413_transpose_y_1 = const()[name = tensor<string, []>("attn_413_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_413_cast_fp16 = matmul(transpose_x = attn_413_transpose_x_1, transpose_y = attn_413_transpose_y_1, x = q_1039_cast_fp16, y = var_12648_cast_fp16)[name = tensor<string, []>("attn_413_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_415_cast_fp16 = softmax(axis = var_21, x = attn_413_cast_fp16)[name = tensor<string, []>("attn_415_cast_fp16")];
            tensor<bool, []> var_12655_transpose_x_0 = const()[name = tensor<string, []>("op_12655_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12655_transpose_y_0 = const()[name = tensor<string, []>("op_12655_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12655_cast_fp16 = matmul(transpose_x = var_12655_transpose_x_0, transpose_y = var_12655_transpose_y_0, x = attn_415_cast_fp16, y = v_623_cast_fp16)[name = tensor<string, []>("op_12655_cast_fp16")];
            tensor<int32, [1]> var_12656_axes_0 = const()[name = tensor<string, []>("op_12656_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12656_cast_fp16 = squeeze(axes = var_12656_axes_0, x = var_12655_cast_fp16)[name = tensor<string, []>("op_12656_cast_fp16")];
            tensor<string, []> var_12662_pad_type_0 = const()[name = tensor<string, []>("op_12662_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12662_strides_0 = const()[name = tensor<string, []>("op_12662_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12662_pad_0 = const()[name = tensor<string, []>("op_12662_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12662_dilations_0 = const()[name = tensor<string, []>("op_12662_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12662_groups_0 = const()[name = tensor<string, []>("op_12662_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117391808)))];
            tensor<fp16, [1, 64, 1, 197]> var_12662_cast_fp16 = conv(dilations = var_12662_dilations_0, groups = var_12662_groups_0, pad = var_12662_pad_0, pad_type = var_12662_pad_type_0, strides = var_12662_strides_0, weight = model_blocks_8_attn_q_projs_8_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12662_cast_fp16")];
            tensor<int32, [1]> var_12663_axes_0 = const()[name = tensor<string, []>("op_12663_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12663_cast_fp16 = squeeze(axes = var_12663_axes_0, x = var_12662_cast_fp16)[name = tensor<string, []>("op_12663_cast_fp16")];
            tensor<int32, [3]> q_1041_perm_0 = const()[name = tensor<string, []>("q_1041_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12670_pad_type_0 = const()[name = tensor<string, []>("op_12670_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12670_strides_0 = const()[name = tensor<string, []>("op_12670_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12670_pad_0 = const()[name = tensor<string, []>("op_12670_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12670_dilations_0 = const()[name = tensor<string, []>("op_12670_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12670_groups_0 = const()[name = tensor<string, []>("op_12670_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117490176)))];
            tensor<fp16, [1, 64, 1, 197]> var_12670_cast_fp16 = conv(dilations = var_12670_dilations_0, groups = var_12670_groups_0, pad = var_12670_pad_0, pad_type = var_12670_pad_type_0, strides = var_12670_strides_0, weight = model_blocks_8_attn_k_projs_8_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12670_cast_fp16")];
            tensor<int32, [1]> var_12671_axes_0 = const()[name = tensor<string, []>("op_12671_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12671_cast_fp16 = squeeze(axes = var_12671_axes_0, x = var_12670_cast_fp16)[name = tensor<string, []>("op_12671_cast_fp16")];
            tensor<int32, [3]> k_625_perm_0 = const()[name = tensor<string, []>("k_625_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12678_pad_type_0 = const()[name = tensor<string, []>("op_12678_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12678_strides_0 = const()[name = tensor<string, []>("op_12678_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12678_pad_0 = const()[name = tensor<string, []>("op_12678_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12678_dilations_0 = const()[name = tensor<string, []>("op_12678_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12678_groups_0 = const()[name = tensor<string, []>("op_12678_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117588544)))];
            tensor<fp16, [1, 64, 1, 197]> var_12678_cast_fp16 = conv(dilations = var_12678_dilations_0, groups = var_12678_groups_0, pad = var_12678_pad_0, pad_type = var_12678_pad_type_0, strides = var_12678_strides_0, weight = model_blocks_8_attn_v_projs_8_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12678_cast_fp16")];
            tensor<int32, [1]> var_12679_axes_0 = const()[name = tensor<string, []>("op_12679_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12679_cast_fp16 = squeeze(axes = var_12679_axes_0, x = var_12678_cast_fp16)[name = tensor<string, []>("op_12679_cast_fp16")];
            tensor<int32, [3]> v_625_perm_0 = const()[name = tensor<string, []>("v_625_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117686912)))];
            tensor<fp16, [1, 197, 64]> q_1041_cast_fp16 = transpose(perm = q_1041_perm_0, x = var_12663_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 197, 64]> q_1043_cast_fp16 = add(x = q_1041_cast_fp16, y = model_blocks_8_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_1043_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117687104)))];
            tensor<fp16, [1, 197, 64]> v_625_cast_fp16 = transpose(perm = v_625_perm_0, x = var_12679_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [1, 197, 64]> v_627_cast_fp16 = add(x = v_625_cast_fp16, y = model_blocks_8_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_627_cast_fp16")];
            tensor<int32, [1]> q_1045_axes_0 = const()[name = tensor<string, []>("q_1045_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1045_cast_fp16 = expand_dims(axes = q_1045_axes_0, x = q_1043_cast_fp16)[name = tensor<string, []>("q_1045_cast_fp16")];
            tensor<int32, [1]> k_627_axes_0 = const()[name = tensor<string, []>("k_627_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_625_cast_fp16 = transpose(perm = k_625_perm_0, x = var_12671_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [1, 1, 197, 64]> k_627_cast_fp16 = expand_dims(axes = k_627_axes_0, x = k_625_cast_fp16)[name = tensor<string, []>("k_627_cast_fp16")];
            tensor<int32, [1]> v_629_axes_0 = const()[name = tensor<string, []>("v_629_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_629_cast_fp16 = expand_dims(axes = v_629_axes_0, x = v_627_cast_fp16)[name = tensor<string, []>("v_629_cast_fp16")];
            tensor<int32, [4]> var_12688_begin_0 = const()[name = tensor<string, []>("op_12688_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12688_end_0 = const()[name = tensor<string, []>("op_12688_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12688_end_mask_0 = const()[name = tensor<string, []>("op_12688_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12688_cast_fp16 = slice_by_index(begin = var_12688_begin_0, end = var_12688_end_0, end_mask = var_12688_end_mask_0, x = q_1045_cast_fp16)[name = tensor<string, []>("op_12688_cast_fp16")];
            tensor<int32, [4]> var_12692_begin_0 = const()[name = tensor<string, []>("op_12692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12692_end_0 = const()[name = tensor<string, []>("op_12692_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12692_end_mask_0 = const()[name = tensor<string, []>("op_12692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12692_cast_fp16 = slice_by_index(begin = var_12692_begin_0, end = var_12692_end_0, end_mask = var_12692_end_mask_0, x = k_627_cast_fp16)[name = tensor<string, []>("op_12692_cast_fp16")];
            tensor<int32, [4]> var_12696_begin_0 = const()[name = tensor<string, []>("op_12696_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12696_end_0 = const()[name = tensor<string, []>("op_12696_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12696_end_mask_0 = const()[name = tensor<string, []>("op_12696_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12696_cast_fp16 = slice_by_index(begin = var_12696_begin_0, end = var_12696_end_0, end_mask = var_12696_end_mask_0, x = q_1045_cast_fp16)[name = tensor<string, []>("op_12696_cast_fp16")];
            tensor<int32, [4]> var_12700_begin_0 = const()[name = tensor<string, []>("op_12700_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12700_end_0 = const()[name = tensor<string, []>("op_12700_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12700_end_mask_0 = const()[name = tensor<string, []>("op_12700_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12700_cast_fp16 = slice_by_index(begin = var_12700_begin_0, end = var_12700_end_0, end_mask = var_12700_end_mask_0, x = k_627_cast_fp16)[name = tensor<string, []>("op_12700_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12702_cast_fp16 = mul(x = var_12696_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12702_cast_fp16")];
            tensor<int32, [5]> var_12710 = const()[name = tensor<string, []>("op_12710"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1419_cast_fp16 = reshape(shape = var_12710, x = var_12696_cast_fp16)[name = tensor<string, []>("x_1419_cast_fp16")];
            tensor<int32, [2]> var_12712_split_sizes_0 = const()[name = tensor<string, []>("op_12712_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12712_axis_0 = const()[name = tensor<string, []>("op_12712_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12712_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12712_cast_fp16_1 = split(axis = var_12712_axis_0, split_sizes = var_12712_split_sizes_0, x = x_1419_cast_fp16)[name = tensor<string, []>("op_12712_cast_fp16")];
            tensor<int32, [1]> squeeze_416_axes_0 = const()[name = tensor<string, []>("squeeze_416_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_416_cast_fp16 = squeeze(axes = squeeze_416_axes_0, x = var_12712_cast_fp16_0)[name = tensor<string, []>("squeeze_416_cast_fp16")];
            tensor<int32, [1]> squeeze_417_axes_0 = const()[name = tensor<string, []>("squeeze_417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_417_cast_fp16 = squeeze(axes = squeeze_417_axes_0, x = var_12712_cast_fp16_1)[name = tensor<string, []>("squeeze_417_cast_fp16")];
            tensor<fp16, []> const_1703_promoted_to_fp16 = const()[name = tensor<string, []>("const_1703_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12715_cast_fp16 = mul(x = squeeze_417_cast_fp16, y = const_1703_promoted_to_fp16)[name = tensor<string, []>("op_12715_cast_fp16")];
            tensor<int32, []> x_1421_axis_0 = const()[name = tensor<string, []>("x_1421_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1421_cast_fp16 = stack(axis = x_1421_axis_0, values = (var_12715_cast_fp16, squeeze_416_cast_fp16))[name = tensor<string, []>("x_1421_cast_fp16")];
            tensor<int32, [4]> var_12721 = const()[name = tensor<string, []>("op_12721"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12722_cast_fp16 = reshape(shape = var_12721, x = x_1421_cast_fp16)[name = tensor<string, []>("op_12722_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12723_cast_fp16 = mul(x = var_12722_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12723_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_209_cast_fp16 = add(x = var_12702_cast_fp16, y = var_12723_cast_fp16)[name = tensor<string, []>("q_patches_209_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12725_cast_fp16 = mul(x = var_12700_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12725_cast_fp16")];
            tensor<int32, [5]> var_12733 = const()[name = tensor<string, []>("op_12733"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1425_cast_fp16 = reshape(shape = var_12733, x = var_12700_cast_fp16)[name = tensor<string, []>("x_1425_cast_fp16")];
            tensor<int32, [2]> var_12735_split_sizes_0 = const()[name = tensor<string, []>("op_12735_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12735_axis_0 = const()[name = tensor<string, []>("op_12735_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12735_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12735_cast_fp16_1 = split(axis = var_12735_axis_0, split_sizes = var_12735_split_sizes_0, x = x_1425_cast_fp16)[name = tensor<string, []>("op_12735_cast_fp16")];
            tensor<int32, [1]> squeeze_418_axes_0 = const()[name = tensor<string, []>("squeeze_418_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_418_cast_fp16 = squeeze(axes = squeeze_418_axes_0, x = var_12735_cast_fp16_0)[name = tensor<string, []>("squeeze_418_cast_fp16")];
            tensor<int32, [1]> squeeze_419_axes_0 = const()[name = tensor<string, []>("squeeze_419_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_419_cast_fp16 = squeeze(axes = squeeze_419_axes_0, x = var_12735_cast_fp16_1)[name = tensor<string, []>("squeeze_419_cast_fp16")];
            tensor<fp16, []> const_1711_promoted_to_fp16 = const()[name = tensor<string, []>("const_1711_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12738_cast_fp16 = mul(x = squeeze_419_cast_fp16, y = const_1711_promoted_to_fp16)[name = tensor<string, []>("op_12738_cast_fp16")];
            tensor<int32, []> x_1427_axis_0 = const()[name = tensor<string, []>("x_1427_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1427_cast_fp16 = stack(axis = x_1427_axis_0, values = (var_12738_cast_fp16, squeeze_418_cast_fp16))[name = tensor<string, []>("x_1427_cast_fp16")];
            tensor<int32, [4]> var_12744 = const()[name = tensor<string, []>("op_12744"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12745_cast_fp16 = reshape(shape = var_12744, x = x_1427_cast_fp16)[name = tensor<string, []>("op_12745_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12746_cast_fp16 = mul(x = var_12745_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12746_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_209_cast_fp16 = add(x = var_12725_cast_fp16, y = var_12746_cast_fp16)[name = tensor<string, []>("k_patches_209_cast_fp16")];
            tensor<bool, []> var_12749_interleave_0 = const()[name = tensor<string, []>("op_12749_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12749_cast_fp16 = concat(axis = var_20, interleave = var_12749_interleave_0, values = (var_12688_cast_fp16, q_patches_209_cast_fp16))[name = tensor<string, []>("op_12749_cast_fp16")];
            tensor<bool, []> var_12752_interleave_0 = const()[name = tensor<string, []>("op_12752_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12752_cast_fp16 = concat(axis = var_20, interleave = var_12752_interleave_0, values = (var_12692_cast_fp16, k_patches_209_cast_fp16))[name = tensor<string, []>("op_12752_cast_fp16")];
            tensor<fp16, []> var_12754_to_fp16 = const()[name = tensor<string, []>("op_12754_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1049_cast_fp16 = mul(x = var_12749_cast_fp16, y = var_12754_to_fp16)[name = tensor<string, []>("q_1049_cast_fp16")];
            tensor<bool, []> attn_417_transpose_x_1 = const()[name = tensor<string, []>("attn_417_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_417_transpose_y_1 = const()[name = tensor<string, []>("attn_417_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_417_cast_fp16 = matmul(transpose_x = attn_417_transpose_x_1, transpose_y = attn_417_transpose_y_1, x = q_1049_cast_fp16, y = var_12752_cast_fp16)[name = tensor<string, []>("attn_417_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_419_cast_fp16 = softmax(axis = var_21, x = attn_417_cast_fp16)[name = tensor<string, []>("attn_419_cast_fp16")];
            tensor<bool, []> var_12759_transpose_x_0 = const()[name = tensor<string, []>("op_12759_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12759_transpose_y_0 = const()[name = tensor<string, []>("op_12759_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12759_cast_fp16 = matmul(transpose_x = var_12759_transpose_x_0, transpose_y = var_12759_transpose_y_0, x = attn_419_cast_fp16, y = v_629_cast_fp16)[name = tensor<string, []>("op_12759_cast_fp16")];
            tensor<int32, [1]> var_12760_axes_0 = const()[name = tensor<string, []>("op_12760_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12760_cast_fp16 = squeeze(axes = var_12760_axes_0, x = var_12759_cast_fp16)[name = tensor<string, []>("op_12760_cast_fp16")];
            tensor<string, []> var_12766_pad_type_0 = const()[name = tensor<string, []>("op_12766_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12766_strides_0 = const()[name = tensor<string, []>("op_12766_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12766_pad_0 = const()[name = tensor<string, []>("op_12766_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12766_dilations_0 = const()[name = tensor<string, []>("op_12766_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12766_groups_0 = const()[name = tensor<string, []>("op_12766_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117687296)))];
            tensor<fp16, [1, 64, 1, 197]> var_12766_cast_fp16 = conv(dilations = var_12766_dilations_0, groups = var_12766_groups_0, pad = var_12766_pad_0, pad_type = var_12766_pad_type_0, strides = var_12766_strides_0, weight = model_blocks_8_attn_q_projs_9_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12766_cast_fp16")];
            tensor<int32, [1]> var_12767_axes_0 = const()[name = tensor<string, []>("op_12767_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12767_cast_fp16 = squeeze(axes = var_12767_axes_0, x = var_12766_cast_fp16)[name = tensor<string, []>("op_12767_cast_fp16")];
            tensor<int32, [3]> q_1051_perm_0 = const()[name = tensor<string, []>("q_1051_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12774_pad_type_0 = const()[name = tensor<string, []>("op_12774_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12774_strides_0 = const()[name = tensor<string, []>("op_12774_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12774_pad_0 = const()[name = tensor<string, []>("op_12774_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12774_dilations_0 = const()[name = tensor<string, []>("op_12774_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12774_groups_0 = const()[name = tensor<string, []>("op_12774_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117785664)))];
            tensor<fp16, [1, 64, 1, 197]> var_12774_cast_fp16 = conv(dilations = var_12774_dilations_0, groups = var_12774_groups_0, pad = var_12774_pad_0, pad_type = var_12774_pad_type_0, strides = var_12774_strides_0, weight = model_blocks_8_attn_k_projs_9_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12774_cast_fp16")];
            tensor<int32, [1]> var_12775_axes_0 = const()[name = tensor<string, []>("op_12775_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12775_cast_fp16 = squeeze(axes = var_12775_axes_0, x = var_12774_cast_fp16)[name = tensor<string, []>("op_12775_cast_fp16")];
            tensor<int32, [3]> k_631_perm_0 = const()[name = tensor<string, []>("k_631_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12782_pad_type_0 = const()[name = tensor<string, []>("op_12782_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12782_strides_0 = const()[name = tensor<string, []>("op_12782_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12782_pad_0 = const()[name = tensor<string, []>("op_12782_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12782_dilations_0 = const()[name = tensor<string, []>("op_12782_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12782_groups_0 = const()[name = tensor<string, []>("op_12782_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117884032)))];
            tensor<fp16, [1, 64, 1, 197]> var_12782_cast_fp16 = conv(dilations = var_12782_dilations_0, groups = var_12782_groups_0, pad = var_12782_pad_0, pad_type = var_12782_pad_type_0, strides = var_12782_strides_0, weight = model_blocks_8_attn_v_projs_9_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12782_cast_fp16")];
            tensor<int32, [1]> var_12783_axes_0 = const()[name = tensor<string, []>("op_12783_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12783_cast_fp16 = squeeze(axes = var_12783_axes_0, x = var_12782_cast_fp16)[name = tensor<string, []>("op_12783_cast_fp16")];
            tensor<int32, [3]> v_631_perm_0 = const()[name = tensor<string, []>("v_631_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117982400)))];
            tensor<fp16, [1, 197, 64]> q_1051_cast_fp16 = transpose(perm = q_1051_perm_0, x = var_12767_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 197, 64]> q_1053_cast_fp16 = add(x = q_1051_cast_fp16, y = model_blocks_8_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_1053_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117982592)))];
            tensor<fp16, [1, 197, 64]> v_631_cast_fp16 = transpose(perm = v_631_perm_0, x = var_12783_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1, 197, 64]> v_633_cast_fp16 = add(x = v_631_cast_fp16, y = model_blocks_8_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_633_cast_fp16")];
            tensor<int32, [1]> q_1055_axes_0 = const()[name = tensor<string, []>("q_1055_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1055_cast_fp16 = expand_dims(axes = q_1055_axes_0, x = q_1053_cast_fp16)[name = tensor<string, []>("q_1055_cast_fp16")];
            tensor<int32, [1]> k_633_axes_0 = const()[name = tensor<string, []>("k_633_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_631_cast_fp16 = transpose(perm = k_631_perm_0, x = var_12775_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [1, 1, 197, 64]> k_633_cast_fp16 = expand_dims(axes = k_633_axes_0, x = k_631_cast_fp16)[name = tensor<string, []>("k_633_cast_fp16")];
            tensor<int32, [1]> v_635_axes_0 = const()[name = tensor<string, []>("v_635_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_635_cast_fp16 = expand_dims(axes = v_635_axes_0, x = v_633_cast_fp16)[name = tensor<string, []>("v_635_cast_fp16")];
            tensor<int32, [4]> var_12792_begin_0 = const()[name = tensor<string, []>("op_12792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12792_end_0 = const()[name = tensor<string, []>("op_12792_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12792_end_mask_0 = const()[name = tensor<string, []>("op_12792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12792_cast_fp16 = slice_by_index(begin = var_12792_begin_0, end = var_12792_end_0, end_mask = var_12792_end_mask_0, x = q_1055_cast_fp16)[name = tensor<string, []>("op_12792_cast_fp16")];
            tensor<int32, [4]> var_12796_begin_0 = const()[name = tensor<string, []>("op_12796_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12796_end_0 = const()[name = tensor<string, []>("op_12796_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12796_end_mask_0 = const()[name = tensor<string, []>("op_12796_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12796_cast_fp16 = slice_by_index(begin = var_12796_begin_0, end = var_12796_end_0, end_mask = var_12796_end_mask_0, x = k_633_cast_fp16)[name = tensor<string, []>("op_12796_cast_fp16")];
            tensor<int32, [4]> var_12800_begin_0 = const()[name = tensor<string, []>("op_12800_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12800_end_0 = const()[name = tensor<string, []>("op_12800_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12800_end_mask_0 = const()[name = tensor<string, []>("op_12800_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12800_cast_fp16 = slice_by_index(begin = var_12800_begin_0, end = var_12800_end_0, end_mask = var_12800_end_mask_0, x = q_1055_cast_fp16)[name = tensor<string, []>("op_12800_cast_fp16")];
            tensor<int32, [4]> var_12804_begin_0 = const()[name = tensor<string, []>("op_12804_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12804_end_0 = const()[name = tensor<string, []>("op_12804_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12804_end_mask_0 = const()[name = tensor<string, []>("op_12804_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12804_cast_fp16 = slice_by_index(begin = var_12804_begin_0, end = var_12804_end_0, end_mask = var_12804_end_mask_0, x = k_633_cast_fp16)[name = tensor<string, []>("op_12804_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12806_cast_fp16 = mul(x = var_12800_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12806_cast_fp16")];
            tensor<int32, [5]> var_12814 = const()[name = tensor<string, []>("op_12814"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1431_cast_fp16 = reshape(shape = var_12814, x = var_12800_cast_fp16)[name = tensor<string, []>("x_1431_cast_fp16")];
            tensor<int32, [2]> var_12816_split_sizes_0 = const()[name = tensor<string, []>("op_12816_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12816_axis_0 = const()[name = tensor<string, []>("op_12816_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12816_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12816_cast_fp16_1 = split(axis = var_12816_axis_0, split_sizes = var_12816_split_sizes_0, x = x_1431_cast_fp16)[name = tensor<string, []>("op_12816_cast_fp16")];
            tensor<int32, [1]> squeeze_420_axes_0 = const()[name = tensor<string, []>("squeeze_420_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_420_cast_fp16 = squeeze(axes = squeeze_420_axes_0, x = var_12816_cast_fp16_0)[name = tensor<string, []>("squeeze_420_cast_fp16")];
            tensor<int32, [1]> squeeze_421_axes_0 = const()[name = tensor<string, []>("squeeze_421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_421_cast_fp16 = squeeze(axes = squeeze_421_axes_0, x = var_12816_cast_fp16_1)[name = tensor<string, []>("squeeze_421_cast_fp16")];
            tensor<fp16, []> const_1719_promoted_to_fp16 = const()[name = tensor<string, []>("const_1719_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12819_cast_fp16 = mul(x = squeeze_421_cast_fp16, y = const_1719_promoted_to_fp16)[name = tensor<string, []>("op_12819_cast_fp16")];
            tensor<int32, []> x_1433_axis_0 = const()[name = tensor<string, []>("x_1433_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1433_cast_fp16 = stack(axis = x_1433_axis_0, values = (var_12819_cast_fp16, squeeze_420_cast_fp16))[name = tensor<string, []>("x_1433_cast_fp16")];
            tensor<int32, [4]> var_12825 = const()[name = tensor<string, []>("op_12825"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12826_cast_fp16 = reshape(shape = var_12825, x = x_1433_cast_fp16)[name = tensor<string, []>("op_12826_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12827_cast_fp16 = mul(x = var_12826_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12827_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_211_cast_fp16 = add(x = var_12806_cast_fp16, y = var_12827_cast_fp16)[name = tensor<string, []>("q_patches_211_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12829_cast_fp16 = mul(x = var_12804_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12829_cast_fp16")];
            tensor<int32, [5]> var_12837 = const()[name = tensor<string, []>("op_12837"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1437_cast_fp16 = reshape(shape = var_12837, x = var_12804_cast_fp16)[name = tensor<string, []>("x_1437_cast_fp16")];
            tensor<int32, [2]> var_12839_split_sizes_0 = const()[name = tensor<string, []>("op_12839_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12839_axis_0 = const()[name = tensor<string, []>("op_12839_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12839_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12839_cast_fp16_1 = split(axis = var_12839_axis_0, split_sizes = var_12839_split_sizes_0, x = x_1437_cast_fp16)[name = tensor<string, []>("op_12839_cast_fp16")];
            tensor<int32, [1]> squeeze_422_axes_0 = const()[name = tensor<string, []>("squeeze_422_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_422_cast_fp16 = squeeze(axes = squeeze_422_axes_0, x = var_12839_cast_fp16_0)[name = tensor<string, []>("squeeze_422_cast_fp16")];
            tensor<int32, [1]> squeeze_423_axes_0 = const()[name = tensor<string, []>("squeeze_423_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_423_cast_fp16 = squeeze(axes = squeeze_423_axes_0, x = var_12839_cast_fp16_1)[name = tensor<string, []>("squeeze_423_cast_fp16")];
            tensor<fp16, []> const_1727_promoted_to_fp16 = const()[name = tensor<string, []>("const_1727_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12842_cast_fp16 = mul(x = squeeze_423_cast_fp16, y = const_1727_promoted_to_fp16)[name = tensor<string, []>("op_12842_cast_fp16")];
            tensor<int32, []> x_1439_axis_0 = const()[name = tensor<string, []>("x_1439_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1439_cast_fp16 = stack(axis = x_1439_axis_0, values = (var_12842_cast_fp16, squeeze_422_cast_fp16))[name = tensor<string, []>("x_1439_cast_fp16")];
            tensor<int32, [4]> var_12848 = const()[name = tensor<string, []>("op_12848"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12849_cast_fp16 = reshape(shape = var_12848, x = x_1439_cast_fp16)[name = tensor<string, []>("op_12849_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12850_cast_fp16 = mul(x = var_12849_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12850_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_211_cast_fp16 = add(x = var_12829_cast_fp16, y = var_12850_cast_fp16)[name = tensor<string, []>("k_patches_211_cast_fp16")];
            tensor<bool, []> var_12853_interleave_0 = const()[name = tensor<string, []>("op_12853_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12853_cast_fp16 = concat(axis = var_20, interleave = var_12853_interleave_0, values = (var_12792_cast_fp16, q_patches_211_cast_fp16))[name = tensor<string, []>("op_12853_cast_fp16")];
            tensor<bool, []> var_12856_interleave_0 = const()[name = tensor<string, []>("op_12856_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12856_cast_fp16 = concat(axis = var_20, interleave = var_12856_interleave_0, values = (var_12796_cast_fp16, k_patches_211_cast_fp16))[name = tensor<string, []>("op_12856_cast_fp16")];
            tensor<fp16, []> var_12858_to_fp16 = const()[name = tensor<string, []>("op_12858_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1059_cast_fp16 = mul(x = var_12853_cast_fp16, y = var_12858_to_fp16)[name = tensor<string, []>("q_1059_cast_fp16")];
            tensor<bool, []> attn_421_transpose_x_1 = const()[name = tensor<string, []>("attn_421_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_421_transpose_y_1 = const()[name = tensor<string, []>("attn_421_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_421_cast_fp16 = matmul(transpose_x = attn_421_transpose_x_1, transpose_y = attn_421_transpose_y_1, x = q_1059_cast_fp16, y = var_12856_cast_fp16)[name = tensor<string, []>("attn_421_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_423_cast_fp16 = softmax(axis = var_21, x = attn_421_cast_fp16)[name = tensor<string, []>("attn_423_cast_fp16")];
            tensor<bool, []> var_12863_transpose_x_0 = const()[name = tensor<string, []>("op_12863_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12863_transpose_y_0 = const()[name = tensor<string, []>("op_12863_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12863_cast_fp16 = matmul(transpose_x = var_12863_transpose_x_0, transpose_y = var_12863_transpose_y_0, x = attn_423_cast_fp16, y = v_635_cast_fp16)[name = tensor<string, []>("op_12863_cast_fp16")];
            tensor<int32, [1]> var_12864_axes_0 = const()[name = tensor<string, []>("op_12864_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12864_cast_fp16 = squeeze(axes = var_12864_axes_0, x = var_12863_cast_fp16)[name = tensor<string, []>("op_12864_cast_fp16")];
            tensor<string, []> var_12870_pad_type_0 = const()[name = tensor<string, []>("op_12870_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12870_strides_0 = const()[name = tensor<string, []>("op_12870_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12870_pad_0 = const()[name = tensor<string, []>("op_12870_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12870_dilations_0 = const()[name = tensor<string, []>("op_12870_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12870_groups_0 = const()[name = tensor<string, []>("op_12870_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117982784)))];
            tensor<fp16, [1, 64, 1, 197]> var_12870_cast_fp16 = conv(dilations = var_12870_dilations_0, groups = var_12870_groups_0, pad = var_12870_pad_0, pad_type = var_12870_pad_type_0, strides = var_12870_strides_0, weight = model_blocks_8_attn_q_projs_10_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12870_cast_fp16")];
            tensor<int32, [1]> var_12871_axes_0 = const()[name = tensor<string, []>("op_12871_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12871_cast_fp16 = squeeze(axes = var_12871_axes_0, x = var_12870_cast_fp16)[name = tensor<string, []>("op_12871_cast_fp16")];
            tensor<int32, [3]> q_1061_perm_0 = const()[name = tensor<string, []>("q_1061_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12878_pad_type_0 = const()[name = tensor<string, []>("op_12878_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12878_strides_0 = const()[name = tensor<string, []>("op_12878_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12878_pad_0 = const()[name = tensor<string, []>("op_12878_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12878_dilations_0 = const()[name = tensor<string, []>("op_12878_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12878_groups_0 = const()[name = tensor<string, []>("op_12878_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118081152)))];
            tensor<fp16, [1, 64, 1, 197]> var_12878_cast_fp16 = conv(dilations = var_12878_dilations_0, groups = var_12878_groups_0, pad = var_12878_pad_0, pad_type = var_12878_pad_type_0, strides = var_12878_strides_0, weight = model_blocks_8_attn_k_projs_10_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12878_cast_fp16")];
            tensor<int32, [1]> var_12879_axes_0 = const()[name = tensor<string, []>("op_12879_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12879_cast_fp16 = squeeze(axes = var_12879_axes_0, x = var_12878_cast_fp16)[name = tensor<string, []>("op_12879_cast_fp16")];
            tensor<int32, [3]> k_637_perm_0 = const()[name = tensor<string, []>("k_637_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12886_pad_type_0 = const()[name = tensor<string, []>("op_12886_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12886_strides_0 = const()[name = tensor<string, []>("op_12886_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12886_pad_0 = const()[name = tensor<string, []>("op_12886_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12886_dilations_0 = const()[name = tensor<string, []>("op_12886_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12886_groups_0 = const()[name = tensor<string, []>("op_12886_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118179520)))];
            tensor<fp16, [1, 64, 1, 197]> var_12886_cast_fp16 = conv(dilations = var_12886_dilations_0, groups = var_12886_groups_0, pad = var_12886_pad_0, pad_type = var_12886_pad_type_0, strides = var_12886_strides_0, weight = model_blocks_8_attn_v_projs_10_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12886_cast_fp16")];
            tensor<int32, [1]> var_12887_axes_0 = const()[name = tensor<string, []>("op_12887_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12887_cast_fp16 = squeeze(axes = var_12887_axes_0, x = var_12886_cast_fp16)[name = tensor<string, []>("op_12887_cast_fp16")];
            tensor<int32, [3]> v_637_perm_0 = const()[name = tensor<string, []>("v_637_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118277888)))];
            tensor<fp16, [1, 197, 64]> q_1061_cast_fp16 = transpose(perm = q_1061_perm_0, x = var_12871_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [1, 197, 64]> q_1063_cast_fp16 = add(x = q_1061_cast_fp16, y = model_blocks_8_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_1063_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118278080)))];
            tensor<fp16, [1, 197, 64]> v_637_cast_fp16 = transpose(perm = v_637_perm_0, x = var_12887_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [1, 197, 64]> v_639_cast_fp16 = add(x = v_637_cast_fp16, y = model_blocks_8_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_639_cast_fp16")];
            tensor<int32, [1]> q_1065_axes_0 = const()[name = tensor<string, []>("q_1065_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1065_cast_fp16 = expand_dims(axes = q_1065_axes_0, x = q_1063_cast_fp16)[name = tensor<string, []>("q_1065_cast_fp16")];
            tensor<int32, [1]> k_639_axes_0 = const()[name = tensor<string, []>("k_639_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_637_cast_fp16 = transpose(perm = k_637_perm_0, x = var_12879_cast_fp16)[name = tensor<string, []>("transpose_139")];
            tensor<fp16, [1, 1, 197, 64]> k_639_cast_fp16 = expand_dims(axes = k_639_axes_0, x = k_637_cast_fp16)[name = tensor<string, []>("k_639_cast_fp16")];
            tensor<int32, [1]> v_641_axes_0 = const()[name = tensor<string, []>("v_641_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_641_cast_fp16 = expand_dims(axes = v_641_axes_0, x = v_639_cast_fp16)[name = tensor<string, []>("v_641_cast_fp16")];
            tensor<int32, [4]> var_12896_begin_0 = const()[name = tensor<string, []>("op_12896_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12896_end_0 = const()[name = tensor<string, []>("op_12896_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12896_end_mask_0 = const()[name = tensor<string, []>("op_12896_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12896_cast_fp16 = slice_by_index(begin = var_12896_begin_0, end = var_12896_end_0, end_mask = var_12896_end_mask_0, x = q_1065_cast_fp16)[name = tensor<string, []>("op_12896_cast_fp16")];
            tensor<int32, [4]> var_12900_begin_0 = const()[name = tensor<string, []>("op_12900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12900_end_0 = const()[name = tensor<string, []>("op_12900_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_12900_end_mask_0 = const()[name = tensor<string, []>("op_12900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_12900_cast_fp16 = slice_by_index(begin = var_12900_begin_0, end = var_12900_end_0, end_mask = var_12900_end_mask_0, x = k_639_cast_fp16)[name = tensor<string, []>("op_12900_cast_fp16")];
            tensor<int32, [4]> var_12904_begin_0 = const()[name = tensor<string, []>("op_12904_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12904_end_0 = const()[name = tensor<string, []>("op_12904_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12904_end_mask_0 = const()[name = tensor<string, []>("op_12904_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12904_cast_fp16 = slice_by_index(begin = var_12904_begin_0, end = var_12904_end_0, end_mask = var_12904_end_mask_0, x = q_1065_cast_fp16)[name = tensor<string, []>("op_12904_cast_fp16")];
            tensor<int32, [4]> var_12908_begin_0 = const()[name = tensor<string, []>("op_12908_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12908_end_0 = const()[name = tensor<string, []>("op_12908_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_12908_end_mask_0 = const()[name = tensor<string, []>("op_12908_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_12908_cast_fp16 = slice_by_index(begin = var_12908_begin_0, end = var_12908_end_0, end_mask = var_12908_end_mask_0, x = k_639_cast_fp16)[name = tensor<string, []>("op_12908_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12910_cast_fp16 = mul(x = var_12904_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12910_cast_fp16")];
            tensor<int32, [5]> var_12918 = const()[name = tensor<string, []>("op_12918"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1443_cast_fp16 = reshape(shape = var_12918, x = var_12904_cast_fp16)[name = tensor<string, []>("x_1443_cast_fp16")];
            tensor<int32, [2]> var_12920_split_sizes_0 = const()[name = tensor<string, []>("op_12920_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12920_axis_0 = const()[name = tensor<string, []>("op_12920_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12920_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12920_cast_fp16_1 = split(axis = var_12920_axis_0, split_sizes = var_12920_split_sizes_0, x = x_1443_cast_fp16)[name = tensor<string, []>("op_12920_cast_fp16")];
            tensor<int32, [1]> squeeze_424_axes_0 = const()[name = tensor<string, []>("squeeze_424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_424_cast_fp16 = squeeze(axes = squeeze_424_axes_0, x = var_12920_cast_fp16_0)[name = tensor<string, []>("squeeze_424_cast_fp16")];
            tensor<int32, [1]> squeeze_425_axes_0 = const()[name = tensor<string, []>("squeeze_425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_425_cast_fp16 = squeeze(axes = squeeze_425_axes_0, x = var_12920_cast_fp16_1)[name = tensor<string, []>("squeeze_425_cast_fp16")];
            tensor<fp16, []> const_1735_promoted_to_fp16 = const()[name = tensor<string, []>("const_1735_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12923_cast_fp16 = mul(x = squeeze_425_cast_fp16, y = const_1735_promoted_to_fp16)[name = tensor<string, []>("op_12923_cast_fp16")];
            tensor<int32, []> x_1445_axis_0 = const()[name = tensor<string, []>("x_1445_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1445_cast_fp16 = stack(axis = x_1445_axis_0, values = (var_12923_cast_fp16, squeeze_424_cast_fp16))[name = tensor<string, []>("x_1445_cast_fp16")];
            tensor<int32, [4]> var_12929 = const()[name = tensor<string, []>("op_12929"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12930_cast_fp16 = reshape(shape = var_12929, x = x_1445_cast_fp16)[name = tensor<string, []>("op_12930_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12931_cast_fp16 = mul(x = var_12930_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12931_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_213_cast_fp16 = add(x = var_12910_cast_fp16, y = var_12931_cast_fp16)[name = tensor<string, []>("q_patches_213_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12933_cast_fp16 = mul(x = var_12908_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_12933_cast_fp16")];
            tensor<int32, [5]> var_12941 = const()[name = tensor<string, []>("op_12941"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1449_cast_fp16 = reshape(shape = var_12941, x = var_12908_cast_fp16)[name = tensor<string, []>("x_1449_cast_fp16")];
            tensor<int32, [2]> var_12943_split_sizes_0 = const()[name = tensor<string, []>("op_12943_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12943_axis_0 = const()[name = tensor<string, []>("op_12943_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_12943_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_12943_cast_fp16_1 = split(axis = var_12943_axis_0, split_sizes = var_12943_split_sizes_0, x = x_1449_cast_fp16)[name = tensor<string, []>("op_12943_cast_fp16")];
            tensor<int32, [1]> squeeze_426_axes_0 = const()[name = tensor<string, []>("squeeze_426_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_426_cast_fp16 = squeeze(axes = squeeze_426_axes_0, x = var_12943_cast_fp16_0)[name = tensor<string, []>("squeeze_426_cast_fp16")];
            tensor<int32, [1]> squeeze_427_axes_0 = const()[name = tensor<string, []>("squeeze_427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_427_cast_fp16 = squeeze(axes = squeeze_427_axes_0, x = var_12943_cast_fp16_1)[name = tensor<string, []>("squeeze_427_cast_fp16")];
            tensor<fp16, []> const_1743_promoted_to_fp16 = const()[name = tensor<string, []>("const_1743_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_12946_cast_fp16 = mul(x = squeeze_427_cast_fp16, y = const_1743_promoted_to_fp16)[name = tensor<string, []>("op_12946_cast_fp16")];
            tensor<int32, []> x_1451_axis_0 = const()[name = tensor<string, []>("x_1451_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1451_cast_fp16 = stack(axis = x_1451_axis_0, values = (var_12946_cast_fp16, squeeze_426_cast_fp16))[name = tensor<string, []>("x_1451_cast_fp16")];
            tensor<int32, [4]> var_12952 = const()[name = tensor<string, []>("op_12952"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_12953_cast_fp16 = reshape(shape = var_12952, x = x_1451_cast_fp16)[name = tensor<string, []>("op_12953_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_12954_cast_fp16 = mul(x = var_12953_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_12954_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_213_cast_fp16 = add(x = var_12933_cast_fp16, y = var_12954_cast_fp16)[name = tensor<string, []>("k_patches_213_cast_fp16")];
            tensor<bool, []> var_12957_interleave_0 = const()[name = tensor<string, []>("op_12957_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12957_cast_fp16 = concat(axis = var_20, interleave = var_12957_interleave_0, values = (var_12896_cast_fp16, q_patches_213_cast_fp16))[name = tensor<string, []>("op_12957_cast_fp16")];
            tensor<bool, []> var_12960_interleave_0 = const()[name = tensor<string, []>("op_12960_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12960_cast_fp16 = concat(axis = var_20, interleave = var_12960_interleave_0, values = (var_12900_cast_fp16, k_patches_213_cast_fp16))[name = tensor<string, []>("op_12960_cast_fp16")];
            tensor<fp16, []> var_12962_to_fp16 = const()[name = tensor<string, []>("op_12962_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1069_cast_fp16 = mul(x = var_12957_cast_fp16, y = var_12962_to_fp16)[name = tensor<string, []>("q_1069_cast_fp16")];
            tensor<bool, []> attn_425_transpose_x_1 = const()[name = tensor<string, []>("attn_425_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_425_transpose_y_1 = const()[name = tensor<string, []>("attn_425_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_425_cast_fp16 = matmul(transpose_x = attn_425_transpose_x_1, transpose_y = attn_425_transpose_y_1, x = q_1069_cast_fp16, y = var_12960_cast_fp16)[name = tensor<string, []>("attn_425_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_427_cast_fp16 = softmax(axis = var_21, x = attn_425_cast_fp16)[name = tensor<string, []>("attn_427_cast_fp16")];
            tensor<bool, []> var_12967_transpose_x_0 = const()[name = tensor<string, []>("op_12967_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_12967_transpose_y_0 = const()[name = tensor<string, []>("op_12967_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_12967_cast_fp16 = matmul(transpose_x = var_12967_transpose_x_0, transpose_y = var_12967_transpose_y_0, x = attn_427_cast_fp16, y = v_641_cast_fp16)[name = tensor<string, []>("op_12967_cast_fp16")];
            tensor<int32, [1]> var_12968_axes_0 = const()[name = tensor<string, []>("op_12968_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_12968_cast_fp16 = squeeze(axes = var_12968_axes_0, x = var_12967_cast_fp16)[name = tensor<string, []>("op_12968_cast_fp16")];
            tensor<string, []> var_12974_pad_type_0 = const()[name = tensor<string, []>("op_12974_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12974_strides_0 = const()[name = tensor<string, []>("op_12974_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12974_pad_0 = const()[name = tensor<string, []>("op_12974_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12974_dilations_0 = const()[name = tensor<string, []>("op_12974_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12974_groups_0 = const()[name = tensor<string, []>("op_12974_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118278272)))];
            tensor<fp16, [1, 64, 1, 197]> var_12974_cast_fp16 = conv(dilations = var_12974_dilations_0, groups = var_12974_groups_0, pad = var_12974_pad_0, pad_type = var_12974_pad_type_0, strides = var_12974_strides_0, weight = model_blocks_8_attn_q_projs_11_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12974_cast_fp16")];
            tensor<int32, [1]> var_12975_axes_0 = const()[name = tensor<string, []>("op_12975_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12975_cast_fp16 = squeeze(axes = var_12975_axes_0, x = var_12974_cast_fp16)[name = tensor<string, []>("op_12975_cast_fp16")];
            tensor<int32, [3]> q_1071_perm_0 = const()[name = tensor<string, []>("q_1071_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12982_pad_type_0 = const()[name = tensor<string, []>("op_12982_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12982_strides_0 = const()[name = tensor<string, []>("op_12982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12982_pad_0 = const()[name = tensor<string, []>("op_12982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12982_dilations_0 = const()[name = tensor<string, []>("op_12982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12982_groups_0 = const()[name = tensor<string, []>("op_12982_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118376640)))];
            tensor<fp16, [1, 64, 1, 197]> var_12982_cast_fp16 = conv(dilations = var_12982_dilations_0, groups = var_12982_groups_0, pad = var_12982_pad_0, pad_type = var_12982_pad_type_0, strides = var_12982_strides_0, weight = model_blocks_8_attn_k_projs_11_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12982_cast_fp16")];
            tensor<int32, [1]> var_12983_axes_0 = const()[name = tensor<string, []>("op_12983_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12983_cast_fp16 = squeeze(axes = var_12983_axes_0, x = var_12982_cast_fp16)[name = tensor<string, []>("op_12983_cast_fp16")];
            tensor<int32, [3]> k_643_perm_0 = const()[name = tensor<string, []>("k_643_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_12990_pad_type_0 = const()[name = tensor<string, []>("op_12990_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_12990_strides_0 = const()[name = tensor<string, []>("op_12990_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_12990_pad_0 = const()[name = tensor<string, []>("op_12990_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_12990_dilations_0 = const()[name = tensor<string, []>("op_12990_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_12990_groups_0 = const()[name = tensor<string, []>("op_12990_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_8_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118475008)))];
            tensor<fp16, [1, 64, 1, 197]> var_12990_cast_fp16 = conv(dilations = var_12990_dilations_0, groups = var_12990_groups_0, pad = var_12990_pad_0, pad_type = var_12990_pad_type_0, strides = var_12990_strides_0, weight = model_blocks_8_attn_v_projs_11_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("op_12990_cast_fp16")];
            tensor<int32, [1]> var_12991_axes_0 = const()[name = tensor<string, []>("op_12991_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_12991_cast_fp16 = squeeze(axes = var_12991_axes_0, x = var_12990_cast_fp16)[name = tensor<string, []>("op_12991_cast_fp16")];
            tensor<int32, [3]> v_643_perm_0 = const()[name = tensor<string, []>("v_643_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_8_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118573376)))];
            tensor<fp16, [1, 197, 64]> q_1071_cast_fp16 = transpose(perm = q_1071_perm_0, x = var_12975_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [1, 197, 64]> q_1073_cast_fp16 = add(x = q_1071_cast_fp16, y = model_blocks_8_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_1073_cast_fp16")];
            tensor<fp16, [64]> model_blocks_8_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118573568)))];
            tensor<fp16, [1, 197, 64]> v_643_cast_fp16 = transpose(perm = v_643_perm_0, x = var_12991_cast_fp16)[name = tensor<string, []>("transpose_135")];
            tensor<fp16, [1, 197, 64]> v_645_cast_fp16 = add(x = v_643_cast_fp16, y = model_blocks_8_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_645_cast_fp16")];
            tensor<int32, [1]> q_1075_axes_0 = const()[name = tensor<string, []>("q_1075_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1075_cast_fp16 = expand_dims(axes = q_1075_axes_0, x = q_1073_cast_fp16)[name = tensor<string, []>("q_1075_cast_fp16")];
            tensor<int32, [1]> k_645_axes_0 = const()[name = tensor<string, []>("k_645_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_643_cast_fp16 = transpose(perm = k_643_perm_0, x = var_12983_cast_fp16)[name = tensor<string, []>("transpose_136")];
            tensor<fp16, [1, 1, 197, 64]> k_645_cast_fp16 = expand_dims(axes = k_645_axes_0, x = k_643_cast_fp16)[name = tensor<string, []>("k_645_cast_fp16")];
            tensor<int32, [1]> v_647_axes_0 = const()[name = tensor<string, []>("v_647_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_647_cast_fp16 = expand_dims(axes = v_647_axes_0, x = v_645_cast_fp16)[name = tensor<string, []>("v_647_cast_fp16")];
            tensor<int32, [4]> var_13000_begin_0 = const()[name = tensor<string, []>("op_13000_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13000_end_0 = const()[name = tensor<string, []>("op_13000_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13000_end_mask_0 = const()[name = tensor<string, []>("op_13000_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13000_cast_fp16 = slice_by_index(begin = var_13000_begin_0, end = var_13000_end_0, end_mask = var_13000_end_mask_0, x = q_1075_cast_fp16)[name = tensor<string, []>("op_13000_cast_fp16")];
            tensor<int32, [4]> var_13004_begin_0 = const()[name = tensor<string, []>("op_13004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13004_end_0 = const()[name = tensor<string, []>("op_13004_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13004_end_mask_0 = const()[name = tensor<string, []>("op_13004_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13004_cast_fp16 = slice_by_index(begin = var_13004_begin_0, end = var_13004_end_0, end_mask = var_13004_end_mask_0, x = k_645_cast_fp16)[name = tensor<string, []>("op_13004_cast_fp16")];
            tensor<int32, [4]> var_13008_begin_0 = const()[name = tensor<string, []>("op_13008_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13008_end_0 = const()[name = tensor<string, []>("op_13008_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13008_end_mask_0 = const()[name = tensor<string, []>("op_13008_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13008_cast_fp16 = slice_by_index(begin = var_13008_begin_0, end = var_13008_end_0, end_mask = var_13008_end_mask_0, x = q_1075_cast_fp16)[name = tensor<string, []>("op_13008_cast_fp16")];
            tensor<int32, [4]> var_13012_begin_0 = const()[name = tensor<string, []>("op_13012_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13012_end_0 = const()[name = tensor<string, []>("op_13012_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13012_end_mask_0 = const()[name = tensor<string, []>("op_13012_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13012_cast_fp16 = slice_by_index(begin = var_13012_begin_0, end = var_13012_end_0, end_mask = var_13012_end_mask_0, x = k_645_cast_fp16)[name = tensor<string, []>("op_13012_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13014_cast_fp16 = mul(x = var_13008_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13014_cast_fp16")];
            tensor<int32, [5]> var_13022 = const()[name = tensor<string, []>("op_13022"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1455_cast_fp16 = reshape(shape = var_13022, x = var_13008_cast_fp16)[name = tensor<string, []>("x_1455_cast_fp16")];
            tensor<int32, [2]> var_13024_split_sizes_0 = const()[name = tensor<string, []>("op_13024_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13024_axis_0 = const()[name = tensor<string, []>("op_13024_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13024_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13024_cast_fp16_1 = split(axis = var_13024_axis_0, split_sizes = var_13024_split_sizes_0, x = x_1455_cast_fp16)[name = tensor<string, []>("op_13024_cast_fp16")];
            tensor<int32, [1]> squeeze_428_axes_0 = const()[name = tensor<string, []>("squeeze_428_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_428_cast_fp16 = squeeze(axes = squeeze_428_axes_0, x = var_13024_cast_fp16_0)[name = tensor<string, []>("squeeze_428_cast_fp16")];
            tensor<int32, [1]> squeeze_429_axes_0 = const()[name = tensor<string, []>("squeeze_429_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_429_cast_fp16 = squeeze(axes = squeeze_429_axes_0, x = var_13024_cast_fp16_1)[name = tensor<string, []>("squeeze_429_cast_fp16")];
            tensor<fp16, []> const_1751_promoted_to_fp16 = const()[name = tensor<string, []>("const_1751_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13027_cast_fp16 = mul(x = squeeze_429_cast_fp16, y = const_1751_promoted_to_fp16)[name = tensor<string, []>("op_13027_cast_fp16")];
            tensor<int32, []> x_1457_axis_0 = const()[name = tensor<string, []>("x_1457_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1457_cast_fp16 = stack(axis = x_1457_axis_0, values = (var_13027_cast_fp16, squeeze_428_cast_fp16))[name = tensor<string, []>("x_1457_cast_fp16")];
            tensor<int32, [4]> var_13033 = const()[name = tensor<string, []>("op_13033"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13034_cast_fp16 = reshape(shape = var_13033, x = x_1457_cast_fp16)[name = tensor<string, []>("op_13034_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13035_cast_fp16 = mul(x = var_13034_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13035_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_215_cast_fp16 = add(x = var_13014_cast_fp16, y = var_13035_cast_fp16)[name = tensor<string, []>("q_patches_215_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13037_cast_fp16 = mul(x = var_13012_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13037_cast_fp16")];
            tensor<int32, [5]> var_13045 = const()[name = tensor<string, []>("op_13045"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1461_cast_fp16 = reshape(shape = var_13045, x = var_13012_cast_fp16)[name = tensor<string, []>("x_1461_cast_fp16")];
            tensor<int32, [2]> var_13047_split_sizes_0 = const()[name = tensor<string, []>("op_13047_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13047_axis_0 = const()[name = tensor<string, []>("op_13047_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13047_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13047_cast_fp16_1 = split(axis = var_13047_axis_0, split_sizes = var_13047_split_sizes_0, x = x_1461_cast_fp16)[name = tensor<string, []>("op_13047_cast_fp16")];
            tensor<int32, [1]> squeeze_430_axes_0 = const()[name = tensor<string, []>("squeeze_430_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_430_cast_fp16 = squeeze(axes = squeeze_430_axes_0, x = var_13047_cast_fp16_0)[name = tensor<string, []>("squeeze_430_cast_fp16")];
            tensor<int32, [1]> squeeze_431_axes_0 = const()[name = tensor<string, []>("squeeze_431_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_431_cast_fp16 = squeeze(axes = squeeze_431_axes_0, x = var_13047_cast_fp16_1)[name = tensor<string, []>("squeeze_431_cast_fp16")];
            tensor<fp16, []> const_1759_promoted_to_fp16 = const()[name = tensor<string, []>("const_1759_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13050_cast_fp16 = mul(x = squeeze_431_cast_fp16, y = const_1759_promoted_to_fp16)[name = tensor<string, []>("op_13050_cast_fp16")];
            tensor<int32, []> x_1463_axis_0 = const()[name = tensor<string, []>("x_1463_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1463_cast_fp16 = stack(axis = x_1463_axis_0, values = (var_13050_cast_fp16, squeeze_430_cast_fp16))[name = tensor<string, []>("x_1463_cast_fp16")];
            tensor<int32, [4]> var_13056 = const()[name = tensor<string, []>("op_13056"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13057_cast_fp16 = reshape(shape = var_13056, x = x_1463_cast_fp16)[name = tensor<string, []>("op_13057_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13058_cast_fp16 = mul(x = var_13057_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13058_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_215_cast_fp16 = add(x = var_13037_cast_fp16, y = var_13058_cast_fp16)[name = tensor<string, []>("k_patches_215_cast_fp16")];
            tensor<bool, []> var_13061_interleave_0 = const()[name = tensor<string, []>("op_13061_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13061_cast_fp16 = concat(axis = var_20, interleave = var_13061_interleave_0, values = (var_13000_cast_fp16, q_patches_215_cast_fp16))[name = tensor<string, []>("op_13061_cast_fp16")];
            tensor<bool, []> var_13064_interleave_0 = const()[name = tensor<string, []>("op_13064_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13064_cast_fp16 = concat(axis = var_20, interleave = var_13064_interleave_0, values = (var_13004_cast_fp16, k_patches_215_cast_fp16))[name = tensor<string, []>("op_13064_cast_fp16")];
            tensor<fp16, []> var_13066_to_fp16 = const()[name = tensor<string, []>("op_13066_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1079_cast_fp16 = mul(x = var_13061_cast_fp16, y = var_13066_to_fp16)[name = tensor<string, []>("q_1079_cast_fp16")];
            tensor<bool, []> attn_429_transpose_x_1 = const()[name = tensor<string, []>("attn_429_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_429_transpose_y_1 = const()[name = tensor<string, []>("attn_429_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_429_cast_fp16 = matmul(transpose_x = attn_429_transpose_x_1, transpose_y = attn_429_transpose_y_1, x = q_1079_cast_fp16, y = var_13064_cast_fp16)[name = tensor<string, []>("attn_429_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_431_cast_fp16 = softmax(axis = var_21, x = attn_429_cast_fp16)[name = tensor<string, []>("attn_431_cast_fp16")];
            tensor<bool, []> var_13071_transpose_x_0 = const()[name = tensor<string, []>("op_13071_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13071_transpose_y_0 = const()[name = tensor<string, []>("op_13071_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13071_cast_fp16 = matmul(transpose_x = var_13071_transpose_x_0, transpose_y = var_13071_transpose_y_0, x = attn_431_cast_fp16, y = v_647_cast_fp16)[name = tensor<string, []>("op_13071_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_17_cast_fp16 = squeeze(axes = out_17_axes_0, x = var_13071_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<bool, []> input_149_interleave_0 = const()[name = tensor<string, []>("input_149_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_149_cast_fp16 = concat(axis = var_21, interleave = input_149_interleave_0, values = (var_11928_cast_fp16, var_12032_cast_fp16, var_12136_cast_fp16, var_12240_cast_fp16, var_12344_cast_fp16, var_12448_cast_fp16, var_12552_cast_fp16, var_12656_cast_fp16, var_12760_cast_fp16, var_12864_cast_fp16, var_12968_cast_fp16, out_17_cast_fp16))[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<int32, [1]> x_1465_axes_0 = const()[name = tensor<string, []>("x_1465_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_8_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118573760)))];
            tensor<fp16, [768]> model_blocks_8_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118575360)))];
            tensor<fp16, [1, 197, 768]> x_1465_cast_fp16 = layer_norm(axes = x_1465_axes_0, beta = model_blocks_8_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_8_attn_inner_attn_ln_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("x_1465_cast_fp16")];
            tensor<int32, [4]> var_13079 = const()[name = tensor<string, []>("op_13079"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1467_cast_fp16 = reshape(shape = var_13079, x = x_1465_cast_fp16)[name = tensor<string, []>("x_1467_cast_fp16")];
            tensor<int32, [4]> input_151_perm_0 = const()[name = tensor<string, []>("input_151_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1469_pad_type_0 = const()[name = tensor<string, []>("x_1469_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1469_strides_0 = const()[name = tensor<string, []>("x_1469_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1469_pad_0 = const()[name = tensor<string, []>("x_1469_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1469_dilations_0 = const()[name = tensor<string, []>("x_1469_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1469_groups_0 = const()[name = tensor<string, []>("x_1469_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_8_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118576960)))];
            tensor<fp16, [768]> model_blocks_8_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119756672)))];
            tensor<fp16, [1, 768, 1, 197]> input_151_cast_fp16 = transpose(perm = input_151_perm_0, x = x_1467_cast_fp16)[name = tensor<string, []>("transpose_134")];
            tensor<fp16, [1, 768, 1, 197]> x_1469_cast_fp16 = conv(bias = model_blocks_8_attn_proj_bias_to_fp16, dilations = x_1469_dilations_0, groups = x_1469_groups_0, pad = x_1469_pad_0, pad_type = x_1469_pad_type_0, strides = x_1469_strides_0, weight = model_blocks_8_attn_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_1469_cast_fp16")];
            tensor<int32, [4]> x_1471_perm_0 = const()[name = tensor<string, []>("x_1471_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_13090 = const()[name = tensor<string, []>("op_13090"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1471_cast_fp16 = transpose(perm = x_1471_perm_0, x = x_1469_cast_fp16)[name = tensor<string, []>("transpose_133")];
            tensor<fp16, [1, 197, 768]> var_13091_cast_fp16 = reshape(shape = var_13090, x = x_1471_cast_fp16)[name = tensor<string, []>("op_13091_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_153_cast_fp16 = add(x = input_145_cast_fp16, y = var_13091_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<int32, [1]> x_1473_axes_0 = const()[name = tensor<string, []>("x_1473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119758272)))];
            tensor<fp16, [768]> model_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119759872)))];
            tensor<fp16, [1, 197, 768]> x_1473_cast_fp16 = layer_norm(axes = x_1473_axes_0, beta = model_blocks_8_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_8_norm2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("x_1473_cast_fp16")];
            tensor<int32, [4]> var_13103 = const()[name = tensor<string, []>("op_13103"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1475_cast_fp16 = reshape(shape = var_13103, x = x_1473_cast_fp16)[name = tensor<string, []>("x_1475_cast_fp16")];
            tensor<int32, [4]> input_155_perm_0 = const()[name = tensor<string, []>("input_155_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_157_strides_0 = const()[name = tensor<string, []>("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_157_dilations_0 = const()[name = tensor<string, []>("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_157_groups_0 = const()[name = tensor<string, []>("input_157_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_8_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119761472)))];
            tensor<fp16, [2048]> model_blocks_8_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122907264)))];
            tensor<fp16, [1, 768, 1, 197]> input_155_cast_fp16 = transpose(perm = input_155_perm_0, x = x_1475_cast_fp16)[name = tensor<string, []>("transpose_132")];
            tensor<fp16, [1, 2048, 1, 197]> input_157_cast_fp16 = conv(bias = model_blocks_8_mlp_w1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = model_blocks_8_mlp_w1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<string, []> x2_449_pad_type_0 = const()[name = tensor<string, []>("x2_449_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_449_strides_0 = const()[name = tensor<string, []>("x2_449_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_449_pad_0 = const()[name = tensor<string, []>("x2_449_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_449_dilations_0 = const()[name = tensor<string, []>("x2_449_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_449_groups_0 = const()[name = tensor<string, []>("x2_449_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_8_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122911424)))];
            tensor<fp16, [2048]> model_blocks_8_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126057216)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_449_cast_fp16 = conv(bias = model_blocks_8_mlp_w2_bias_to_fp16, dilations = x2_449_dilations_0, groups = x2_449_groups_0, pad = x2_449_pad_0, pad_type = x2_449_pad_type_0, strides = x2_449_strides_0, weight = model_blocks_8_mlp_w2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("x2_449_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_13120_cast_fp16 = silu(x = input_157_cast_fp16)[name = tensor<string, []>("op_13120_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_65_cast_fp16 = mul(x = var_13120_cast_fp16, y = x2_449_cast_fp16)[name = tensor<string, []>("hidden_65_cast_fp16")];
            tensor<int32, [4]> hidden_67_perm_0 = const()[name = tensor<string, []>("hidden_67_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_13123 = const()[name = tensor<string, []>("op_13123"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_67_cast_fp16 = transpose(perm = hidden_67_perm_0, x = hidden_65_cast_fp16)[name = tensor<string, []>("transpose_131")];
            tensor<fp16, [1, 197, 2048]> input_159_cast_fp16 = reshape(shape = var_13123, x = hidden_67_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<int32, [1]> hidden_69_axes_0 = const()[name = tensor<string, []>("hidden_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_8_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126061376)))];
            tensor<fp16, [2048]> model_blocks_8_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126065536)))];
            tensor<fp16, [1, 197, 2048]> hidden_69_cast_fp16 = layer_norm(axes = hidden_69_axes_0, beta = model_blocks_8_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_8_mlp_ffn_ln_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("hidden_69_cast_fp16")];
            tensor<int32, [4]> var_13129 = const()[name = tensor<string, []>("op_13129"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_71_cast_fp16 = reshape(shape = var_13129, x = hidden_69_cast_fp16)[name = tensor<string, []>("hidden_71_cast_fp16")];
            tensor<int32, [4]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1477_pad_type_0 = const()[name = tensor<string, []>("x_1477_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1477_strides_0 = const()[name = tensor<string, []>("x_1477_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1477_pad_0 = const()[name = tensor<string, []>("x_1477_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1477_dilations_0 = const()[name = tensor<string, []>("x_1477_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1477_groups_0 = const()[name = tensor<string, []>("x_1477_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_8_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126069696)))];
            tensor<fp16, [768]> model_blocks_8_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_8_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129215488)))];
            tensor<fp16, [1, 2048, 1, 197]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = hidden_71_cast_fp16)[name = tensor<string, []>("transpose_130")];
            tensor<fp16, [1, 768, 1, 197]> x_1477_cast_fp16 = conv(bias = model_blocks_8_mlp_w3_bias_to_fp16, dilations = x_1477_dilations_0, groups = x_1477_groups_0, pad = x_1477_pad_0, pad_type = x_1477_pad_type_0, strides = x_1477_strides_0, weight = model_blocks_8_mlp_w3_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("x_1477_cast_fp16")];
            tensor<int32, [4]> x_1479_perm_0 = const()[name = tensor<string, []>("x_1479_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_13140 = const()[name = tensor<string, []>("op_13140"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1479_cast_fp16 = transpose(perm = x_1479_perm_0, x = x_1477_cast_fp16)[name = tensor<string, []>("transpose_129")];
            tensor<fp16, [1, 197, 768]> var_13141_cast_fp16 = reshape(shape = var_13140, x = x_1479_cast_fp16)[name = tensor<string, []>("op_13141_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_163_cast_fp16 = add(x = input_153_cast_fp16, y = var_13141_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [1]> x_1481_axes_0 = const()[name = tensor<string, []>("x_1481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129217088)))];
            tensor<fp16, [768]> model_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129218688)))];
            tensor<fp16, [1, 197, 768]> x_1481_cast_fp16 = layer_norm(axes = x_1481_axes_0, beta = model_blocks_9_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_9_norm1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_1481_cast_fp16")];
            tensor<int32, [4]> var_13275 = const()[name = tensor<string, []>("op_13275"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1483_cast_fp16 = reshape(shape = var_13275, x = x_1481_cast_fp16)[name = tensor<string, []>("x_1483_cast_fp16")];
            tensor<int32, [4]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_13283_pad_type_0 = const()[name = tensor<string, []>("op_13283_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13283_strides_0 = const()[name = tensor<string, []>("op_13283_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13283_pad_0 = const()[name = tensor<string, []>("op_13283_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13283_dilations_0 = const()[name = tensor<string, []>("op_13283_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13283_groups_0 = const()[name = tensor<string, []>("op_13283_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129220288)))];
            tensor<fp16, [1, 768, 1, 197]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_1483_cast_fp16)[name = tensor<string, []>("transpose_128")];
            tensor<fp16, [1, 64, 1, 197]> var_13283_cast_fp16 = conv(dilations = var_13283_dilations_0, groups = var_13283_groups_0, pad = var_13283_pad_0, pad_type = var_13283_pad_type_0, strides = var_13283_strides_0, weight = model_blocks_9_attn_q_projs_0_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13283_cast_fp16")];
            tensor<int32, [1]> var_13284_axes_0 = const()[name = tensor<string, []>("op_13284_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13284_cast_fp16 = squeeze(axes = var_13284_axes_0, x = var_13283_cast_fp16)[name = tensor<string, []>("op_13284_cast_fp16")];
            tensor<int32, [3]> q_1081_perm_0 = const()[name = tensor<string, []>("q_1081_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13291_pad_type_0 = const()[name = tensor<string, []>("op_13291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13291_strides_0 = const()[name = tensor<string, []>("op_13291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13291_pad_0 = const()[name = tensor<string, []>("op_13291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13291_dilations_0 = const()[name = tensor<string, []>("op_13291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13291_groups_0 = const()[name = tensor<string, []>("op_13291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129318656)))];
            tensor<fp16, [1, 64, 1, 197]> var_13291_cast_fp16 = conv(dilations = var_13291_dilations_0, groups = var_13291_groups_0, pad = var_13291_pad_0, pad_type = var_13291_pad_type_0, strides = var_13291_strides_0, weight = model_blocks_9_attn_k_projs_0_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13291_cast_fp16")];
            tensor<int32, [1]> var_13292_axes_0 = const()[name = tensor<string, []>("op_13292_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13292_cast_fp16 = squeeze(axes = var_13292_axes_0, x = var_13291_cast_fp16)[name = tensor<string, []>("op_13292_cast_fp16")];
            tensor<int32, [3]> k_649_perm_0 = const()[name = tensor<string, []>("k_649_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13299_pad_type_0 = const()[name = tensor<string, []>("op_13299_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13299_strides_0 = const()[name = tensor<string, []>("op_13299_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13299_pad_0 = const()[name = tensor<string, []>("op_13299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13299_dilations_0 = const()[name = tensor<string, []>("op_13299_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13299_groups_0 = const()[name = tensor<string, []>("op_13299_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129417024)))];
            tensor<fp16, [1, 64, 1, 197]> var_13299_cast_fp16 = conv(dilations = var_13299_dilations_0, groups = var_13299_groups_0, pad = var_13299_pad_0, pad_type = var_13299_pad_type_0, strides = var_13299_strides_0, weight = model_blocks_9_attn_v_projs_0_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13299_cast_fp16")];
            tensor<int32, [1]> var_13300_axes_0 = const()[name = tensor<string, []>("op_13300_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13300_cast_fp16 = squeeze(axes = var_13300_axes_0, x = var_13299_cast_fp16)[name = tensor<string, []>("op_13300_cast_fp16")];
            tensor<int32, [3]> v_649_perm_0 = const()[name = tensor<string, []>("v_649_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129515392)))];
            tensor<fp16, [1, 197, 64]> q_1081_cast_fp16 = transpose(perm = q_1081_perm_0, x = var_13284_cast_fp16)[name = tensor<string, []>("transpose_127")];
            tensor<fp16, [1, 197, 64]> q_1083_cast_fp16 = add(x = q_1081_cast_fp16, y = model_blocks_9_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_1083_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129515584)))];
            tensor<fp16, [1, 197, 64]> v_649_cast_fp16 = transpose(perm = v_649_perm_0, x = var_13300_cast_fp16)[name = tensor<string, []>("transpose_125")];
            tensor<fp16, [1, 197, 64]> v_651_cast_fp16 = add(x = v_649_cast_fp16, y = model_blocks_9_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_651_cast_fp16")];
            tensor<int32, [1]> q_1085_axes_0 = const()[name = tensor<string, []>("q_1085_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1085_cast_fp16 = expand_dims(axes = q_1085_axes_0, x = q_1083_cast_fp16)[name = tensor<string, []>("q_1085_cast_fp16")];
            tensor<int32, [1]> k_651_axes_0 = const()[name = tensor<string, []>("k_651_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_649_cast_fp16 = transpose(perm = k_649_perm_0, x = var_13292_cast_fp16)[name = tensor<string, []>("transpose_126")];
            tensor<fp16, [1, 1, 197, 64]> k_651_cast_fp16 = expand_dims(axes = k_651_axes_0, x = k_649_cast_fp16)[name = tensor<string, []>("k_651_cast_fp16")];
            tensor<int32, [1]> v_653_axes_0 = const()[name = tensor<string, []>("v_653_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_653_cast_fp16 = expand_dims(axes = v_653_axes_0, x = v_651_cast_fp16)[name = tensor<string, []>("v_653_cast_fp16")];
            tensor<int32, [4]> var_13309_begin_0 = const()[name = tensor<string, []>("op_13309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13309_end_0 = const()[name = tensor<string, []>("op_13309_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13309_end_mask_0 = const()[name = tensor<string, []>("op_13309_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13309_cast_fp16 = slice_by_index(begin = var_13309_begin_0, end = var_13309_end_0, end_mask = var_13309_end_mask_0, x = q_1085_cast_fp16)[name = tensor<string, []>("op_13309_cast_fp16")];
            tensor<int32, [4]> var_13313_begin_0 = const()[name = tensor<string, []>("op_13313_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13313_end_0 = const()[name = tensor<string, []>("op_13313_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13313_end_mask_0 = const()[name = tensor<string, []>("op_13313_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13313_cast_fp16 = slice_by_index(begin = var_13313_begin_0, end = var_13313_end_0, end_mask = var_13313_end_mask_0, x = k_651_cast_fp16)[name = tensor<string, []>("op_13313_cast_fp16")];
            tensor<int32, [4]> var_13317_begin_0 = const()[name = tensor<string, []>("op_13317_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13317_end_0 = const()[name = tensor<string, []>("op_13317_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13317_end_mask_0 = const()[name = tensor<string, []>("op_13317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13317_cast_fp16 = slice_by_index(begin = var_13317_begin_0, end = var_13317_end_0, end_mask = var_13317_end_mask_0, x = q_1085_cast_fp16)[name = tensor<string, []>("op_13317_cast_fp16")];
            tensor<int32, [4]> var_13321_begin_0 = const()[name = tensor<string, []>("op_13321_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13321_end_0 = const()[name = tensor<string, []>("op_13321_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13321_end_mask_0 = const()[name = tensor<string, []>("op_13321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13321_cast_fp16 = slice_by_index(begin = var_13321_begin_0, end = var_13321_end_0, end_mask = var_13321_end_mask_0, x = k_651_cast_fp16)[name = tensor<string, []>("op_13321_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13323_cast_fp16 = mul(x = var_13317_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13323_cast_fp16")];
            tensor<int32, [5]> var_13331 = const()[name = tensor<string, []>("op_13331"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1487_cast_fp16 = reshape(shape = var_13331, x = var_13317_cast_fp16)[name = tensor<string, []>("x_1487_cast_fp16")];
            tensor<int32, [2]> var_13333_split_sizes_0 = const()[name = tensor<string, []>("op_13333_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13333_axis_0 = const()[name = tensor<string, []>("op_13333_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13333_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13333_cast_fp16_1 = split(axis = var_13333_axis_0, split_sizes = var_13333_split_sizes_0, x = x_1487_cast_fp16)[name = tensor<string, []>("op_13333_cast_fp16")];
            tensor<int32, [1]> squeeze_432_axes_0 = const()[name = tensor<string, []>("squeeze_432_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_432_cast_fp16 = squeeze(axes = squeeze_432_axes_0, x = var_13333_cast_fp16_0)[name = tensor<string, []>("squeeze_432_cast_fp16")];
            tensor<int32, [1]> squeeze_433_axes_0 = const()[name = tensor<string, []>("squeeze_433_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_433_cast_fp16 = squeeze(axes = squeeze_433_axes_0, x = var_13333_cast_fp16_1)[name = tensor<string, []>("squeeze_433_cast_fp16")];
            tensor<fp16, []> const_1771_promoted_to_fp16 = const()[name = tensor<string, []>("const_1771_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13336_cast_fp16 = mul(x = squeeze_433_cast_fp16, y = const_1771_promoted_to_fp16)[name = tensor<string, []>("op_13336_cast_fp16")];
            tensor<int32, []> x_1489_axis_0 = const()[name = tensor<string, []>("x_1489_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1489_cast_fp16 = stack(axis = x_1489_axis_0, values = (var_13336_cast_fp16, squeeze_432_cast_fp16))[name = tensor<string, []>("x_1489_cast_fp16")];
            tensor<int32, [4]> var_13342 = const()[name = tensor<string, []>("op_13342"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13343_cast_fp16 = reshape(shape = var_13342, x = x_1489_cast_fp16)[name = tensor<string, []>("op_13343_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13344_cast_fp16 = mul(x = var_13343_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13344_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_217_cast_fp16 = add(x = var_13323_cast_fp16, y = var_13344_cast_fp16)[name = tensor<string, []>("q_patches_217_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13346_cast_fp16 = mul(x = var_13321_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13346_cast_fp16")];
            tensor<int32, [5]> var_13354 = const()[name = tensor<string, []>("op_13354"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1493_cast_fp16 = reshape(shape = var_13354, x = var_13321_cast_fp16)[name = tensor<string, []>("x_1493_cast_fp16")];
            tensor<int32, [2]> var_13356_split_sizes_0 = const()[name = tensor<string, []>("op_13356_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13356_axis_0 = const()[name = tensor<string, []>("op_13356_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13356_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13356_cast_fp16_1 = split(axis = var_13356_axis_0, split_sizes = var_13356_split_sizes_0, x = x_1493_cast_fp16)[name = tensor<string, []>("op_13356_cast_fp16")];
            tensor<int32, [1]> squeeze_434_axes_0 = const()[name = tensor<string, []>("squeeze_434_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_434_cast_fp16 = squeeze(axes = squeeze_434_axes_0, x = var_13356_cast_fp16_0)[name = tensor<string, []>("squeeze_434_cast_fp16")];
            tensor<int32, [1]> squeeze_435_axes_0 = const()[name = tensor<string, []>("squeeze_435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_435_cast_fp16 = squeeze(axes = squeeze_435_axes_0, x = var_13356_cast_fp16_1)[name = tensor<string, []>("squeeze_435_cast_fp16")];
            tensor<fp16, []> const_1779_promoted_to_fp16 = const()[name = tensor<string, []>("const_1779_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13359_cast_fp16 = mul(x = squeeze_435_cast_fp16, y = const_1779_promoted_to_fp16)[name = tensor<string, []>("op_13359_cast_fp16")];
            tensor<int32, []> x_1495_axis_0 = const()[name = tensor<string, []>("x_1495_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1495_cast_fp16 = stack(axis = x_1495_axis_0, values = (var_13359_cast_fp16, squeeze_434_cast_fp16))[name = tensor<string, []>("x_1495_cast_fp16")];
            tensor<int32, [4]> var_13365 = const()[name = tensor<string, []>("op_13365"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13366_cast_fp16 = reshape(shape = var_13365, x = x_1495_cast_fp16)[name = tensor<string, []>("op_13366_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13367_cast_fp16 = mul(x = var_13366_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13367_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_217_cast_fp16 = add(x = var_13346_cast_fp16, y = var_13367_cast_fp16)[name = tensor<string, []>("k_patches_217_cast_fp16")];
            tensor<bool, []> var_13370_interleave_0 = const()[name = tensor<string, []>("op_13370_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13370_cast_fp16 = concat(axis = var_20, interleave = var_13370_interleave_0, values = (var_13309_cast_fp16, q_patches_217_cast_fp16))[name = tensor<string, []>("op_13370_cast_fp16")];
            tensor<bool, []> var_13373_interleave_0 = const()[name = tensor<string, []>("op_13373_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13373_cast_fp16 = concat(axis = var_20, interleave = var_13373_interleave_0, values = (var_13313_cast_fp16, k_patches_217_cast_fp16))[name = tensor<string, []>("op_13373_cast_fp16")];
            tensor<fp16, []> var_13375_to_fp16 = const()[name = tensor<string, []>("op_13375_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1089_cast_fp16 = mul(x = var_13370_cast_fp16, y = var_13375_to_fp16)[name = tensor<string, []>("q_1089_cast_fp16")];
            tensor<bool, []> attn_433_transpose_x_1 = const()[name = tensor<string, []>("attn_433_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_433_transpose_y_1 = const()[name = tensor<string, []>("attn_433_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_433_cast_fp16 = matmul(transpose_x = attn_433_transpose_x_1, transpose_y = attn_433_transpose_y_1, x = q_1089_cast_fp16, y = var_13373_cast_fp16)[name = tensor<string, []>("attn_433_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_435_cast_fp16 = softmax(axis = var_21, x = attn_433_cast_fp16)[name = tensor<string, []>("attn_435_cast_fp16")];
            tensor<bool, []> var_13380_transpose_x_0 = const()[name = tensor<string, []>("op_13380_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13380_transpose_y_0 = const()[name = tensor<string, []>("op_13380_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13380_cast_fp16 = matmul(transpose_x = var_13380_transpose_x_0, transpose_y = var_13380_transpose_y_0, x = attn_435_cast_fp16, y = v_653_cast_fp16)[name = tensor<string, []>("op_13380_cast_fp16")];
            tensor<int32, [1]> var_13381_axes_0 = const()[name = tensor<string, []>("op_13381_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13381_cast_fp16 = squeeze(axes = var_13381_axes_0, x = var_13380_cast_fp16)[name = tensor<string, []>("op_13381_cast_fp16")];
            tensor<string, []> var_13387_pad_type_0 = const()[name = tensor<string, []>("op_13387_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13387_strides_0 = const()[name = tensor<string, []>("op_13387_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13387_pad_0 = const()[name = tensor<string, []>("op_13387_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13387_dilations_0 = const()[name = tensor<string, []>("op_13387_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13387_groups_0 = const()[name = tensor<string, []>("op_13387_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129515776)))];
            tensor<fp16, [1, 64, 1, 197]> var_13387_cast_fp16 = conv(dilations = var_13387_dilations_0, groups = var_13387_groups_0, pad = var_13387_pad_0, pad_type = var_13387_pad_type_0, strides = var_13387_strides_0, weight = model_blocks_9_attn_q_projs_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13387_cast_fp16")];
            tensor<int32, [1]> var_13388_axes_0 = const()[name = tensor<string, []>("op_13388_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13388_cast_fp16 = squeeze(axes = var_13388_axes_0, x = var_13387_cast_fp16)[name = tensor<string, []>("op_13388_cast_fp16")];
            tensor<int32, [3]> q_1091_perm_0 = const()[name = tensor<string, []>("q_1091_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13395_pad_type_0 = const()[name = tensor<string, []>("op_13395_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13395_strides_0 = const()[name = tensor<string, []>("op_13395_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13395_pad_0 = const()[name = tensor<string, []>("op_13395_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13395_dilations_0 = const()[name = tensor<string, []>("op_13395_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13395_groups_0 = const()[name = tensor<string, []>("op_13395_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129614144)))];
            tensor<fp16, [1, 64, 1, 197]> var_13395_cast_fp16 = conv(dilations = var_13395_dilations_0, groups = var_13395_groups_0, pad = var_13395_pad_0, pad_type = var_13395_pad_type_0, strides = var_13395_strides_0, weight = model_blocks_9_attn_k_projs_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13395_cast_fp16")];
            tensor<int32, [1]> var_13396_axes_0 = const()[name = tensor<string, []>("op_13396_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13396_cast_fp16 = squeeze(axes = var_13396_axes_0, x = var_13395_cast_fp16)[name = tensor<string, []>("op_13396_cast_fp16")];
            tensor<int32, [3]> k_655_perm_0 = const()[name = tensor<string, []>("k_655_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13403_pad_type_0 = const()[name = tensor<string, []>("op_13403_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13403_strides_0 = const()[name = tensor<string, []>("op_13403_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13403_pad_0 = const()[name = tensor<string, []>("op_13403_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13403_dilations_0 = const()[name = tensor<string, []>("op_13403_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13403_groups_0 = const()[name = tensor<string, []>("op_13403_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129712512)))];
            tensor<fp16, [1, 64, 1, 197]> var_13403_cast_fp16 = conv(dilations = var_13403_dilations_0, groups = var_13403_groups_0, pad = var_13403_pad_0, pad_type = var_13403_pad_type_0, strides = var_13403_strides_0, weight = model_blocks_9_attn_v_projs_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13403_cast_fp16")];
            tensor<int32, [1]> var_13404_axes_0 = const()[name = tensor<string, []>("op_13404_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13404_cast_fp16 = squeeze(axes = var_13404_axes_0, x = var_13403_cast_fp16)[name = tensor<string, []>("op_13404_cast_fp16")];
            tensor<int32, [3]> v_655_perm_0 = const()[name = tensor<string, []>("v_655_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129810880)))];
            tensor<fp16, [1, 197, 64]> q_1091_cast_fp16 = transpose(perm = q_1091_perm_0, x = var_13388_cast_fp16)[name = tensor<string, []>("transpose_124")];
            tensor<fp16, [1, 197, 64]> q_1093_cast_fp16 = add(x = q_1091_cast_fp16, y = model_blocks_9_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_1093_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129811072)))];
            tensor<fp16, [1, 197, 64]> v_655_cast_fp16 = transpose(perm = v_655_perm_0, x = var_13404_cast_fp16)[name = tensor<string, []>("transpose_122")];
            tensor<fp16, [1, 197, 64]> v_657_cast_fp16 = add(x = v_655_cast_fp16, y = model_blocks_9_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_657_cast_fp16")];
            tensor<int32, [1]> q_1095_axes_0 = const()[name = tensor<string, []>("q_1095_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1095_cast_fp16 = expand_dims(axes = q_1095_axes_0, x = q_1093_cast_fp16)[name = tensor<string, []>("q_1095_cast_fp16")];
            tensor<int32, [1]> k_657_axes_0 = const()[name = tensor<string, []>("k_657_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_655_cast_fp16 = transpose(perm = k_655_perm_0, x = var_13396_cast_fp16)[name = tensor<string, []>("transpose_123")];
            tensor<fp16, [1, 1, 197, 64]> k_657_cast_fp16 = expand_dims(axes = k_657_axes_0, x = k_655_cast_fp16)[name = tensor<string, []>("k_657_cast_fp16")];
            tensor<int32, [1]> v_659_axes_0 = const()[name = tensor<string, []>("v_659_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_659_cast_fp16 = expand_dims(axes = v_659_axes_0, x = v_657_cast_fp16)[name = tensor<string, []>("v_659_cast_fp16")];
            tensor<int32, [4]> var_13413_begin_0 = const()[name = tensor<string, []>("op_13413_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13413_end_0 = const()[name = tensor<string, []>("op_13413_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13413_end_mask_0 = const()[name = tensor<string, []>("op_13413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13413_cast_fp16 = slice_by_index(begin = var_13413_begin_0, end = var_13413_end_0, end_mask = var_13413_end_mask_0, x = q_1095_cast_fp16)[name = tensor<string, []>("op_13413_cast_fp16")];
            tensor<int32, [4]> var_13417_begin_0 = const()[name = tensor<string, []>("op_13417_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13417_end_0 = const()[name = tensor<string, []>("op_13417_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13417_end_mask_0 = const()[name = tensor<string, []>("op_13417_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13417_cast_fp16 = slice_by_index(begin = var_13417_begin_0, end = var_13417_end_0, end_mask = var_13417_end_mask_0, x = k_657_cast_fp16)[name = tensor<string, []>("op_13417_cast_fp16")];
            tensor<int32, [4]> var_13421_begin_0 = const()[name = tensor<string, []>("op_13421_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13421_end_0 = const()[name = tensor<string, []>("op_13421_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13421_end_mask_0 = const()[name = tensor<string, []>("op_13421_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13421_cast_fp16 = slice_by_index(begin = var_13421_begin_0, end = var_13421_end_0, end_mask = var_13421_end_mask_0, x = q_1095_cast_fp16)[name = tensor<string, []>("op_13421_cast_fp16")];
            tensor<int32, [4]> var_13425_begin_0 = const()[name = tensor<string, []>("op_13425_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13425_end_0 = const()[name = tensor<string, []>("op_13425_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13425_end_mask_0 = const()[name = tensor<string, []>("op_13425_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13425_cast_fp16 = slice_by_index(begin = var_13425_begin_0, end = var_13425_end_0, end_mask = var_13425_end_mask_0, x = k_657_cast_fp16)[name = tensor<string, []>("op_13425_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13427_cast_fp16 = mul(x = var_13421_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13427_cast_fp16")];
            tensor<int32, [5]> var_13435 = const()[name = tensor<string, []>("op_13435"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1499_cast_fp16 = reshape(shape = var_13435, x = var_13421_cast_fp16)[name = tensor<string, []>("x_1499_cast_fp16")];
            tensor<int32, [2]> var_13437_split_sizes_0 = const()[name = tensor<string, []>("op_13437_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13437_axis_0 = const()[name = tensor<string, []>("op_13437_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13437_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13437_cast_fp16_1 = split(axis = var_13437_axis_0, split_sizes = var_13437_split_sizes_0, x = x_1499_cast_fp16)[name = tensor<string, []>("op_13437_cast_fp16")];
            tensor<int32, [1]> squeeze_436_axes_0 = const()[name = tensor<string, []>("squeeze_436_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_436_cast_fp16 = squeeze(axes = squeeze_436_axes_0, x = var_13437_cast_fp16_0)[name = tensor<string, []>("squeeze_436_cast_fp16")];
            tensor<int32, [1]> squeeze_437_axes_0 = const()[name = tensor<string, []>("squeeze_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_437_cast_fp16 = squeeze(axes = squeeze_437_axes_0, x = var_13437_cast_fp16_1)[name = tensor<string, []>("squeeze_437_cast_fp16")];
            tensor<fp16, []> const_1787_promoted_to_fp16 = const()[name = tensor<string, []>("const_1787_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13440_cast_fp16 = mul(x = squeeze_437_cast_fp16, y = const_1787_promoted_to_fp16)[name = tensor<string, []>("op_13440_cast_fp16")];
            tensor<int32, []> x_1501_axis_0 = const()[name = tensor<string, []>("x_1501_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1501_cast_fp16 = stack(axis = x_1501_axis_0, values = (var_13440_cast_fp16, squeeze_436_cast_fp16))[name = tensor<string, []>("x_1501_cast_fp16")];
            tensor<int32, [4]> var_13446 = const()[name = tensor<string, []>("op_13446"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13447_cast_fp16 = reshape(shape = var_13446, x = x_1501_cast_fp16)[name = tensor<string, []>("op_13447_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13448_cast_fp16 = mul(x = var_13447_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13448_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_219_cast_fp16 = add(x = var_13427_cast_fp16, y = var_13448_cast_fp16)[name = tensor<string, []>("q_patches_219_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13450_cast_fp16 = mul(x = var_13425_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13450_cast_fp16")];
            tensor<int32, [5]> var_13458 = const()[name = tensor<string, []>("op_13458"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1505_cast_fp16 = reshape(shape = var_13458, x = var_13425_cast_fp16)[name = tensor<string, []>("x_1505_cast_fp16")];
            tensor<int32, [2]> var_13460_split_sizes_0 = const()[name = tensor<string, []>("op_13460_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13460_axis_0 = const()[name = tensor<string, []>("op_13460_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13460_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13460_cast_fp16_1 = split(axis = var_13460_axis_0, split_sizes = var_13460_split_sizes_0, x = x_1505_cast_fp16)[name = tensor<string, []>("op_13460_cast_fp16")];
            tensor<int32, [1]> squeeze_438_axes_0 = const()[name = tensor<string, []>("squeeze_438_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_438_cast_fp16 = squeeze(axes = squeeze_438_axes_0, x = var_13460_cast_fp16_0)[name = tensor<string, []>("squeeze_438_cast_fp16")];
            tensor<int32, [1]> squeeze_439_axes_0 = const()[name = tensor<string, []>("squeeze_439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_439_cast_fp16 = squeeze(axes = squeeze_439_axes_0, x = var_13460_cast_fp16_1)[name = tensor<string, []>("squeeze_439_cast_fp16")];
            tensor<fp16, []> const_1795_promoted_to_fp16 = const()[name = tensor<string, []>("const_1795_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13463_cast_fp16 = mul(x = squeeze_439_cast_fp16, y = const_1795_promoted_to_fp16)[name = tensor<string, []>("op_13463_cast_fp16")];
            tensor<int32, []> x_1507_axis_0 = const()[name = tensor<string, []>("x_1507_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1507_cast_fp16 = stack(axis = x_1507_axis_0, values = (var_13463_cast_fp16, squeeze_438_cast_fp16))[name = tensor<string, []>("x_1507_cast_fp16")];
            tensor<int32, [4]> var_13469 = const()[name = tensor<string, []>("op_13469"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13470_cast_fp16 = reshape(shape = var_13469, x = x_1507_cast_fp16)[name = tensor<string, []>("op_13470_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13471_cast_fp16 = mul(x = var_13470_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13471_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_219_cast_fp16 = add(x = var_13450_cast_fp16, y = var_13471_cast_fp16)[name = tensor<string, []>("k_patches_219_cast_fp16")];
            tensor<bool, []> var_13474_interleave_0 = const()[name = tensor<string, []>("op_13474_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13474_cast_fp16 = concat(axis = var_20, interleave = var_13474_interleave_0, values = (var_13413_cast_fp16, q_patches_219_cast_fp16))[name = tensor<string, []>("op_13474_cast_fp16")];
            tensor<bool, []> var_13477_interleave_0 = const()[name = tensor<string, []>("op_13477_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13477_cast_fp16 = concat(axis = var_20, interleave = var_13477_interleave_0, values = (var_13417_cast_fp16, k_patches_219_cast_fp16))[name = tensor<string, []>("op_13477_cast_fp16")];
            tensor<fp16, []> var_13479_to_fp16 = const()[name = tensor<string, []>("op_13479_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1099_cast_fp16 = mul(x = var_13474_cast_fp16, y = var_13479_to_fp16)[name = tensor<string, []>("q_1099_cast_fp16")];
            tensor<bool, []> attn_437_transpose_x_1 = const()[name = tensor<string, []>("attn_437_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_437_transpose_y_1 = const()[name = tensor<string, []>("attn_437_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_437_cast_fp16 = matmul(transpose_x = attn_437_transpose_x_1, transpose_y = attn_437_transpose_y_1, x = q_1099_cast_fp16, y = var_13477_cast_fp16)[name = tensor<string, []>("attn_437_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_439_cast_fp16 = softmax(axis = var_21, x = attn_437_cast_fp16)[name = tensor<string, []>("attn_439_cast_fp16")];
            tensor<bool, []> var_13484_transpose_x_0 = const()[name = tensor<string, []>("op_13484_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13484_transpose_y_0 = const()[name = tensor<string, []>("op_13484_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13484_cast_fp16 = matmul(transpose_x = var_13484_transpose_x_0, transpose_y = var_13484_transpose_y_0, x = attn_439_cast_fp16, y = v_659_cast_fp16)[name = tensor<string, []>("op_13484_cast_fp16")];
            tensor<int32, [1]> var_13485_axes_0 = const()[name = tensor<string, []>("op_13485_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13485_cast_fp16 = squeeze(axes = var_13485_axes_0, x = var_13484_cast_fp16)[name = tensor<string, []>("op_13485_cast_fp16")];
            tensor<string, []> var_13491_pad_type_0 = const()[name = tensor<string, []>("op_13491_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13491_strides_0 = const()[name = tensor<string, []>("op_13491_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13491_pad_0 = const()[name = tensor<string, []>("op_13491_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13491_dilations_0 = const()[name = tensor<string, []>("op_13491_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13491_groups_0 = const()[name = tensor<string, []>("op_13491_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129811264)))];
            tensor<fp16, [1, 64, 1, 197]> var_13491_cast_fp16 = conv(dilations = var_13491_dilations_0, groups = var_13491_groups_0, pad = var_13491_pad_0, pad_type = var_13491_pad_type_0, strides = var_13491_strides_0, weight = model_blocks_9_attn_q_projs_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13491_cast_fp16")];
            tensor<int32, [1]> var_13492_axes_0 = const()[name = tensor<string, []>("op_13492_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13492_cast_fp16 = squeeze(axes = var_13492_axes_0, x = var_13491_cast_fp16)[name = tensor<string, []>("op_13492_cast_fp16")];
            tensor<int32, [3]> q_1101_perm_0 = const()[name = tensor<string, []>("q_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13499_pad_type_0 = const()[name = tensor<string, []>("op_13499_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13499_strides_0 = const()[name = tensor<string, []>("op_13499_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13499_pad_0 = const()[name = tensor<string, []>("op_13499_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13499_dilations_0 = const()[name = tensor<string, []>("op_13499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13499_groups_0 = const()[name = tensor<string, []>("op_13499_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129909632)))];
            tensor<fp16, [1, 64, 1, 197]> var_13499_cast_fp16 = conv(dilations = var_13499_dilations_0, groups = var_13499_groups_0, pad = var_13499_pad_0, pad_type = var_13499_pad_type_0, strides = var_13499_strides_0, weight = model_blocks_9_attn_k_projs_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13499_cast_fp16")];
            tensor<int32, [1]> var_13500_axes_0 = const()[name = tensor<string, []>("op_13500_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13500_cast_fp16 = squeeze(axes = var_13500_axes_0, x = var_13499_cast_fp16)[name = tensor<string, []>("op_13500_cast_fp16")];
            tensor<int32, [3]> k_661_perm_0 = const()[name = tensor<string, []>("k_661_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13507_pad_type_0 = const()[name = tensor<string, []>("op_13507_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13507_strides_0 = const()[name = tensor<string, []>("op_13507_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13507_pad_0 = const()[name = tensor<string, []>("op_13507_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13507_dilations_0 = const()[name = tensor<string, []>("op_13507_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13507_groups_0 = const()[name = tensor<string, []>("op_13507_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130008000)))];
            tensor<fp16, [1, 64, 1, 197]> var_13507_cast_fp16 = conv(dilations = var_13507_dilations_0, groups = var_13507_groups_0, pad = var_13507_pad_0, pad_type = var_13507_pad_type_0, strides = var_13507_strides_0, weight = model_blocks_9_attn_v_projs_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13507_cast_fp16")];
            tensor<int32, [1]> var_13508_axes_0 = const()[name = tensor<string, []>("op_13508_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13508_cast_fp16 = squeeze(axes = var_13508_axes_0, x = var_13507_cast_fp16)[name = tensor<string, []>("op_13508_cast_fp16")];
            tensor<int32, [3]> v_661_perm_0 = const()[name = tensor<string, []>("v_661_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130106368)))];
            tensor<fp16, [1, 197, 64]> q_1101_cast_fp16 = transpose(perm = q_1101_perm_0, x = var_13492_cast_fp16)[name = tensor<string, []>("transpose_121")];
            tensor<fp16, [1, 197, 64]> q_1103_cast_fp16 = add(x = q_1101_cast_fp16, y = model_blocks_9_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_1103_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130106560)))];
            tensor<fp16, [1, 197, 64]> v_661_cast_fp16 = transpose(perm = v_661_perm_0, x = var_13508_cast_fp16)[name = tensor<string, []>("transpose_119")];
            tensor<fp16, [1, 197, 64]> v_663_cast_fp16 = add(x = v_661_cast_fp16, y = model_blocks_9_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_663_cast_fp16")];
            tensor<int32, [1]> q_1105_axes_0 = const()[name = tensor<string, []>("q_1105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1105_cast_fp16 = expand_dims(axes = q_1105_axes_0, x = q_1103_cast_fp16)[name = tensor<string, []>("q_1105_cast_fp16")];
            tensor<int32, [1]> k_663_axes_0 = const()[name = tensor<string, []>("k_663_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_661_cast_fp16 = transpose(perm = k_661_perm_0, x = var_13500_cast_fp16)[name = tensor<string, []>("transpose_120")];
            tensor<fp16, [1, 1, 197, 64]> k_663_cast_fp16 = expand_dims(axes = k_663_axes_0, x = k_661_cast_fp16)[name = tensor<string, []>("k_663_cast_fp16")];
            tensor<int32, [1]> v_665_axes_0 = const()[name = tensor<string, []>("v_665_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_665_cast_fp16 = expand_dims(axes = v_665_axes_0, x = v_663_cast_fp16)[name = tensor<string, []>("v_665_cast_fp16")];
            tensor<int32, [4]> var_13517_begin_0 = const()[name = tensor<string, []>("op_13517_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13517_end_0 = const()[name = tensor<string, []>("op_13517_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13517_end_mask_0 = const()[name = tensor<string, []>("op_13517_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13517_cast_fp16 = slice_by_index(begin = var_13517_begin_0, end = var_13517_end_0, end_mask = var_13517_end_mask_0, x = q_1105_cast_fp16)[name = tensor<string, []>("op_13517_cast_fp16")];
            tensor<int32, [4]> var_13521_begin_0 = const()[name = tensor<string, []>("op_13521_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13521_end_0 = const()[name = tensor<string, []>("op_13521_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13521_end_mask_0 = const()[name = tensor<string, []>("op_13521_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13521_cast_fp16 = slice_by_index(begin = var_13521_begin_0, end = var_13521_end_0, end_mask = var_13521_end_mask_0, x = k_663_cast_fp16)[name = tensor<string, []>("op_13521_cast_fp16")];
            tensor<int32, [4]> var_13525_begin_0 = const()[name = tensor<string, []>("op_13525_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13525_end_0 = const()[name = tensor<string, []>("op_13525_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13525_end_mask_0 = const()[name = tensor<string, []>("op_13525_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13525_cast_fp16 = slice_by_index(begin = var_13525_begin_0, end = var_13525_end_0, end_mask = var_13525_end_mask_0, x = q_1105_cast_fp16)[name = tensor<string, []>("op_13525_cast_fp16")];
            tensor<int32, [4]> var_13529_begin_0 = const()[name = tensor<string, []>("op_13529_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13529_end_0 = const()[name = tensor<string, []>("op_13529_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13529_end_mask_0 = const()[name = tensor<string, []>("op_13529_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13529_cast_fp16 = slice_by_index(begin = var_13529_begin_0, end = var_13529_end_0, end_mask = var_13529_end_mask_0, x = k_663_cast_fp16)[name = tensor<string, []>("op_13529_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13531_cast_fp16 = mul(x = var_13525_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13531_cast_fp16")];
            tensor<int32, [5]> var_13539 = const()[name = tensor<string, []>("op_13539"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1511_cast_fp16 = reshape(shape = var_13539, x = var_13525_cast_fp16)[name = tensor<string, []>("x_1511_cast_fp16")];
            tensor<int32, [2]> var_13541_split_sizes_0 = const()[name = tensor<string, []>("op_13541_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13541_axis_0 = const()[name = tensor<string, []>("op_13541_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13541_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13541_cast_fp16_1 = split(axis = var_13541_axis_0, split_sizes = var_13541_split_sizes_0, x = x_1511_cast_fp16)[name = tensor<string, []>("op_13541_cast_fp16")];
            tensor<int32, [1]> squeeze_440_axes_0 = const()[name = tensor<string, []>("squeeze_440_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_440_cast_fp16 = squeeze(axes = squeeze_440_axes_0, x = var_13541_cast_fp16_0)[name = tensor<string, []>("squeeze_440_cast_fp16")];
            tensor<int32, [1]> squeeze_441_axes_0 = const()[name = tensor<string, []>("squeeze_441_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_441_cast_fp16 = squeeze(axes = squeeze_441_axes_0, x = var_13541_cast_fp16_1)[name = tensor<string, []>("squeeze_441_cast_fp16")];
            tensor<fp16, []> const_1803_promoted_to_fp16 = const()[name = tensor<string, []>("const_1803_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13544_cast_fp16 = mul(x = squeeze_441_cast_fp16, y = const_1803_promoted_to_fp16)[name = tensor<string, []>("op_13544_cast_fp16")];
            tensor<int32, []> x_1513_axis_0 = const()[name = tensor<string, []>("x_1513_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1513_cast_fp16 = stack(axis = x_1513_axis_0, values = (var_13544_cast_fp16, squeeze_440_cast_fp16))[name = tensor<string, []>("x_1513_cast_fp16")];
            tensor<int32, [4]> var_13550 = const()[name = tensor<string, []>("op_13550"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13551_cast_fp16 = reshape(shape = var_13550, x = x_1513_cast_fp16)[name = tensor<string, []>("op_13551_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13552_cast_fp16 = mul(x = var_13551_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13552_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_221_cast_fp16 = add(x = var_13531_cast_fp16, y = var_13552_cast_fp16)[name = tensor<string, []>("q_patches_221_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13554_cast_fp16 = mul(x = var_13529_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13554_cast_fp16")];
            tensor<int32, [5]> var_13562 = const()[name = tensor<string, []>("op_13562"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1517_cast_fp16 = reshape(shape = var_13562, x = var_13529_cast_fp16)[name = tensor<string, []>("x_1517_cast_fp16")];
            tensor<int32, [2]> var_13564_split_sizes_0 = const()[name = tensor<string, []>("op_13564_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13564_axis_0 = const()[name = tensor<string, []>("op_13564_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13564_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13564_cast_fp16_1 = split(axis = var_13564_axis_0, split_sizes = var_13564_split_sizes_0, x = x_1517_cast_fp16)[name = tensor<string, []>("op_13564_cast_fp16")];
            tensor<int32, [1]> squeeze_442_axes_0 = const()[name = tensor<string, []>("squeeze_442_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_442_cast_fp16 = squeeze(axes = squeeze_442_axes_0, x = var_13564_cast_fp16_0)[name = tensor<string, []>("squeeze_442_cast_fp16")];
            tensor<int32, [1]> squeeze_443_axes_0 = const()[name = tensor<string, []>("squeeze_443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_443_cast_fp16 = squeeze(axes = squeeze_443_axes_0, x = var_13564_cast_fp16_1)[name = tensor<string, []>("squeeze_443_cast_fp16")];
            tensor<fp16, []> const_1811_promoted_to_fp16 = const()[name = tensor<string, []>("const_1811_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13567_cast_fp16 = mul(x = squeeze_443_cast_fp16, y = const_1811_promoted_to_fp16)[name = tensor<string, []>("op_13567_cast_fp16")];
            tensor<int32, []> x_1519_axis_0 = const()[name = tensor<string, []>("x_1519_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1519_cast_fp16 = stack(axis = x_1519_axis_0, values = (var_13567_cast_fp16, squeeze_442_cast_fp16))[name = tensor<string, []>("x_1519_cast_fp16")];
            tensor<int32, [4]> var_13573 = const()[name = tensor<string, []>("op_13573"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13574_cast_fp16 = reshape(shape = var_13573, x = x_1519_cast_fp16)[name = tensor<string, []>("op_13574_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13575_cast_fp16 = mul(x = var_13574_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13575_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_221_cast_fp16 = add(x = var_13554_cast_fp16, y = var_13575_cast_fp16)[name = tensor<string, []>("k_patches_221_cast_fp16")];
            tensor<bool, []> var_13578_interleave_0 = const()[name = tensor<string, []>("op_13578_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13578_cast_fp16 = concat(axis = var_20, interleave = var_13578_interleave_0, values = (var_13517_cast_fp16, q_patches_221_cast_fp16))[name = tensor<string, []>("op_13578_cast_fp16")];
            tensor<bool, []> var_13581_interleave_0 = const()[name = tensor<string, []>("op_13581_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13581_cast_fp16 = concat(axis = var_20, interleave = var_13581_interleave_0, values = (var_13521_cast_fp16, k_patches_221_cast_fp16))[name = tensor<string, []>("op_13581_cast_fp16")];
            tensor<fp16, []> var_13583_to_fp16 = const()[name = tensor<string, []>("op_13583_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1109_cast_fp16 = mul(x = var_13578_cast_fp16, y = var_13583_to_fp16)[name = tensor<string, []>("q_1109_cast_fp16")];
            tensor<bool, []> attn_441_transpose_x_1 = const()[name = tensor<string, []>("attn_441_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_441_transpose_y_1 = const()[name = tensor<string, []>("attn_441_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_441_cast_fp16 = matmul(transpose_x = attn_441_transpose_x_1, transpose_y = attn_441_transpose_y_1, x = q_1109_cast_fp16, y = var_13581_cast_fp16)[name = tensor<string, []>("attn_441_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_443_cast_fp16 = softmax(axis = var_21, x = attn_441_cast_fp16)[name = tensor<string, []>("attn_443_cast_fp16")];
            tensor<bool, []> var_13588_transpose_x_0 = const()[name = tensor<string, []>("op_13588_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13588_transpose_y_0 = const()[name = tensor<string, []>("op_13588_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13588_cast_fp16 = matmul(transpose_x = var_13588_transpose_x_0, transpose_y = var_13588_transpose_y_0, x = attn_443_cast_fp16, y = v_665_cast_fp16)[name = tensor<string, []>("op_13588_cast_fp16")];
            tensor<int32, [1]> var_13589_axes_0 = const()[name = tensor<string, []>("op_13589_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13589_cast_fp16 = squeeze(axes = var_13589_axes_0, x = var_13588_cast_fp16)[name = tensor<string, []>("op_13589_cast_fp16")];
            tensor<string, []> var_13595_pad_type_0 = const()[name = tensor<string, []>("op_13595_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13595_strides_0 = const()[name = tensor<string, []>("op_13595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13595_pad_0 = const()[name = tensor<string, []>("op_13595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13595_dilations_0 = const()[name = tensor<string, []>("op_13595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13595_groups_0 = const()[name = tensor<string, []>("op_13595_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130106752)))];
            tensor<fp16, [1, 64, 1, 197]> var_13595_cast_fp16 = conv(dilations = var_13595_dilations_0, groups = var_13595_groups_0, pad = var_13595_pad_0, pad_type = var_13595_pad_type_0, strides = var_13595_strides_0, weight = model_blocks_9_attn_q_projs_3_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13595_cast_fp16")];
            tensor<int32, [1]> var_13596_axes_0 = const()[name = tensor<string, []>("op_13596_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13596_cast_fp16 = squeeze(axes = var_13596_axes_0, x = var_13595_cast_fp16)[name = tensor<string, []>("op_13596_cast_fp16")];
            tensor<int32, [3]> q_1111_perm_0 = const()[name = tensor<string, []>("q_1111_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13603_pad_type_0 = const()[name = tensor<string, []>("op_13603_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13603_strides_0 = const()[name = tensor<string, []>("op_13603_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13603_pad_0 = const()[name = tensor<string, []>("op_13603_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13603_dilations_0 = const()[name = tensor<string, []>("op_13603_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13603_groups_0 = const()[name = tensor<string, []>("op_13603_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130205120)))];
            tensor<fp16, [1, 64, 1, 197]> var_13603_cast_fp16 = conv(dilations = var_13603_dilations_0, groups = var_13603_groups_0, pad = var_13603_pad_0, pad_type = var_13603_pad_type_0, strides = var_13603_strides_0, weight = model_blocks_9_attn_k_projs_3_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13603_cast_fp16")];
            tensor<int32, [1]> var_13604_axes_0 = const()[name = tensor<string, []>("op_13604_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13604_cast_fp16 = squeeze(axes = var_13604_axes_0, x = var_13603_cast_fp16)[name = tensor<string, []>("op_13604_cast_fp16")];
            tensor<int32, [3]> k_667_perm_0 = const()[name = tensor<string, []>("k_667_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13611_pad_type_0 = const()[name = tensor<string, []>("op_13611_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13611_strides_0 = const()[name = tensor<string, []>("op_13611_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13611_pad_0 = const()[name = tensor<string, []>("op_13611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13611_dilations_0 = const()[name = tensor<string, []>("op_13611_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13611_groups_0 = const()[name = tensor<string, []>("op_13611_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130303488)))];
            tensor<fp16, [1, 64, 1, 197]> var_13611_cast_fp16 = conv(dilations = var_13611_dilations_0, groups = var_13611_groups_0, pad = var_13611_pad_0, pad_type = var_13611_pad_type_0, strides = var_13611_strides_0, weight = model_blocks_9_attn_v_projs_3_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13611_cast_fp16")];
            tensor<int32, [1]> var_13612_axes_0 = const()[name = tensor<string, []>("op_13612_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13612_cast_fp16 = squeeze(axes = var_13612_axes_0, x = var_13611_cast_fp16)[name = tensor<string, []>("op_13612_cast_fp16")];
            tensor<int32, [3]> v_667_perm_0 = const()[name = tensor<string, []>("v_667_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130401856)))];
            tensor<fp16, [1, 197, 64]> q_1111_cast_fp16 = transpose(perm = q_1111_perm_0, x = var_13596_cast_fp16)[name = tensor<string, []>("transpose_118")];
            tensor<fp16, [1, 197, 64]> q_1113_cast_fp16 = add(x = q_1111_cast_fp16, y = model_blocks_9_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_1113_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130402048)))];
            tensor<fp16, [1, 197, 64]> v_667_cast_fp16 = transpose(perm = v_667_perm_0, x = var_13612_cast_fp16)[name = tensor<string, []>("transpose_116")];
            tensor<fp16, [1, 197, 64]> v_669_cast_fp16 = add(x = v_667_cast_fp16, y = model_blocks_9_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_669_cast_fp16")];
            tensor<int32, [1]> q_1115_axes_0 = const()[name = tensor<string, []>("q_1115_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1115_cast_fp16 = expand_dims(axes = q_1115_axes_0, x = q_1113_cast_fp16)[name = tensor<string, []>("q_1115_cast_fp16")];
            tensor<int32, [1]> k_669_axes_0 = const()[name = tensor<string, []>("k_669_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_667_cast_fp16 = transpose(perm = k_667_perm_0, x = var_13604_cast_fp16)[name = tensor<string, []>("transpose_117")];
            tensor<fp16, [1, 1, 197, 64]> k_669_cast_fp16 = expand_dims(axes = k_669_axes_0, x = k_667_cast_fp16)[name = tensor<string, []>("k_669_cast_fp16")];
            tensor<int32, [1]> v_671_axes_0 = const()[name = tensor<string, []>("v_671_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_671_cast_fp16 = expand_dims(axes = v_671_axes_0, x = v_669_cast_fp16)[name = tensor<string, []>("v_671_cast_fp16")];
            tensor<int32, [4]> var_13621_begin_0 = const()[name = tensor<string, []>("op_13621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13621_end_0 = const()[name = tensor<string, []>("op_13621_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13621_end_mask_0 = const()[name = tensor<string, []>("op_13621_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13621_cast_fp16 = slice_by_index(begin = var_13621_begin_0, end = var_13621_end_0, end_mask = var_13621_end_mask_0, x = q_1115_cast_fp16)[name = tensor<string, []>("op_13621_cast_fp16")];
            tensor<int32, [4]> var_13625_begin_0 = const()[name = tensor<string, []>("op_13625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13625_end_0 = const()[name = tensor<string, []>("op_13625_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13625_end_mask_0 = const()[name = tensor<string, []>("op_13625_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13625_cast_fp16 = slice_by_index(begin = var_13625_begin_0, end = var_13625_end_0, end_mask = var_13625_end_mask_0, x = k_669_cast_fp16)[name = tensor<string, []>("op_13625_cast_fp16")];
            tensor<int32, [4]> var_13629_begin_0 = const()[name = tensor<string, []>("op_13629_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13629_end_0 = const()[name = tensor<string, []>("op_13629_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13629_end_mask_0 = const()[name = tensor<string, []>("op_13629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13629_cast_fp16 = slice_by_index(begin = var_13629_begin_0, end = var_13629_end_0, end_mask = var_13629_end_mask_0, x = q_1115_cast_fp16)[name = tensor<string, []>("op_13629_cast_fp16")];
            tensor<int32, [4]> var_13633_begin_0 = const()[name = tensor<string, []>("op_13633_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13633_end_0 = const()[name = tensor<string, []>("op_13633_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13633_end_mask_0 = const()[name = tensor<string, []>("op_13633_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13633_cast_fp16 = slice_by_index(begin = var_13633_begin_0, end = var_13633_end_0, end_mask = var_13633_end_mask_0, x = k_669_cast_fp16)[name = tensor<string, []>("op_13633_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13635_cast_fp16 = mul(x = var_13629_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13635_cast_fp16")];
            tensor<int32, [5]> var_13643 = const()[name = tensor<string, []>("op_13643"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1523_cast_fp16 = reshape(shape = var_13643, x = var_13629_cast_fp16)[name = tensor<string, []>("x_1523_cast_fp16")];
            tensor<int32, [2]> var_13645_split_sizes_0 = const()[name = tensor<string, []>("op_13645_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13645_axis_0 = const()[name = tensor<string, []>("op_13645_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13645_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13645_cast_fp16_1 = split(axis = var_13645_axis_0, split_sizes = var_13645_split_sizes_0, x = x_1523_cast_fp16)[name = tensor<string, []>("op_13645_cast_fp16")];
            tensor<int32, [1]> squeeze_444_axes_0 = const()[name = tensor<string, []>("squeeze_444_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_444_cast_fp16 = squeeze(axes = squeeze_444_axes_0, x = var_13645_cast_fp16_0)[name = tensor<string, []>("squeeze_444_cast_fp16")];
            tensor<int32, [1]> squeeze_445_axes_0 = const()[name = tensor<string, []>("squeeze_445_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_445_cast_fp16 = squeeze(axes = squeeze_445_axes_0, x = var_13645_cast_fp16_1)[name = tensor<string, []>("squeeze_445_cast_fp16")];
            tensor<fp16, []> const_1819_promoted_to_fp16 = const()[name = tensor<string, []>("const_1819_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13648_cast_fp16 = mul(x = squeeze_445_cast_fp16, y = const_1819_promoted_to_fp16)[name = tensor<string, []>("op_13648_cast_fp16")];
            tensor<int32, []> x_1525_axis_0 = const()[name = tensor<string, []>("x_1525_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1525_cast_fp16 = stack(axis = x_1525_axis_0, values = (var_13648_cast_fp16, squeeze_444_cast_fp16))[name = tensor<string, []>("x_1525_cast_fp16")];
            tensor<int32, [4]> var_13654 = const()[name = tensor<string, []>("op_13654"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13655_cast_fp16 = reshape(shape = var_13654, x = x_1525_cast_fp16)[name = tensor<string, []>("op_13655_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13656_cast_fp16 = mul(x = var_13655_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13656_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_223_cast_fp16 = add(x = var_13635_cast_fp16, y = var_13656_cast_fp16)[name = tensor<string, []>("q_patches_223_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13658_cast_fp16 = mul(x = var_13633_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13658_cast_fp16")];
            tensor<int32, [5]> var_13666 = const()[name = tensor<string, []>("op_13666"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1529_cast_fp16 = reshape(shape = var_13666, x = var_13633_cast_fp16)[name = tensor<string, []>("x_1529_cast_fp16")];
            tensor<int32, [2]> var_13668_split_sizes_0 = const()[name = tensor<string, []>("op_13668_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13668_axis_0 = const()[name = tensor<string, []>("op_13668_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13668_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13668_cast_fp16_1 = split(axis = var_13668_axis_0, split_sizes = var_13668_split_sizes_0, x = x_1529_cast_fp16)[name = tensor<string, []>("op_13668_cast_fp16")];
            tensor<int32, [1]> squeeze_446_axes_0 = const()[name = tensor<string, []>("squeeze_446_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_446_cast_fp16 = squeeze(axes = squeeze_446_axes_0, x = var_13668_cast_fp16_0)[name = tensor<string, []>("squeeze_446_cast_fp16")];
            tensor<int32, [1]> squeeze_447_axes_0 = const()[name = tensor<string, []>("squeeze_447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_447_cast_fp16 = squeeze(axes = squeeze_447_axes_0, x = var_13668_cast_fp16_1)[name = tensor<string, []>("squeeze_447_cast_fp16")];
            tensor<fp16, []> const_1827_promoted_to_fp16 = const()[name = tensor<string, []>("const_1827_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13671_cast_fp16 = mul(x = squeeze_447_cast_fp16, y = const_1827_promoted_to_fp16)[name = tensor<string, []>("op_13671_cast_fp16")];
            tensor<int32, []> x_1531_axis_0 = const()[name = tensor<string, []>("x_1531_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1531_cast_fp16 = stack(axis = x_1531_axis_0, values = (var_13671_cast_fp16, squeeze_446_cast_fp16))[name = tensor<string, []>("x_1531_cast_fp16")];
            tensor<int32, [4]> var_13677 = const()[name = tensor<string, []>("op_13677"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13678_cast_fp16 = reshape(shape = var_13677, x = x_1531_cast_fp16)[name = tensor<string, []>("op_13678_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13679_cast_fp16 = mul(x = var_13678_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13679_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_223_cast_fp16 = add(x = var_13658_cast_fp16, y = var_13679_cast_fp16)[name = tensor<string, []>("k_patches_223_cast_fp16")];
            tensor<bool, []> var_13682_interleave_0 = const()[name = tensor<string, []>("op_13682_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13682_cast_fp16 = concat(axis = var_20, interleave = var_13682_interleave_0, values = (var_13621_cast_fp16, q_patches_223_cast_fp16))[name = tensor<string, []>("op_13682_cast_fp16")];
            tensor<bool, []> var_13685_interleave_0 = const()[name = tensor<string, []>("op_13685_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13685_cast_fp16 = concat(axis = var_20, interleave = var_13685_interleave_0, values = (var_13625_cast_fp16, k_patches_223_cast_fp16))[name = tensor<string, []>("op_13685_cast_fp16")];
            tensor<fp16, []> var_13687_to_fp16 = const()[name = tensor<string, []>("op_13687_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1119_cast_fp16 = mul(x = var_13682_cast_fp16, y = var_13687_to_fp16)[name = tensor<string, []>("q_1119_cast_fp16")];
            tensor<bool, []> attn_445_transpose_x_1 = const()[name = tensor<string, []>("attn_445_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_445_transpose_y_1 = const()[name = tensor<string, []>("attn_445_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_445_cast_fp16 = matmul(transpose_x = attn_445_transpose_x_1, transpose_y = attn_445_transpose_y_1, x = q_1119_cast_fp16, y = var_13685_cast_fp16)[name = tensor<string, []>("attn_445_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_447_cast_fp16 = softmax(axis = var_21, x = attn_445_cast_fp16)[name = tensor<string, []>("attn_447_cast_fp16")];
            tensor<bool, []> var_13692_transpose_x_0 = const()[name = tensor<string, []>("op_13692_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13692_transpose_y_0 = const()[name = tensor<string, []>("op_13692_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13692_cast_fp16 = matmul(transpose_x = var_13692_transpose_x_0, transpose_y = var_13692_transpose_y_0, x = attn_447_cast_fp16, y = v_671_cast_fp16)[name = tensor<string, []>("op_13692_cast_fp16")];
            tensor<int32, [1]> var_13693_axes_0 = const()[name = tensor<string, []>("op_13693_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13693_cast_fp16 = squeeze(axes = var_13693_axes_0, x = var_13692_cast_fp16)[name = tensor<string, []>("op_13693_cast_fp16")];
            tensor<string, []> var_13699_pad_type_0 = const()[name = tensor<string, []>("op_13699_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13699_strides_0 = const()[name = tensor<string, []>("op_13699_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13699_pad_0 = const()[name = tensor<string, []>("op_13699_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13699_dilations_0 = const()[name = tensor<string, []>("op_13699_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13699_groups_0 = const()[name = tensor<string, []>("op_13699_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130402240)))];
            tensor<fp16, [1, 64, 1, 197]> var_13699_cast_fp16 = conv(dilations = var_13699_dilations_0, groups = var_13699_groups_0, pad = var_13699_pad_0, pad_type = var_13699_pad_type_0, strides = var_13699_strides_0, weight = model_blocks_9_attn_q_projs_4_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13699_cast_fp16")];
            tensor<int32, [1]> var_13700_axes_0 = const()[name = tensor<string, []>("op_13700_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13700_cast_fp16 = squeeze(axes = var_13700_axes_0, x = var_13699_cast_fp16)[name = tensor<string, []>("op_13700_cast_fp16")];
            tensor<int32, [3]> q_1121_perm_0 = const()[name = tensor<string, []>("q_1121_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13707_pad_type_0 = const()[name = tensor<string, []>("op_13707_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13707_strides_0 = const()[name = tensor<string, []>("op_13707_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13707_pad_0 = const()[name = tensor<string, []>("op_13707_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13707_dilations_0 = const()[name = tensor<string, []>("op_13707_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13707_groups_0 = const()[name = tensor<string, []>("op_13707_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130500608)))];
            tensor<fp16, [1, 64, 1, 197]> var_13707_cast_fp16 = conv(dilations = var_13707_dilations_0, groups = var_13707_groups_0, pad = var_13707_pad_0, pad_type = var_13707_pad_type_0, strides = var_13707_strides_0, weight = model_blocks_9_attn_k_projs_4_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13707_cast_fp16")];
            tensor<int32, [1]> var_13708_axes_0 = const()[name = tensor<string, []>("op_13708_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13708_cast_fp16 = squeeze(axes = var_13708_axes_0, x = var_13707_cast_fp16)[name = tensor<string, []>("op_13708_cast_fp16")];
            tensor<int32, [3]> k_673_perm_0 = const()[name = tensor<string, []>("k_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13715_pad_type_0 = const()[name = tensor<string, []>("op_13715_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13715_strides_0 = const()[name = tensor<string, []>("op_13715_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13715_pad_0 = const()[name = tensor<string, []>("op_13715_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13715_dilations_0 = const()[name = tensor<string, []>("op_13715_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13715_groups_0 = const()[name = tensor<string, []>("op_13715_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130598976)))];
            tensor<fp16, [1, 64, 1, 197]> var_13715_cast_fp16 = conv(dilations = var_13715_dilations_0, groups = var_13715_groups_0, pad = var_13715_pad_0, pad_type = var_13715_pad_type_0, strides = var_13715_strides_0, weight = model_blocks_9_attn_v_projs_4_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13715_cast_fp16")];
            tensor<int32, [1]> var_13716_axes_0 = const()[name = tensor<string, []>("op_13716_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13716_cast_fp16 = squeeze(axes = var_13716_axes_0, x = var_13715_cast_fp16)[name = tensor<string, []>("op_13716_cast_fp16")];
            tensor<int32, [3]> v_673_perm_0 = const()[name = tensor<string, []>("v_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130697344)))];
            tensor<fp16, [1, 197, 64]> q_1121_cast_fp16 = transpose(perm = q_1121_perm_0, x = var_13700_cast_fp16)[name = tensor<string, []>("transpose_115")];
            tensor<fp16, [1, 197, 64]> q_1123_cast_fp16 = add(x = q_1121_cast_fp16, y = model_blocks_9_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_1123_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130697536)))];
            tensor<fp16, [1, 197, 64]> v_673_cast_fp16 = transpose(perm = v_673_perm_0, x = var_13716_cast_fp16)[name = tensor<string, []>("transpose_113")];
            tensor<fp16, [1, 197, 64]> v_675_cast_fp16 = add(x = v_673_cast_fp16, y = model_blocks_9_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_675_cast_fp16")];
            tensor<int32, [1]> q_1125_axes_0 = const()[name = tensor<string, []>("q_1125_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1125_cast_fp16 = expand_dims(axes = q_1125_axes_0, x = q_1123_cast_fp16)[name = tensor<string, []>("q_1125_cast_fp16")];
            tensor<int32, [1]> k_675_axes_0 = const()[name = tensor<string, []>("k_675_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_673_cast_fp16 = transpose(perm = k_673_perm_0, x = var_13708_cast_fp16)[name = tensor<string, []>("transpose_114")];
            tensor<fp16, [1, 1, 197, 64]> k_675_cast_fp16 = expand_dims(axes = k_675_axes_0, x = k_673_cast_fp16)[name = tensor<string, []>("k_675_cast_fp16")];
            tensor<int32, [1]> v_677_axes_0 = const()[name = tensor<string, []>("v_677_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_677_cast_fp16 = expand_dims(axes = v_677_axes_0, x = v_675_cast_fp16)[name = tensor<string, []>("v_677_cast_fp16")];
            tensor<int32, [4]> var_13725_begin_0 = const()[name = tensor<string, []>("op_13725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13725_end_0 = const()[name = tensor<string, []>("op_13725_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13725_end_mask_0 = const()[name = tensor<string, []>("op_13725_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13725_cast_fp16 = slice_by_index(begin = var_13725_begin_0, end = var_13725_end_0, end_mask = var_13725_end_mask_0, x = q_1125_cast_fp16)[name = tensor<string, []>("op_13725_cast_fp16")];
            tensor<int32, [4]> var_13729_begin_0 = const()[name = tensor<string, []>("op_13729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13729_end_0 = const()[name = tensor<string, []>("op_13729_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13729_end_mask_0 = const()[name = tensor<string, []>("op_13729_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13729_cast_fp16 = slice_by_index(begin = var_13729_begin_0, end = var_13729_end_0, end_mask = var_13729_end_mask_0, x = k_675_cast_fp16)[name = tensor<string, []>("op_13729_cast_fp16")];
            tensor<int32, [4]> var_13733_begin_0 = const()[name = tensor<string, []>("op_13733_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13733_end_0 = const()[name = tensor<string, []>("op_13733_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13733_end_mask_0 = const()[name = tensor<string, []>("op_13733_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13733_cast_fp16 = slice_by_index(begin = var_13733_begin_0, end = var_13733_end_0, end_mask = var_13733_end_mask_0, x = q_1125_cast_fp16)[name = tensor<string, []>("op_13733_cast_fp16")];
            tensor<int32, [4]> var_13737_begin_0 = const()[name = tensor<string, []>("op_13737_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13737_end_0 = const()[name = tensor<string, []>("op_13737_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13737_end_mask_0 = const()[name = tensor<string, []>("op_13737_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13737_cast_fp16 = slice_by_index(begin = var_13737_begin_0, end = var_13737_end_0, end_mask = var_13737_end_mask_0, x = k_675_cast_fp16)[name = tensor<string, []>("op_13737_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13739_cast_fp16 = mul(x = var_13733_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13739_cast_fp16")];
            tensor<int32, [5]> var_13747 = const()[name = tensor<string, []>("op_13747"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1535_cast_fp16 = reshape(shape = var_13747, x = var_13733_cast_fp16)[name = tensor<string, []>("x_1535_cast_fp16")];
            tensor<int32, [2]> var_13749_split_sizes_0 = const()[name = tensor<string, []>("op_13749_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13749_axis_0 = const()[name = tensor<string, []>("op_13749_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13749_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13749_cast_fp16_1 = split(axis = var_13749_axis_0, split_sizes = var_13749_split_sizes_0, x = x_1535_cast_fp16)[name = tensor<string, []>("op_13749_cast_fp16")];
            tensor<int32, [1]> squeeze_448_axes_0 = const()[name = tensor<string, []>("squeeze_448_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_448_cast_fp16 = squeeze(axes = squeeze_448_axes_0, x = var_13749_cast_fp16_0)[name = tensor<string, []>("squeeze_448_cast_fp16")];
            tensor<int32, [1]> squeeze_449_axes_0 = const()[name = tensor<string, []>("squeeze_449_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_449_cast_fp16 = squeeze(axes = squeeze_449_axes_0, x = var_13749_cast_fp16_1)[name = tensor<string, []>("squeeze_449_cast_fp16")];
            tensor<fp16, []> const_1835_promoted_to_fp16 = const()[name = tensor<string, []>("const_1835_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13752_cast_fp16 = mul(x = squeeze_449_cast_fp16, y = const_1835_promoted_to_fp16)[name = tensor<string, []>("op_13752_cast_fp16")];
            tensor<int32, []> x_1537_axis_0 = const()[name = tensor<string, []>("x_1537_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1537_cast_fp16 = stack(axis = x_1537_axis_0, values = (var_13752_cast_fp16, squeeze_448_cast_fp16))[name = tensor<string, []>("x_1537_cast_fp16")];
            tensor<int32, [4]> var_13758 = const()[name = tensor<string, []>("op_13758"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13759_cast_fp16 = reshape(shape = var_13758, x = x_1537_cast_fp16)[name = tensor<string, []>("op_13759_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13760_cast_fp16 = mul(x = var_13759_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13760_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_225_cast_fp16 = add(x = var_13739_cast_fp16, y = var_13760_cast_fp16)[name = tensor<string, []>("q_patches_225_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13762_cast_fp16 = mul(x = var_13737_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13762_cast_fp16")];
            tensor<int32, [5]> var_13770 = const()[name = tensor<string, []>("op_13770"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1541_cast_fp16 = reshape(shape = var_13770, x = var_13737_cast_fp16)[name = tensor<string, []>("x_1541_cast_fp16")];
            tensor<int32, [2]> var_13772_split_sizes_0 = const()[name = tensor<string, []>("op_13772_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13772_axis_0 = const()[name = tensor<string, []>("op_13772_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13772_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13772_cast_fp16_1 = split(axis = var_13772_axis_0, split_sizes = var_13772_split_sizes_0, x = x_1541_cast_fp16)[name = tensor<string, []>("op_13772_cast_fp16")];
            tensor<int32, [1]> squeeze_450_axes_0 = const()[name = tensor<string, []>("squeeze_450_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_450_cast_fp16 = squeeze(axes = squeeze_450_axes_0, x = var_13772_cast_fp16_0)[name = tensor<string, []>("squeeze_450_cast_fp16")];
            tensor<int32, [1]> squeeze_451_axes_0 = const()[name = tensor<string, []>("squeeze_451_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_451_cast_fp16 = squeeze(axes = squeeze_451_axes_0, x = var_13772_cast_fp16_1)[name = tensor<string, []>("squeeze_451_cast_fp16")];
            tensor<fp16, []> const_1843_promoted_to_fp16 = const()[name = tensor<string, []>("const_1843_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13775_cast_fp16 = mul(x = squeeze_451_cast_fp16, y = const_1843_promoted_to_fp16)[name = tensor<string, []>("op_13775_cast_fp16")];
            tensor<int32, []> x_1543_axis_0 = const()[name = tensor<string, []>("x_1543_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1543_cast_fp16 = stack(axis = x_1543_axis_0, values = (var_13775_cast_fp16, squeeze_450_cast_fp16))[name = tensor<string, []>("x_1543_cast_fp16")];
            tensor<int32, [4]> var_13781 = const()[name = tensor<string, []>("op_13781"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13782_cast_fp16 = reshape(shape = var_13781, x = x_1543_cast_fp16)[name = tensor<string, []>("op_13782_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13783_cast_fp16 = mul(x = var_13782_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13783_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_225_cast_fp16 = add(x = var_13762_cast_fp16, y = var_13783_cast_fp16)[name = tensor<string, []>("k_patches_225_cast_fp16")];
            tensor<bool, []> var_13786_interleave_0 = const()[name = tensor<string, []>("op_13786_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13786_cast_fp16 = concat(axis = var_20, interleave = var_13786_interleave_0, values = (var_13725_cast_fp16, q_patches_225_cast_fp16))[name = tensor<string, []>("op_13786_cast_fp16")];
            tensor<bool, []> var_13789_interleave_0 = const()[name = tensor<string, []>("op_13789_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13789_cast_fp16 = concat(axis = var_20, interleave = var_13789_interleave_0, values = (var_13729_cast_fp16, k_patches_225_cast_fp16))[name = tensor<string, []>("op_13789_cast_fp16")];
            tensor<fp16, []> var_13791_to_fp16 = const()[name = tensor<string, []>("op_13791_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1129_cast_fp16 = mul(x = var_13786_cast_fp16, y = var_13791_to_fp16)[name = tensor<string, []>("q_1129_cast_fp16")];
            tensor<bool, []> attn_449_transpose_x_1 = const()[name = tensor<string, []>("attn_449_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_449_transpose_y_1 = const()[name = tensor<string, []>("attn_449_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_449_cast_fp16 = matmul(transpose_x = attn_449_transpose_x_1, transpose_y = attn_449_transpose_y_1, x = q_1129_cast_fp16, y = var_13789_cast_fp16)[name = tensor<string, []>("attn_449_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_451_cast_fp16 = softmax(axis = var_21, x = attn_449_cast_fp16)[name = tensor<string, []>("attn_451_cast_fp16")];
            tensor<bool, []> var_13796_transpose_x_0 = const()[name = tensor<string, []>("op_13796_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13796_transpose_y_0 = const()[name = tensor<string, []>("op_13796_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13796_cast_fp16 = matmul(transpose_x = var_13796_transpose_x_0, transpose_y = var_13796_transpose_y_0, x = attn_451_cast_fp16, y = v_677_cast_fp16)[name = tensor<string, []>("op_13796_cast_fp16")];
            tensor<int32, [1]> var_13797_axes_0 = const()[name = tensor<string, []>("op_13797_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13797_cast_fp16 = squeeze(axes = var_13797_axes_0, x = var_13796_cast_fp16)[name = tensor<string, []>("op_13797_cast_fp16")];
            tensor<string, []> var_13803_pad_type_0 = const()[name = tensor<string, []>("op_13803_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13803_strides_0 = const()[name = tensor<string, []>("op_13803_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13803_pad_0 = const()[name = tensor<string, []>("op_13803_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13803_dilations_0 = const()[name = tensor<string, []>("op_13803_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13803_groups_0 = const()[name = tensor<string, []>("op_13803_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130697728)))];
            tensor<fp16, [1, 64, 1, 197]> var_13803_cast_fp16 = conv(dilations = var_13803_dilations_0, groups = var_13803_groups_0, pad = var_13803_pad_0, pad_type = var_13803_pad_type_0, strides = var_13803_strides_0, weight = model_blocks_9_attn_q_projs_5_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13803_cast_fp16")];
            tensor<int32, [1]> var_13804_axes_0 = const()[name = tensor<string, []>("op_13804_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13804_cast_fp16 = squeeze(axes = var_13804_axes_0, x = var_13803_cast_fp16)[name = tensor<string, []>("op_13804_cast_fp16")];
            tensor<int32, [3]> q_1131_perm_0 = const()[name = tensor<string, []>("q_1131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13811_pad_type_0 = const()[name = tensor<string, []>("op_13811_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13811_strides_0 = const()[name = tensor<string, []>("op_13811_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13811_pad_0 = const()[name = tensor<string, []>("op_13811_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13811_dilations_0 = const()[name = tensor<string, []>("op_13811_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13811_groups_0 = const()[name = tensor<string, []>("op_13811_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130796096)))];
            tensor<fp16, [1, 64, 1, 197]> var_13811_cast_fp16 = conv(dilations = var_13811_dilations_0, groups = var_13811_groups_0, pad = var_13811_pad_0, pad_type = var_13811_pad_type_0, strides = var_13811_strides_0, weight = model_blocks_9_attn_k_projs_5_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13811_cast_fp16")];
            tensor<int32, [1]> var_13812_axes_0 = const()[name = tensor<string, []>("op_13812_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13812_cast_fp16 = squeeze(axes = var_13812_axes_0, x = var_13811_cast_fp16)[name = tensor<string, []>("op_13812_cast_fp16")];
            tensor<int32, [3]> k_679_perm_0 = const()[name = tensor<string, []>("k_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13819_pad_type_0 = const()[name = tensor<string, []>("op_13819_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13819_strides_0 = const()[name = tensor<string, []>("op_13819_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13819_pad_0 = const()[name = tensor<string, []>("op_13819_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13819_dilations_0 = const()[name = tensor<string, []>("op_13819_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13819_groups_0 = const()[name = tensor<string, []>("op_13819_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130894464)))];
            tensor<fp16, [1, 64, 1, 197]> var_13819_cast_fp16 = conv(dilations = var_13819_dilations_0, groups = var_13819_groups_0, pad = var_13819_pad_0, pad_type = var_13819_pad_type_0, strides = var_13819_strides_0, weight = model_blocks_9_attn_v_projs_5_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13819_cast_fp16")];
            tensor<int32, [1]> var_13820_axes_0 = const()[name = tensor<string, []>("op_13820_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13820_cast_fp16 = squeeze(axes = var_13820_axes_0, x = var_13819_cast_fp16)[name = tensor<string, []>("op_13820_cast_fp16")];
            tensor<int32, [3]> v_679_perm_0 = const()[name = tensor<string, []>("v_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130992832)))];
            tensor<fp16, [1, 197, 64]> q_1131_cast_fp16 = transpose(perm = q_1131_perm_0, x = var_13804_cast_fp16)[name = tensor<string, []>("transpose_112")];
            tensor<fp16, [1, 197, 64]> q_1133_cast_fp16 = add(x = q_1131_cast_fp16, y = model_blocks_9_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_1133_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130993024)))];
            tensor<fp16, [1, 197, 64]> v_679_cast_fp16 = transpose(perm = v_679_perm_0, x = var_13820_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 197, 64]> v_681_cast_fp16 = add(x = v_679_cast_fp16, y = model_blocks_9_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_681_cast_fp16")];
            tensor<int32, [1]> q_1135_axes_0 = const()[name = tensor<string, []>("q_1135_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1135_cast_fp16 = expand_dims(axes = q_1135_axes_0, x = q_1133_cast_fp16)[name = tensor<string, []>("q_1135_cast_fp16")];
            tensor<int32, [1]> k_681_axes_0 = const()[name = tensor<string, []>("k_681_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_679_cast_fp16 = transpose(perm = k_679_perm_0, x = var_13812_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 1, 197, 64]> k_681_cast_fp16 = expand_dims(axes = k_681_axes_0, x = k_679_cast_fp16)[name = tensor<string, []>("k_681_cast_fp16")];
            tensor<int32, [1]> v_683_axes_0 = const()[name = tensor<string, []>("v_683_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_683_cast_fp16 = expand_dims(axes = v_683_axes_0, x = v_681_cast_fp16)[name = tensor<string, []>("v_683_cast_fp16")];
            tensor<int32, [4]> var_13829_begin_0 = const()[name = tensor<string, []>("op_13829_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13829_end_0 = const()[name = tensor<string, []>("op_13829_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13829_end_mask_0 = const()[name = tensor<string, []>("op_13829_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13829_cast_fp16 = slice_by_index(begin = var_13829_begin_0, end = var_13829_end_0, end_mask = var_13829_end_mask_0, x = q_1135_cast_fp16)[name = tensor<string, []>("op_13829_cast_fp16")];
            tensor<int32, [4]> var_13833_begin_0 = const()[name = tensor<string, []>("op_13833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13833_end_0 = const()[name = tensor<string, []>("op_13833_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13833_end_mask_0 = const()[name = tensor<string, []>("op_13833_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13833_cast_fp16 = slice_by_index(begin = var_13833_begin_0, end = var_13833_end_0, end_mask = var_13833_end_mask_0, x = k_681_cast_fp16)[name = tensor<string, []>("op_13833_cast_fp16")];
            tensor<int32, [4]> var_13837_begin_0 = const()[name = tensor<string, []>("op_13837_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13837_end_0 = const()[name = tensor<string, []>("op_13837_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13837_end_mask_0 = const()[name = tensor<string, []>("op_13837_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13837_cast_fp16 = slice_by_index(begin = var_13837_begin_0, end = var_13837_end_0, end_mask = var_13837_end_mask_0, x = q_1135_cast_fp16)[name = tensor<string, []>("op_13837_cast_fp16")];
            tensor<int32, [4]> var_13841_begin_0 = const()[name = tensor<string, []>("op_13841_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13841_end_0 = const()[name = tensor<string, []>("op_13841_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13841_end_mask_0 = const()[name = tensor<string, []>("op_13841_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13841_cast_fp16 = slice_by_index(begin = var_13841_begin_0, end = var_13841_end_0, end_mask = var_13841_end_mask_0, x = k_681_cast_fp16)[name = tensor<string, []>("op_13841_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13843_cast_fp16 = mul(x = var_13837_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13843_cast_fp16")];
            tensor<int32, [5]> var_13851 = const()[name = tensor<string, []>("op_13851"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1547_cast_fp16 = reshape(shape = var_13851, x = var_13837_cast_fp16)[name = tensor<string, []>("x_1547_cast_fp16")];
            tensor<int32, [2]> var_13853_split_sizes_0 = const()[name = tensor<string, []>("op_13853_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13853_axis_0 = const()[name = tensor<string, []>("op_13853_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13853_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13853_cast_fp16_1 = split(axis = var_13853_axis_0, split_sizes = var_13853_split_sizes_0, x = x_1547_cast_fp16)[name = tensor<string, []>("op_13853_cast_fp16")];
            tensor<int32, [1]> squeeze_452_axes_0 = const()[name = tensor<string, []>("squeeze_452_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_452_cast_fp16 = squeeze(axes = squeeze_452_axes_0, x = var_13853_cast_fp16_0)[name = tensor<string, []>("squeeze_452_cast_fp16")];
            tensor<int32, [1]> squeeze_453_axes_0 = const()[name = tensor<string, []>("squeeze_453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_453_cast_fp16 = squeeze(axes = squeeze_453_axes_0, x = var_13853_cast_fp16_1)[name = tensor<string, []>("squeeze_453_cast_fp16")];
            tensor<fp16, []> const_1851_promoted_to_fp16 = const()[name = tensor<string, []>("const_1851_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13856_cast_fp16 = mul(x = squeeze_453_cast_fp16, y = const_1851_promoted_to_fp16)[name = tensor<string, []>("op_13856_cast_fp16")];
            tensor<int32, []> x_1549_axis_0 = const()[name = tensor<string, []>("x_1549_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1549_cast_fp16 = stack(axis = x_1549_axis_0, values = (var_13856_cast_fp16, squeeze_452_cast_fp16))[name = tensor<string, []>("x_1549_cast_fp16")];
            tensor<int32, [4]> var_13862 = const()[name = tensor<string, []>("op_13862"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13863_cast_fp16 = reshape(shape = var_13862, x = x_1549_cast_fp16)[name = tensor<string, []>("op_13863_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13864_cast_fp16 = mul(x = var_13863_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13864_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_227_cast_fp16 = add(x = var_13843_cast_fp16, y = var_13864_cast_fp16)[name = tensor<string, []>("q_patches_227_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13866_cast_fp16 = mul(x = var_13841_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13866_cast_fp16")];
            tensor<int32, [5]> var_13874 = const()[name = tensor<string, []>("op_13874"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1553_cast_fp16 = reshape(shape = var_13874, x = var_13841_cast_fp16)[name = tensor<string, []>("x_1553_cast_fp16")];
            tensor<int32, [2]> var_13876_split_sizes_0 = const()[name = tensor<string, []>("op_13876_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13876_axis_0 = const()[name = tensor<string, []>("op_13876_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13876_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13876_cast_fp16_1 = split(axis = var_13876_axis_0, split_sizes = var_13876_split_sizes_0, x = x_1553_cast_fp16)[name = tensor<string, []>("op_13876_cast_fp16")];
            tensor<int32, [1]> squeeze_454_axes_0 = const()[name = tensor<string, []>("squeeze_454_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_454_cast_fp16 = squeeze(axes = squeeze_454_axes_0, x = var_13876_cast_fp16_0)[name = tensor<string, []>("squeeze_454_cast_fp16")];
            tensor<int32, [1]> squeeze_455_axes_0 = const()[name = tensor<string, []>("squeeze_455_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_455_cast_fp16 = squeeze(axes = squeeze_455_axes_0, x = var_13876_cast_fp16_1)[name = tensor<string, []>("squeeze_455_cast_fp16")];
            tensor<fp16, []> const_1859_promoted_to_fp16 = const()[name = tensor<string, []>("const_1859_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13879_cast_fp16 = mul(x = squeeze_455_cast_fp16, y = const_1859_promoted_to_fp16)[name = tensor<string, []>("op_13879_cast_fp16")];
            tensor<int32, []> x_1555_axis_0 = const()[name = tensor<string, []>("x_1555_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1555_cast_fp16 = stack(axis = x_1555_axis_0, values = (var_13879_cast_fp16, squeeze_454_cast_fp16))[name = tensor<string, []>("x_1555_cast_fp16")];
            tensor<int32, [4]> var_13885 = const()[name = tensor<string, []>("op_13885"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13886_cast_fp16 = reshape(shape = var_13885, x = x_1555_cast_fp16)[name = tensor<string, []>("op_13886_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13887_cast_fp16 = mul(x = var_13886_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13887_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_227_cast_fp16 = add(x = var_13866_cast_fp16, y = var_13887_cast_fp16)[name = tensor<string, []>("k_patches_227_cast_fp16")];
            tensor<bool, []> var_13890_interleave_0 = const()[name = tensor<string, []>("op_13890_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13890_cast_fp16 = concat(axis = var_20, interleave = var_13890_interleave_0, values = (var_13829_cast_fp16, q_patches_227_cast_fp16))[name = tensor<string, []>("op_13890_cast_fp16")];
            tensor<bool, []> var_13893_interleave_0 = const()[name = tensor<string, []>("op_13893_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13893_cast_fp16 = concat(axis = var_20, interleave = var_13893_interleave_0, values = (var_13833_cast_fp16, k_patches_227_cast_fp16))[name = tensor<string, []>("op_13893_cast_fp16")];
            tensor<fp16, []> var_13895_to_fp16 = const()[name = tensor<string, []>("op_13895_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1139_cast_fp16 = mul(x = var_13890_cast_fp16, y = var_13895_to_fp16)[name = tensor<string, []>("q_1139_cast_fp16")];
            tensor<bool, []> attn_453_transpose_x_1 = const()[name = tensor<string, []>("attn_453_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_453_transpose_y_1 = const()[name = tensor<string, []>("attn_453_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_453_cast_fp16 = matmul(transpose_x = attn_453_transpose_x_1, transpose_y = attn_453_transpose_y_1, x = q_1139_cast_fp16, y = var_13893_cast_fp16)[name = tensor<string, []>("attn_453_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_455_cast_fp16 = softmax(axis = var_21, x = attn_453_cast_fp16)[name = tensor<string, []>("attn_455_cast_fp16")];
            tensor<bool, []> var_13900_transpose_x_0 = const()[name = tensor<string, []>("op_13900_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_13900_transpose_y_0 = const()[name = tensor<string, []>("op_13900_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13900_cast_fp16 = matmul(transpose_x = var_13900_transpose_x_0, transpose_y = var_13900_transpose_y_0, x = attn_455_cast_fp16, y = v_683_cast_fp16)[name = tensor<string, []>("op_13900_cast_fp16")];
            tensor<int32, [1]> var_13901_axes_0 = const()[name = tensor<string, []>("op_13901_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_13901_cast_fp16 = squeeze(axes = var_13901_axes_0, x = var_13900_cast_fp16)[name = tensor<string, []>("op_13901_cast_fp16")];
            tensor<string, []> var_13907_pad_type_0 = const()[name = tensor<string, []>("op_13907_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13907_strides_0 = const()[name = tensor<string, []>("op_13907_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13907_pad_0 = const()[name = tensor<string, []>("op_13907_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13907_dilations_0 = const()[name = tensor<string, []>("op_13907_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13907_groups_0 = const()[name = tensor<string, []>("op_13907_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130993216)))];
            tensor<fp16, [1, 64, 1, 197]> var_13907_cast_fp16 = conv(dilations = var_13907_dilations_0, groups = var_13907_groups_0, pad = var_13907_pad_0, pad_type = var_13907_pad_type_0, strides = var_13907_strides_0, weight = model_blocks_9_attn_q_projs_6_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13907_cast_fp16")];
            tensor<int32, [1]> var_13908_axes_0 = const()[name = tensor<string, []>("op_13908_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13908_cast_fp16 = squeeze(axes = var_13908_axes_0, x = var_13907_cast_fp16)[name = tensor<string, []>("op_13908_cast_fp16")];
            tensor<int32, [3]> q_1141_perm_0 = const()[name = tensor<string, []>("q_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13915_pad_type_0 = const()[name = tensor<string, []>("op_13915_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13915_strides_0 = const()[name = tensor<string, []>("op_13915_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13915_pad_0 = const()[name = tensor<string, []>("op_13915_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13915_dilations_0 = const()[name = tensor<string, []>("op_13915_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13915_groups_0 = const()[name = tensor<string, []>("op_13915_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131091584)))];
            tensor<fp16, [1, 64, 1, 197]> var_13915_cast_fp16 = conv(dilations = var_13915_dilations_0, groups = var_13915_groups_0, pad = var_13915_pad_0, pad_type = var_13915_pad_type_0, strides = var_13915_strides_0, weight = model_blocks_9_attn_k_projs_6_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13915_cast_fp16")];
            tensor<int32, [1]> var_13916_axes_0 = const()[name = tensor<string, []>("op_13916_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13916_cast_fp16 = squeeze(axes = var_13916_axes_0, x = var_13915_cast_fp16)[name = tensor<string, []>("op_13916_cast_fp16")];
            tensor<int32, [3]> k_685_perm_0 = const()[name = tensor<string, []>("k_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_13923_pad_type_0 = const()[name = tensor<string, []>("op_13923_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_13923_strides_0 = const()[name = tensor<string, []>("op_13923_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_13923_pad_0 = const()[name = tensor<string, []>("op_13923_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_13923_dilations_0 = const()[name = tensor<string, []>("op_13923_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13923_groups_0 = const()[name = tensor<string, []>("op_13923_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131189952)))];
            tensor<fp16, [1, 64, 1, 197]> var_13923_cast_fp16 = conv(dilations = var_13923_dilations_0, groups = var_13923_groups_0, pad = var_13923_pad_0, pad_type = var_13923_pad_type_0, strides = var_13923_strides_0, weight = model_blocks_9_attn_v_projs_6_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_13923_cast_fp16")];
            tensor<int32, [1]> var_13924_axes_0 = const()[name = tensor<string, []>("op_13924_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_13924_cast_fp16 = squeeze(axes = var_13924_axes_0, x = var_13923_cast_fp16)[name = tensor<string, []>("op_13924_cast_fp16")];
            tensor<int32, [3]> v_685_perm_0 = const()[name = tensor<string, []>("v_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131288320)))];
            tensor<fp16, [1, 197, 64]> q_1141_cast_fp16 = transpose(perm = q_1141_perm_0, x = var_13908_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [1, 197, 64]> q_1143_cast_fp16 = add(x = q_1141_cast_fp16, y = model_blocks_9_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_1143_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131288512)))];
            tensor<fp16, [1, 197, 64]> v_685_cast_fp16 = transpose(perm = v_685_perm_0, x = var_13924_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 197, 64]> v_687_cast_fp16 = add(x = v_685_cast_fp16, y = model_blocks_9_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_687_cast_fp16")];
            tensor<int32, [1]> q_1145_axes_0 = const()[name = tensor<string, []>("q_1145_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1145_cast_fp16 = expand_dims(axes = q_1145_axes_0, x = q_1143_cast_fp16)[name = tensor<string, []>("q_1145_cast_fp16")];
            tensor<int32, [1]> k_687_axes_0 = const()[name = tensor<string, []>("k_687_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_685_cast_fp16 = transpose(perm = k_685_perm_0, x = var_13916_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [1, 1, 197, 64]> k_687_cast_fp16 = expand_dims(axes = k_687_axes_0, x = k_685_cast_fp16)[name = tensor<string, []>("k_687_cast_fp16")];
            tensor<int32, [1]> v_689_axes_0 = const()[name = tensor<string, []>("v_689_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_689_cast_fp16 = expand_dims(axes = v_689_axes_0, x = v_687_cast_fp16)[name = tensor<string, []>("v_689_cast_fp16")];
            tensor<int32, [4]> var_13933_begin_0 = const()[name = tensor<string, []>("op_13933_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13933_end_0 = const()[name = tensor<string, []>("op_13933_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13933_end_mask_0 = const()[name = tensor<string, []>("op_13933_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13933_cast_fp16 = slice_by_index(begin = var_13933_begin_0, end = var_13933_end_0, end_mask = var_13933_end_mask_0, x = q_1145_cast_fp16)[name = tensor<string, []>("op_13933_cast_fp16")];
            tensor<int32, [4]> var_13937_begin_0 = const()[name = tensor<string, []>("op_13937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13937_end_0 = const()[name = tensor<string, []>("op_13937_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_13937_end_mask_0 = const()[name = tensor<string, []>("op_13937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_13937_cast_fp16 = slice_by_index(begin = var_13937_begin_0, end = var_13937_end_0, end_mask = var_13937_end_mask_0, x = k_687_cast_fp16)[name = tensor<string, []>("op_13937_cast_fp16")];
            tensor<int32, [4]> var_13941_begin_0 = const()[name = tensor<string, []>("op_13941_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13941_end_0 = const()[name = tensor<string, []>("op_13941_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13941_end_mask_0 = const()[name = tensor<string, []>("op_13941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13941_cast_fp16 = slice_by_index(begin = var_13941_begin_0, end = var_13941_end_0, end_mask = var_13941_end_mask_0, x = q_1145_cast_fp16)[name = tensor<string, []>("op_13941_cast_fp16")];
            tensor<int32, [4]> var_13945_begin_0 = const()[name = tensor<string, []>("op_13945_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13945_end_0 = const()[name = tensor<string, []>("op_13945_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_13945_end_mask_0 = const()[name = tensor<string, []>("op_13945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_13945_cast_fp16 = slice_by_index(begin = var_13945_begin_0, end = var_13945_end_0, end_mask = var_13945_end_mask_0, x = k_687_cast_fp16)[name = tensor<string, []>("op_13945_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13947_cast_fp16 = mul(x = var_13941_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13947_cast_fp16")];
            tensor<int32, [5]> var_13955 = const()[name = tensor<string, []>("op_13955"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1559_cast_fp16 = reshape(shape = var_13955, x = var_13941_cast_fp16)[name = tensor<string, []>("x_1559_cast_fp16")];
            tensor<int32, [2]> var_13957_split_sizes_0 = const()[name = tensor<string, []>("op_13957_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13957_axis_0 = const()[name = tensor<string, []>("op_13957_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13957_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13957_cast_fp16_1 = split(axis = var_13957_axis_0, split_sizes = var_13957_split_sizes_0, x = x_1559_cast_fp16)[name = tensor<string, []>("op_13957_cast_fp16")];
            tensor<int32, [1]> squeeze_456_axes_0 = const()[name = tensor<string, []>("squeeze_456_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_456_cast_fp16 = squeeze(axes = squeeze_456_axes_0, x = var_13957_cast_fp16_0)[name = tensor<string, []>("squeeze_456_cast_fp16")];
            tensor<int32, [1]> squeeze_457_axes_0 = const()[name = tensor<string, []>("squeeze_457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_457_cast_fp16 = squeeze(axes = squeeze_457_axes_0, x = var_13957_cast_fp16_1)[name = tensor<string, []>("squeeze_457_cast_fp16")];
            tensor<fp16, []> const_1867_promoted_to_fp16 = const()[name = tensor<string, []>("const_1867_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13960_cast_fp16 = mul(x = squeeze_457_cast_fp16, y = const_1867_promoted_to_fp16)[name = tensor<string, []>("op_13960_cast_fp16")];
            tensor<int32, []> x_1561_axis_0 = const()[name = tensor<string, []>("x_1561_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1561_cast_fp16 = stack(axis = x_1561_axis_0, values = (var_13960_cast_fp16, squeeze_456_cast_fp16))[name = tensor<string, []>("x_1561_cast_fp16")];
            tensor<int32, [4]> var_13966 = const()[name = tensor<string, []>("op_13966"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13967_cast_fp16 = reshape(shape = var_13966, x = x_1561_cast_fp16)[name = tensor<string, []>("op_13967_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13968_cast_fp16 = mul(x = var_13967_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13968_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_229_cast_fp16 = add(x = var_13947_cast_fp16, y = var_13968_cast_fp16)[name = tensor<string, []>("q_patches_229_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13970_cast_fp16 = mul(x = var_13945_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_13970_cast_fp16")];
            tensor<int32, [5]> var_13978 = const()[name = tensor<string, []>("op_13978"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1565_cast_fp16 = reshape(shape = var_13978, x = var_13945_cast_fp16)[name = tensor<string, []>("x_1565_cast_fp16")];
            tensor<int32, [2]> var_13980_split_sizes_0 = const()[name = tensor<string, []>("op_13980_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_13980_axis_0 = const()[name = tensor<string, []>("op_13980_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_13980_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_13980_cast_fp16_1 = split(axis = var_13980_axis_0, split_sizes = var_13980_split_sizes_0, x = x_1565_cast_fp16)[name = tensor<string, []>("op_13980_cast_fp16")];
            tensor<int32, [1]> squeeze_458_axes_0 = const()[name = tensor<string, []>("squeeze_458_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_458_cast_fp16 = squeeze(axes = squeeze_458_axes_0, x = var_13980_cast_fp16_0)[name = tensor<string, []>("squeeze_458_cast_fp16")];
            tensor<int32, [1]> squeeze_459_axes_0 = const()[name = tensor<string, []>("squeeze_459_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_459_cast_fp16 = squeeze(axes = squeeze_459_axes_0, x = var_13980_cast_fp16_1)[name = tensor<string, []>("squeeze_459_cast_fp16")];
            tensor<fp16, []> const_1875_promoted_to_fp16 = const()[name = tensor<string, []>("const_1875_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_13983_cast_fp16 = mul(x = squeeze_459_cast_fp16, y = const_1875_promoted_to_fp16)[name = tensor<string, []>("op_13983_cast_fp16")];
            tensor<int32, []> x_1567_axis_0 = const()[name = tensor<string, []>("x_1567_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1567_cast_fp16 = stack(axis = x_1567_axis_0, values = (var_13983_cast_fp16, squeeze_458_cast_fp16))[name = tensor<string, []>("x_1567_cast_fp16")];
            tensor<int32, [4]> var_13989 = const()[name = tensor<string, []>("op_13989"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_13990_cast_fp16 = reshape(shape = var_13989, x = x_1567_cast_fp16)[name = tensor<string, []>("op_13990_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_13991_cast_fp16 = mul(x = var_13990_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_13991_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_229_cast_fp16 = add(x = var_13970_cast_fp16, y = var_13991_cast_fp16)[name = tensor<string, []>("k_patches_229_cast_fp16")];
            tensor<bool, []> var_13994_interleave_0 = const()[name = tensor<string, []>("op_13994_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13994_cast_fp16 = concat(axis = var_20, interleave = var_13994_interleave_0, values = (var_13933_cast_fp16, q_patches_229_cast_fp16))[name = tensor<string, []>("op_13994_cast_fp16")];
            tensor<bool, []> var_13997_interleave_0 = const()[name = tensor<string, []>("op_13997_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_13997_cast_fp16 = concat(axis = var_20, interleave = var_13997_interleave_0, values = (var_13937_cast_fp16, k_patches_229_cast_fp16))[name = tensor<string, []>("op_13997_cast_fp16")];
            tensor<fp16, []> var_13999_to_fp16 = const()[name = tensor<string, []>("op_13999_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1149_cast_fp16 = mul(x = var_13994_cast_fp16, y = var_13999_to_fp16)[name = tensor<string, []>("q_1149_cast_fp16")];
            tensor<bool, []> attn_457_transpose_x_1 = const()[name = tensor<string, []>("attn_457_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_457_transpose_y_1 = const()[name = tensor<string, []>("attn_457_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_457_cast_fp16 = matmul(transpose_x = attn_457_transpose_x_1, transpose_y = attn_457_transpose_y_1, x = q_1149_cast_fp16, y = var_13997_cast_fp16)[name = tensor<string, []>("attn_457_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_459_cast_fp16 = softmax(axis = var_21, x = attn_457_cast_fp16)[name = tensor<string, []>("attn_459_cast_fp16")];
            tensor<bool, []> var_14004_transpose_x_0 = const()[name = tensor<string, []>("op_14004_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14004_transpose_y_0 = const()[name = tensor<string, []>("op_14004_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14004_cast_fp16 = matmul(transpose_x = var_14004_transpose_x_0, transpose_y = var_14004_transpose_y_0, x = attn_459_cast_fp16, y = v_689_cast_fp16)[name = tensor<string, []>("op_14004_cast_fp16")];
            tensor<int32, [1]> var_14005_axes_0 = const()[name = tensor<string, []>("op_14005_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14005_cast_fp16 = squeeze(axes = var_14005_axes_0, x = var_14004_cast_fp16)[name = tensor<string, []>("op_14005_cast_fp16")];
            tensor<string, []> var_14011_pad_type_0 = const()[name = tensor<string, []>("op_14011_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14011_strides_0 = const()[name = tensor<string, []>("op_14011_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14011_pad_0 = const()[name = tensor<string, []>("op_14011_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14011_dilations_0 = const()[name = tensor<string, []>("op_14011_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14011_groups_0 = const()[name = tensor<string, []>("op_14011_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131288704)))];
            tensor<fp16, [1, 64, 1, 197]> var_14011_cast_fp16 = conv(dilations = var_14011_dilations_0, groups = var_14011_groups_0, pad = var_14011_pad_0, pad_type = var_14011_pad_type_0, strides = var_14011_strides_0, weight = model_blocks_9_attn_q_projs_7_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14011_cast_fp16")];
            tensor<int32, [1]> var_14012_axes_0 = const()[name = tensor<string, []>("op_14012_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14012_cast_fp16 = squeeze(axes = var_14012_axes_0, x = var_14011_cast_fp16)[name = tensor<string, []>("op_14012_cast_fp16")];
            tensor<int32, [3]> q_1151_perm_0 = const()[name = tensor<string, []>("q_1151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14019_pad_type_0 = const()[name = tensor<string, []>("op_14019_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14019_strides_0 = const()[name = tensor<string, []>("op_14019_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14019_pad_0 = const()[name = tensor<string, []>("op_14019_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14019_dilations_0 = const()[name = tensor<string, []>("op_14019_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14019_groups_0 = const()[name = tensor<string, []>("op_14019_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131387072)))];
            tensor<fp16, [1, 64, 1, 197]> var_14019_cast_fp16 = conv(dilations = var_14019_dilations_0, groups = var_14019_groups_0, pad = var_14019_pad_0, pad_type = var_14019_pad_type_0, strides = var_14019_strides_0, weight = model_blocks_9_attn_k_projs_7_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14019_cast_fp16")];
            tensor<int32, [1]> var_14020_axes_0 = const()[name = tensor<string, []>("op_14020_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14020_cast_fp16 = squeeze(axes = var_14020_axes_0, x = var_14019_cast_fp16)[name = tensor<string, []>("op_14020_cast_fp16")];
            tensor<int32, [3]> k_691_perm_0 = const()[name = tensor<string, []>("k_691_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14027_pad_type_0 = const()[name = tensor<string, []>("op_14027_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14027_strides_0 = const()[name = tensor<string, []>("op_14027_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14027_pad_0 = const()[name = tensor<string, []>("op_14027_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14027_dilations_0 = const()[name = tensor<string, []>("op_14027_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14027_groups_0 = const()[name = tensor<string, []>("op_14027_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131485440)))];
            tensor<fp16, [1, 64, 1, 197]> var_14027_cast_fp16 = conv(dilations = var_14027_dilations_0, groups = var_14027_groups_0, pad = var_14027_pad_0, pad_type = var_14027_pad_type_0, strides = var_14027_strides_0, weight = model_blocks_9_attn_v_projs_7_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14027_cast_fp16")];
            tensor<int32, [1]> var_14028_axes_0 = const()[name = tensor<string, []>("op_14028_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14028_cast_fp16 = squeeze(axes = var_14028_axes_0, x = var_14027_cast_fp16)[name = tensor<string, []>("op_14028_cast_fp16")];
            tensor<int32, [3]> v_691_perm_0 = const()[name = tensor<string, []>("v_691_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131583808)))];
            tensor<fp16, [1, 197, 64]> q_1151_cast_fp16 = transpose(perm = q_1151_perm_0, x = var_14012_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [1, 197, 64]> q_1153_cast_fp16 = add(x = q_1151_cast_fp16, y = model_blocks_9_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_1153_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131584000)))];
            tensor<fp16, [1, 197, 64]> v_691_cast_fp16 = transpose(perm = v_691_perm_0, x = var_14028_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [1, 197, 64]> v_693_cast_fp16 = add(x = v_691_cast_fp16, y = model_blocks_9_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_693_cast_fp16")];
            tensor<int32, [1]> q_1155_axes_0 = const()[name = tensor<string, []>("q_1155_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1155_cast_fp16 = expand_dims(axes = q_1155_axes_0, x = q_1153_cast_fp16)[name = tensor<string, []>("q_1155_cast_fp16")];
            tensor<int32, [1]> k_693_axes_0 = const()[name = tensor<string, []>("k_693_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_691_cast_fp16 = transpose(perm = k_691_perm_0, x = var_14020_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [1, 1, 197, 64]> k_693_cast_fp16 = expand_dims(axes = k_693_axes_0, x = k_691_cast_fp16)[name = tensor<string, []>("k_693_cast_fp16")];
            tensor<int32, [1]> v_695_axes_0 = const()[name = tensor<string, []>("v_695_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_695_cast_fp16 = expand_dims(axes = v_695_axes_0, x = v_693_cast_fp16)[name = tensor<string, []>("v_695_cast_fp16")];
            tensor<int32, [4]> var_14037_begin_0 = const()[name = tensor<string, []>("op_14037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14037_end_0 = const()[name = tensor<string, []>("op_14037_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14037_end_mask_0 = const()[name = tensor<string, []>("op_14037_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14037_cast_fp16 = slice_by_index(begin = var_14037_begin_0, end = var_14037_end_0, end_mask = var_14037_end_mask_0, x = q_1155_cast_fp16)[name = tensor<string, []>("op_14037_cast_fp16")];
            tensor<int32, [4]> var_14041_begin_0 = const()[name = tensor<string, []>("op_14041_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14041_end_0 = const()[name = tensor<string, []>("op_14041_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14041_end_mask_0 = const()[name = tensor<string, []>("op_14041_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14041_cast_fp16 = slice_by_index(begin = var_14041_begin_0, end = var_14041_end_0, end_mask = var_14041_end_mask_0, x = k_693_cast_fp16)[name = tensor<string, []>("op_14041_cast_fp16")];
            tensor<int32, [4]> var_14045_begin_0 = const()[name = tensor<string, []>("op_14045_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14045_end_0 = const()[name = tensor<string, []>("op_14045_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14045_end_mask_0 = const()[name = tensor<string, []>("op_14045_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14045_cast_fp16 = slice_by_index(begin = var_14045_begin_0, end = var_14045_end_0, end_mask = var_14045_end_mask_0, x = q_1155_cast_fp16)[name = tensor<string, []>("op_14045_cast_fp16")];
            tensor<int32, [4]> var_14049_begin_0 = const()[name = tensor<string, []>("op_14049_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14049_end_0 = const()[name = tensor<string, []>("op_14049_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14049_end_mask_0 = const()[name = tensor<string, []>("op_14049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14049_cast_fp16 = slice_by_index(begin = var_14049_begin_0, end = var_14049_end_0, end_mask = var_14049_end_mask_0, x = k_693_cast_fp16)[name = tensor<string, []>("op_14049_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14051_cast_fp16 = mul(x = var_14045_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14051_cast_fp16")];
            tensor<int32, [5]> var_14059 = const()[name = tensor<string, []>("op_14059"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1571_cast_fp16 = reshape(shape = var_14059, x = var_14045_cast_fp16)[name = tensor<string, []>("x_1571_cast_fp16")];
            tensor<int32, [2]> var_14061_split_sizes_0 = const()[name = tensor<string, []>("op_14061_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14061_axis_0 = const()[name = tensor<string, []>("op_14061_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14061_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14061_cast_fp16_1 = split(axis = var_14061_axis_0, split_sizes = var_14061_split_sizes_0, x = x_1571_cast_fp16)[name = tensor<string, []>("op_14061_cast_fp16")];
            tensor<int32, [1]> squeeze_460_axes_0 = const()[name = tensor<string, []>("squeeze_460_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_460_cast_fp16 = squeeze(axes = squeeze_460_axes_0, x = var_14061_cast_fp16_0)[name = tensor<string, []>("squeeze_460_cast_fp16")];
            tensor<int32, [1]> squeeze_461_axes_0 = const()[name = tensor<string, []>("squeeze_461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_461_cast_fp16 = squeeze(axes = squeeze_461_axes_0, x = var_14061_cast_fp16_1)[name = tensor<string, []>("squeeze_461_cast_fp16")];
            tensor<fp16, []> const_1883_promoted_to_fp16 = const()[name = tensor<string, []>("const_1883_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14064_cast_fp16 = mul(x = squeeze_461_cast_fp16, y = const_1883_promoted_to_fp16)[name = tensor<string, []>("op_14064_cast_fp16")];
            tensor<int32, []> x_1573_axis_0 = const()[name = tensor<string, []>("x_1573_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1573_cast_fp16 = stack(axis = x_1573_axis_0, values = (var_14064_cast_fp16, squeeze_460_cast_fp16))[name = tensor<string, []>("x_1573_cast_fp16")];
            tensor<int32, [4]> var_14070 = const()[name = tensor<string, []>("op_14070"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14071_cast_fp16 = reshape(shape = var_14070, x = x_1573_cast_fp16)[name = tensor<string, []>("op_14071_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14072_cast_fp16 = mul(x = var_14071_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14072_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_231_cast_fp16 = add(x = var_14051_cast_fp16, y = var_14072_cast_fp16)[name = tensor<string, []>("q_patches_231_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14074_cast_fp16 = mul(x = var_14049_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14074_cast_fp16")];
            tensor<int32, [5]> var_14082 = const()[name = tensor<string, []>("op_14082"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1577_cast_fp16 = reshape(shape = var_14082, x = var_14049_cast_fp16)[name = tensor<string, []>("x_1577_cast_fp16")];
            tensor<int32, [2]> var_14084_split_sizes_0 = const()[name = tensor<string, []>("op_14084_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14084_axis_0 = const()[name = tensor<string, []>("op_14084_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14084_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14084_cast_fp16_1 = split(axis = var_14084_axis_0, split_sizes = var_14084_split_sizes_0, x = x_1577_cast_fp16)[name = tensor<string, []>("op_14084_cast_fp16")];
            tensor<int32, [1]> squeeze_462_axes_0 = const()[name = tensor<string, []>("squeeze_462_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_462_cast_fp16 = squeeze(axes = squeeze_462_axes_0, x = var_14084_cast_fp16_0)[name = tensor<string, []>("squeeze_462_cast_fp16")];
            tensor<int32, [1]> squeeze_463_axes_0 = const()[name = tensor<string, []>("squeeze_463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_463_cast_fp16 = squeeze(axes = squeeze_463_axes_0, x = var_14084_cast_fp16_1)[name = tensor<string, []>("squeeze_463_cast_fp16")];
            tensor<fp16, []> const_1891_promoted_to_fp16 = const()[name = tensor<string, []>("const_1891_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14087_cast_fp16 = mul(x = squeeze_463_cast_fp16, y = const_1891_promoted_to_fp16)[name = tensor<string, []>("op_14087_cast_fp16")];
            tensor<int32, []> x_1579_axis_0 = const()[name = tensor<string, []>("x_1579_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1579_cast_fp16 = stack(axis = x_1579_axis_0, values = (var_14087_cast_fp16, squeeze_462_cast_fp16))[name = tensor<string, []>("x_1579_cast_fp16")];
            tensor<int32, [4]> var_14093 = const()[name = tensor<string, []>("op_14093"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14094_cast_fp16 = reshape(shape = var_14093, x = x_1579_cast_fp16)[name = tensor<string, []>("op_14094_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14095_cast_fp16 = mul(x = var_14094_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14095_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_231_cast_fp16 = add(x = var_14074_cast_fp16, y = var_14095_cast_fp16)[name = tensor<string, []>("k_patches_231_cast_fp16")];
            tensor<bool, []> var_14098_interleave_0 = const()[name = tensor<string, []>("op_14098_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14098_cast_fp16 = concat(axis = var_20, interleave = var_14098_interleave_0, values = (var_14037_cast_fp16, q_patches_231_cast_fp16))[name = tensor<string, []>("op_14098_cast_fp16")];
            tensor<bool, []> var_14101_interleave_0 = const()[name = tensor<string, []>("op_14101_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14101_cast_fp16 = concat(axis = var_20, interleave = var_14101_interleave_0, values = (var_14041_cast_fp16, k_patches_231_cast_fp16))[name = tensor<string, []>("op_14101_cast_fp16")];
            tensor<fp16, []> var_14103_to_fp16 = const()[name = tensor<string, []>("op_14103_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1159_cast_fp16 = mul(x = var_14098_cast_fp16, y = var_14103_to_fp16)[name = tensor<string, []>("q_1159_cast_fp16")];
            tensor<bool, []> attn_461_transpose_x_1 = const()[name = tensor<string, []>("attn_461_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_461_transpose_y_1 = const()[name = tensor<string, []>("attn_461_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_461_cast_fp16 = matmul(transpose_x = attn_461_transpose_x_1, transpose_y = attn_461_transpose_y_1, x = q_1159_cast_fp16, y = var_14101_cast_fp16)[name = tensor<string, []>("attn_461_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_463_cast_fp16 = softmax(axis = var_21, x = attn_461_cast_fp16)[name = tensor<string, []>("attn_463_cast_fp16")];
            tensor<bool, []> var_14108_transpose_x_0 = const()[name = tensor<string, []>("op_14108_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14108_transpose_y_0 = const()[name = tensor<string, []>("op_14108_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14108_cast_fp16 = matmul(transpose_x = var_14108_transpose_x_0, transpose_y = var_14108_transpose_y_0, x = attn_463_cast_fp16, y = v_695_cast_fp16)[name = tensor<string, []>("op_14108_cast_fp16")];
            tensor<int32, [1]> var_14109_axes_0 = const()[name = tensor<string, []>("op_14109_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14109_cast_fp16 = squeeze(axes = var_14109_axes_0, x = var_14108_cast_fp16)[name = tensor<string, []>("op_14109_cast_fp16")];
            tensor<string, []> var_14115_pad_type_0 = const()[name = tensor<string, []>("op_14115_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14115_strides_0 = const()[name = tensor<string, []>("op_14115_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14115_pad_0 = const()[name = tensor<string, []>("op_14115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14115_dilations_0 = const()[name = tensor<string, []>("op_14115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14115_groups_0 = const()[name = tensor<string, []>("op_14115_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131584192)))];
            tensor<fp16, [1, 64, 1, 197]> var_14115_cast_fp16 = conv(dilations = var_14115_dilations_0, groups = var_14115_groups_0, pad = var_14115_pad_0, pad_type = var_14115_pad_type_0, strides = var_14115_strides_0, weight = model_blocks_9_attn_q_projs_8_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14115_cast_fp16")];
            tensor<int32, [1]> var_14116_axes_0 = const()[name = tensor<string, []>("op_14116_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14116_cast_fp16 = squeeze(axes = var_14116_axes_0, x = var_14115_cast_fp16)[name = tensor<string, []>("op_14116_cast_fp16")];
            tensor<int32, [3]> q_1161_perm_0 = const()[name = tensor<string, []>("q_1161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14123_pad_type_0 = const()[name = tensor<string, []>("op_14123_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14123_strides_0 = const()[name = tensor<string, []>("op_14123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14123_pad_0 = const()[name = tensor<string, []>("op_14123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14123_dilations_0 = const()[name = tensor<string, []>("op_14123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14123_groups_0 = const()[name = tensor<string, []>("op_14123_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131682560)))];
            tensor<fp16, [1, 64, 1, 197]> var_14123_cast_fp16 = conv(dilations = var_14123_dilations_0, groups = var_14123_groups_0, pad = var_14123_pad_0, pad_type = var_14123_pad_type_0, strides = var_14123_strides_0, weight = model_blocks_9_attn_k_projs_8_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14123_cast_fp16")];
            tensor<int32, [1]> var_14124_axes_0 = const()[name = tensor<string, []>("op_14124_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14124_cast_fp16 = squeeze(axes = var_14124_axes_0, x = var_14123_cast_fp16)[name = tensor<string, []>("op_14124_cast_fp16")];
            tensor<int32, [3]> k_697_perm_0 = const()[name = tensor<string, []>("k_697_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14131_pad_type_0 = const()[name = tensor<string, []>("op_14131_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14131_strides_0 = const()[name = tensor<string, []>("op_14131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14131_pad_0 = const()[name = tensor<string, []>("op_14131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14131_dilations_0 = const()[name = tensor<string, []>("op_14131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14131_groups_0 = const()[name = tensor<string, []>("op_14131_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131780928)))];
            tensor<fp16, [1, 64, 1, 197]> var_14131_cast_fp16 = conv(dilations = var_14131_dilations_0, groups = var_14131_groups_0, pad = var_14131_pad_0, pad_type = var_14131_pad_type_0, strides = var_14131_strides_0, weight = model_blocks_9_attn_v_projs_8_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14131_cast_fp16")];
            tensor<int32, [1]> var_14132_axes_0 = const()[name = tensor<string, []>("op_14132_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14132_cast_fp16 = squeeze(axes = var_14132_axes_0, x = var_14131_cast_fp16)[name = tensor<string, []>("op_14132_cast_fp16")];
            tensor<int32, [3]> v_697_perm_0 = const()[name = tensor<string, []>("v_697_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131879296)))];
            tensor<fp16, [1, 197, 64]> q_1161_cast_fp16 = transpose(perm = q_1161_perm_0, x = var_14116_cast_fp16)[name = tensor<string, []>("transpose_103")];
            tensor<fp16, [1, 197, 64]> q_1163_cast_fp16 = add(x = q_1161_cast_fp16, y = model_blocks_9_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_1163_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131879488)))];
            tensor<fp16, [1, 197, 64]> v_697_cast_fp16 = transpose(perm = v_697_perm_0, x = var_14132_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [1, 197, 64]> v_699_cast_fp16 = add(x = v_697_cast_fp16, y = model_blocks_9_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_699_cast_fp16")];
            tensor<int32, [1]> q_1165_axes_0 = const()[name = tensor<string, []>("q_1165_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1165_cast_fp16 = expand_dims(axes = q_1165_axes_0, x = q_1163_cast_fp16)[name = tensor<string, []>("q_1165_cast_fp16")];
            tensor<int32, [1]> k_699_axes_0 = const()[name = tensor<string, []>("k_699_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_697_cast_fp16 = transpose(perm = k_697_perm_0, x = var_14124_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [1, 1, 197, 64]> k_699_cast_fp16 = expand_dims(axes = k_699_axes_0, x = k_697_cast_fp16)[name = tensor<string, []>("k_699_cast_fp16")];
            tensor<int32, [1]> v_701_axes_0 = const()[name = tensor<string, []>("v_701_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_701_cast_fp16 = expand_dims(axes = v_701_axes_0, x = v_699_cast_fp16)[name = tensor<string, []>("v_701_cast_fp16")];
            tensor<int32, [4]> var_14141_begin_0 = const()[name = tensor<string, []>("op_14141_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14141_end_0 = const()[name = tensor<string, []>("op_14141_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14141_end_mask_0 = const()[name = tensor<string, []>("op_14141_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14141_cast_fp16 = slice_by_index(begin = var_14141_begin_0, end = var_14141_end_0, end_mask = var_14141_end_mask_0, x = q_1165_cast_fp16)[name = tensor<string, []>("op_14141_cast_fp16")];
            tensor<int32, [4]> var_14145_begin_0 = const()[name = tensor<string, []>("op_14145_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14145_end_0 = const()[name = tensor<string, []>("op_14145_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14145_end_mask_0 = const()[name = tensor<string, []>("op_14145_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14145_cast_fp16 = slice_by_index(begin = var_14145_begin_0, end = var_14145_end_0, end_mask = var_14145_end_mask_0, x = k_699_cast_fp16)[name = tensor<string, []>("op_14145_cast_fp16")];
            tensor<int32, [4]> var_14149_begin_0 = const()[name = tensor<string, []>("op_14149_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14149_end_0 = const()[name = tensor<string, []>("op_14149_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14149_end_mask_0 = const()[name = tensor<string, []>("op_14149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14149_cast_fp16 = slice_by_index(begin = var_14149_begin_0, end = var_14149_end_0, end_mask = var_14149_end_mask_0, x = q_1165_cast_fp16)[name = tensor<string, []>("op_14149_cast_fp16")];
            tensor<int32, [4]> var_14153_begin_0 = const()[name = tensor<string, []>("op_14153_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14153_end_0 = const()[name = tensor<string, []>("op_14153_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14153_end_mask_0 = const()[name = tensor<string, []>("op_14153_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14153_cast_fp16 = slice_by_index(begin = var_14153_begin_0, end = var_14153_end_0, end_mask = var_14153_end_mask_0, x = k_699_cast_fp16)[name = tensor<string, []>("op_14153_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14155_cast_fp16 = mul(x = var_14149_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14155_cast_fp16")];
            tensor<int32, [5]> var_14163 = const()[name = tensor<string, []>("op_14163"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1583_cast_fp16 = reshape(shape = var_14163, x = var_14149_cast_fp16)[name = tensor<string, []>("x_1583_cast_fp16")];
            tensor<int32, [2]> var_14165_split_sizes_0 = const()[name = tensor<string, []>("op_14165_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14165_axis_0 = const()[name = tensor<string, []>("op_14165_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14165_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14165_cast_fp16_1 = split(axis = var_14165_axis_0, split_sizes = var_14165_split_sizes_0, x = x_1583_cast_fp16)[name = tensor<string, []>("op_14165_cast_fp16")];
            tensor<int32, [1]> squeeze_464_axes_0 = const()[name = tensor<string, []>("squeeze_464_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_464_cast_fp16 = squeeze(axes = squeeze_464_axes_0, x = var_14165_cast_fp16_0)[name = tensor<string, []>("squeeze_464_cast_fp16")];
            tensor<int32, [1]> squeeze_465_axes_0 = const()[name = tensor<string, []>("squeeze_465_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_465_cast_fp16 = squeeze(axes = squeeze_465_axes_0, x = var_14165_cast_fp16_1)[name = tensor<string, []>("squeeze_465_cast_fp16")];
            tensor<fp16, []> const_1899_promoted_to_fp16 = const()[name = tensor<string, []>("const_1899_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14168_cast_fp16 = mul(x = squeeze_465_cast_fp16, y = const_1899_promoted_to_fp16)[name = tensor<string, []>("op_14168_cast_fp16")];
            tensor<int32, []> x_1585_axis_0 = const()[name = tensor<string, []>("x_1585_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1585_cast_fp16 = stack(axis = x_1585_axis_0, values = (var_14168_cast_fp16, squeeze_464_cast_fp16))[name = tensor<string, []>("x_1585_cast_fp16")];
            tensor<int32, [4]> var_14174 = const()[name = tensor<string, []>("op_14174"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14175_cast_fp16 = reshape(shape = var_14174, x = x_1585_cast_fp16)[name = tensor<string, []>("op_14175_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14176_cast_fp16 = mul(x = var_14175_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14176_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_233_cast_fp16 = add(x = var_14155_cast_fp16, y = var_14176_cast_fp16)[name = tensor<string, []>("q_patches_233_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14178_cast_fp16 = mul(x = var_14153_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14178_cast_fp16")];
            tensor<int32, [5]> var_14186 = const()[name = tensor<string, []>("op_14186"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1589_cast_fp16 = reshape(shape = var_14186, x = var_14153_cast_fp16)[name = tensor<string, []>("x_1589_cast_fp16")];
            tensor<int32, [2]> var_14188_split_sizes_0 = const()[name = tensor<string, []>("op_14188_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14188_axis_0 = const()[name = tensor<string, []>("op_14188_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14188_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14188_cast_fp16_1 = split(axis = var_14188_axis_0, split_sizes = var_14188_split_sizes_0, x = x_1589_cast_fp16)[name = tensor<string, []>("op_14188_cast_fp16")];
            tensor<int32, [1]> squeeze_466_axes_0 = const()[name = tensor<string, []>("squeeze_466_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_466_cast_fp16 = squeeze(axes = squeeze_466_axes_0, x = var_14188_cast_fp16_0)[name = tensor<string, []>("squeeze_466_cast_fp16")];
            tensor<int32, [1]> squeeze_467_axes_0 = const()[name = tensor<string, []>("squeeze_467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_467_cast_fp16 = squeeze(axes = squeeze_467_axes_0, x = var_14188_cast_fp16_1)[name = tensor<string, []>("squeeze_467_cast_fp16")];
            tensor<fp16, []> const_1907_promoted_to_fp16 = const()[name = tensor<string, []>("const_1907_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14191_cast_fp16 = mul(x = squeeze_467_cast_fp16, y = const_1907_promoted_to_fp16)[name = tensor<string, []>("op_14191_cast_fp16")];
            tensor<int32, []> x_1591_axis_0 = const()[name = tensor<string, []>("x_1591_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1591_cast_fp16 = stack(axis = x_1591_axis_0, values = (var_14191_cast_fp16, squeeze_466_cast_fp16))[name = tensor<string, []>("x_1591_cast_fp16")];
            tensor<int32, [4]> var_14197 = const()[name = tensor<string, []>("op_14197"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14198_cast_fp16 = reshape(shape = var_14197, x = x_1591_cast_fp16)[name = tensor<string, []>("op_14198_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14199_cast_fp16 = mul(x = var_14198_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14199_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_233_cast_fp16 = add(x = var_14178_cast_fp16, y = var_14199_cast_fp16)[name = tensor<string, []>("k_patches_233_cast_fp16")];
            tensor<bool, []> var_14202_interleave_0 = const()[name = tensor<string, []>("op_14202_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14202_cast_fp16 = concat(axis = var_20, interleave = var_14202_interleave_0, values = (var_14141_cast_fp16, q_patches_233_cast_fp16))[name = tensor<string, []>("op_14202_cast_fp16")];
            tensor<bool, []> var_14205_interleave_0 = const()[name = tensor<string, []>("op_14205_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14205_cast_fp16 = concat(axis = var_20, interleave = var_14205_interleave_0, values = (var_14145_cast_fp16, k_patches_233_cast_fp16))[name = tensor<string, []>("op_14205_cast_fp16")];
            tensor<fp16, []> var_14207_to_fp16 = const()[name = tensor<string, []>("op_14207_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1169_cast_fp16 = mul(x = var_14202_cast_fp16, y = var_14207_to_fp16)[name = tensor<string, []>("q_1169_cast_fp16")];
            tensor<bool, []> attn_465_transpose_x_1 = const()[name = tensor<string, []>("attn_465_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_465_transpose_y_1 = const()[name = tensor<string, []>("attn_465_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_465_cast_fp16 = matmul(transpose_x = attn_465_transpose_x_1, transpose_y = attn_465_transpose_y_1, x = q_1169_cast_fp16, y = var_14205_cast_fp16)[name = tensor<string, []>("attn_465_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_467_cast_fp16 = softmax(axis = var_21, x = attn_465_cast_fp16)[name = tensor<string, []>("attn_467_cast_fp16")];
            tensor<bool, []> var_14212_transpose_x_0 = const()[name = tensor<string, []>("op_14212_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14212_transpose_y_0 = const()[name = tensor<string, []>("op_14212_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14212_cast_fp16 = matmul(transpose_x = var_14212_transpose_x_0, transpose_y = var_14212_transpose_y_0, x = attn_467_cast_fp16, y = v_701_cast_fp16)[name = tensor<string, []>("op_14212_cast_fp16")];
            tensor<int32, [1]> var_14213_axes_0 = const()[name = tensor<string, []>("op_14213_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14213_cast_fp16 = squeeze(axes = var_14213_axes_0, x = var_14212_cast_fp16)[name = tensor<string, []>("op_14213_cast_fp16")];
            tensor<string, []> var_14219_pad_type_0 = const()[name = tensor<string, []>("op_14219_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14219_strides_0 = const()[name = tensor<string, []>("op_14219_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14219_pad_0 = const()[name = tensor<string, []>("op_14219_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14219_dilations_0 = const()[name = tensor<string, []>("op_14219_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14219_groups_0 = const()[name = tensor<string, []>("op_14219_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131879680)))];
            tensor<fp16, [1, 64, 1, 197]> var_14219_cast_fp16 = conv(dilations = var_14219_dilations_0, groups = var_14219_groups_0, pad = var_14219_pad_0, pad_type = var_14219_pad_type_0, strides = var_14219_strides_0, weight = model_blocks_9_attn_q_projs_9_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14219_cast_fp16")];
            tensor<int32, [1]> var_14220_axes_0 = const()[name = tensor<string, []>("op_14220_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14220_cast_fp16 = squeeze(axes = var_14220_axes_0, x = var_14219_cast_fp16)[name = tensor<string, []>("op_14220_cast_fp16")];
            tensor<int32, [3]> q_1171_perm_0 = const()[name = tensor<string, []>("q_1171_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14227_pad_type_0 = const()[name = tensor<string, []>("op_14227_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14227_strides_0 = const()[name = tensor<string, []>("op_14227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14227_pad_0 = const()[name = tensor<string, []>("op_14227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14227_dilations_0 = const()[name = tensor<string, []>("op_14227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14227_groups_0 = const()[name = tensor<string, []>("op_14227_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131978048)))];
            tensor<fp16, [1, 64, 1, 197]> var_14227_cast_fp16 = conv(dilations = var_14227_dilations_0, groups = var_14227_groups_0, pad = var_14227_pad_0, pad_type = var_14227_pad_type_0, strides = var_14227_strides_0, weight = model_blocks_9_attn_k_projs_9_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14227_cast_fp16")];
            tensor<int32, [1]> var_14228_axes_0 = const()[name = tensor<string, []>("op_14228_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14228_cast_fp16 = squeeze(axes = var_14228_axes_0, x = var_14227_cast_fp16)[name = tensor<string, []>("op_14228_cast_fp16")];
            tensor<int32, [3]> k_703_perm_0 = const()[name = tensor<string, []>("k_703_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14235_pad_type_0 = const()[name = tensor<string, []>("op_14235_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14235_strides_0 = const()[name = tensor<string, []>("op_14235_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14235_pad_0 = const()[name = tensor<string, []>("op_14235_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14235_dilations_0 = const()[name = tensor<string, []>("op_14235_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14235_groups_0 = const()[name = tensor<string, []>("op_14235_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132076416)))];
            tensor<fp16, [1, 64, 1, 197]> var_14235_cast_fp16 = conv(dilations = var_14235_dilations_0, groups = var_14235_groups_0, pad = var_14235_pad_0, pad_type = var_14235_pad_type_0, strides = var_14235_strides_0, weight = model_blocks_9_attn_v_projs_9_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14235_cast_fp16")];
            tensor<int32, [1]> var_14236_axes_0 = const()[name = tensor<string, []>("op_14236_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14236_cast_fp16 = squeeze(axes = var_14236_axes_0, x = var_14235_cast_fp16)[name = tensor<string, []>("op_14236_cast_fp16")];
            tensor<int32, [3]> v_703_perm_0 = const()[name = tensor<string, []>("v_703_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132174784)))];
            tensor<fp16, [1, 197, 64]> q_1171_cast_fp16 = transpose(perm = q_1171_perm_0, x = var_14220_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [1, 197, 64]> q_1173_cast_fp16 = add(x = q_1171_cast_fp16, y = model_blocks_9_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_1173_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132174976)))];
            tensor<fp16, [1, 197, 64]> v_703_cast_fp16 = transpose(perm = v_703_perm_0, x = var_14236_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 197, 64]> v_705_cast_fp16 = add(x = v_703_cast_fp16, y = model_blocks_9_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_705_cast_fp16")];
            tensor<int32, [1]> q_1175_axes_0 = const()[name = tensor<string, []>("q_1175_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1175_cast_fp16 = expand_dims(axes = q_1175_axes_0, x = q_1173_cast_fp16)[name = tensor<string, []>("q_1175_cast_fp16")];
            tensor<int32, [1]> k_705_axes_0 = const()[name = tensor<string, []>("k_705_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_703_cast_fp16 = transpose(perm = k_703_perm_0, x = var_14228_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [1, 1, 197, 64]> k_705_cast_fp16 = expand_dims(axes = k_705_axes_0, x = k_703_cast_fp16)[name = tensor<string, []>("k_705_cast_fp16")];
            tensor<int32, [1]> v_707_axes_0 = const()[name = tensor<string, []>("v_707_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_707_cast_fp16 = expand_dims(axes = v_707_axes_0, x = v_705_cast_fp16)[name = tensor<string, []>("v_707_cast_fp16")];
            tensor<int32, [4]> var_14245_begin_0 = const()[name = tensor<string, []>("op_14245_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14245_end_0 = const()[name = tensor<string, []>("op_14245_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14245_end_mask_0 = const()[name = tensor<string, []>("op_14245_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14245_cast_fp16 = slice_by_index(begin = var_14245_begin_0, end = var_14245_end_0, end_mask = var_14245_end_mask_0, x = q_1175_cast_fp16)[name = tensor<string, []>("op_14245_cast_fp16")];
            tensor<int32, [4]> var_14249_begin_0 = const()[name = tensor<string, []>("op_14249_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14249_end_0 = const()[name = tensor<string, []>("op_14249_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14249_end_mask_0 = const()[name = tensor<string, []>("op_14249_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14249_cast_fp16 = slice_by_index(begin = var_14249_begin_0, end = var_14249_end_0, end_mask = var_14249_end_mask_0, x = k_705_cast_fp16)[name = tensor<string, []>("op_14249_cast_fp16")];
            tensor<int32, [4]> var_14253_begin_0 = const()[name = tensor<string, []>("op_14253_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14253_end_0 = const()[name = tensor<string, []>("op_14253_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14253_end_mask_0 = const()[name = tensor<string, []>("op_14253_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14253_cast_fp16 = slice_by_index(begin = var_14253_begin_0, end = var_14253_end_0, end_mask = var_14253_end_mask_0, x = q_1175_cast_fp16)[name = tensor<string, []>("op_14253_cast_fp16")];
            tensor<int32, [4]> var_14257_begin_0 = const()[name = tensor<string, []>("op_14257_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14257_end_0 = const()[name = tensor<string, []>("op_14257_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14257_end_mask_0 = const()[name = tensor<string, []>("op_14257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14257_cast_fp16 = slice_by_index(begin = var_14257_begin_0, end = var_14257_end_0, end_mask = var_14257_end_mask_0, x = k_705_cast_fp16)[name = tensor<string, []>("op_14257_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14259_cast_fp16 = mul(x = var_14253_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14259_cast_fp16")];
            tensor<int32, [5]> var_14267 = const()[name = tensor<string, []>("op_14267"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1595_cast_fp16 = reshape(shape = var_14267, x = var_14253_cast_fp16)[name = tensor<string, []>("x_1595_cast_fp16")];
            tensor<int32, [2]> var_14269_split_sizes_0 = const()[name = tensor<string, []>("op_14269_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14269_axis_0 = const()[name = tensor<string, []>("op_14269_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14269_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14269_cast_fp16_1 = split(axis = var_14269_axis_0, split_sizes = var_14269_split_sizes_0, x = x_1595_cast_fp16)[name = tensor<string, []>("op_14269_cast_fp16")];
            tensor<int32, [1]> squeeze_468_axes_0 = const()[name = tensor<string, []>("squeeze_468_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_468_cast_fp16 = squeeze(axes = squeeze_468_axes_0, x = var_14269_cast_fp16_0)[name = tensor<string, []>("squeeze_468_cast_fp16")];
            tensor<int32, [1]> squeeze_469_axes_0 = const()[name = tensor<string, []>("squeeze_469_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_469_cast_fp16 = squeeze(axes = squeeze_469_axes_0, x = var_14269_cast_fp16_1)[name = tensor<string, []>("squeeze_469_cast_fp16")];
            tensor<fp16, []> const_1915_promoted_to_fp16 = const()[name = tensor<string, []>("const_1915_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14272_cast_fp16 = mul(x = squeeze_469_cast_fp16, y = const_1915_promoted_to_fp16)[name = tensor<string, []>("op_14272_cast_fp16")];
            tensor<int32, []> x_1597_axis_0 = const()[name = tensor<string, []>("x_1597_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1597_cast_fp16 = stack(axis = x_1597_axis_0, values = (var_14272_cast_fp16, squeeze_468_cast_fp16))[name = tensor<string, []>("x_1597_cast_fp16")];
            tensor<int32, [4]> var_14278 = const()[name = tensor<string, []>("op_14278"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14279_cast_fp16 = reshape(shape = var_14278, x = x_1597_cast_fp16)[name = tensor<string, []>("op_14279_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14280_cast_fp16 = mul(x = var_14279_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14280_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_235_cast_fp16 = add(x = var_14259_cast_fp16, y = var_14280_cast_fp16)[name = tensor<string, []>("q_patches_235_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14282_cast_fp16 = mul(x = var_14257_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14282_cast_fp16")];
            tensor<int32, [5]> var_14290 = const()[name = tensor<string, []>("op_14290"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1601_cast_fp16 = reshape(shape = var_14290, x = var_14257_cast_fp16)[name = tensor<string, []>("x_1601_cast_fp16")];
            tensor<int32, [2]> var_14292_split_sizes_0 = const()[name = tensor<string, []>("op_14292_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14292_axis_0 = const()[name = tensor<string, []>("op_14292_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14292_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14292_cast_fp16_1 = split(axis = var_14292_axis_0, split_sizes = var_14292_split_sizes_0, x = x_1601_cast_fp16)[name = tensor<string, []>("op_14292_cast_fp16")];
            tensor<int32, [1]> squeeze_470_axes_0 = const()[name = tensor<string, []>("squeeze_470_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_470_cast_fp16 = squeeze(axes = squeeze_470_axes_0, x = var_14292_cast_fp16_0)[name = tensor<string, []>("squeeze_470_cast_fp16")];
            tensor<int32, [1]> squeeze_471_axes_0 = const()[name = tensor<string, []>("squeeze_471_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_471_cast_fp16 = squeeze(axes = squeeze_471_axes_0, x = var_14292_cast_fp16_1)[name = tensor<string, []>("squeeze_471_cast_fp16")];
            tensor<fp16, []> const_1923_promoted_to_fp16 = const()[name = tensor<string, []>("const_1923_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14295_cast_fp16 = mul(x = squeeze_471_cast_fp16, y = const_1923_promoted_to_fp16)[name = tensor<string, []>("op_14295_cast_fp16")];
            tensor<int32, []> x_1603_axis_0 = const()[name = tensor<string, []>("x_1603_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1603_cast_fp16 = stack(axis = x_1603_axis_0, values = (var_14295_cast_fp16, squeeze_470_cast_fp16))[name = tensor<string, []>("x_1603_cast_fp16")];
            tensor<int32, [4]> var_14301 = const()[name = tensor<string, []>("op_14301"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14302_cast_fp16 = reshape(shape = var_14301, x = x_1603_cast_fp16)[name = tensor<string, []>("op_14302_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14303_cast_fp16 = mul(x = var_14302_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14303_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_235_cast_fp16 = add(x = var_14282_cast_fp16, y = var_14303_cast_fp16)[name = tensor<string, []>("k_patches_235_cast_fp16")];
            tensor<bool, []> var_14306_interleave_0 = const()[name = tensor<string, []>("op_14306_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14306_cast_fp16 = concat(axis = var_20, interleave = var_14306_interleave_0, values = (var_14245_cast_fp16, q_patches_235_cast_fp16))[name = tensor<string, []>("op_14306_cast_fp16")];
            tensor<bool, []> var_14309_interleave_0 = const()[name = tensor<string, []>("op_14309_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14309_cast_fp16 = concat(axis = var_20, interleave = var_14309_interleave_0, values = (var_14249_cast_fp16, k_patches_235_cast_fp16))[name = tensor<string, []>("op_14309_cast_fp16")];
            tensor<fp16, []> var_14311_to_fp16 = const()[name = tensor<string, []>("op_14311_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1179_cast_fp16 = mul(x = var_14306_cast_fp16, y = var_14311_to_fp16)[name = tensor<string, []>("q_1179_cast_fp16")];
            tensor<bool, []> attn_469_transpose_x_1 = const()[name = tensor<string, []>("attn_469_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_469_transpose_y_1 = const()[name = tensor<string, []>("attn_469_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_469_cast_fp16 = matmul(transpose_x = attn_469_transpose_x_1, transpose_y = attn_469_transpose_y_1, x = q_1179_cast_fp16, y = var_14309_cast_fp16)[name = tensor<string, []>("attn_469_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_471_cast_fp16 = softmax(axis = var_21, x = attn_469_cast_fp16)[name = tensor<string, []>("attn_471_cast_fp16")];
            tensor<bool, []> var_14316_transpose_x_0 = const()[name = tensor<string, []>("op_14316_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14316_transpose_y_0 = const()[name = tensor<string, []>("op_14316_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14316_cast_fp16 = matmul(transpose_x = var_14316_transpose_x_0, transpose_y = var_14316_transpose_y_0, x = attn_471_cast_fp16, y = v_707_cast_fp16)[name = tensor<string, []>("op_14316_cast_fp16")];
            tensor<int32, [1]> var_14317_axes_0 = const()[name = tensor<string, []>("op_14317_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14317_cast_fp16 = squeeze(axes = var_14317_axes_0, x = var_14316_cast_fp16)[name = tensor<string, []>("op_14317_cast_fp16")];
            tensor<string, []> var_14323_pad_type_0 = const()[name = tensor<string, []>("op_14323_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14323_strides_0 = const()[name = tensor<string, []>("op_14323_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14323_pad_0 = const()[name = tensor<string, []>("op_14323_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14323_dilations_0 = const()[name = tensor<string, []>("op_14323_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14323_groups_0 = const()[name = tensor<string, []>("op_14323_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132175168)))];
            tensor<fp16, [1, 64, 1, 197]> var_14323_cast_fp16 = conv(dilations = var_14323_dilations_0, groups = var_14323_groups_0, pad = var_14323_pad_0, pad_type = var_14323_pad_type_0, strides = var_14323_strides_0, weight = model_blocks_9_attn_q_projs_10_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14323_cast_fp16")];
            tensor<int32, [1]> var_14324_axes_0 = const()[name = tensor<string, []>("op_14324_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14324_cast_fp16 = squeeze(axes = var_14324_axes_0, x = var_14323_cast_fp16)[name = tensor<string, []>("op_14324_cast_fp16")];
            tensor<int32, [3]> q_1181_perm_0 = const()[name = tensor<string, []>("q_1181_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14331_pad_type_0 = const()[name = tensor<string, []>("op_14331_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14331_strides_0 = const()[name = tensor<string, []>("op_14331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14331_pad_0 = const()[name = tensor<string, []>("op_14331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14331_dilations_0 = const()[name = tensor<string, []>("op_14331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14331_groups_0 = const()[name = tensor<string, []>("op_14331_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132273536)))];
            tensor<fp16, [1, 64, 1, 197]> var_14331_cast_fp16 = conv(dilations = var_14331_dilations_0, groups = var_14331_groups_0, pad = var_14331_pad_0, pad_type = var_14331_pad_type_0, strides = var_14331_strides_0, weight = model_blocks_9_attn_k_projs_10_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14331_cast_fp16")];
            tensor<int32, [1]> var_14332_axes_0 = const()[name = tensor<string, []>("op_14332_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14332_cast_fp16 = squeeze(axes = var_14332_axes_0, x = var_14331_cast_fp16)[name = tensor<string, []>("op_14332_cast_fp16")];
            tensor<int32, [3]> k_709_perm_0 = const()[name = tensor<string, []>("k_709_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14339_pad_type_0 = const()[name = tensor<string, []>("op_14339_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14339_strides_0 = const()[name = tensor<string, []>("op_14339_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14339_pad_0 = const()[name = tensor<string, []>("op_14339_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14339_dilations_0 = const()[name = tensor<string, []>("op_14339_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14339_groups_0 = const()[name = tensor<string, []>("op_14339_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132371904)))];
            tensor<fp16, [1, 64, 1, 197]> var_14339_cast_fp16 = conv(dilations = var_14339_dilations_0, groups = var_14339_groups_0, pad = var_14339_pad_0, pad_type = var_14339_pad_type_0, strides = var_14339_strides_0, weight = model_blocks_9_attn_v_projs_10_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14339_cast_fp16")];
            tensor<int32, [1]> var_14340_axes_0 = const()[name = tensor<string, []>("op_14340_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14340_cast_fp16 = squeeze(axes = var_14340_axes_0, x = var_14339_cast_fp16)[name = tensor<string, []>("op_14340_cast_fp16")];
            tensor<int32, [3]> v_709_perm_0 = const()[name = tensor<string, []>("v_709_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132470272)))];
            tensor<fp16, [1, 197, 64]> q_1181_cast_fp16 = transpose(perm = q_1181_perm_0, x = var_14324_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [1, 197, 64]> q_1183_cast_fp16 = add(x = q_1181_cast_fp16, y = model_blocks_9_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_1183_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132470464)))];
            tensor<fp16, [1, 197, 64]> v_709_cast_fp16 = transpose(perm = v_709_perm_0, x = var_14340_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [1, 197, 64]> v_711_cast_fp16 = add(x = v_709_cast_fp16, y = model_blocks_9_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_711_cast_fp16")];
            tensor<int32, [1]> q_1185_axes_0 = const()[name = tensor<string, []>("q_1185_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1185_cast_fp16 = expand_dims(axes = q_1185_axes_0, x = q_1183_cast_fp16)[name = tensor<string, []>("q_1185_cast_fp16")];
            tensor<int32, [1]> k_711_axes_0 = const()[name = tensor<string, []>("k_711_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_709_cast_fp16 = transpose(perm = k_709_perm_0, x = var_14332_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [1, 1, 197, 64]> k_711_cast_fp16 = expand_dims(axes = k_711_axes_0, x = k_709_cast_fp16)[name = tensor<string, []>("k_711_cast_fp16")];
            tensor<int32, [1]> v_713_axes_0 = const()[name = tensor<string, []>("v_713_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_713_cast_fp16 = expand_dims(axes = v_713_axes_0, x = v_711_cast_fp16)[name = tensor<string, []>("v_713_cast_fp16")];
            tensor<int32, [4]> var_14349_begin_0 = const()[name = tensor<string, []>("op_14349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14349_end_0 = const()[name = tensor<string, []>("op_14349_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14349_end_mask_0 = const()[name = tensor<string, []>("op_14349_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14349_cast_fp16 = slice_by_index(begin = var_14349_begin_0, end = var_14349_end_0, end_mask = var_14349_end_mask_0, x = q_1185_cast_fp16)[name = tensor<string, []>("op_14349_cast_fp16")];
            tensor<int32, [4]> var_14353_begin_0 = const()[name = tensor<string, []>("op_14353_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14353_end_0 = const()[name = tensor<string, []>("op_14353_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14353_end_mask_0 = const()[name = tensor<string, []>("op_14353_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14353_cast_fp16 = slice_by_index(begin = var_14353_begin_0, end = var_14353_end_0, end_mask = var_14353_end_mask_0, x = k_711_cast_fp16)[name = tensor<string, []>("op_14353_cast_fp16")];
            tensor<int32, [4]> var_14357_begin_0 = const()[name = tensor<string, []>("op_14357_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14357_end_0 = const()[name = tensor<string, []>("op_14357_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14357_end_mask_0 = const()[name = tensor<string, []>("op_14357_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14357_cast_fp16 = slice_by_index(begin = var_14357_begin_0, end = var_14357_end_0, end_mask = var_14357_end_mask_0, x = q_1185_cast_fp16)[name = tensor<string, []>("op_14357_cast_fp16")];
            tensor<int32, [4]> var_14361_begin_0 = const()[name = tensor<string, []>("op_14361_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14361_end_0 = const()[name = tensor<string, []>("op_14361_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14361_end_mask_0 = const()[name = tensor<string, []>("op_14361_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14361_cast_fp16 = slice_by_index(begin = var_14361_begin_0, end = var_14361_end_0, end_mask = var_14361_end_mask_0, x = k_711_cast_fp16)[name = tensor<string, []>("op_14361_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14363_cast_fp16 = mul(x = var_14357_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14363_cast_fp16")];
            tensor<int32, [5]> var_14371 = const()[name = tensor<string, []>("op_14371"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1607_cast_fp16 = reshape(shape = var_14371, x = var_14357_cast_fp16)[name = tensor<string, []>("x_1607_cast_fp16")];
            tensor<int32, [2]> var_14373_split_sizes_0 = const()[name = tensor<string, []>("op_14373_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14373_axis_0 = const()[name = tensor<string, []>("op_14373_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14373_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14373_cast_fp16_1 = split(axis = var_14373_axis_0, split_sizes = var_14373_split_sizes_0, x = x_1607_cast_fp16)[name = tensor<string, []>("op_14373_cast_fp16")];
            tensor<int32, [1]> squeeze_472_axes_0 = const()[name = tensor<string, []>("squeeze_472_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_472_cast_fp16 = squeeze(axes = squeeze_472_axes_0, x = var_14373_cast_fp16_0)[name = tensor<string, []>("squeeze_472_cast_fp16")];
            tensor<int32, [1]> squeeze_473_axes_0 = const()[name = tensor<string, []>("squeeze_473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_473_cast_fp16 = squeeze(axes = squeeze_473_axes_0, x = var_14373_cast_fp16_1)[name = tensor<string, []>("squeeze_473_cast_fp16")];
            tensor<fp16, []> const_1931_promoted_to_fp16 = const()[name = tensor<string, []>("const_1931_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14376_cast_fp16 = mul(x = squeeze_473_cast_fp16, y = const_1931_promoted_to_fp16)[name = tensor<string, []>("op_14376_cast_fp16")];
            tensor<int32, []> x_1609_axis_0 = const()[name = tensor<string, []>("x_1609_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1609_cast_fp16 = stack(axis = x_1609_axis_0, values = (var_14376_cast_fp16, squeeze_472_cast_fp16))[name = tensor<string, []>("x_1609_cast_fp16")];
            tensor<int32, [4]> var_14382 = const()[name = tensor<string, []>("op_14382"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14383_cast_fp16 = reshape(shape = var_14382, x = x_1609_cast_fp16)[name = tensor<string, []>("op_14383_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14384_cast_fp16 = mul(x = var_14383_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14384_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_237_cast_fp16 = add(x = var_14363_cast_fp16, y = var_14384_cast_fp16)[name = tensor<string, []>("q_patches_237_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14386_cast_fp16 = mul(x = var_14361_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14386_cast_fp16")];
            tensor<int32, [5]> var_14394 = const()[name = tensor<string, []>("op_14394"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1613_cast_fp16 = reshape(shape = var_14394, x = var_14361_cast_fp16)[name = tensor<string, []>("x_1613_cast_fp16")];
            tensor<int32, [2]> var_14396_split_sizes_0 = const()[name = tensor<string, []>("op_14396_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14396_axis_0 = const()[name = tensor<string, []>("op_14396_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14396_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14396_cast_fp16_1 = split(axis = var_14396_axis_0, split_sizes = var_14396_split_sizes_0, x = x_1613_cast_fp16)[name = tensor<string, []>("op_14396_cast_fp16")];
            tensor<int32, [1]> squeeze_474_axes_0 = const()[name = tensor<string, []>("squeeze_474_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_474_cast_fp16 = squeeze(axes = squeeze_474_axes_0, x = var_14396_cast_fp16_0)[name = tensor<string, []>("squeeze_474_cast_fp16")];
            tensor<int32, [1]> squeeze_475_axes_0 = const()[name = tensor<string, []>("squeeze_475_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_475_cast_fp16 = squeeze(axes = squeeze_475_axes_0, x = var_14396_cast_fp16_1)[name = tensor<string, []>("squeeze_475_cast_fp16")];
            tensor<fp16, []> const_1939_promoted_to_fp16 = const()[name = tensor<string, []>("const_1939_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14399_cast_fp16 = mul(x = squeeze_475_cast_fp16, y = const_1939_promoted_to_fp16)[name = tensor<string, []>("op_14399_cast_fp16")];
            tensor<int32, []> x_1615_axis_0 = const()[name = tensor<string, []>("x_1615_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1615_cast_fp16 = stack(axis = x_1615_axis_0, values = (var_14399_cast_fp16, squeeze_474_cast_fp16))[name = tensor<string, []>("x_1615_cast_fp16")];
            tensor<int32, [4]> var_14405 = const()[name = tensor<string, []>("op_14405"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14406_cast_fp16 = reshape(shape = var_14405, x = x_1615_cast_fp16)[name = tensor<string, []>("op_14406_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14407_cast_fp16 = mul(x = var_14406_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14407_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_237_cast_fp16 = add(x = var_14386_cast_fp16, y = var_14407_cast_fp16)[name = tensor<string, []>("k_patches_237_cast_fp16")];
            tensor<bool, []> var_14410_interleave_0 = const()[name = tensor<string, []>("op_14410_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14410_cast_fp16 = concat(axis = var_20, interleave = var_14410_interleave_0, values = (var_14349_cast_fp16, q_patches_237_cast_fp16))[name = tensor<string, []>("op_14410_cast_fp16")];
            tensor<bool, []> var_14413_interleave_0 = const()[name = tensor<string, []>("op_14413_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14413_cast_fp16 = concat(axis = var_20, interleave = var_14413_interleave_0, values = (var_14353_cast_fp16, k_patches_237_cast_fp16))[name = tensor<string, []>("op_14413_cast_fp16")];
            tensor<fp16, []> var_14415_to_fp16 = const()[name = tensor<string, []>("op_14415_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1189_cast_fp16 = mul(x = var_14410_cast_fp16, y = var_14415_to_fp16)[name = tensor<string, []>("q_1189_cast_fp16")];
            tensor<bool, []> attn_473_transpose_x_1 = const()[name = tensor<string, []>("attn_473_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_473_transpose_y_1 = const()[name = tensor<string, []>("attn_473_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_473_cast_fp16 = matmul(transpose_x = attn_473_transpose_x_1, transpose_y = attn_473_transpose_y_1, x = q_1189_cast_fp16, y = var_14413_cast_fp16)[name = tensor<string, []>("attn_473_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_475_cast_fp16 = softmax(axis = var_21, x = attn_473_cast_fp16)[name = tensor<string, []>("attn_475_cast_fp16")];
            tensor<bool, []> var_14420_transpose_x_0 = const()[name = tensor<string, []>("op_14420_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14420_transpose_y_0 = const()[name = tensor<string, []>("op_14420_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14420_cast_fp16 = matmul(transpose_x = var_14420_transpose_x_0, transpose_y = var_14420_transpose_y_0, x = attn_475_cast_fp16, y = v_713_cast_fp16)[name = tensor<string, []>("op_14420_cast_fp16")];
            tensor<int32, [1]> var_14421_axes_0 = const()[name = tensor<string, []>("op_14421_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14421_cast_fp16 = squeeze(axes = var_14421_axes_0, x = var_14420_cast_fp16)[name = tensor<string, []>("op_14421_cast_fp16")];
            tensor<string, []> var_14427_pad_type_0 = const()[name = tensor<string, []>("op_14427_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14427_strides_0 = const()[name = tensor<string, []>("op_14427_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14427_pad_0 = const()[name = tensor<string, []>("op_14427_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14427_dilations_0 = const()[name = tensor<string, []>("op_14427_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14427_groups_0 = const()[name = tensor<string, []>("op_14427_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132470656)))];
            tensor<fp16, [1, 64, 1, 197]> var_14427_cast_fp16 = conv(dilations = var_14427_dilations_0, groups = var_14427_groups_0, pad = var_14427_pad_0, pad_type = var_14427_pad_type_0, strides = var_14427_strides_0, weight = model_blocks_9_attn_q_projs_11_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14427_cast_fp16")];
            tensor<int32, [1]> var_14428_axes_0 = const()[name = tensor<string, []>("op_14428_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14428_cast_fp16 = squeeze(axes = var_14428_axes_0, x = var_14427_cast_fp16)[name = tensor<string, []>("op_14428_cast_fp16")];
            tensor<int32, [3]> q_1191_perm_0 = const()[name = tensor<string, []>("q_1191_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14435_pad_type_0 = const()[name = tensor<string, []>("op_14435_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14435_strides_0 = const()[name = tensor<string, []>("op_14435_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14435_pad_0 = const()[name = tensor<string, []>("op_14435_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14435_dilations_0 = const()[name = tensor<string, []>("op_14435_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14435_groups_0 = const()[name = tensor<string, []>("op_14435_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132569024)))];
            tensor<fp16, [1, 64, 1, 197]> var_14435_cast_fp16 = conv(dilations = var_14435_dilations_0, groups = var_14435_groups_0, pad = var_14435_pad_0, pad_type = var_14435_pad_type_0, strides = var_14435_strides_0, weight = model_blocks_9_attn_k_projs_11_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14435_cast_fp16")];
            tensor<int32, [1]> var_14436_axes_0 = const()[name = tensor<string, []>("op_14436_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14436_cast_fp16 = squeeze(axes = var_14436_axes_0, x = var_14435_cast_fp16)[name = tensor<string, []>("op_14436_cast_fp16")];
            tensor<int32, [3]> k_715_perm_0 = const()[name = tensor<string, []>("k_715_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14443_pad_type_0 = const()[name = tensor<string, []>("op_14443_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14443_strides_0 = const()[name = tensor<string, []>("op_14443_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14443_pad_0 = const()[name = tensor<string, []>("op_14443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14443_dilations_0 = const()[name = tensor<string, []>("op_14443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14443_groups_0 = const()[name = tensor<string, []>("op_14443_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_9_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132667392)))];
            tensor<fp16, [1, 64, 1, 197]> var_14443_cast_fp16 = conv(dilations = var_14443_dilations_0, groups = var_14443_groups_0, pad = var_14443_pad_0, pad_type = var_14443_pad_type_0, strides = var_14443_strides_0, weight = model_blocks_9_attn_v_projs_11_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_14443_cast_fp16")];
            tensor<int32, [1]> var_14444_axes_0 = const()[name = tensor<string, []>("op_14444_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14444_cast_fp16 = squeeze(axes = var_14444_axes_0, x = var_14443_cast_fp16)[name = tensor<string, []>("op_14444_cast_fp16")];
            tensor<int32, [3]> v_715_perm_0 = const()[name = tensor<string, []>("v_715_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_9_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132765760)))];
            tensor<fp16, [1, 197, 64]> q_1191_cast_fp16 = transpose(perm = q_1191_perm_0, x = var_14428_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 197, 64]> q_1193_cast_fp16 = add(x = q_1191_cast_fp16, y = model_blocks_9_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_1193_cast_fp16")];
            tensor<fp16, [64]> model_blocks_9_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132765952)))];
            tensor<fp16, [1, 197, 64]> v_715_cast_fp16 = transpose(perm = v_715_perm_0, x = var_14444_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [1, 197, 64]> v_717_cast_fp16 = add(x = v_715_cast_fp16, y = model_blocks_9_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_717_cast_fp16")];
            tensor<int32, [1]> q_1195_axes_0 = const()[name = tensor<string, []>("q_1195_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1195_cast_fp16 = expand_dims(axes = q_1195_axes_0, x = q_1193_cast_fp16)[name = tensor<string, []>("q_1195_cast_fp16")];
            tensor<int32, [1]> k_717_axes_0 = const()[name = tensor<string, []>("k_717_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_715_cast_fp16 = transpose(perm = k_715_perm_0, x = var_14436_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [1, 1, 197, 64]> k_717_cast_fp16 = expand_dims(axes = k_717_axes_0, x = k_715_cast_fp16)[name = tensor<string, []>("k_717_cast_fp16")];
            tensor<int32, [1]> v_719_axes_0 = const()[name = tensor<string, []>("v_719_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_719_cast_fp16 = expand_dims(axes = v_719_axes_0, x = v_717_cast_fp16)[name = tensor<string, []>("v_719_cast_fp16")];
            tensor<int32, [4]> var_14453_begin_0 = const()[name = tensor<string, []>("op_14453_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14453_end_0 = const()[name = tensor<string, []>("op_14453_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14453_end_mask_0 = const()[name = tensor<string, []>("op_14453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14453_cast_fp16 = slice_by_index(begin = var_14453_begin_0, end = var_14453_end_0, end_mask = var_14453_end_mask_0, x = q_1195_cast_fp16)[name = tensor<string, []>("op_14453_cast_fp16")];
            tensor<int32, [4]> var_14457_begin_0 = const()[name = tensor<string, []>("op_14457_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14457_end_0 = const()[name = tensor<string, []>("op_14457_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14457_end_mask_0 = const()[name = tensor<string, []>("op_14457_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14457_cast_fp16 = slice_by_index(begin = var_14457_begin_0, end = var_14457_end_0, end_mask = var_14457_end_mask_0, x = k_717_cast_fp16)[name = tensor<string, []>("op_14457_cast_fp16")];
            tensor<int32, [4]> var_14461_begin_0 = const()[name = tensor<string, []>("op_14461_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14461_end_0 = const()[name = tensor<string, []>("op_14461_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14461_end_mask_0 = const()[name = tensor<string, []>("op_14461_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14461_cast_fp16 = slice_by_index(begin = var_14461_begin_0, end = var_14461_end_0, end_mask = var_14461_end_mask_0, x = q_1195_cast_fp16)[name = tensor<string, []>("op_14461_cast_fp16")];
            tensor<int32, [4]> var_14465_begin_0 = const()[name = tensor<string, []>("op_14465_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14465_end_0 = const()[name = tensor<string, []>("op_14465_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14465_end_mask_0 = const()[name = tensor<string, []>("op_14465_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14465_cast_fp16 = slice_by_index(begin = var_14465_begin_0, end = var_14465_end_0, end_mask = var_14465_end_mask_0, x = k_717_cast_fp16)[name = tensor<string, []>("op_14465_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14467_cast_fp16 = mul(x = var_14461_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14467_cast_fp16")];
            tensor<int32, [5]> var_14475 = const()[name = tensor<string, []>("op_14475"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1619_cast_fp16 = reshape(shape = var_14475, x = var_14461_cast_fp16)[name = tensor<string, []>("x_1619_cast_fp16")];
            tensor<int32, [2]> var_14477_split_sizes_0 = const()[name = tensor<string, []>("op_14477_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14477_axis_0 = const()[name = tensor<string, []>("op_14477_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14477_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14477_cast_fp16_1 = split(axis = var_14477_axis_0, split_sizes = var_14477_split_sizes_0, x = x_1619_cast_fp16)[name = tensor<string, []>("op_14477_cast_fp16")];
            tensor<int32, [1]> squeeze_476_axes_0 = const()[name = tensor<string, []>("squeeze_476_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_476_cast_fp16 = squeeze(axes = squeeze_476_axes_0, x = var_14477_cast_fp16_0)[name = tensor<string, []>("squeeze_476_cast_fp16")];
            tensor<int32, [1]> squeeze_477_axes_0 = const()[name = tensor<string, []>("squeeze_477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_477_cast_fp16 = squeeze(axes = squeeze_477_axes_0, x = var_14477_cast_fp16_1)[name = tensor<string, []>("squeeze_477_cast_fp16")];
            tensor<fp16, []> const_1947_promoted_to_fp16 = const()[name = tensor<string, []>("const_1947_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14480_cast_fp16 = mul(x = squeeze_477_cast_fp16, y = const_1947_promoted_to_fp16)[name = tensor<string, []>("op_14480_cast_fp16")];
            tensor<int32, []> x_1621_axis_0 = const()[name = tensor<string, []>("x_1621_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1621_cast_fp16 = stack(axis = x_1621_axis_0, values = (var_14480_cast_fp16, squeeze_476_cast_fp16))[name = tensor<string, []>("x_1621_cast_fp16")];
            tensor<int32, [4]> var_14486 = const()[name = tensor<string, []>("op_14486"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14487_cast_fp16 = reshape(shape = var_14486, x = x_1621_cast_fp16)[name = tensor<string, []>("op_14487_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14488_cast_fp16 = mul(x = var_14487_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14488_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_239_cast_fp16 = add(x = var_14467_cast_fp16, y = var_14488_cast_fp16)[name = tensor<string, []>("q_patches_239_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14490_cast_fp16 = mul(x = var_14465_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14490_cast_fp16")];
            tensor<int32, [5]> var_14498 = const()[name = tensor<string, []>("op_14498"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1625_cast_fp16 = reshape(shape = var_14498, x = var_14465_cast_fp16)[name = tensor<string, []>("x_1625_cast_fp16")];
            tensor<int32, [2]> var_14500_split_sizes_0 = const()[name = tensor<string, []>("op_14500_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14500_axis_0 = const()[name = tensor<string, []>("op_14500_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14500_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14500_cast_fp16_1 = split(axis = var_14500_axis_0, split_sizes = var_14500_split_sizes_0, x = x_1625_cast_fp16)[name = tensor<string, []>("op_14500_cast_fp16")];
            tensor<int32, [1]> squeeze_478_axes_0 = const()[name = tensor<string, []>("squeeze_478_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_478_cast_fp16 = squeeze(axes = squeeze_478_axes_0, x = var_14500_cast_fp16_0)[name = tensor<string, []>("squeeze_478_cast_fp16")];
            tensor<int32, [1]> squeeze_479_axes_0 = const()[name = tensor<string, []>("squeeze_479_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_479_cast_fp16 = squeeze(axes = squeeze_479_axes_0, x = var_14500_cast_fp16_1)[name = tensor<string, []>("squeeze_479_cast_fp16")];
            tensor<fp16, []> const_1955_promoted_to_fp16 = const()[name = tensor<string, []>("const_1955_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14503_cast_fp16 = mul(x = squeeze_479_cast_fp16, y = const_1955_promoted_to_fp16)[name = tensor<string, []>("op_14503_cast_fp16")];
            tensor<int32, []> x_1627_axis_0 = const()[name = tensor<string, []>("x_1627_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1627_cast_fp16 = stack(axis = x_1627_axis_0, values = (var_14503_cast_fp16, squeeze_478_cast_fp16))[name = tensor<string, []>("x_1627_cast_fp16")];
            tensor<int32, [4]> var_14509 = const()[name = tensor<string, []>("op_14509"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14510_cast_fp16 = reshape(shape = var_14509, x = x_1627_cast_fp16)[name = tensor<string, []>("op_14510_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14511_cast_fp16 = mul(x = var_14510_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14511_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_239_cast_fp16 = add(x = var_14490_cast_fp16, y = var_14511_cast_fp16)[name = tensor<string, []>("k_patches_239_cast_fp16")];
            tensor<bool, []> var_14514_interleave_0 = const()[name = tensor<string, []>("op_14514_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14514_cast_fp16 = concat(axis = var_20, interleave = var_14514_interleave_0, values = (var_14453_cast_fp16, q_patches_239_cast_fp16))[name = tensor<string, []>("op_14514_cast_fp16")];
            tensor<bool, []> var_14517_interleave_0 = const()[name = tensor<string, []>("op_14517_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14517_cast_fp16 = concat(axis = var_20, interleave = var_14517_interleave_0, values = (var_14457_cast_fp16, k_patches_239_cast_fp16))[name = tensor<string, []>("op_14517_cast_fp16")];
            tensor<fp16, []> var_14519_to_fp16 = const()[name = tensor<string, []>("op_14519_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1199_cast_fp16 = mul(x = var_14514_cast_fp16, y = var_14519_to_fp16)[name = tensor<string, []>("q_1199_cast_fp16")];
            tensor<bool, []> attn_477_transpose_x_1 = const()[name = tensor<string, []>("attn_477_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_477_transpose_y_1 = const()[name = tensor<string, []>("attn_477_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_477_cast_fp16 = matmul(transpose_x = attn_477_transpose_x_1, transpose_y = attn_477_transpose_y_1, x = q_1199_cast_fp16, y = var_14517_cast_fp16)[name = tensor<string, []>("attn_477_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_479_cast_fp16 = softmax(axis = var_21, x = attn_477_cast_fp16)[name = tensor<string, []>("attn_479_cast_fp16")];
            tensor<bool, []> var_14524_transpose_x_0 = const()[name = tensor<string, []>("op_14524_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14524_transpose_y_0 = const()[name = tensor<string, []>("op_14524_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14524_cast_fp16 = matmul(transpose_x = var_14524_transpose_x_0, transpose_y = var_14524_transpose_y_0, x = attn_479_cast_fp16, y = v_719_cast_fp16)[name = tensor<string, []>("op_14524_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_19_cast_fp16 = squeeze(axes = out_19_axes_0, x = var_14524_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<bool, []> input_167_interleave_0 = const()[name = tensor<string, []>("input_167_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_167_cast_fp16 = concat(axis = var_21, interleave = input_167_interleave_0, values = (var_13381_cast_fp16, var_13485_cast_fp16, var_13589_cast_fp16, var_13693_cast_fp16, var_13797_cast_fp16, var_13901_cast_fp16, var_14005_cast_fp16, var_14109_cast_fp16, var_14213_cast_fp16, var_14317_cast_fp16, var_14421_cast_fp16, out_19_cast_fp16))[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<int32, [1]> x_1629_axes_0 = const()[name = tensor<string, []>("x_1629_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_9_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132766144)))];
            tensor<fp16, [768]> model_blocks_9_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132767744)))];
            tensor<fp16, [1, 197, 768]> x_1629_cast_fp16 = layer_norm(axes = x_1629_axes_0, beta = model_blocks_9_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_9_attn_inner_attn_ln_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("x_1629_cast_fp16")];
            tensor<int32, [4]> var_14532 = const()[name = tensor<string, []>("op_14532"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1631_cast_fp16 = reshape(shape = var_14532, x = x_1629_cast_fp16)[name = tensor<string, []>("x_1631_cast_fp16")];
            tensor<int32, [4]> input_169_perm_0 = const()[name = tensor<string, []>("input_169_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1633_pad_type_0 = const()[name = tensor<string, []>("x_1633_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1633_strides_0 = const()[name = tensor<string, []>("x_1633_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1633_pad_0 = const()[name = tensor<string, []>("x_1633_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1633_dilations_0 = const()[name = tensor<string, []>("x_1633_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1633_groups_0 = const()[name = tensor<string, []>("x_1633_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_9_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132769344)))];
            tensor<fp16, [768]> model_blocks_9_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133949056)))];
            tensor<fp16, [1, 768, 1, 197]> input_169_cast_fp16 = transpose(perm = input_169_perm_0, x = x_1631_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [1, 768, 1, 197]> x_1633_cast_fp16 = conv(bias = model_blocks_9_attn_proj_bias_to_fp16, dilations = x_1633_dilations_0, groups = x_1633_groups_0, pad = x_1633_pad_0, pad_type = x_1633_pad_type_0, strides = x_1633_strides_0, weight = model_blocks_9_attn_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("x_1633_cast_fp16")];
            tensor<int32, [4]> x_1635_perm_0 = const()[name = tensor<string, []>("x_1635_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_14543 = const()[name = tensor<string, []>("op_14543"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1635_cast_fp16 = transpose(perm = x_1635_perm_0, x = x_1633_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 197, 768]> var_14544_cast_fp16 = reshape(shape = var_14543, x = x_1635_cast_fp16)[name = tensor<string, []>("op_14544_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_171_cast_fp16 = add(x = input_163_cast_fp16, y = var_14544_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<int32, [1]> x_1637_axes_0 = const()[name = tensor<string, []>("x_1637_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133950656)))];
            tensor<fp16, [768]> model_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133952256)))];
            tensor<fp16, [1, 197, 768]> x_1637_cast_fp16 = layer_norm(axes = x_1637_axes_0, beta = model_blocks_9_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_9_norm2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("x_1637_cast_fp16")];
            tensor<int32, [4]> var_14556 = const()[name = tensor<string, []>("op_14556"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1639_cast_fp16 = reshape(shape = var_14556, x = x_1637_cast_fp16)[name = tensor<string, []>("x_1639_cast_fp16")];
            tensor<int32, [4]> input_173_perm_0 = const()[name = tensor<string, []>("input_173_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_175_pad_type_0 = const()[name = tensor<string, []>("input_175_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_175_strides_0 = const()[name = tensor<string, []>("input_175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_175_pad_0 = const()[name = tensor<string, []>("input_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_175_dilations_0 = const()[name = tensor<string, []>("input_175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_175_groups_0 = const()[name = tensor<string, []>("input_175_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_9_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133953856)))];
            tensor<fp16, [2048]> model_blocks_9_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137099648)))];
            tensor<fp16, [1, 768, 1, 197]> input_173_cast_fp16 = transpose(perm = input_173_perm_0, x = x_1639_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [1, 2048, 1, 197]> input_175_cast_fp16 = conv(bias = model_blocks_9_mlp_w1_bias_to_fp16, dilations = input_175_dilations_0, groups = input_175_groups_0, pad = input_175_pad_0, pad_type = input_175_pad_type_0, strides = input_175_strides_0, weight = model_blocks_9_mlp_w1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<string, []> x2_499_pad_type_0 = const()[name = tensor<string, []>("x2_499_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_499_strides_0 = const()[name = tensor<string, []>("x2_499_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_499_pad_0 = const()[name = tensor<string, []>("x2_499_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_499_dilations_0 = const()[name = tensor<string, []>("x2_499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_499_groups_0 = const()[name = tensor<string, []>("x2_499_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_9_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137103808)))];
            tensor<fp16, [2048]> model_blocks_9_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140249600)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_499_cast_fp16 = conv(bias = model_blocks_9_mlp_w2_bias_to_fp16, dilations = x2_499_dilations_0, groups = x2_499_groups_0, pad = x2_499_pad_0, pad_type = x2_499_pad_type_0, strides = x2_499_strides_0, weight = model_blocks_9_mlp_w2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("x2_499_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_14573_cast_fp16 = silu(x = input_175_cast_fp16)[name = tensor<string, []>("op_14573_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_73_cast_fp16 = mul(x = var_14573_cast_fp16, y = x2_499_cast_fp16)[name = tensor<string, []>("hidden_73_cast_fp16")];
            tensor<int32, [4]> hidden_75_perm_0 = const()[name = tensor<string, []>("hidden_75_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_14576 = const()[name = tensor<string, []>("op_14576"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_75_cast_fp16 = transpose(perm = hidden_75_perm_0, x = hidden_73_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [1, 197, 2048]> input_177_cast_fp16 = reshape(shape = var_14576, x = hidden_75_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<int32, [1]> hidden_77_axes_0 = const()[name = tensor<string, []>("hidden_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_9_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140253760)))];
            tensor<fp16, [2048]> model_blocks_9_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140257920)))];
            tensor<fp16, [1, 197, 2048]> hidden_77_cast_fp16 = layer_norm(axes = hidden_77_axes_0, beta = model_blocks_9_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_9_mlp_ffn_ln_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("hidden_77_cast_fp16")];
            tensor<int32, [4]> var_14582 = const()[name = tensor<string, []>("op_14582"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_79_cast_fp16 = reshape(shape = var_14582, x = hidden_77_cast_fp16)[name = tensor<string, []>("hidden_79_cast_fp16")];
            tensor<int32, [4]> input_179_perm_0 = const()[name = tensor<string, []>("input_179_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1641_pad_type_0 = const()[name = tensor<string, []>("x_1641_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1641_strides_0 = const()[name = tensor<string, []>("x_1641_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1641_pad_0 = const()[name = tensor<string, []>("x_1641_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1641_dilations_0 = const()[name = tensor<string, []>("x_1641_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1641_groups_0 = const()[name = tensor<string, []>("x_1641_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_9_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140262080)))];
            tensor<fp16, [768]> model_blocks_9_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_9_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143407872)))];
            tensor<fp16, [1, 2048, 1, 197]> input_179_cast_fp16 = transpose(perm = input_179_perm_0, x = hidden_79_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [1, 768, 1, 197]> x_1641_cast_fp16 = conv(bias = model_blocks_9_mlp_w3_bias_to_fp16, dilations = x_1641_dilations_0, groups = x_1641_groups_0, pad = x_1641_pad_0, pad_type = x_1641_pad_type_0, strides = x_1641_strides_0, weight = model_blocks_9_mlp_w3_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("x_1641_cast_fp16")];
            tensor<int32, [4]> x_1643_perm_0 = const()[name = tensor<string, []>("x_1643_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_14593 = const()[name = tensor<string, []>("op_14593"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1643_cast_fp16 = transpose(perm = x_1643_perm_0, x = x_1641_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [1, 197, 768]> var_14594_cast_fp16 = reshape(shape = var_14593, x = x_1643_cast_fp16)[name = tensor<string, []>("op_14594_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_181_cast_fp16 = add(x = input_171_cast_fp16, y = var_14594_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<int32, [1]> x_1645_axes_0 = const()[name = tensor<string, []>("x_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_10_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143409472)))];
            tensor<fp16, [768]> model_blocks_10_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143411072)))];
            tensor<fp16, [1, 197, 768]> x_1645_cast_fp16 = layer_norm(axes = x_1645_axes_0, beta = model_blocks_10_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_10_norm1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("x_1645_cast_fp16")];
            tensor<int32, [4]> var_14728 = const()[name = tensor<string, []>("op_14728"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1647_cast_fp16 = reshape(shape = var_14728, x = x_1645_cast_fp16)[name = tensor<string, []>("x_1647_cast_fp16")];
            tensor<int32, [4]> input_183_perm_0 = const()[name = tensor<string, []>("input_183_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_14736_pad_type_0 = const()[name = tensor<string, []>("op_14736_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14736_strides_0 = const()[name = tensor<string, []>("op_14736_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14736_pad_0 = const()[name = tensor<string, []>("op_14736_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14736_dilations_0 = const()[name = tensor<string, []>("op_14736_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14736_groups_0 = const()[name = tensor<string, []>("op_14736_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143412672)))];
            tensor<fp16, [1, 768, 1, 197]> input_183_cast_fp16 = transpose(perm = input_183_perm_0, x = x_1647_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp16, [1, 64, 1, 197]> var_14736_cast_fp16 = conv(dilations = var_14736_dilations_0, groups = var_14736_groups_0, pad = var_14736_pad_0, pad_type = var_14736_pad_type_0, strides = var_14736_strides_0, weight = model_blocks_10_attn_q_projs_0_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14736_cast_fp16")];
            tensor<int32, [1]> var_14737_axes_0 = const()[name = tensor<string, []>("op_14737_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14737_cast_fp16 = squeeze(axes = var_14737_axes_0, x = var_14736_cast_fp16)[name = tensor<string, []>("op_14737_cast_fp16")];
            tensor<int32, [3]> q_1201_perm_0 = const()[name = tensor<string, []>("q_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14744_pad_type_0 = const()[name = tensor<string, []>("op_14744_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14744_strides_0 = const()[name = tensor<string, []>("op_14744_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14744_pad_0 = const()[name = tensor<string, []>("op_14744_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14744_dilations_0 = const()[name = tensor<string, []>("op_14744_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14744_groups_0 = const()[name = tensor<string, []>("op_14744_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143511040)))];
            tensor<fp16, [1, 64, 1, 197]> var_14744_cast_fp16 = conv(dilations = var_14744_dilations_0, groups = var_14744_groups_0, pad = var_14744_pad_0, pad_type = var_14744_pad_type_0, strides = var_14744_strides_0, weight = model_blocks_10_attn_k_projs_0_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14744_cast_fp16")];
            tensor<int32, [1]> var_14745_axes_0 = const()[name = tensor<string, []>("op_14745_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14745_cast_fp16 = squeeze(axes = var_14745_axes_0, x = var_14744_cast_fp16)[name = tensor<string, []>("op_14745_cast_fp16")];
            tensor<int32, [3]> k_721_perm_0 = const()[name = tensor<string, []>("k_721_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14752_pad_type_0 = const()[name = tensor<string, []>("op_14752_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14752_strides_0 = const()[name = tensor<string, []>("op_14752_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14752_pad_0 = const()[name = tensor<string, []>("op_14752_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14752_dilations_0 = const()[name = tensor<string, []>("op_14752_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14752_groups_0 = const()[name = tensor<string, []>("op_14752_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143609408)))];
            tensor<fp16, [1, 64, 1, 197]> var_14752_cast_fp16 = conv(dilations = var_14752_dilations_0, groups = var_14752_groups_0, pad = var_14752_pad_0, pad_type = var_14752_pad_type_0, strides = var_14752_strides_0, weight = model_blocks_10_attn_v_projs_0_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14752_cast_fp16")];
            tensor<int32, [1]> var_14753_axes_0 = const()[name = tensor<string, []>("op_14753_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14753_cast_fp16 = squeeze(axes = var_14753_axes_0, x = var_14752_cast_fp16)[name = tensor<string, []>("op_14753_cast_fp16")];
            tensor<int32, [3]> v_721_perm_0 = const()[name = tensor<string, []>("v_721_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143707776)))];
            tensor<fp16, [1, 197, 64]> q_1201_cast_fp16 = transpose(perm = q_1201_perm_0, x = var_14737_cast_fp16)[name = tensor<string, []>("transpose_84")];
            tensor<fp16, [1, 197, 64]> q_1203_cast_fp16 = add(x = q_1201_cast_fp16, y = model_blocks_10_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_1203_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143707968)))];
            tensor<fp16, [1, 197, 64]> v_721_cast_fp16 = transpose(perm = v_721_perm_0, x = var_14753_cast_fp16)[name = tensor<string, []>("transpose_82")];
            tensor<fp16, [1, 197, 64]> v_723_cast_fp16 = add(x = v_721_cast_fp16, y = model_blocks_10_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_723_cast_fp16")];
            tensor<int32, [1]> q_1205_axes_0 = const()[name = tensor<string, []>("q_1205_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1205_cast_fp16 = expand_dims(axes = q_1205_axes_0, x = q_1203_cast_fp16)[name = tensor<string, []>("q_1205_cast_fp16")];
            tensor<int32, [1]> k_723_axes_0 = const()[name = tensor<string, []>("k_723_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_721_cast_fp16 = transpose(perm = k_721_perm_0, x = var_14745_cast_fp16)[name = tensor<string, []>("transpose_83")];
            tensor<fp16, [1, 1, 197, 64]> k_723_cast_fp16 = expand_dims(axes = k_723_axes_0, x = k_721_cast_fp16)[name = tensor<string, []>("k_723_cast_fp16")];
            tensor<int32, [1]> v_725_axes_0 = const()[name = tensor<string, []>("v_725_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_725_cast_fp16 = expand_dims(axes = v_725_axes_0, x = v_723_cast_fp16)[name = tensor<string, []>("v_725_cast_fp16")];
            tensor<int32, [4]> var_14762_begin_0 = const()[name = tensor<string, []>("op_14762_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14762_end_0 = const()[name = tensor<string, []>("op_14762_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14762_end_mask_0 = const()[name = tensor<string, []>("op_14762_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14762_cast_fp16 = slice_by_index(begin = var_14762_begin_0, end = var_14762_end_0, end_mask = var_14762_end_mask_0, x = q_1205_cast_fp16)[name = tensor<string, []>("op_14762_cast_fp16")];
            tensor<int32, [4]> var_14766_begin_0 = const()[name = tensor<string, []>("op_14766_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14766_end_0 = const()[name = tensor<string, []>("op_14766_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14766_end_mask_0 = const()[name = tensor<string, []>("op_14766_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14766_cast_fp16 = slice_by_index(begin = var_14766_begin_0, end = var_14766_end_0, end_mask = var_14766_end_mask_0, x = k_723_cast_fp16)[name = tensor<string, []>("op_14766_cast_fp16")];
            tensor<int32, [4]> var_14770_begin_0 = const()[name = tensor<string, []>("op_14770_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14770_end_0 = const()[name = tensor<string, []>("op_14770_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14770_end_mask_0 = const()[name = tensor<string, []>("op_14770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14770_cast_fp16 = slice_by_index(begin = var_14770_begin_0, end = var_14770_end_0, end_mask = var_14770_end_mask_0, x = q_1205_cast_fp16)[name = tensor<string, []>("op_14770_cast_fp16")];
            tensor<int32, [4]> var_14774_begin_0 = const()[name = tensor<string, []>("op_14774_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14774_end_0 = const()[name = tensor<string, []>("op_14774_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14774_end_mask_0 = const()[name = tensor<string, []>("op_14774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14774_cast_fp16 = slice_by_index(begin = var_14774_begin_0, end = var_14774_end_0, end_mask = var_14774_end_mask_0, x = k_723_cast_fp16)[name = tensor<string, []>("op_14774_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14776_cast_fp16 = mul(x = var_14770_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14776_cast_fp16")];
            tensor<int32, [5]> var_14784 = const()[name = tensor<string, []>("op_14784"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1651_cast_fp16 = reshape(shape = var_14784, x = var_14770_cast_fp16)[name = tensor<string, []>("x_1651_cast_fp16")];
            tensor<int32, [2]> var_14786_split_sizes_0 = const()[name = tensor<string, []>("op_14786_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14786_axis_0 = const()[name = tensor<string, []>("op_14786_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14786_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14786_cast_fp16_1 = split(axis = var_14786_axis_0, split_sizes = var_14786_split_sizes_0, x = x_1651_cast_fp16)[name = tensor<string, []>("op_14786_cast_fp16")];
            tensor<int32, [1]> squeeze_480_axes_0 = const()[name = tensor<string, []>("squeeze_480_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_480_cast_fp16 = squeeze(axes = squeeze_480_axes_0, x = var_14786_cast_fp16_0)[name = tensor<string, []>("squeeze_480_cast_fp16")];
            tensor<int32, [1]> squeeze_481_axes_0 = const()[name = tensor<string, []>("squeeze_481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_481_cast_fp16 = squeeze(axes = squeeze_481_axes_0, x = var_14786_cast_fp16_1)[name = tensor<string, []>("squeeze_481_cast_fp16")];
            tensor<fp16, []> const_1967_promoted_to_fp16 = const()[name = tensor<string, []>("const_1967_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14789_cast_fp16 = mul(x = squeeze_481_cast_fp16, y = const_1967_promoted_to_fp16)[name = tensor<string, []>("op_14789_cast_fp16")];
            tensor<int32, []> x_1653_axis_0 = const()[name = tensor<string, []>("x_1653_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1653_cast_fp16 = stack(axis = x_1653_axis_0, values = (var_14789_cast_fp16, squeeze_480_cast_fp16))[name = tensor<string, []>("x_1653_cast_fp16")];
            tensor<int32, [4]> var_14795 = const()[name = tensor<string, []>("op_14795"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14796_cast_fp16 = reshape(shape = var_14795, x = x_1653_cast_fp16)[name = tensor<string, []>("op_14796_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14797_cast_fp16 = mul(x = var_14796_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14797_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_241_cast_fp16 = add(x = var_14776_cast_fp16, y = var_14797_cast_fp16)[name = tensor<string, []>("q_patches_241_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14799_cast_fp16 = mul(x = var_14774_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14799_cast_fp16")];
            tensor<int32, [5]> var_14807 = const()[name = tensor<string, []>("op_14807"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1657_cast_fp16 = reshape(shape = var_14807, x = var_14774_cast_fp16)[name = tensor<string, []>("x_1657_cast_fp16")];
            tensor<int32, [2]> var_14809_split_sizes_0 = const()[name = tensor<string, []>("op_14809_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14809_axis_0 = const()[name = tensor<string, []>("op_14809_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14809_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14809_cast_fp16_1 = split(axis = var_14809_axis_0, split_sizes = var_14809_split_sizes_0, x = x_1657_cast_fp16)[name = tensor<string, []>("op_14809_cast_fp16")];
            tensor<int32, [1]> squeeze_482_axes_0 = const()[name = tensor<string, []>("squeeze_482_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_482_cast_fp16 = squeeze(axes = squeeze_482_axes_0, x = var_14809_cast_fp16_0)[name = tensor<string, []>("squeeze_482_cast_fp16")];
            tensor<int32, [1]> squeeze_483_axes_0 = const()[name = tensor<string, []>("squeeze_483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_483_cast_fp16 = squeeze(axes = squeeze_483_axes_0, x = var_14809_cast_fp16_1)[name = tensor<string, []>("squeeze_483_cast_fp16")];
            tensor<fp16, []> const_1975_promoted_to_fp16 = const()[name = tensor<string, []>("const_1975_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14812_cast_fp16 = mul(x = squeeze_483_cast_fp16, y = const_1975_promoted_to_fp16)[name = tensor<string, []>("op_14812_cast_fp16")];
            tensor<int32, []> x_1659_axis_0 = const()[name = tensor<string, []>("x_1659_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1659_cast_fp16 = stack(axis = x_1659_axis_0, values = (var_14812_cast_fp16, squeeze_482_cast_fp16))[name = tensor<string, []>("x_1659_cast_fp16")];
            tensor<int32, [4]> var_14818 = const()[name = tensor<string, []>("op_14818"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14819_cast_fp16 = reshape(shape = var_14818, x = x_1659_cast_fp16)[name = tensor<string, []>("op_14819_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14820_cast_fp16 = mul(x = var_14819_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14820_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_241_cast_fp16 = add(x = var_14799_cast_fp16, y = var_14820_cast_fp16)[name = tensor<string, []>("k_patches_241_cast_fp16")];
            tensor<bool, []> var_14823_interleave_0 = const()[name = tensor<string, []>("op_14823_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14823_cast_fp16 = concat(axis = var_20, interleave = var_14823_interleave_0, values = (var_14762_cast_fp16, q_patches_241_cast_fp16))[name = tensor<string, []>("op_14823_cast_fp16")];
            tensor<bool, []> var_14826_interleave_0 = const()[name = tensor<string, []>("op_14826_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14826_cast_fp16 = concat(axis = var_20, interleave = var_14826_interleave_0, values = (var_14766_cast_fp16, k_patches_241_cast_fp16))[name = tensor<string, []>("op_14826_cast_fp16")];
            tensor<fp16, []> var_14828_to_fp16 = const()[name = tensor<string, []>("op_14828_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1209_cast_fp16 = mul(x = var_14823_cast_fp16, y = var_14828_to_fp16)[name = tensor<string, []>("q_1209_cast_fp16")];
            tensor<bool, []> attn_481_transpose_x_1 = const()[name = tensor<string, []>("attn_481_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_481_transpose_y_1 = const()[name = tensor<string, []>("attn_481_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_481_cast_fp16 = matmul(transpose_x = attn_481_transpose_x_1, transpose_y = attn_481_transpose_y_1, x = q_1209_cast_fp16, y = var_14826_cast_fp16)[name = tensor<string, []>("attn_481_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_483_cast_fp16 = softmax(axis = var_21, x = attn_481_cast_fp16)[name = tensor<string, []>("attn_483_cast_fp16")];
            tensor<bool, []> var_14833_transpose_x_0 = const()[name = tensor<string, []>("op_14833_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14833_transpose_y_0 = const()[name = tensor<string, []>("op_14833_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14833_cast_fp16 = matmul(transpose_x = var_14833_transpose_x_0, transpose_y = var_14833_transpose_y_0, x = attn_483_cast_fp16, y = v_725_cast_fp16)[name = tensor<string, []>("op_14833_cast_fp16")];
            tensor<int32, [1]> var_14834_axes_0 = const()[name = tensor<string, []>("op_14834_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14834_cast_fp16 = squeeze(axes = var_14834_axes_0, x = var_14833_cast_fp16)[name = tensor<string, []>("op_14834_cast_fp16")];
            tensor<string, []> var_14840_pad_type_0 = const()[name = tensor<string, []>("op_14840_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14840_strides_0 = const()[name = tensor<string, []>("op_14840_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14840_pad_0 = const()[name = tensor<string, []>("op_14840_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14840_dilations_0 = const()[name = tensor<string, []>("op_14840_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14840_groups_0 = const()[name = tensor<string, []>("op_14840_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143708160)))];
            tensor<fp16, [1, 64, 1, 197]> var_14840_cast_fp16 = conv(dilations = var_14840_dilations_0, groups = var_14840_groups_0, pad = var_14840_pad_0, pad_type = var_14840_pad_type_0, strides = var_14840_strides_0, weight = model_blocks_10_attn_q_projs_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14840_cast_fp16")];
            tensor<int32, [1]> var_14841_axes_0 = const()[name = tensor<string, []>("op_14841_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14841_cast_fp16 = squeeze(axes = var_14841_axes_0, x = var_14840_cast_fp16)[name = tensor<string, []>("op_14841_cast_fp16")];
            tensor<int32, [3]> q_1211_perm_0 = const()[name = tensor<string, []>("q_1211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14848_pad_type_0 = const()[name = tensor<string, []>("op_14848_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14848_strides_0 = const()[name = tensor<string, []>("op_14848_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14848_pad_0 = const()[name = tensor<string, []>("op_14848_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14848_dilations_0 = const()[name = tensor<string, []>("op_14848_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14848_groups_0 = const()[name = tensor<string, []>("op_14848_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143806528)))];
            tensor<fp16, [1, 64, 1, 197]> var_14848_cast_fp16 = conv(dilations = var_14848_dilations_0, groups = var_14848_groups_0, pad = var_14848_pad_0, pad_type = var_14848_pad_type_0, strides = var_14848_strides_0, weight = model_blocks_10_attn_k_projs_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14848_cast_fp16")];
            tensor<int32, [1]> var_14849_axes_0 = const()[name = tensor<string, []>("op_14849_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14849_cast_fp16 = squeeze(axes = var_14849_axes_0, x = var_14848_cast_fp16)[name = tensor<string, []>("op_14849_cast_fp16")];
            tensor<int32, [3]> k_727_perm_0 = const()[name = tensor<string, []>("k_727_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14856_pad_type_0 = const()[name = tensor<string, []>("op_14856_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14856_strides_0 = const()[name = tensor<string, []>("op_14856_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14856_pad_0 = const()[name = tensor<string, []>("op_14856_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14856_dilations_0 = const()[name = tensor<string, []>("op_14856_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14856_groups_0 = const()[name = tensor<string, []>("op_14856_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143904896)))];
            tensor<fp16, [1, 64, 1, 197]> var_14856_cast_fp16 = conv(dilations = var_14856_dilations_0, groups = var_14856_groups_0, pad = var_14856_pad_0, pad_type = var_14856_pad_type_0, strides = var_14856_strides_0, weight = model_blocks_10_attn_v_projs_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14856_cast_fp16")];
            tensor<int32, [1]> var_14857_axes_0 = const()[name = tensor<string, []>("op_14857_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14857_cast_fp16 = squeeze(axes = var_14857_axes_0, x = var_14856_cast_fp16)[name = tensor<string, []>("op_14857_cast_fp16")];
            tensor<int32, [3]> v_727_perm_0 = const()[name = tensor<string, []>("v_727_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144003264)))];
            tensor<fp16, [1, 197, 64]> q_1211_cast_fp16 = transpose(perm = q_1211_perm_0, x = var_14841_cast_fp16)[name = tensor<string, []>("transpose_81")];
            tensor<fp16, [1, 197, 64]> q_1213_cast_fp16 = add(x = q_1211_cast_fp16, y = model_blocks_10_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_1213_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144003456)))];
            tensor<fp16, [1, 197, 64]> v_727_cast_fp16 = transpose(perm = v_727_perm_0, x = var_14857_cast_fp16)[name = tensor<string, []>("transpose_79")];
            tensor<fp16, [1, 197, 64]> v_729_cast_fp16 = add(x = v_727_cast_fp16, y = model_blocks_10_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_729_cast_fp16")];
            tensor<int32, [1]> q_1215_axes_0 = const()[name = tensor<string, []>("q_1215_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1215_cast_fp16 = expand_dims(axes = q_1215_axes_0, x = q_1213_cast_fp16)[name = tensor<string, []>("q_1215_cast_fp16")];
            tensor<int32, [1]> k_729_axes_0 = const()[name = tensor<string, []>("k_729_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_727_cast_fp16 = transpose(perm = k_727_perm_0, x = var_14849_cast_fp16)[name = tensor<string, []>("transpose_80")];
            tensor<fp16, [1, 1, 197, 64]> k_729_cast_fp16 = expand_dims(axes = k_729_axes_0, x = k_727_cast_fp16)[name = tensor<string, []>("k_729_cast_fp16")];
            tensor<int32, [1]> v_731_axes_0 = const()[name = tensor<string, []>("v_731_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_731_cast_fp16 = expand_dims(axes = v_731_axes_0, x = v_729_cast_fp16)[name = tensor<string, []>("v_731_cast_fp16")];
            tensor<int32, [4]> var_14866_begin_0 = const()[name = tensor<string, []>("op_14866_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14866_end_0 = const()[name = tensor<string, []>("op_14866_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14866_end_mask_0 = const()[name = tensor<string, []>("op_14866_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14866_cast_fp16 = slice_by_index(begin = var_14866_begin_0, end = var_14866_end_0, end_mask = var_14866_end_mask_0, x = q_1215_cast_fp16)[name = tensor<string, []>("op_14866_cast_fp16")];
            tensor<int32, [4]> var_14870_begin_0 = const()[name = tensor<string, []>("op_14870_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14870_end_0 = const()[name = tensor<string, []>("op_14870_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14870_end_mask_0 = const()[name = tensor<string, []>("op_14870_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14870_cast_fp16 = slice_by_index(begin = var_14870_begin_0, end = var_14870_end_0, end_mask = var_14870_end_mask_0, x = k_729_cast_fp16)[name = tensor<string, []>("op_14870_cast_fp16")];
            tensor<int32, [4]> var_14874_begin_0 = const()[name = tensor<string, []>("op_14874_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14874_end_0 = const()[name = tensor<string, []>("op_14874_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14874_end_mask_0 = const()[name = tensor<string, []>("op_14874_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14874_cast_fp16 = slice_by_index(begin = var_14874_begin_0, end = var_14874_end_0, end_mask = var_14874_end_mask_0, x = q_1215_cast_fp16)[name = tensor<string, []>("op_14874_cast_fp16")];
            tensor<int32, [4]> var_14878_begin_0 = const()[name = tensor<string, []>("op_14878_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14878_end_0 = const()[name = tensor<string, []>("op_14878_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14878_end_mask_0 = const()[name = tensor<string, []>("op_14878_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14878_cast_fp16 = slice_by_index(begin = var_14878_begin_0, end = var_14878_end_0, end_mask = var_14878_end_mask_0, x = k_729_cast_fp16)[name = tensor<string, []>("op_14878_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14880_cast_fp16 = mul(x = var_14874_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14880_cast_fp16")];
            tensor<int32, [5]> var_14888 = const()[name = tensor<string, []>("op_14888"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1663_cast_fp16 = reshape(shape = var_14888, x = var_14874_cast_fp16)[name = tensor<string, []>("x_1663_cast_fp16")];
            tensor<int32, [2]> var_14890_split_sizes_0 = const()[name = tensor<string, []>("op_14890_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14890_axis_0 = const()[name = tensor<string, []>("op_14890_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14890_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14890_cast_fp16_1 = split(axis = var_14890_axis_0, split_sizes = var_14890_split_sizes_0, x = x_1663_cast_fp16)[name = tensor<string, []>("op_14890_cast_fp16")];
            tensor<int32, [1]> squeeze_484_axes_0 = const()[name = tensor<string, []>("squeeze_484_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_484_cast_fp16 = squeeze(axes = squeeze_484_axes_0, x = var_14890_cast_fp16_0)[name = tensor<string, []>("squeeze_484_cast_fp16")];
            tensor<int32, [1]> squeeze_485_axes_0 = const()[name = tensor<string, []>("squeeze_485_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_485_cast_fp16 = squeeze(axes = squeeze_485_axes_0, x = var_14890_cast_fp16_1)[name = tensor<string, []>("squeeze_485_cast_fp16")];
            tensor<fp16, []> const_1983_promoted_to_fp16 = const()[name = tensor<string, []>("const_1983_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14893_cast_fp16 = mul(x = squeeze_485_cast_fp16, y = const_1983_promoted_to_fp16)[name = tensor<string, []>("op_14893_cast_fp16")];
            tensor<int32, []> x_1665_axis_0 = const()[name = tensor<string, []>("x_1665_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1665_cast_fp16 = stack(axis = x_1665_axis_0, values = (var_14893_cast_fp16, squeeze_484_cast_fp16))[name = tensor<string, []>("x_1665_cast_fp16")];
            tensor<int32, [4]> var_14899 = const()[name = tensor<string, []>("op_14899"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14900_cast_fp16 = reshape(shape = var_14899, x = x_1665_cast_fp16)[name = tensor<string, []>("op_14900_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14901_cast_fp16 = mul(x = var_14900_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14901_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_243_cast_fp16 = add(x = var_14880_cast_fp16, y = var_14901_cast_fp16)[name = tensor<string, []>("q_patches_243_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14903_cast_fp16 = mul(x = var_14878_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14903_cast_fp16")];
            tensor<int32, [5]> var_14911 = const()[name = tensor<string, []>("op_14911"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1669_cast_fp16 = reshape(shape = var_14911, x = var_14878_cast_fp16)[name = tensor<string, []>("x_1669_cast_fp16")];
            tensor<int32, [2]> var_14913_split_sizes_0 = const()[name = tensor<string, []>("op_14913_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14913_axis_0 = const()[name = tensor<string, []>("op_14913_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14913_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14913_cast_fp16_1 = split(axis = var_14913_axis_0, split_sizes = var_14913_split_sizes_0, x = x_1669_cast_fp16)[name = tensor<string, []>("op_14913_cast_fp16")];
            tensor<int32, [1]> squeeze_486_axes_0 = const()[name = tensor<string, []>("squeeze_486_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_486_cast_fp16 = squeeze(axes = squeeze_486_axes_0, x = var_14913_cast_fp16_0)[name = tensor<string, []>("squeeze_486_cast_fp16")];
            tensor<int32, [1]> squeeze_487_axes_0 = const()[name = tensor<string, []>("squeeze_487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_487_cast_fp16 = squeeze(axes = squeeze_487_axes_0, x = var_14913_cast_fp16_1)[name = tensor<string, []>("squeeze_487_cast_fp16")];
            tensor<fp16, []> const_1991_promoted_to_fp16 = const()[name = tensor<string, []>("const_1991_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14916_cast_fp16 = mul(x = squeeze_487_cast_fp16, y = const_1991_promoted_to_fp16)[name = tensor<string, []>("op_14916_cast_fp16")];
            tensor<int32, []> x_1671_axis_0 = const()[name = tensor<string, []>("x_1671_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1671_cast_fp16 = stack(axis = x_1671_axis_0, values = (var_14916_cast_fp16, squeeze_486_cast_fp16))[name = tensor<string, []>("x_1671_cast_fp16")];
            tensor<int32, [4]> var_14922 = const()[name = tensor<string, []>("op_14922"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_14923_cast_fp16 = reshape(shape = var_14922, x = x_1671_cast_fp16)[name = tensor<string, []>("op_14923_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14924_cast_fp16 = mul(x = var_14923_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_14924_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_243_cast_fp16 = add(x = var_14903_cast_fp16, y = var_14924_cast_fp16)[name = tensor<string, []>("k_patches_243_cast_fp16")];
            tensor<bool, []> var_14927_interleave_0 = const()[name = tensor<string, []>("op_14927_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14927_cast_fp16 = concat(axis = var_20, interleave = var_14927_interleave_0, values = (var_14866_cast_fp16, q_patches_243_cast_fp16))[name = tensor<string, []>("op_14927_cast_fp16")];
            tensor<bool, []> var_14930_interleave_0 = const()[name = tensor<string, []>("op_14930_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14930_cast_fp16 = concat(axis = var_20, interleave = var_14930_interleave_0, values = (var_14870_cast_fp16, k_patches_243_cast_fp16))[name = tensor<string, []>("op_14930_cast_fp16")];
            tensor<fp16, []> var_14932_to_fp16 = const()[name = tensor<string, []>("op_14932_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1219_cast_fp16 = mul(x = var_14927_cast_fp16, y = var_14932_to_fp16)[name = tensor<string, []>("q_1219_cast_fp16")];
            tensor<bool, []> attn_485_transpose_x_1 = const()[name = tensor<string, []>("attn_485_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_485_transpose_y_1 = const()[name = tensor<string, []>("attn_485_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_485_cast_fp16 = matmul(transpose_x = attn_485_transpose_x_1, transpose_y = attn_485_transpose_y_1, x = q_1219_cast_fp16, y = var_14930_cast_fp16)[name = tensor<string, []>("attn_485_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_487_cast_fp16 = softmax(axis = var_21, x = attn_485_cast_fp16)[name = tensor<string, []>("attn_487_cast_fp16")];
            tensor<bool, []> var_14937_transpose_x_0 = const()[name = tensor<string, []>("op_14937_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_14937_transpose_y_0 = const()[name = tensor<string, []>("op_14937_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_14937_cast_fp16 = matmul(transpose_x = var_14937_transpose_x_0, transpose_y = var_14937_transpose_y_0, x = attn_487_cast_fp16, y = v_731_cast_fp16)[name = tensor<string, []>("op_14937_cast_fp16")];
            tensor<int32, [1]> var_14938_axes_0 = const()[name = tensor<string, []>("op_14938_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_14938_cast_fp16 = squeeze(axes = var_14938_axes_0, x = var_14937_cast_fp16)[name = tensor<string, []>("op_14938_cast_fp16")];
            tensor<string, []> var_14944_pad_type_0 = const()[name = tensor<string, []>("op_14944_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14944_strides_0 = const()[name = tensor<string, []>("op_14944_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14944_pad_0 = const()[name = tensor<string, []>("op_14944_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14944_dilations_0 = const()[name = tensor<string, []>("op_14944_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14944_groups_0 = const()[name = tensor<string, []>("op_14944_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144003648)))];
            tensor<fp16, [1, 64, 1, 197]> var_14944_cast_fp16 = conv(dilations = var_14944_dilations_0, groups = var_14944_groups_0, pad = var_14944_pad_0, pad_type = var_14944_pad_type_0, strides = var_14944_strides_0, weight = model_blocks_10_attn_q_projs_2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14944_cast_fp16")];
            tensor<int32, [1]> var_14945_axes_0 = const()[name = tensor<string, []>("op_14945_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14945_cast_fp16 = squeeze(axes = var_14945_axes_0, x = var_14944_cast_fp16)[name = tensor<string, []>("op_14945_cast_fp16")];
            tensor<int32, [3]> q_1221_perm_0 = const()[name = tensor<string, []>("q_1221_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14952_pad_type_0 = const()[name = tensor<string, []>("op_14952_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14952_strides_0 = const()[name = tensor<string, []>("op_14952_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14952_pad_0 = const()[name = tensor<string, []>("op_14952_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14952_dilations_0 = const()[name = tensor<string, []>("op_14952_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14952_groups_0 = const()[name = tensor<string, []>("op_14952_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144102016)))];
            tensor<fp16, [1, 64, 1, 197]> var_14952_cast_fp16 = conv(dilations = var_14952_dilations_0, groups = var_14952_groups_0, pad = var_14952_pad_0, pad_type = var_14952_pad_type_0, strides = var_14952_strides_0, weight = model_blocks_10_attn_k_projs_2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14952_cast_fp16")];
            tensor<int32, [1]> var_14953_axes_0 = const()[name = tensor<string, []>("op_14953_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14953_cast_fp16 = squeeze(axes = var_14953_axes_0, x = var_14952_cast_fp16)[name = tensor<string, []>("op_14953_cast_fp16")];
            tensor<int32, [3]> k_733_perm_0 = const()[name = tensor<string, []>("k_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_14960_pad_type_0 = const()[name = tensor<string, []>("op_14960_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_14960_strides_0 = const()[name = tensor<string, []>("op_14960_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_14960_pad_0 = const()[name = tensor<string, []>("op_14960_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_14960_dilations_0 = const()[name = tensor<string, []>("op_14960_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14960_groups_0 = const()[name = tensor<string, []>("op_14960_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144200384)))];
            tensor<fp16, [1, 64, 1, 197]> var_14960_cast_fp16 = conv(dilations = var_14960_dilations_0, groups = var_14960_groups_0, pad = var_14960_pad_0, pad_type = var_14960_pad_type_0, strides = var_14960_strides_0, weight = model_blocks_10_attn_v_projs_2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_14960_cast_fp16")];
            tensor<int32, [1]> var_14961_axes_0 = const()[name = tensor<string, []>("op_14961_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_14961_cast_fp16 = squeeze(axes = var_14961_axes_0, x = var_14960_cast_fp16)[name = tensor<string, []>("op_14961_cast_fp16")];
            tensor<int32, [3]> v_733_perm_0 = const()[name = tensor<string, []>("v_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144298752)))];
            tensor<fp16, [1, 197, 64]> q_1221_cast_fp16 = transpose(perm = q_1221_perm_0, x = var_14945_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [1, 197, 64]> q_1223_cast_fp16 = add(x = q_1221_cast_fp16, y = model_blocks_10_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_1223_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144298944)))];
            tensor<fp16, [1, 197, 64]> v_733_cast_fp16 = transpose(perm = v_733_perm_0, x = var_14961_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [1, 197, 64]> v_735_cast_fp16 = add(x = v_733_cast_fp16, y = model_blocks_10_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_735_cast_fp16")];
            tensor<int32, [1]> q_1225_axes_0 = const()[name = tensor<string, []>("q_1225_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1225_cast_fp16 = expand_dims(axes = q_1225_axes_0, x = q_1223_cast_fp16)[name = tensor<string, []>("q_1225_cast_fp16")];
            tensor<int32, [1]> k_735_axes_0 = const()[name = tensor<string, []>("k_735_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_733_cast_fp16 = transpose(perm = k_733_perm_0, x = var_14953_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 1, 197, 64]> k_735_cast_fp16 = expand_dims(axes = k_735_axes_0, x = k_733_cast_fp16)[name = tensor<string, []>("k_735_cast_fp16")];
            tensor<int32, [1]> v_737_axes_0 = const()[name = tensor<string, []>("v_737_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_737_cast_fp16 = expand_dims(axes = v_737_axes_0, x = v_735_cast_fp16)[name = tensor<string, []>("v_737_cast_fp16")];
            tensor<int32, [4]> var_14970_begin_0 = const()[name = tensor<string, []>("op_14970_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14970_end_0 = const()[name = tensor<string, []>("op_14970_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14970_end_mask_0 = const()[name = tensor<string, []>("op_14970_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14970_cast_fp16 = slice_by_index(begin = var_14970_begin_0, end = var_14970_end_0, end_mask = var_14970_end_mask_0, x = q_1225_cast_fp16)[name = tensor<string, []>("op_14970_cast_fp16")];
            tensor<int32, [4]> var_14974_begin_0 = const()[name = tensor<string, []>("op_14974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14974_end_0 = const()[name = tensor<string, []>("op_14974_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_14974_end_mask_0 = const()[name = tensor<string, []>("op_14974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_14974_cast_fp16 = slice_by_index(begin = var_14974_begin_0, end = var_14974_end_0, end_mask = var_14974_end_mask_0, x = k_735_cast_fp16)[name = tensor<string, []>("op_14974_cast_fp16")];
            tensor<int32, [4]> var_14978_begin_0 = const()[name = tensor<string, []>("op_14978_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14978_end_0 = const()[name = tensor<string, []>("op_14978_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14978_end_mask_0 = const()[name = tensor<string, []>("op_14978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14978_cast_fp16 = slice_by_index(begin = var_14978_begin_0, end = var_14978_end_0, end_mask = var_14978_end_mask_0, x = q_1225_cast_fp16)[name = tensor<string, []>("op_14978_cast_fp16")];
            tensor<int32, [4]> var_14982_begin_0 = const()[name = tensor<string, []>("op_14982_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14982_end_0 = const()[name = tensor<string, []>("op_14982_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_14982_end_mask_0 = const()[name = tensor<string, []>("op_14982_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_14982_cast_fp16 = slice_by_index(begin = var_14982_begin_0, end = var_14982_end_0, end_mask = var_14982_end_mask_0, x = k_735_cast_fp16)[name = tensor<string, []>("op_14982_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_14984_cast_fp16 = mul(x = var_14978_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_14984_cast_fp16")];
            tensor<int32, [5]> var_14992 = const()[name = tensor<string, []>("op_14992"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1675_cast_fp16 = reshape(shape = var_14992, x = var_14978_cast_fp16)[name = tensor<string, []>("x_1675_cast_fp16")];
            tensor<int32, [2]> var_14994_split_sizes_0 = const()[name = tensor<string, []>("op_14994_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_14994_axis_0 = const()[name = tensor<string, []>("op_14994_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_14994_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_14994_cast_fp16_1 = split(axis = var_14994_axis_0, split_sizes = var_14994_split_sizes_0, x = x_1675_cast_fp16)[name = tensor<string, []>("op_14994_cast_fp16")];
            tensor<int32, [1]> squeeze_488_axes_0 = const()[name = tensor<string, []>("squeeze_488_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_488_cast_fp16 = squeeze(axes = squeeze_488_axes_0, x = var_14994_cast_fp16_0)[name = tensor<string, []>("squeeze_488_cast_fp16")];
            tensor<int32, [1]> squeeze_489_axes_0 = const()[name = tensor<string, []>("squeeze_489_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_489_cast_fp16 = squeeze(axes = squeeze_489_axes_0, x = var_14994_cast_fp16_1)[name = tensor<string, []>("squeeze_489_cast_fp16")];
            tensor<fp16, []> const_1999_promoted_to_fp16 = const()[name = tensor<string, []>("const_1999_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_14997_cast_fp16 = mul(x = squeeze_489_cast_fp16, y = const_1999_promoted_to_fp16)[name = tensor<string, []>("op_14997_cast_fp16")];
            tensor<int32, []> x_1677_axis_0 = const()[name = tensor<string, []>("x_1677_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1677_cast_fp16 = stack(axis = x_1677_axis_0, values = (var_14997_cast_fp16, squeeze_488_cast_fp16))[name = tensor<string, []>("x_1677_cast_fp16")];
            tensor<int32, [4]> var_15003 = const()[name = tensor<string, []>("op_15003"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15004_cast_fp16 = reshape(shape = var_15003, x = x_1677_cast_fp16)[name = tensor<string, []>("op_15004_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15005_cast_fp16 = mul(x = var_15004_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15005_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_245_cast_fp16 = add(x = var_14984_cast_fp16, y = var_15005_cast_fp16)[name = tensor<string, []>("q_patches_245_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15007_cast_fp16 = mul(x = var_14982_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15007_cast_fp16")];
            tensor<int32, [5]> var_15015 = const()[name = tensor<string, []>("op_15015"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1681_cast_fp16 = reshape(shape = var_15015, x = var_14982_cast_fp16)[name = tensor<string, []>("x_1681_cast_fp16")];
            tensor<int32, [2]> var_15017_split_sizes_0 = const()[name = tensor<string, []>("op_15017_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15017_axis_0 = const()[name = tensor<string, []>("op_15017_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15017_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15017_cast_fp16_1 = split(axis = var_15017_axis_0, split_sizes = var_15017_split_sizes_0, x = x_1681_cast_fp16)[name = tensor<string, []>("op_15017_cast_fp16")];
            tensor<int32, [1]> squeeze_490_axes_0 = const()[name = tensor<string, []>("squeeze_490_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_490_cast_fp16 = squeeze(axes = squeeze_490_axes_0, x = var_15017_cast_fp16_0)[name = tensor<string, []>("squeeze_490_cast_fp16")];
            tensor<int32, [1]> squeeze_491_axes_0 = const()[name = tensor<string, []>("squeeze_491_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_491_cast_fp16 = squeeze(axes = squeeze_491_axes_0, x = var_15017_cast_fp16_1)[name = tensor<string, []>("squeeze_491_cast_fp16")];
            tensor<fp16, []> const_2007_promoted_to_fp16 = const()[name = tensor<string, []>("const_2007_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15020_cast_fp16 = mul(x = squeeze_491_cast_fp16, y = const_2007_promoted_to_fp16)[name = tensor<string, []>("op_15020_cast_fp16")];
            tensor<int32, []> x_1683_axis_0 = const()[name = tensor<string, []>("x_1683_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1683_cast_fp16 = stack(axis = x_1683_axis_0, values = (var_15020_cast_fp16, squeeze_490_cast_fp16))[name = tensor<string, []>("x_1683_cast_fp16")];
            tensor<int32, [4]> var_15026 = const()[name = tensor<string, []>("op_15026"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15027_cast_fp16 = reshape(shape = var_15026, x = x_1683_cast_fp16)[name = tensor<string, []>("op_15027_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15028_cast_fp16 = mul(x = var_15027_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15028_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_245_cast_fp16 = add(x = var_15007_cast_fp16, y = var_15028_cast_fp16)[name = tensor<string, []>("k_patches_245_cast_fp16")];
            tensor<bool, []> var_15031_interleave_0 = const()[name = tensor<string, []>("op_15031_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15031_cast_fp16 = concat(axis = var_20, interleave = var_15031_interleave_0, values = (var_14970_cast_fp16, q_patches_245_cast_fp16))[name = tensor<string, []>("op_15031_cast_fp16")];
            tensor<bool, []> var_15034_interleave_0 = const()[name = tensor<string, []>("op_15034_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15034_cast_fp16 = concat(axis = var_20, interleave = var_15034_interleave_0, values = (var_14974_cast_fp16, k_patches_245_cast_fp16))[name = tensor<string, []>("op_15034_cast_fp16")];
            tensor<fp16, []> var_15036_to_fp16 = const()[name = tensor<string, []>("op_15036_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1229_cast_fp16 = mul(x = var_15031_cast_fp16, y = var_15036_to_fp16)[name = tensor<string, []>("q_1229_cast_fp16")];
            tensor<bool, []> attn_489_transpose_x_1 = const()[name = tensor<string, []>("attn_489_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_489_transpose_y_1 = const()[name = tensor<string, []>("attn_489_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_489_cast_fp16 = matmul(transpose_x = attn_489_transpose_x_1, transpose_y = attn_489_transpose_y_1, x = q_1229_cast_fp16, y = var_15034_cast_fp16)[name = tensor<string, []>("attn_489_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_491_cast_fp16 = softmax(axis = var_21, x = attn_489_cast_fp16)[name = tensor<string, []>("attn_491_cast_fp16")];
            tensor<bool, []> var_15041_transpose_x_0 = const()[name = tensor<string, []>("op_15041_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15041_transpose_y_0 = const()[name = tensor<string, []>("op_15041_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15041_cast_fp16 = matmul(transpose_x = var_15041_transpose_x_0, transpose_y = var_15041_transpose_y_0, x = attn_491_cast_fp16, y = v_737_cast_fp16)[name = tensor<string, []>("op_15041_cast_fp16")];
            tensor<int32, [1]> var_15042_axes_0 = const()[name = tensor<string, []>("op_15042_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15042_cast_fp16 = squeeze(axes = var_15042_axes_0, x = var_15041_cast_fp16)[name = tensor<string, []>("op_15042_cast_fp16")];
            tensor<string, []> var_15048_pad_type_0 = const()[name = tensor<string, []>("op_15048_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15048_strides_0 = const()[name = tensor<string, []>("op_15048_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15048_pad_0 = const()[name = tensor<string, []>("op_15048_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15048_dilations_0 = const()[name = tensor<string, []>("op_15048_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15048_groups_0 = const()[name = tensor<string, []>("op_15048_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144299136)))];
            tensor<fp16, [1, 64, 1, 197]> var_15048_cast_fp16 = conv(dilations = var_15048_dilations_0, groups = var_15048_groups_0, pad = var_15048_pad_0, pad_type = var_15048_pad_type_0, strides = var_15048_strides_0, weight = model_blocks_10_attn_q_projs_3_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15048_cast_fp16")];
            tensor<int32, [1]> var_15049_axes_0 = const()[name = tensor<string, []>("op_15049_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15049_cast_fp16 = squeeze(axes = var_15049_axes_0, x = var_15048_cast_fp16)[name = tensor<string, []>("op_15049_cast_fp16")];
            tensor<int32, [3]> q_1231_perm_0 = const()[name = tensor<string, []>("q_1231_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15056_pad_type_0 = const()[name = tensor<string, []>("op_15056_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15056_strides_0 = const()[name = tensor<string, []>("op_15056_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15056_pad_0 = const()[name = tensor<string, []>("op_15056_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15056_dilations_0 = const()[name = tensor<string, []>("op_15056_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15056_groups_0 = const()[name = tensor<string, []>("op_15056_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144397504)))];
            tensor<fp16, [1, 64, 1, 197]> var_15056_cast_fp16 = conv(dilations = var_15056_dilations_0, groups = var_15056_groups_0, pad = var_15056_pad_0, pad_type = var_15056_pad_type_0, strides = var_15056_strides_0, weight = model_blocks_10_attn_k_projs_3_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15056_cast_fp16")];
            tensor<int32, [1]> var_15057_axes_0 = const()[name = tensor<string, []>("op_15057_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15057_cast_fp16 = squeeze(axes = var_15057_axes_0, x = var_15056_cast_fp16)[name = tensor<string, []>("op_15057_cast_fp16")];
            tensor<int32, [3]> k_739_perm_0 = const()[name = tensor<string, []>("k_739_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15064_pad_type_0 = const()[name = tensor<string, []>("op_15064_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15064_strides_0 = const()[name = tensor<string, []>("op_15064_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15064_pad_0 = const()[name = tensor<string, []>("op_15064_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15064_dilations_0 = const()[name = tensor<string, []>("op_15064_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15064_groups_0 = const()[name = tensor<string, []>("op_15064_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144495872)))];
            tensor<fp16, [1, 64, 1, 197]> var_15064_cast_fp16 = conv(dilations = var_15064_dilations_0, groups = var_15064_groups_0, pad = var_15064_pad_0, pad_type = var_15064_pad_type_0, strides = var_15064_strides_0, weight = model_blocks_10_attn_v_projs_3_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15064_cast_fp16")];
            tensor<int32, [1]> var_15065_axes_0 = const()[name = tensor<string, []>("op_15065_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15065_cast_fp16 = squeeze(axes = var_15065_axes_0, x = var_15064_cast_fp16)[name = tensor<string, []>("op_15065_cast_fp16")];
            tensor<int32, [3]> v_739_perm_0 = const()[name = tensor<string, []>("v_739_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144594240)))];
            tensor<fp16, [1, 197, 64]> q_1231_cast_fp16 = transpose(perm = q_1231_perm_0, x = var_15049_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [1, 197, 64]> q_1233_cast_fp16 = add(x = q_1231_cast_fp16, y = model_blocks_10_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_1233_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144594432)))];
            tensor<fp16, [1, 197, 64]> v_739_cast_fp16 = transpose(perm = v_739_perm_0, x = var_15065_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [1, 197, 64]> v_741_cast_fp16 = add(x = v_739_cast_fp16, y = model_blocks_10_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_741_cast_fp16")];
            tensor<int32, [1]> q_1235_axes_0 = const()[name = tensor<string, []>("q_1235_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1235_cast_fp16 = expand_dims(axes = q_1235_axes_0, x = q_1233_cast_fp16)[name = tensor<string, []>("q_1235_cast_fp16")];
            tensor<int32, [1]> k_741_axes_0 = const()[name = tensor<string, []>("k_741_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_739_cast_fp16 = transpose(perm = k_739_perm_0, x = var_15057_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [1, 1, 197, 64]> k_741_cast_fp16 = expand_dims(axes = k_741_axes_0, x = k_739_cast_fp16)[name = tensor<string, []>("k_741_cast_fp16")];
            tensor<int32, [1]> v_743_axes_0 = const()[name = tensor<string, []>("v_743_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_743_cast_fp16 = expand_dims(axes = v_743_axes_0, x = v_741_cast_fp16)[name = tensor<string, []>("v_743_cast_fp16")];
            tensor<int32, [4]> var_15074_begin_0 = const()[name = tensor<string, []>("op_15074_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15074_end_0 = const()[name = tensor<string, []>("op_15074_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15074_end_mask_0 = const()[name = tensor<string, []>("op_15074_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15074_cast_fp16 = slice_by_index(begin = var_15074_begin_0, end = var_15074_end_0, end_mask = var_15074_end_mask_0, x = q_1235_cast_fp16)[name = tensor<string, []>("op_15074_cast_fp16")];
            tensor<int32, [4]> var_15078_begin_0 = const()[name = tensor<string, []>("op_15078_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15078_end_0 = const()[name = tensor<string, []>("op_15078_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15078_end_mask_0 = const()[name = tensor<string, []>("op_15078_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15078_cast_fp16 = slice_by_index(begin = var_15078_begin_0, end = var_15078_end_0, end_mask = var_15078_end_mask_0, x = k_741_cast_fp16)[name = tensor<string, []>("op_15078_cast_fp16")];
            tensor<int32, [4]> var_15082_begin_0 = const()[name = tensor<string, []>("op_15082_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15082_end_0 = const()[name = tensor<string, []>("op_15082_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15082_end_mask_0 = const()[name = tensor<string, []>("op_15082_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15082_cast_fp16 = slice_by_index(begin = var_15082_begin_0, end = var_15082_end_0, end_mask = var_15082_end_mask_0, x = q_1235_cast_fp16)[name = tensor<string, []>("op_15082_cast_fp16")];
            tensor<int32, [4]> var_15086_begin_0 = const()[name = tensor<string, []>("op_15086_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15086_end_0 = const()[name = tensor<string, []>("op_15086_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15086_end_mask_0 = const()[name = tensor<string, []>("op_15086_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15086_cast_fp16 = slice_by_index(begin = var_15086_begin_0, end = var_15086_end_0, end_mask = var_15086_end_mask_0, x = k_741_cast_fp16)[name = tensor<string, []>("op_15086_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15088_cast_fp16 = mul(x = var_15082_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15088_cast_fp16")];
            tensor<int32, [5]> var_15096 = const()[name = tensor<string, []>("op_15096"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1687_cast_fp16 = reshape(shape = var_15096, x = var_15082_cast_fp16)[name = tensor<string, []>("x_1687_cast_fp16")];
            tensor<int32, [2]> var_15098_split_sizes_0 = const()[name = tensor<string, []>("op_15098_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15098_axis_0 = const()[name = tensor<string, []>("op_15098_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15098_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15098_cast_fp16_1 = split(axis = var_15098_axis_0, split_sizes = var_15098_split_sizes_0, x = x_1687_cast_fp16)[name = tensor<string, []>("op_15098_cast_fp16")];
            tensor<int32, [1]> squeeze_492_axes_0 = const()[name = tensor<string, []>("squeeze_492_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_492_cast_fp16 = squeeze(axes = squeeze_492_axes_0, x = var_15098_cast_fp16_0)[name = tensor<string, []>("squeeze_492_cast_fp16")];
            tensor<int32, [1]> squeeze_493_axes_0 = const()[name = tensor<string, []>("squeeze_493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_493_cast_fp16 = squeeze(axes = squeeze_493_axes_0, x = var_15098_cast_fp16_1)[name = tensor<string, []>("squeeze_493_cast_fp16")];
            tensor<fp16, []> const_2015_promoted_to_fp16 = const()[name = tensor<string, []>("const_2015_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15101_cast_fp16 = mul(x = squeeze_493_cast_fp16, y = const_2015_promoted_to_fp16)[name = tensor<string, []>("op_15101_cast_fp16")];
            tensor<int32, []> x_1689_axis_0 = const()[name = tensor<string, []>("x_1689_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1689_cast_fp16 = stack(axis = x_1689_axis_0, values = (var_15101_cast_fp16, squeeze_492_cast_fp16))[name = tensor<string, []>("x_1689_cast_fp16")];
            tensor<int32, [4]> var_15107 = const()[name = tensor<string, []>("op_15107"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15108_cast_fp16 = reshape(shape = var_15107, x = x_1689_cast_fp16)[name = tensor<string, []>("op_15108_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15109_cast_fp16 = mul(x = var_15108_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15109_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_247_cast_fp16 = add(x = var_15088_cast_fp16, y = var_15109_cast_fp16)[name = tensor<string, []>("q_patches_247_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15111_cast_fp16 = mul(x = var_15086_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15111_cast_fp16")];
            tensor<int32, [5]> var_15119 = const()[name = tensor<string, []>("op_15119"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1693_cast_fp16 = reshape(shape = var_15119, x = var_15086_cast_fp16)[name = tensor<string, []>("x_1693_cast_fp16")];
            tensor<int32, [2]> var_15121_split_sizes_0 = const()[name = tensor<string, []>("op_15121_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15121_axis_0 = const()[name = tensor<string, []>("op_15121_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15121_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15121_cast_fp16_1 = split(axis = var_15121_axis_0, split_sizes = var_15121_split_sizes_0, x = x_1693_cast_fp16)[name = tensor<string, []>("op_15121_cast_fp16")];
            tensor<int32, [1]> squeeze_494_axes_0 = const()[name = tensor<string, []>("squeeze_494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_494_cast_fp16 = squeeze(axes = squeeze_494_axes_0, x = var_15121_cast_fp16_0)[name = tensor<string, []>("squeeze_494_cast_fp16")];
            tensor<int32, [1]> squeeze_495_axes_0 = const()[name = tensor<string, []>("squeeze_495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_495_cast_fp16 = squeeze(axes = squeeze_495_axes_0, x = var_15121_cast_fp16_1)[name = tensor<string, []>("squeeze_495_cast_fp16")];
            tensor<fp16, []> const_2023_promoted_to_fp16 = const()[name = tensor<string, []>("const_2023_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15124_cast_fp16 = mul(x = squeeze_495_cast_fp16, y = const_2023_promoted_to_fp16)[name = tensor<string, []>("op_15124_cast_fp16")];
            tensor<int32, []> x_1695_axis_0 = const()[name = tensor<string, []>("x_1695_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1695_cast_fp16 = stack(axis = x_1695_axis_0, values = (var_15124_cast_fp16, squeeze_494_cast_fp16))[name = tensor<string, []>("x_1695_cast_fp16")];
            tensor<int32, [4]> var_15130 = const()[name = tensor<string, []>("op_15130"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15131_cast_fp16 = reshape(shape = var_15130, x = x_1695_cast_fp16)[name = tensor<string, []>("op_15131_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15132_cast_fp16 = mul(x = var_15131_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15132_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_247_cast_fp16 = add(x = var_15111_cast_fp16, y = var_15132_cast_fp16)[name = tensor<string, []>("k_patches_247_cast_fp16")];
            tensor<bool, []> var_15135_interleave_0 = const()[name = tensor<string, []>("op_15135_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15135_cast_fp16 = concat(axis = var_20, interleave = var_15135_interleave_0, values = (var_15074_cast_fp16, q_patches_247_cast_fp16))[name = tensor<string, []>("op_15135_cast_fp16")];
            tensor<bool, []> var_15138_interleave_0 = const()[name = tensor<string, []>("op_15138_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15138_cast_fp16 = concat(axis = var_20, interleave = var_15138_interleave_0, values = (var_15078_cast_fp16, k_patches_247_cast_fp16))[name = tensor<string, []>("op_15138_cast_fp16")];
            tensor<fp16, []> var_15140_to_fp16 = const()[name = tensor<string, []>("op_15140_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1239_cast_fp16 = mul(x = var_15135_cast_fp16, y = var_15140_to_fp16)[name = tensor<string, []>("q_1239_cast_fp16")];
            tensor<bool, []> attn_493_transpose_x_1 = const()[name = tensor<string, []>("attn_493_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_493_transpose_y_1 = const()[name = tensor<string, []>("attn_493_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_493_cast_fp16 = matmul(transpose_x = attn_493_transpose_x_1, transpose_y = attn_493_transpose_y_1, x = q_1239_cast_fp16, y = var_15138_cast_fp16)[name = tensor<string, []>("attn_493_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_495_cast_fp16 = softmax(axis = var_21, x = attn_493_cast_fp16)[name = tensor<string, []>("attn_495_cast_fp16")];
            tensor<bool, []> var_15145_transpose_x_0 = const()[name = tensor<string, []>("op_15145_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15145_transpose_y_0 = const()[name = tensor<string, []>("op_15145_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15145_cast_fp16 = matmul(transpose_x = var_15145_transpose_x_0, transpose_y = var_15145_transpose_y_0, x = attn_495_cast_fp16, y = v_743_cast_fp16)[name = tensor<string, []>("op_15145_cast_fp16")];
            tensor<int32, [1]> var_15146_axes_0 = const()[name = tensor<string, []>("op_15146_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15146_cast_fp16 = squeeze(axes = var_15146_axes_0, x = var_15145_cast_fp16)[name = tensor<string, []>("op_15146_cast_fp16")];
            tensor<string, []> var_15152_pad_type_0 = const()[name = tensor<string, []>("op_15152_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15152_strides_0 = const()[name = tensor<string, []>("op_15152_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15152_pad_0 = const()[name = tensor<string, []>("op_15152_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15152_dilations_0 = const()[name = tensor<string, []>("op_15152_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15152_groups_0 = const()[name = tensor<string, []>("op_15152_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144594624)))];
            tensor<fp16, [1, 64, 1, 197]> var_15152_cast_fp16 = conv(dilations = var_15152_dilations_0, groups = var_15152_groups_0, pad = var_15152_pad_0, pad_type = var_15152_pad_type_0, strides = var_15152_strides_0, weight = model_blocks_10_attn_q_projs_4_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15152_cast_fp16")];
            tensor<int32, [1]> var_15153_axes_0 = const()[name = tensor<string, []>("op_15153_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15153_cast_fp16 = squeeze(axes = var_15153_axes_0, x = var_15152_cast_fp16)[name = tensor<string, []>("op_15153_cast_fp16")];
            tensor<int32, [3]> q_1241_perm_0 = const()[name = tensor<string, []>("q_1241_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15160_pad_type_0 = const()[name = tensor<string, []>("op_15160_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15160_strides_0 = const()[name = tensor<string, []>("op_15160_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15160_pad_0 = const()[name = tensor<string, []>("op_15160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15160_dilations_0 = const()[name = tensor<string, []>("op_15160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15160_groups_0 = const()[name = tensor<string, []>("op_15160_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144692992)))];
            tensor<fp16, [1, 64, 1, 197]> var_15160_cast_fp16 = conv(dilations = var_15160_dilations_0, groups = var_15160_groups_0, pad = var_15160_pad_0, pad_type = var_15160_pad_type_0, strides = var_15160_strides_0, weight = model_blocks_10_attn_k_projs_4_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15160_cast_fp16")];
            tensor<int32, [1]> var_15161_axes_0 = const()[name = tensor<string, []>("op_15161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15161_cast_fp16 = squeeze(axes = var_15161_axes_0, x = var_15160_cast_fp16)[name = tensor<string, []>("op_15161_cast_fp16")];
            tensor<int32, [3]> k_745_perm_0 = const()[name = tensor<string, []>("k_745_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15168_pad_type_0 = const()[name = tensor<string, []>("op_15168_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15168_strides_0 = const()[name = tensor<string, []>("op_15168_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15168_pad_0 = const()[name = tensor<string, []>("op_15168_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15168_dilations_0 = const()[name = tensor<string, []>("op_15168_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15168_groups_0 = const()[name = tensor<string, []>("op_15168_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144791360)))];
            tensor<fp16, [1, 64, 1, 197]> var_15168_cast_fp16 = conv(dilations = var_15168_dilations_0, groups = var_15168_groups_0, pad = var_15168_pad_0, pad_type = var_15168_pad_type_0, strides = var_15168_strides_0, weight = model_blocks_10_attn_v_projs_4_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15168_cast_fp16")];
            tensor<int32, [1]> var_15169_axes_0 = const()[name = tensor<string, []>("op_15169_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15169_cast_fp16 = squeeze(axes = var_15169_axes_0, x = var_15168_cast_fp16)[name = tensor<string, []>("op_15169_cast_fp16")];
            tensor<int32, [3]> v_745_perm_0 = const()[name = tensor<string, []>("v_745_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144889728)))];
            tensor<fp16, [1, 197, 64]> q_1241_cast_fp16 = transpose(perm = q_1241_perm_0, x = var_15153_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 197, 64]> q_1243_cast_fp16 = add(x = q_1241_cast_fp16, y = model_blocks_10_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_1243_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144889920)))];
            tensor<fp16, [1, 197, 64]> v_745_cast_fp16 = transpose(perm = v_745_perm_0, x = var_15169_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [1, 197, 64]> v_747_cast_fp16 = add(x = v_745_cast_fp16, y = model_blocks_10_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_747_cast_fp16")];
            tensor<int32, [1]> q_1245_axes_0 = const()[name = tensor<string, []>("q_1245_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1245_cast_fp16 = expand_dims(axes = q_1245_axes_0, x = q_1243_cast_fp16)[name = tensor<string, []>("q_1245_cast_fp16")];
            tensor<int32, [1]> k_747_axes_0 = const()[name = tensor<string, []>("k_747_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_745_cast_fp16 = transpose(perm = k_745_perm_0, x = var_15161_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [1, 1, 197, 64]> k_747_cast_fp16 = expand_dims(axes = k_747_axes_0, x = k_745_cast_fp16)[name = tensor<string, []>("k_747_cast_fp16")];
            tensor<int32, [1]> v_749_axes_0 = const()[name = tensor<string, []>("v_749_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_749_cast_fp16 = expand_dims(axes = v_749_axes_0, x = v_747_cast_fp16)[name = tensor<string, []>("v_749_cast_fp16")];
            tensor<int32, [4]> var_15178_begin_0 = const()[name = tensor<string, []>("op_15178_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15178_end_0 = const()[name = tensor<string, []>("op_15178_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15178_end_mask_0 = const()[name = tensor<string, []>("op_15178_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15178_cast_fp16 = slice_by_index(begin = var_15178_begin_0, end = var_15178_end_0, end_mask = var_15178_end_mask_0, x = q_1245_cast_fp16)[name = tensor<string, []>("op_15178_cast_fp16")];
            tensor<int32, [4]> var_15182_begin_0 = const()[name = tensor<string, []>("op_15182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15182_end_0 = const()[name = tensor<string, []>("op_15182_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15182_end_mask_0 = const()[name = tensor<string, []>("op_15182_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15182_cast_fp16 = slice_by_index(begin = var_15182_begin_0, end = var_15182_end_0, end_mask = var_15182_end_mask_0, x = k_747_cast_fp16)[name = tensor<string, []>("op_15182_cast_fp16")];
            tensor<int32, [4]> var_15186_begin_0 = const()[name = tensor<string, []>("op_15186_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15186_end_0 = const()[name = tensor<string, []>("op_15186_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15186_end_mask_0 = const()[name = tensor<string, []>("op_15186_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15186_cast_fp16 = slice_by_index(begin = var_15186_begin_0, end = var_15186_end_0, end_mask = var_15186_end_mask_0, x = q_1245_cast_fp16)[name = tensor<string, []>("op_15186_cast_fp16")];
            tensor<int32, [4]> var_15190_begin_0 = const()[name = tensor<string, []>("op_15190_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15190_end_0 = const()[name = tensor<string, []>("op_15190_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15190_end_mask_0 = const()[name = tensor<string, []>("op_15190_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15190_cast_fp16 = slice_by_index(begin = var_15190_begin_0, end = var_15190_end_0, end_mask = var_15190_end_mask_0, x = k_747_cast_fp16)[name = tensor<string, []>("op_15190_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15192_cast_fp16 = mul(x = var_15186_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15192_cast_fp16")];
            tensor<int32, [5]> var_15200 = const()[name = tensor<string, []>("op_15200"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1699_cast_fp16 = reshape(shape = var_15200, x = var_15186_cast_fp16)[name = tensor<string, []>("x_1699_cast_fp16")];
            tensor<int32, [2]> var_15202_split_sizes_0 = const()[name = tensor<string, []>("op_15202_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15202_axis_0 = const()[name = tensor<string, []>("op_15202_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15202_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15202_cast_fp16_1 = split(axis = var_15202_axis_0, split_sizes = var_15202_split_sizes_0, x = x_1699_cast_fp16)[name = tensor<string, []>("op_15202_cast_fp16")];
            tensor<int32, [1]> squeeze_496_axes_0 = const()[name = tensor<string, []>("squeeze_496_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_496_cast_fp16 = squeeze(axes = squeeze_496_axes_0, x = var_15202_cast_fp16_0)[name = tensor<string, []>("squeeze_496_cast_fp16")];
            tensor<int32, [1]> squeeze_497_axes_0 = const()[name = tensor<string, []>("squeeze_497_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_497_cast_fp16 = squeeze(axes = squeeze_497_axes_0, x = var_15202_cast_fp16_1)[name = tensor<string, []>("squeeze_497_cast_fp16")];
            tensor<fp16, []> const_2031_promoted_to_fp16 = const()[name = tensor<string, []>("const_2031_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15205_cast_fp16 = mul(x = squeeze_497_cast_fp16, y = const_2031_promoted_to_fp16)[name = tensor<string, []>("op_15205_cast_fp16")];
            tensor<int32, []> x_1701_axis_0 = const()[name = tensor<string, []>("x_1701_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1701_cast_fp16 = stack(axis = x_1701_axis_0, values = (var_15205_cast_fp16, squeeze_496_cast_fp16))[name = tensor<string, []>("x_1701_cast_fp16")];
            tensor<int32, [4]> var_15211 = const()[name = tensor<string, []>("op_15211"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15212_cast_fp16 = reshape(shape = var_15211, x = x_1701_cast_fp16)[name = tensor<string, []>("op_15212_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15213_cast_fp16 = mul(x = var_15212_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15213_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_249_cast_fp16 = add(x = var_15192_cast_fp16, y = var_15213_cast_fp16)[name = tensor<string, []>("q_patches_249_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15215_cast_fp16 = mul(x = var_15190_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15215_cast_fp16")];
            tensor<int32, [5]> var_15223 = const()[name = tensor<string, []>("op_15223"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1705_cast_fp16 = reshape(shape = var_15223, x = var_15190_cast_fp16)[name = tensor<string, []>("x_1705_cast_fp16")];
            tensor<int32, [2]> var_15225_split_sizes_0 = const()[name = tensor<string, []>("op_15225_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15225_axis_0 = const()[name = tensor<string, []>("op_15225_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15225_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15225_cast_fp16_1 = split(axis = var_15225_axis_0, split_sizes = var_15225_split_sizes_0, x = x_1705_cast_fp16)[name = tensor<string, []>("op_15225_cast_fp16")];
            tensor<int32, [1]> squeeze_498_axes_0 = const()[name = tensor<string, []>("squeeze_498_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_498_cast_fp16 = squeeze(axes = squeeze_498_axes_0, x = var_15225_cast_fp16_0)[name = tensor<string, []>("squeeze_498_cast_fp16")];
            tensor<int32, [1]> squeeze_499_axes_0 = const()[name = tensor<string, []>("squeeze_499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_499_cast_fp16 = squeeze(axes = squeeze_499_axes_0, x = var_15225_cast_fp16_1)[name = tensor<string, []>("squeeze_499_cast_fp16")];
            tensor<fp16, []> const_2039_promoted_to_fp16 = const()[name = tensor<string, []>("const_2039_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15228_cast_fp16 = mul(x = squeeze_499_cast_fp16, y = const_2039_promoted_to_fp16)[name = tensor<string, []>("op_15228_cast_fp16")];
            tensor<int32, []> x_1707_axis_0 = const()[name = tensor<string, []>("x_1707_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1707_cast_fp16 = stack(axis = x_1707_axis_0, values = (var_15228_cast_fp16, squeeze_498_cast_fp16))[name = tensor<string, []>("x_1707_cast_fp16")];
            tensor<int32, [4]> var_15234 = const()[name = tensor<string, []>("op_15234"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15235_cast_fp16 = reshape(shape = var_15234, x = x_1707_cast_fp16)[name = tensor<string, []>("op_15235_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15236_cast_fp16 = mul(x = var_15235_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15236_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_249_cast_fp16 = add(x = var_15215_cast_fp16, y = var_15236_cast_fp16)[name = tensor<string, []>("k_patches_249_cast_fp16")];
            tensor<bool, []> var_15239_interleave_0 = const()[name = tensor<string, []>("op_15239_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15239_cast_fp16 = concat(axis = var_20, interleave = var_15239_interleave_0, values = (var_15178_cast_fp16, q_patches_249_cast_fp16))[name = tensor<string, []>("op_15239_cast_fp16")];
            tensor<bool, []> var_15242_interleave_0 = const()[name = tensor<string, []>("op_15242_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15242_cast_fp16 = concat(axis = var_20, interleave = var_15242_interleave_0, values = (var_15182_cast_fp16, k_patches_249_cast_fp16))[name = tensor<string, []>("op_15242_cast_fp16")];
            tensor<fp16, []> var_15244_to_fp16 = const()[name = tensor<string, []>("op_15244_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1249_cast_fp16 = mul(x = var_15239_cast_fp16, y = var_15244_to_fp16)[name = tensor<string, []>("q_1249_cast_fp16")];
            tensor<bool, []> attn_497_transpose_x_1 = const()[name = tensor<string, []>("attn_497_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_497_transpose_y_1 = const()[name = tensor<string, []>("attn_497_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_497_cast_fp16 = matmul(transpose_x = attn_497_transpose_x_1, transpose_y = attn_497_transpose_y_1, x = q_1249_cast_fp16, y = var_15242_cast_fp16)[name = tensor<string, []>("attn_497_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_499_cast_fp16 = softmax(axis = var_21, x = attn_497_cast_fp16)[name = tensor<string, []>("attn_499_cast_fp16")];
            tensor<bool, []> var_15249_transpose_x_0 = const()[name = tensor<string, []>("op_15249_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15249_transpose_y_0 = const()[name = tensor<string, []>("op_15249_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15249_cast_fp16 = matmul(transpose_x = var_15249_transpose_x_0, transpose_y = var_15249_transpose_y_0, x = attn_499_cast_fp16, y = v_749_cast_fp16)[name = tensor<string, []>("op_15249_cast_fp16")];
            tensor<int32, [1]> var_15250_axes_0 = const()[name = tensor<string, []>("op_15250_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15250_cast_fp16 = squeeze(axes = var_15250_axes_0, x = var_15249_cast_fp16)[name = tensor<string, []>("op_15250_cast_fp16")];
            tensor<string, []> var_15256_pad_type_0 = const()[name = tensor<string, []>("op_15256_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15256_strides_0 = const()[name = tensor<string, []>("op_15256_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15256_pad_0 = const()[name = tensor<string, []>("op_15256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15256_dilations_0 = const()[name = tensor<string, []>("op_15256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15256_groups_0 = const()[name = tensor<string, []>("op_15256_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144890112)))];
            tensor<fp16, [1, 64, 1, 197]> var_15256_cast_fp16 = conv(dilations = var_15256_dilations_0, groups = var_15256_groups_0, pad = var_15256_pad_0, pad_type = var_15256_pad_type_0, strides = var_15256_strides_0, weight = model_blocks_10_attn_q_projs_5_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15256_cast_fp16")];
            tensor<int32, [1]> var_15257_axes_0 = const()[name = tensor<string, []>("op_15257_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15257_cast_fp16 = squeeze(axes = var_15257_axes_0, x = var_15256_cast_fp16)[name = tensor<string, []>("op_15257_cast_fp16")];
            tensor<int32, [3]> q_1251_perm_0 = const()[name = tensor<string, []>("q_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15264_pad_type_0 = const()[name = tensor<string, []>("op_15264_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15264_strides_0 = const()[name = tensor<string, []>("op_15264_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15264_pad_0 = const()[name = tensor<string, []>("op_15264_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15264_dilations_0 = const()[name = tensor<string, []>("op_15264_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15264_groups_0 = const()[name = tensor<string, []>("op_15264_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144988480)))];
            tensor<fp16, [1, 64, 1, 197]> var_15264_cast_fp16 = conv(dilations = var_15264_dilations_0, groups = var_15264_groups_0, pad = var_15264_pad_0, pad_type = var_15264_pad_type_0, strides = var_15264_strides_0, weight = model_blocks_10_attn_k_projs_5_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15264_cast_fp16")];
            tensor<int32, [1]> var_15265_axes_0 = const()[name = tensor<string, []>("op_15265_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15265_cast_fp16 = squeeze(axes = var_15265_axes_0, x = var_15264_cast_fp16)[name = tensor<string, []>("op_15265_cast_fp16")];
            tensor<int32, [3]> k_751_perm_0 = const()[name = tensor<string, []>("k_751_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15272_pad_type_0 = const()[name = tensor<string, []>("op_15272_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15272_strides_0 = const()[name = tensor<string, []>("op_15272_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15272_pad_0 = const()[name = tensor<string, []>("op_15272_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15272_dilations_0 = const()[name = tensor<string, []>("op_15272_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15272_groups_0 = const()[name = tensor<string, []>("op_15272_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145086848)))];
            tensor<fp16, [1, 64, 1, 197]> var_15272_cast_fp16 = conv(dilations = var_15272_dilations_0, groups = var_15272_groups_0, pad = var_15272_pad_0, pad_type = var_15272_pad_type_0, strides = var_15272_strides_0, weight = model_blocks_10_attn_v_projs_5_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15272_cast_fp16")];
            tensor<int32, [1]> var_15273_axes_0 = const()[name = tensor<string, []>("op_15273_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15273_cast_fp16 = squeeze(axes = var_15273_axes_0, x = var_15272_cast_fp16)[name = tensor<string, []>("op_15273_cast_fp16")];
            tensor<int32, [3]> v_751_perm_0 = const()[name = tensor<string, []>("v_751_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145185216)))];
            tensor<fp16, [1, 197, 64]> q_1251_cast_fp16 = transpose(perm = q_1251_perm_0, x = var_15257_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [1, 197, 64]> q_1253_cast_fp16 = add(x = q_1251_cast_fp16, y = model_blocks_10_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_1253_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145185408)))];
            tensor<fp16, [1, 197, 64]> v_751_cast_fp16 = transpose(perm = v_751_perm_0, x = var_15273_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [1, 197, 64]> v_753_cast_fp16 = add(x = v_751_cast_fp16, y = model_blocks_10_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_753_cast_fp16")];
            tensor<int32, [1]> q_1255_axes_0 = const()[name = tensor<string, []>("q_1255_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1255_cast_fp16 = expand_dims(axes = q_1255_axes_0, x = q_1253_cast_fp16)[name = tensor<string, []>("q_1255_cast_fp16")];
            tensor<int32, [1]> k_753_axes_0 = const()[name = tensor<string, []>("k_753_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_751_cast_fp16 = transpose(perm = k_751_perm_0, x = var_15265_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 1, 197, 64]> k_753_cast_fp16 = expand_dims(axes = k_753_axes_0, x = k_751_cast_fp16)[name = tensor<string, []>("k_753_cast_fp16")];
            tensor<int32, [1]> v_755_axes_0 = const()[name = tensor<string, []>("v_755_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_755_cast_fp16 = expand_dims(axes = v_755_axes_0, x = v_753_cast_fp16)[name = tensor<string, []>("v_755_cast_fp16")];
            tensor<int32, [4]> var_15282_begin_0 = const()[name = tensor<string, []>("op_15282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15282_end_0 = const()[name = tensor<string, []>("op_15282_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15282_end_mask_0 = const()[name = tensor<string, []>("op_15282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15282_cast_fp16 = slice_by_index(begin = var_15282_begin_0, end = var_15282_end_0, end_mask = var_15282_end_mask_0, x = q_1255_cast_fp16)[name = tensor<string, []>("op_15282_cast_fp16")];
            tensor<int32, [4]> var_15286_begin_0 = const()[name = tensor<string, []>("op_15286_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15286_end_0 = const()[name = tensor<string, []>("op_15286_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15286_end_mask_0 = const()[name = tensor<string, []>("op_15286_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15286_cast_fp16 = slice_by_index(begin = var_15286_begin_0, end = var_15286_end_0, end_mask = var_15286_end_mask_0, x = k_753_cast_fp16)[name = tensor<string, []>("op_15286_cast_fp16")];
            tensor<int32, [4]> var_15290_begin_0 = const()[name = tensor<string, []>("op_15290_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15290_end_0 = const()[name = tensor<string, []>("op_15290_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15290_end_mask_0 = const()[name = tensor<string, []>("op_15290_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15290_cast_fp16 = slice_by_index(begin = var_15290_begin_0, end = var_15290_end_0, end_mask = var_15290_end_mask_0, x = q_1255_cast_fp16)[name = tensor<string, []>("op_15290_cast_fp16")];
            tensor<int32, [4]> var_15294_begin_0 = const()[name = tensor<string, []>("op_15294_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15294_end_0 = const()[name = tensor<string, []>("op_15294_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15294_end_mask_0 = const()[name = tensor<string, []>("op_15294_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15294_cast_fp16 = slice_by_index(begin = var_15294_begin_0, end = var_15294_end_0, end_mask = var_15294_end_mask_0, x = k_753_cast_fp16)[name = tensor<string, []>("op_15294_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15296_cast_fp16 = mul(x = var_15290_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15296_cast_fp16")];
            tensor<int32, [5]> var_15304 = const()[name = tensor<string, []>("op_15304"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1711_cast_fp16 = reshape(shape = var_15304, x = var_15290_cast_fp16)[name = tensor<string, []>("x_1711_cast_fp16")];
            tensor<int32, [2]> var_15306_split_sizes_0 = const()[name = tensor<string, []>("op_15306_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15306_axis_0 = const()[name = tensor<string, []>("op_15306_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15306_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15306_cast_fp16_1 = split(axis = var_15306_axis_0, split_sizes = var_15306_split_sizes_0, x = x_1711_cast_fp16)[name = tensor<string, []>("op_15306_cast_fp16")];
            tensor<int32, [1]> squeeze_500_axes_0 = const()[name = tensor<string, []>("squeeze_500_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_500_cast_fp16 = squeeze(axes = squeeze_500_axes_0, x = var_15306_cast_fp16_0)[name = tensor<string, []>("squeeze_500_cast_fp16")];
            tensor<int32, [1]> squeeze_501_axes_0 = const()[name = tensor<string, []>("squeeze_501_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_501_cast_fp16 = squeeze(axes = squeeze_501_axes_0, x = var_15306_cast_fp16_1)[name = tensor<string, []>("squeeze_501_cast_fp16")];
            tensor<fp16, []> const_2047_promoted_to_fp16 = const()[name = tensor<string, []>("const_2047_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15309_cast_fp16 = mul(x = squeeze_501_cast_fp16, y = const_2047_promoted_to_fp16)[name = tensor<string, []>("op_15309_cast_fp16")];
            tensor<int32, []> x_1713_axis_0 = const()[name = tensor<string, []>("x_1713_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1713_cast_fp16 = stack(axis = x_1713_axis_0, values = (var_15309_cast_fp16, squeeze_500_cast_fp16))[name = tensor<string, []>("x_1713_cast_fp16")];
            tensor<int32, [4]> var_15315 = const()[name = tensor<string, []>("op_15315"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15316_cast_fp16 = reshape(shape = var_15315, x = x_1713_cast_fp16)[name = tensor<string, []>("op_15316_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15317_cast_fp16 = mul(x = var_15316_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15317_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_251_cast_fp16 = add(x = var_15296_cast_fp16, y = var_15317_cast_fp16)[name = tensor<string, []>("q_patches_251_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15319_cast_fp16 = mul(x = var_15294_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15319_cast_fp16")];
            tensor<int32, [5]> var_15327 = const()[name = tensor<string, []>("op_15327"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1717_cast_fp16 = reshape(shape = var_15327, x = var_15294_cast_fp16)[name = tensor<string, []>("x_1717_cast_fp16")];
            tensor<int32, [2]> var_15329_split_sizes_0 = const()[name = tensor<string, []>("op_15329_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15329_axis_0 = const()[name = tensor<string, []>("op_15329_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15329_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15329_cast_fp16_1 = split(axis = var_15329_axis_0, split_sizes = var_15329_split_sizes_0, x = x_1717_cast_fp16)[name = tensor<string, []>("op_15329_cast_fp16")];
            tensor<int32, [1]> squeeze_502_axes_0 = const()[name = tensor<string, []>("squeeze_502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_502_cast_fp16 = squeeze(axes = squeeze_502_axes_0, x = var_15329_cast_fp16_0)[name = tensor<string, []>("squeeze_502_cast_fp16")];
            tensor<int32, [1]> squeeze_503_axes_0 = const()[name = tensor<string, []>("squeeze_503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_503_cast_fp16 = squeeze(axes = squeeze_503_axes_0, x = var_15329_cast_fp16_1)[name = tensor<string, []>("squeeze_503_cast_fp16")];
            tensor<fp16, []> const_2055_promoted_to_fp16 = const()[name = tensor<string, []>("const_2055_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15332_cast_fp16 = mul(x = squeeze_503_cast_fp16, y = const_2055_promoted_to_fp16)[name = tensor<string, []>("op_15332_cast_fp16")];
            tensor<int32, []> x_1719_axis_0 = const()[name = tensor<string, []>("x_1719_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1719_cast_fp16 = stack(axis = x_1719_axis_0, values = (var_15332_cast_fp16, squeeze_502_cast_fp16))[name = tensor<string, []>("x_1719_cast_fp16")];
            tensor<int32, [4]> var_15338 = const()[name = tensor<string, []>("op_15338"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15339_cast_fp16 = reshape(shape = var_15338, x = x_1719_cast_fp16)[name = tensor<string, []>("op_15339_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15340_cast_fp16 = mul(x = var_15339_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15340_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_251_cast_fp16 = add(x = var_15319_cast_fp16, y = var_15340_cast_fp16)[name = tensor<string, []>("k_patches_251_cast_fp16")];
            tensor<bool, []> var_15343_interleave_0 = const()[name = tensor<string, []>("op_15343_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15343_cast_fp16 = concat(axis = var_20, interleave = var_15343_interleave_0, values = (var_15282_cast_fp16, q_patches_251_cast_fp16))[name = tensor<string, []>("op_15343_cast_fp16")];
            tensor<bool, []> var_15346_interleave_0 = const()[name = tensor<string, []>("op_15346_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15346_cast_fp16 = concat(axis = var_20, interleave = var_15346_interleave_0, values = (var_15286_cast_fp16, k_patches_251_cast_fp16))[name = tensor<string, []>("op_15346_cast_fp16")];
            tensor<fp16, []> var_15348_to_fp16 = const()[name = tensor<string, []>("op_15348_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1259_cast_fp16 = mul(x = var_15343_cast_fp16, y = var_15348_to_fp16)[name = tensor<string, []>("q_1259_cast_fp16")];
            tensor<bool, []> attn_501_transpose_x_1 = const()[name = tensor<string, []>("attn_501_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_501_transpose_y_1 = const()[name = tensor<string, []>("attn_501_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_501_cast_fp16 = matmul(transpose_x = attn_501_transpose_x_1, transpose_y = attn_501_transpose_y_1, x = q_1259_cast_fp16, y = var_15346_cast_fp16)[name = tensor<string, []>("attn_501_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_503_cast_fp16 = softmax(axis = var_21, x = attn_501_cast_fp16)[name = tensor<string, []>("attn_503_cast_fp16")];
            tensor<bool, []> var_15353_transpose_x_0 = const()[name = tensor<string, []>("op_15353_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15353_transpose_y_0 = const()[name = tensor<string, []>("op_15353_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15353_cast_fp16 = matmul(transpose_x = var_15353_transpose_x_0, transpose_y = var_15353_transpose_y_0, x = attn_503_cast_fp16, y = v_755_cast_fp16)[name = tensor<string, []>("op_15353_cast_fp16")];
            tensor<int32, [1]> var_15354_axes_0 = const()[name = tensor<string, []>("op_15354_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15354_cast_fp16 = squeeze(axes = var_15354_axes_0, x = var_15353_cast_fp16)[name = tensor<string, []>("op_15354_cast_fp16")];
            tensor<string, []> var_15360_pad_type_0 = const()[name = tensor<string, []>("op_15360_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15360_strides_0 = const()[name = tensor<string, []>("op_15360_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15360_pad_0 = const()[name = tensor<string, []>("op_15360_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15360_dilations_0 = const()[name = tensor<string, []>("op_15360_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15360_groups_0 = const()[name = tensor<string, []>("op_15360_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145185600)))];
            tensor<fp16, [1, 64, 1, 197]> var_15360_cast_fp16 = conv(dilations = var_15360_dilations_0, groups = var_15360_groups_0, pad = var_15360_pad_0, pad_type = var_15360_pad_type_0, strides = var_15360_strides_0, weight = model_blocks_10_attn_q_projs_6_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15360_cast_fp16")];
            tensor<int32, [1]> var_15361_axes_0 = const()[name = tensor<string, []>("op_15361_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15361_cast_fp16 = squeeze(axes = var_15361_axes_0, x = var_15360_cast_fp16)[name = tensor<string, []>("op_15361_cast_fp16")];
            tensor<int32, [3]> q_1261_perm_0 = const()[name = tensor<string, []>("q_1261_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15368_pad_type_0 = const()[name = tensor<string, []>("op_15368_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15368_strides_0 = const()[name = tensor<string, []>("op_15368_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15368_pad_0 = const()[name = tensor<string, []>("op_15368_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15368_dilations_0 = const()[name = tensor<string, []>("op_15368_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15368_groups_0 = const()[name = tensor<string, []>("op_15368_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145283968)))];
            tensor<fp16, [1, 64, 1, 197]> var_15368_cast_fp16 = conv(dilations = var_15368_dilations_0, groups = var_15368_groups_0, pad = var_15368_pad_0, pad_type = var_15368_pad_type_0, strides = var_15368_strides_0, weight = model_blocks_10_attn_k_projs_6_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15368_cast_fp16")];
            tensor<int32, [1]> var_15369_axes_0 = const()[name = tensor<string, []>("op_15369_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15369_cast_fp16 = squeeze(axes = var_15369_axes_0, x = var_15368_cast_fp16)[name = tensor<string, []>("op_15369_cast_fp16")];
            tensor<int32, [3]> k_757_perm_0 = const()[name = tensor<string, []>("k_757_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15376_pad_type_0 = const()[name = tensor<string, []>("op_15376_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15376_strides_0 = const()[name = tensor<string, []>("op_15376_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15376_pad_0 = const()[name = tensor<string, []>("op_15376_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15376_dilations_0 = const()[name = tensor<string, []>("op_15376_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15376_groups_0 = const()[name = tensor<string, []>("op_15376_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145382336)))];
            tensor<fp16, [1, 64, 1, 197]> var_15376_cast_fp16 = conv(dilations = var_15376_dilations_0, groups = var_15376_groups_0, pad = var_15376_pad_0, pad_type = var_15376_pad_type_0, strides = var_15376_strides_0, weight = model_blocks_10_attn_v_projs_6_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15376_cast_fp16")];
            tensor<int32, [1]> var_15377_axes_0 = const()[name = tensor<string, []>("op_15377_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15377_cast_fp16 = squeeze(axes = var_15377_axes_0, x = var_15376_cast_fp16)[name = tensor<string, []>("op_15377_cast_fp16")];
            tensor<int32, [3]> v_757_perm_0 = const()[name = tensor<string, []>("v_757_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145480704)))];
            tensor<fp16, [1, 197, 64]> q_1261_cast_fp16 = transpose(perm = q_1261_perm_0, x = var_15361_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [1, 197, 64]> q_1263_cast_fp16 = add(x = q_1261_cast_fp16, y = model_blocks_10_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_1263_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145480896)))];
            tensor<fp16, [1, 197, 64]> v_757_cast_fp16 = transpose(perm = v_757_perm_0, x = var_15377_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 197, 64]> v_759_cast_fp16 = add(x = v_757_cast_fp16, y = model_blocks_10_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_759_cast_fp16")];
            tensor<int32, [1]> q_1265_axes_0 = const()[name = tensor<string, []>("q_1265_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1265_cast_fp16 = expand_dims(axes = q_1265_axes_0, x = q_1263_cast_fp16)[name = tensor<string, []>("q_1265_cast_fp16")];
            tensor<int32, [1]> k_759_axes_0 = const()[name = tensor<string, []>("k_759_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_757_cast_fp16 = transpose(perm = k_757_perm_0, x = var_15369_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [1, 1, 197, 64]> k_759_cast_fp16 = expand_dims(axes = k_759_axes_0, x = k_757_cast_fp16)[name = tensor<string, []>("k_759_cast_fp16")];
            tensor<int32, [1]> v_761_axes_0 = const()[name = tensor<string, []>("v_761_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_761_cast_fp16 = expand_dims(axes = v_761_axes_0, x = v_759_cast_fp16)[name = tensor<string, []>("v_761_cast_fp16")];
            tensor<int32, [4]> var_15386_begin_0 = const()[name = tensor<string, []>("op_15386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15386_end_0 = const()[name = tensor<string, []>("op_15386_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15386_end_mask_0 = const()[name = tensor<string, []>("op_15386_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15386_cast_fp16 = slice_by_index(begin = var_15386_begin_0, end = var_15386_end_0, end_mask = var_15386_end_mask_0, x = q_1265_cast_fp16)[name = tensor<string, []>("op_15386_cast_fp16")];
            tensor<int32, [4]> var_15390_begin_0 = const()[name = tensor<string, []>("op_15390_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15390_end_0 = const()[name = tensor<string, []>("op_15390_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15390_end_mask_0 = const()[name = tensor<string, []>("op_15390_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15390_cast_fp16 = slice_by_index(begin = var_15390_begin_0, end = var_15390_end_0, end_mask = var_15390_end_mask_0, x = k_759_cast_fp16)[name = tensor<string, []>("op_15390_cast_fp16")];
            tensor<int32, [4]> var_15394_begin_0 = const()[name = tensor<string, []>("op_15394_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15394_end_0 = const()[name = tensor<string, []>("op_15394_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15394_end_mask_0 = const()[name = tensor<string, []>("op_15394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15394_cast_fp16 = slice_by_index(begin = var_15394_begin_0, end = var_15394_end_0, end_mask = var_15394_end_mask_0, x = q_1265_cast_fp16)[name = tensor<string, []>("op_15394_cast_fp16")];
            tensor<int32, [4]> var_15398_begin_0 = const()[name = tensor<string, []>("op_15398_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15398_end_0 = const()[name = tensor<string, []>("op_15398_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15398_end_mask_0 = const()[name = tensor<string, []>("op_15398_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15398_cast_fp16 = slice_by_index(begin = var_15398_begin_0, end = var_15398_end_0, end_mask = var_15398_end_mask_0, x = k_759_cast_fp16)[name = tensor<string, []>("op_15398_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15400_cast_fp16 = mul(x = var_15394_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15400_cast_fp16")];
            tensor<int32, [5]> var_15408 = const()[name = tensor<string, []>("op_15408"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1723_cast_fp16 = reshape(shape = var_15408, x = var_15394_cast_fp16)[name = tensor<string, []>("x_1723_cast_fp16")];
            tensor<int32, [2]> var_15410_split_sizes_0 = const()[name = tensor<string, []>("op_15410_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15410_axis_0 = const()[name = tensor<string, []>("op_15410_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15410_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15410_cast_fp16_1 = split(axis = var_15410_axis_0, split_sizes = var_15410_split_sizes_0, x = x_1723_cast_fp16)[name = tensor<string, []>("op_15410_cast_fp16")];
            tensor<int32, [1]> squeeze_504_axes_0 = const()[name = tensor<string, []>("squeeze_504_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_504_cast_fp16 = squeeze(axes = squeeze_504_axes_0, x = var_15410_cast_fp16_0)[name = tensor<string, []>("squeeze_504_cast_fp16")];
            tensor<int32, [1]> squeeze_505_axes_0 = const()[name = tensor<string, []>("squeeze_505_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_505_cast_fp16 = squeeze(axes = squeeze_505_axes_0, x = var_15410_cast_fp16_1)[name = tensor<string, []>("squeeze_505_cast_fp16")];
            tensor<fp16, []> const_2063_promoted_to_fp16 = const()[name = tensor<string, []>("const_2063_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15413_cast_fp16 = mul(x = squeeze_505_cast_fp16, y = const_2063_promoted_to_fp16)[name = tensor<string, []>("op_15413_cast_fp16")];
            tensor<int32, []> x_1725_axis_0 = const()[name = tensor<string, []>("x_1725_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1725_cast_fp16 = stack(axis = x_1725_axis_0, values = (var_15413_cast_fp16, squeeze_504_cast_fp16))[name = tensor<string, []>("x_1725_cast_fp16")];
            tensor<int32, [4]> var_15419 = const()[name = tensor<string, []>("op_15419"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15420_cast_fp16 = reshape(shape = var_15419, x = x_1725_cast_fp16)[name = tensor<string, []>("op_15420_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15421_cast_fp16 = mul(x = var_15420_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15421_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_253_cast_fp16 = add(x = var_15400_cast_fp16, y = var_15421_cast_fp16)[name = tensor<string, []>("q_patches_253_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15423_cast_fp16 = mul(x = var_15398_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15423_cast_fp16")];
            tensor<int32, [5]> var_15431 = const()[name = tensor<string, []>("op_15431"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1729_cast_fp16 = reshape(shape = var_15431, x = var_15398_cast_fp16)[name = tensor<string, []>("x_1729_cast_fp16")];
            tensor<int32, [2]> var_15433_split_sizes_0 = const()[name = tensor<string, []>("op_15433_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15433_axis_0 = const()[name = tensor<string, []>("op_15433_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15433_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15433_cast_fp16_1 = split(axis = var_15433_axis_0, split_sizes = var_15433_split_sizes_0, x = x_1729_cast_fp16)[name = tensor<string, []>("op_15433_cast_fp16")];
            tensor<int32, [1]> squeeze_506_axes_0 = const()[name = tensor<string, []>("squeeze_506_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_506_cast_fp16 = squeeze(axes = squeeze_506_axes_0, x = var_15433_cast_fp16_0)[name = tensor<string, []>("squeeze_506_cast_fp16")];
            tensor<int32, [1]> squeeze_507_axes_0 = const()[name = tensor<string, []>("squeeze_507_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_507_cast_fp16 = squeeze(axes = squeeze_507_axes_0, x = var_15433_cast_fp16_1)[name = tensor<string, []>("squeeze_507_cast_fp16")];
            tensor<fp16, []> const_2071_promoted_to_fp16 = const()[name = tensor<string, []>("const_2071_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15436_cast_fp16 = mul(x = squeeze_507_cast_fp16, y = const_2071_promoted_to_fp16)[name = tensor<string, []>("op_15436_cast_fp16")];
            tensor<int32, []> x_1731_axis_0 = const()[name = tensor<string, []>("x_1731_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1731_cast_fp16 = stack(axis = x_1731_axis_0, values = (var_15436_cast_fp16, squeeze_506_cast_fp16))[name = tensor<string, []>("x_1731_cast_fp16")];
            tensor<int32, [4]> var_15442 = const()[name = tensor<string, []>("op_15442"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15443_cast_fp16 = reshape(shape = var_15442, x = x_1731_cast_fp16)[name = tensor<string, []>("op_15443_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15444_cast_fp16 = mul(x = var_15443_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15444_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_253_cast_fp16 = add(x = var_15423_cast_fp16, y = var_15444_cast_fp16)[name = tensor<string, []>("k_patches_253_cast_fp16")];
            tensor<bool, []> var_15447_interleave_0 = const()[name = tensor<string, []>("op_15447_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15447_cast_fp16 = concat(axis = var_20, interleave = var_15447_interleave_0, values = (var_15386_cast_fp16, q_patches_253_cast_fp16))[name = tensor<string, []>("op_15447_cast_fp16")];
            tensor<bool, []> var_15450_interleave_0 = const()[name = tensor<string, []>("op_15450_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15450_cast_fp16 = concat(axis = var_20, interleave = var_15450_interleave_0, values = (var_15390_cast_fp16, k_patches_253_cast_fp16))[name = tensor<string, []>("op_15450_cast_fp16")];
            tensor<fp16, []> var_15452_to_fp16 = const()[name = tensor<string, []>("op_15452_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1269_cast_fp16 = mul(x = var_15447_cast_fp16, y = var_15452_to_fp16)[name = tensor<string, []>("q_1269_cast_fp16")];
            tensor<bool, []> attn_505_transpose_x_1 = const()[name = tensor<string, []>("attn_505_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_505_transpose_y_1 = const()[name = tensor<string, []>("attn_505_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_505_cast_fp16 = matmul(transpose_x = attn_505_transpose_x_1, transpose_y = attn_505_transpose_y_1, x = q_1269_cast_fp16, y = var_15450_cast_fp16)[name = tensor<string, []>("attn_505_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_507_cast_fp16 = softmax(axis = var_21, x = attn_505_cast_fp16)[name = tensor<string, []>("attn_507_cast_fp16")];
            tensor<bool, []> var_15457_transpose_x_0 = const()[name = tensor<string, []>("op_15457_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15457_transpose_y_0 = const()[name = tensor<string, []>("op_15457_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15457_cast_fp16 = matmul(transpose_x = var_15457_transpose_x_0, transpose_y = var_15457_transpose_y_0, x = attn_507_cast_fp16, y = v_761_cast_fp16)[name = tensor<string, []>("op_15457_cast_fp16")];
            tensor<int32, [1]> var_15458_axes_0 = const()[name = tensor<string, []>("op_15458_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15458_cast_fp16 = squeeze(axes = var_15458_axes_0, x = var_15457_cast_fp16)[name = tensor<string, []>("op_15458_cast_fp16")];
            tensor<string, []> var_15464_pad_type_0 = const()[name = tensor<string, []>("op_15464_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15464_strides_0 = const()[name = tensor<string, []>("op_15464_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15464_pad_0 = const()[name = tensor<string, []>("op_15464_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15464_dilations_0 = const()[name = tensor<string, []>("op_15464_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15464_groups_0 = const()[name = tensor<string, []>("op_15464_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145481088)))];
            tensor<fp16, [1, 64, 1, 197]> var_15464_cast_fp16 = conv(dilations = var_15464_dilations_0, groups = var_15464_groups_0, pad = var_15464_pad_0, pad_type = var_15464_pad_type_0, strides = var_15464_strides_0, weight = model_blocks_10_attn_q_projs_7_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15464_cast_fp16")];
            tensor<int32, [1]> var_15465_axes_0 = const()[name = tensor<string, []>("op_15465_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15465_cast_fp16 = squeeze(axes = var_15465_axes_0, x = var_15464_cast_fp16)[name = tensor<string, []>("op_15465_cast_fp16")];
            tensor<int32, [3]> q_1271_perm_0 = const()[name = tensor<string, []>("q_1271_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15472_pad_type_0 = const()[name = tensor<string, []>("op_15472_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15472_strides_0 = const()[name = tensor<string, []>("op_15472_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15472_pad_0 = const()[name = tensor<string, []>("op_15472_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15472_dilations_0 = const()[name = tensor<string, []>("op_15472_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15472_groups_0 = const()[name = tensor<string, []>("op_15472_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145579456)))];
            tensor<fp16, [1, 64, 1, 197]> var_15472_cast_fp16 = conv(dilations = var_15472_dilations_0, groups = var_15472_groups_0, pad = var_15472_pad_0, pad_type = var_15472_pad_type_0, strides = var_15472_strides_0, weight = model_blocks_10_attn_k_projs_7_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15472_cast_fp16")];
            tensor<int32, [1]> var_15473_axes_0 = const()[name = tensor<string, []>("op_15473_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15473_cast_fp16 = squeeze(axes = var_15473_axes_0, x = var_15472_cast_fp16)[name = tensor<string, []>("op_15473_cast_fp16")];
            tensor<int32, [3]> k_763_perm_0 = const()[name = tensor<string, []>("k_763_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15480_pad_type_0 = const()[name = tensor<string, []>("op_15480_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15480_strides_0 = const()[name = tensor<string, []>("op_15480_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15480_pad_0 = const()[name = tensor<string, []>("op_15480_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15480_dilations_0 = const()[name = tensor<string, []>("op_15480_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15480_groups_0 = const()[name = tensor<string, []>("op_15480_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145677824)))];
            tensor<fp16, [1, 64, 1, 197]> var_15480_cast_fp16 = conv(dilations = var_15480_dilations_0, groups = var_15480_groups_0, pad = var_15480_pad_0, pad_type = var_15480_pad_type_0, strides = var_15480_strides_0, weight = model_blocks_10_attn_v_projs_7_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15480_cast_fp16")];
            tensor<int32, [1]> var_15481_axes_0 = const()[name = tensor<string, []>("op_15481_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15481_cast_fp16 = squeeze(axes = var_15481_axes_0, x = var_15480_cast_fp16)[name = tensor<string, []>("op_15481_cast_fp16")];
            tensor<int32, [3]> v_763_perm_0 = const()[name = tensor<string, []>("v_763_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145776192)))];
            tensor<fp16, [1, 197, 64]> q_1271_cast_fp16 = transpose(perm = q_1271_perm_0, x = var_15465_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [1, 197, 64]> q_1273_cast_fp16 = add(x = q_1271_cast_fp16, y = model_blocks_10_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_1273_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145776384)))];
            tensor<fp16, [1, 197, 64]> v_763_cast_fp16 = transpose(perm = v_763_perm_0, x = var_15481_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [1, 197, 64]> v_765_cast_fp16 = add(x = v_763_cast_fp16, y = model_blocks_10_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_765_cast_fp16")];
            tensor<int32, [1]> q_1275_axes_0 = const()[name = tensor<string, []>("q_1275_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1275_cast_fp16 = expand_dims(axes = q_1275_axes_0, x = q_1273_cast_fp16)[name = tensor<string, []>("q_1275_cast_fp16")];
            tensor<int32, [1]> k_765_axes_0 = const()[name = tensor<string, []>("k_765_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_763_cast_fp16 = transpose(perm = k_763_perm_0, x = var_15473_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [1, 1, 197, 64]> k_765_cast_fp16 = expand_dims(axes = k_765_axes_0, x = k_763_cast_fp16)[name = tensor<string, []>("k_765_cast_fp16")];
            tensor<int32, [1]> v_767_axes_0 = const()[name = tensor<string, []>("v_767_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_767_cast_fp16 = expand_dims(axes = v_767_axes_0, x = v_765_cast_fp16)[name = tensor<string, []>("v_767_cast_fp16")];
            tensor<int32, [4]> var_15490_begin_0 = const()[name = tensor<string, []>("op_15490_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15490_end_0 = const()[name = tensor<string, []>("op_15490_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15490_end_mask_0 = const()[name = tensor<string, []>("op_15490_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15490_cast_fp16 = slice_by_index(begin = var_15490_begin_0, end = var_15490_end_0, end_mask = var_15490_end_mask_0, x = q_1275_cast_fp16)[name = tensor<string, []>("op_15490_cast_fp16")];
            tensor<int32, [4]> var_15494_begin_0 = const()[name = tensor<string, []>("op_15494_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15494_end_0 = const()[name = tensor<string, []>("op_15494_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15494_end_mask_0 = const()[name = tensor<string, []>("op_15494_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15494_cast_fp16 = slice_by_index(begin = var_15494_begin_0, end = var_15494_end_0, end_mask = var_15494_end_mask_0, x = k_765_cast_fp16)[name = tensor<string, []>("op_15494_cast_fp16")];
            tensor<int32, [4]> var_15498_begin_0 = const()[name = tensor<string, []>("op_15498_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15498_end_0 = const()[name = tensor<string, []>("op_15498_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15498_end_mask_0 = const()[name = tensor<string, []>("op_15498_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15498_cast_fp16 = slice_by_index(begin = var_15498_begin_0, end = var_15498_end_0, end_mask = var_15498_end_mask_0, x = q_1275_cast_fp16)[name = tensor<string, []>("op_15498_cast_fp16")];
            tensor<int32, [4]> var_15502_begin_0 = const()[name = tensor<string, []>("op_15502_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15502_end_0 = const()[name = tensor<string, []>("op_15502_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15502_end_mask_0 = const()[name = tensor<string, []>("op_15502_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15502_cast_fp16 = slice_by_index(begin = var_15502_begin_0, end = var_15502_end_0, end_mask = var_15502_end_mask_0, x = k_765_cast_fp16)[name = tensor<string, []>("op_15502_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15504_cast_fp16 = mul(x = var_15498_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15504_cast_fp16")];
            tensor<int32, [5]> var_15512 = const()[name = tensor<string, []>("op_15512"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1735_cast_fp16 = reshape(shape = var_15512, x = var_15498_cast_fp16)[name = tensor<string, []>("x_1735_cast_fp16")];
            tensor<int32, [2]> var_15514_split_sizes_0 = const()[name = tensor<string, []>("op_15514_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15514_axis_0 = const()[name = tensor<string, []>("op_15514_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15514_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15514_cast_fp16_1 = split(axis = var_15514_axis_0, split_sizes = var_15514_split_sizes_0, x = x_1735_cast_fp16)[name = tensor<string, []>("op_15514_cast_fp16")];
            tensor<int32, [1]> squeeze_508_axes_0 = const()[name = tensor<string, []>("squeeze_508_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_508_cast_fp16 = squeeze(axes = squeeze_508_axes_0, x = var_15514_cast_fp16_0)[name = tensor<string, []>("squeeze_508_cast_fp16")];
            tensor<int32, [1]> squeeze_509_axes_0 = const()[name = tensor<string, []>("squeeze_509_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_509_cast_fp16 = squeeze(axes = squeeze_509_axes_0, x = var_15514_cast_fp16_1)[name = tensor<string, []>("squeeze_509_cast_fp16")];
            tensor<fp16, []> const_2079_promoted_to_fp16 = const()[name = tensor<string, []>("const_2079_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15517_cast_fp16 = mul(x = squeeze_509_cast_fp16, y = const_2079_promoted_to_fp16)[name = tensor<string, []>("op_15517_cast_fp16")];
            tensor<int32, []> x_1737_axis_0 = const()[name = tensor<string, []>("x_1737_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1737_cast_fp16 = stack(axis = x_1737_axis_0, values = (var_15517_cast_fp16, squeeze_508_cast_fp16))[name = tensor<string, []>("x_1737_cast_fp16")];
            tensor<int32, [4]> var_15523 = const()[name = tensor<string, []>("op_15523"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15524_cast_fp16 = reshape(shape = var_15523, x = x_1737_cast_fp16)[name = tensor<string, []>("op_15524_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15525_cast_fp16 = mul(x = var_15524_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15525_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_255_cast_fp16 = add(x = var_15504_cast_fp16, y = var_15525_cast_fp16)[name = tensor<string, []>("q_patches_255_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15527_cast_fp16 = mul(x = var_15502_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15527_cast_fp16")];
            tensor<int32, [5]> var_15535 = const()[name = tensor<string, []>("op_15535"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1741_cast_fp16 = reshape(shape = var_15535, x = var_15502_cast_fp16)[name = tensor<string, []>("x_1741_cast_fp16")];
            tensor<int32, [2]> var_15537_split_sizes_0 = const()[name = tensor<string, []>("op_15537_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15537_axis_0 = const()[name = tensor<string, []>("op_15537_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15537_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15537_cast_fp16_1 = split(axis = var_15537_axis_0, split_sizes = var_15537_split_sizes_0, x = x_1741_cast_fp16)[name = tensor<string, []>("op_15537_cast_fp16")];
            tensor<int32, [1]> squeeze_510_axes_0 = const()[name = tensor<string, []>("squeeze_510_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_510_cast_fp16 = squeeze(axes = squeeze_510_axes_0, x = var_15537_cast_fp16_0)[name = tensor<string, []>("squeeze_510_cast_fp16")];
            tensor<int32, [1]> squeeze_511_axes_0 = const()[name = tensor<string, []>("squeeze_511_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_511_cast_fp16 = squeeze(axes = squeeze_511_axes_0, x = var_15537_cast_fp16_1)[name = tensor<string, []>("squeeze_511_cast_fp16")];
            tensor<fp16, []> const_2087_promoted_to_fp16 = const()[name = tensor<string, []>("const_2087_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15540_cast_fp16 = mul(x = squeeze_511_cast_fp16, y = const_2087_promoted_to_fp16)[name = tensor<string, []>("op_15540_cast_fp16")];
            tensor<int32, []> x_1743_axis_0 = const()[name = tensor<string, []>("x_1743_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1743_cast_fp16 = stack(axis = x_1743_axis_0, values = (var_15540_cast_fp16, squeeze_510_cast_fp16))[name = tensor<string, []>("x_1743_cast_fp16")];
            tensor<int32, [4]> var_15546 = const()[name = tensor<string, []>("op_15546"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15547_cast_fp16 = reshape(shape = var_15546, x = x_1743_cast_fp16)[name = tensor<string, []>("op_15547_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15548_cast_fp16 = mul(x = var_15547_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15548_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_255_cast_fp16 = add(x = var_15527_cast_fp16, y = var_15548_cast_fp16)[name = tensor<string, []>("k_patches_255_cast_fp16")];
            tensor<bool, []> var_15551_interleave_0 = const()[name = tensor<string, []>("op_15551_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15551_cast_fp16 = concat(axis = var_20, interleave = var_15551_interleave_0, values = (var_15490_cast_fp16, q_patches_255_cast_fp16))[name = tensor<string, []>("op_15551_cast_fp16")];
            tensor<bool, []> var_15554_interleave_0 = const()[name = tensor<string, []>("op_15554_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15554_cast_fp16 = concat(axis = var_20, interleave = var_15554_interleave_0, values = (var_15494_cast_fp16, k_patches_255_cast_fp16))[name = tensor<string, []>("op_15554_cast_fp16")];
            tensor<fp16, []> var_15556_to_fp16 = const()[name = tensor<string, []>("op_15556_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1279_cast_fp16 = mul(x = var_15551_cast_fp16, y = var_15556_to_fp16)[name = tensor<string, []>("q_1279_cast_fp16")];
            tensor<bool, []> attn_509_transpose_x_1 = const()[name = tensor<string, []>("attn_509_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_509_transpose_y_1 = const()[name = tensor<string, []>("attn_509_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_509_cast_fp16 = matmul(transpose_x = attn_509_transpose_x_1, transpose_y = attn_509_transpose_y_1, x = q_1279_cast_fp16, y = var_15554_cast_fp16)[name = tensor<string, []>("attn_509_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_511_cast_fp16 = softmax(axis = var_21, x = attn_509_cast_fp16)[name = tensor<string, []>("attn_511_cast_fp16")];
            tensor<bool, []> var_15561_transpose_x_0 = const()[name = tensor<string, []>("op_15561_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15561_transpose_y_0 = const()[name = tensor<string, []>("op_15561_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15561_cast_fp16 = matmul(transpose_x = var_15561_transpose_x_0, transpose_y = var_15561_transpose_y_0, x = attn_511_cast_fp16, y = v_767_cast_fp16)[name = tensor<string, []>("op_15561_cast_fp16")];
            tensor<int32, [1]> var_15562_axes_0 = const()[name = tensor<string, []>("op_15562_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15562_cast_fp16 = squeeze(axes = var_15562_axes_0, x = var_15561_cast_fp16)[name = tensor<string, []>("op_15562_cast_fp16")];
            tensor<string, []> var_15568_pad_type_0 = const()[name = tensor<string, []>("op_15568_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15568_strides_0 = const()[name = tensor<string, []>("op_15568_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15568_pad_0 = const()[name = tensor<string, []>("op_15568_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15568_dilations_0 = const()[name = tensor<string, []>("op_15568_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15568_groups_0 = const()[name = tensor<string, []>("op_15568_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145776576)))];
            tensor<fp16, [1, 64, 1, 197]> var_15568_cast_fp16 = conv(dilations = var_15568_dilations_0, groups = var_15568_groups_0, pad = var_15568_pad_0, pad_type = var_15568_pad_type_0, strides = var_15568_strides_0, weight = model_blocks_10_attn_q_projs_8_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15568_cast_fp16")];
            tensor<int32, [1]> var_15569_axes_0 = const()[name = tensor<string, []>("op_15569_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15569_cast_fp16 = squeeze(axes = var_15569_axes_0, x = var_15568_cast_fp16)[name = tensor<string, []>("op_15569_cast_fp16")];
            tensor<int32, [3]> q_1281_perm_0 = const()[name = tensor<string, []>("q_1281_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15576_pad_type_0 = const()[name = tensor<string, []>("op_15576_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15576_strides_0 = const()[name = tensor<string, []>("op_15576_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15576_pad_0 = const()[name = tensor<string, []>("op_15576_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15576_dilations_0 = const()[name = tensor<string, []>("op_15576_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15576_groups_0 = const()[name = tensor<string, []>("op_15576_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145874944)))];
            tensor<fp16, [1, 64, 1, 197]> var_15576_cast_fp16 = conv(dilations = var_15576_dilations_0, groups = var_15576_groups_0, pad = var_15576_pad_0, pad_type = var_15576_pad_type_0, strides = var_15576_strides_0, weight = model_blocks_10_attn_k_projs_8_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15576_cast_fp16")];
            tensor<int32, [1]> var_15577_axes_0 = const()[name = tensor<string, []>("op_15577_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15577_cast_fp16 = squeeze(axes = var_15577_axes_0, x = var_15576_cast_fp16)[name = tensor<string, []>("op_15577_cast_fp16")];
            tensor<int32, [3]> k_769_perm_0 = const()[name = tensor<string, []>("k_769_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15584_pad_type_0 = const()[name = tensor<string, []>("op_15584_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15584_strides_0 = const()[name = tensor<string, []>("op_15584_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15584_pad_0 = const()[name = tensor<string, []>("op_15584_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15584_dilations_0 = const()[name = tensor<string, []>("op_15584_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15584_groups_0 = const()[name = tensor<string, []>("op_15584_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145973312)))];
            tensor<fp16, [1, 64, 1, 197]> var_15584_cast_fp16 = conv(dilations = var_15584_dilations_0, groups = var_15584_groups_0, pad = var_15584_pad_0, pad_type = var_15584_pad_type_0, strides = var_15584_strides_0, weight = model_blocks_10_attn_v_projs_8_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15584_cast_fp16")];
            tensor<int32, [1]> var_15585_axes_0 = const()[name = tensor<string, []>("op_15585_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15585_cast_fp16 = squeeze(axes = var_15585_axes_0, x = var_15584_cast_fp16)[name = tensor<string, []>("op_15585_cast_fp16")];
            tensor<int32, [3]> v_769_perm_0 = const()[name = tensor<string, []>("v_769_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146071680)))];
            tensor<fp16, [1, 197, 64]> q_1281_cast_fp16 = transpose(perm = q_1281_perm_0, x = var_15569_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [1, 197, 64]> q_1283_cast_fp16 = add(x = q_1281_cast_fp16, y = model_blocks_10_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_1283_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146071872)))];
            tensor<fp16, [1, 197, 64]> v_769_cast_fp16 = transpose(perm = v_769_perm_0, x = var_15585_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [1, 197, 64]> v_771_cast_fp16 = add(x = v_769_cast_fp16, y = model_blocks_10_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_771_cast_fp16")];
            tensor<int32, [1]> q_1285_axes_0 = const()[name = tensor<string, []>("q_1285_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1285_cast_fp16 = expand_dims(axes = q_1285_axes_0, x = q_1283_cast_fp16)[name = tensor<string, []>("q_1285_cast_fp16")];
            tensor<int32, [1]> k_771_axes_0 = const()[name = tensor<string, []>("k_771_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_769_cast_fp16 = transpose(perm = k_769_perm_0, x = var_15577_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 1, 197, 64]> k_771_cast_fp16 = expand_dims(axes = k_771_axes_0, x = k_769_cast_fp16)[name = tensor<string, []>("k_771_cast_fp16")];
            tensor<int32, [1]> v_773_axes_0 = const()[name = tensor<string, []>("v_773_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_773_cast_fp16 = expand_dims(axes = v_773_axes_0, x = v_771_cast_fp16)[name = tensor<string, []>("v_773_cast_fp16")];
            tensor<int32, [4]> var_15594_begin_0 = const()[name = tensor<string, []>("op_15594_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15594_end_0 = const()[name = tensor<string, []>("op_15594_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15594_end_mask_0 = const()[name = tensor<string, []>("op_15594_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15594_cast_fp16 = slice_by_index(begin = var_15594_begin_0, end = var_15594_end_0, end_mask = var_15594_end_mask_0, x = q_1285_cast_fp16)[name = tensor<string, []>("op_15594_cast_fp16")];
            tensor<int32, [4]> var_15598_begin_0 = const()[name = tensor<string, []>("op_15598_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15598_end_0 = const()[name = tensor<string, []>("op_15598_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15598_end_mask_0 = const()[name = tensor<string, []>("op_15598_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15598_cast_fp16 = slice_by_index(begin = var_15598_begin_0, end = var_15598_end_0, end_mask = var_15598_end_mask_0, x = k_771_cast_fp16)[name = tensor<string, []>("op_15598_cast_fp16")];
            tensor<int32, [4]> var_15602_begin_0 = const()[name = tensor<string, []>("op_15602_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15602_end_0 = const()[name = tensor<string, []>("op_15602_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15602_end_mask_0 = const()[name = tensor<string, []>("op_15602_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15602_cast_fp16 = slice_by_index(begin = var_15602_begin_0, end = var_15602_end_0, end_mask = var_15602_end_mask_0, x = q_1285_cast_fp16)[name = tensor<string, []>("op_15602_cast_fp16")];
            tensor<int32, [4]> var_15606_begin_0 = const()[name = tensor<string, []>("op_15606_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15606_end_0 = const()[name = tensor<string, []>("op_15606_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15606_end_mask_0 = const()[name = tensor<string, []>("op_15606_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15606_cast_fp16 = slice_by_index(begin = var_15606_begin_0, end = var_15606_end_0, end_mask = var_15606_end_mask_0, x = k_771_cast_fp16)[name = tensor<string, []>("op_15606_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15608_cast_fp16 = mul(x = var_15602_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15608_cast_fp16")];
            tensor<int32, [5]> var_15616 = const()[name = tensor<string, []>("op_15616"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1747_cast_fp16 = reshape(shape = var_15616, x = var_15602_cast_fp16)[name = tensor<string, []>("x_1747_cast_fp16")];
            tensor<int32, [2]> var_15618_split_sizes_0 = const()[name = tensor<string, []>("op_15618_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15618_axis_0 = const()[name = tensor<string, []>("op_15618_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15618_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15618_cast_fp16_1 = split(axis = var_15618_axis_0, split_sizes = var_15618_split_sizes_0, x = x_1747_cast_fp16)[name = tensor<string, []>("op_15618_cast_fp16")];
            tensor<int32, [1]> squeeze_512_axes_0 = const()[name = tensor<string, []>("squeeze_512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_512_cast_fp16 = squeeze(axes = squeeze_512_axes_0, x = var_15618_cast_fp16_0)[name = tensor<string, []>("squeeze_512_cast_fp16")];
            tensor<int32, [1]> squeeze_513_axes_0 = const()[name = tensor<string, []>("squeeze_513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_513_cast_fp16 = squeeze(axes = squeeze_513_axes_0, x = var_15618_cast_fp16_1)[name = tensor<string, []>("squeeze_513_cast_fp16")];
            tensor<fp16, []> const_2095_promoted_to_fp16 = const()[name = tensor<string, []>("const_2095_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15621_cast_fp16 = mul(x = squeeze_513_cast_fp16, y = const_2095_promoted_to_fp16)[name = tensor<string, []>("op_15621_cast_fp16")];
            tensor<int32, []> x_1749_axis_0 = const()[name = tensor<string, []>("x_1749_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1749_cast_fp16 = stack(axis = x_1749_axis_0, values = (var_15621_cast_fp16, squeeze_512_cast_fp16))[name = tensor<string, []>("x_1749_cast_fp16")];
            tensor<int32, [4]> var_15627 = const()[name = tensor<string, []>("op_15627"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15628_cast_fp16 = reshape(shape = var_15627, x = x_1749_cast_fp16)[name = tensor<string, []>("op_15628_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15629_cast_fp16 = mul(x = var_15628_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15629_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_257_cast_fp16 = add(x = var_15608_cast_fp16, y = var_15629_cast_fp16)[name = tensor<string, []>("q_patches_257_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15631_cast_fp16 = mul(x = var_15606_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15631_cast_fp16")];
            tensor<int32, [5]> var_15639 = const()[name = tensor<string, []>("op_15639"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1753_cast_fp16 = reshape(shape = var_15639, x = var_15606_cast_fp16)[name = tensor<string, []>("x_1753_cast_fp16")];
            tensor<int32, [2]> var_15641_split_sizes_0 = const()[name = tensor<string, []>("op_15641_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15641_axis_0 = const()[name = tensor<string, []>("op_15641_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15641_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15641_cast_fp16_1 = split(axis = var_15641_axis_0, split_sizes = var_15641_split_sizes_0, x = x_1753_cast_fp16)[name = tensor<string, []>("op_15641_cast_fp16")];
            tensor<int32, [1]> squeeze_514_axes_0 = const()[name = tensor<string, []>("squeeze_514_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_514_cast_fp16 = squeeze(axes = squeeze_514_axes_0, x = var_15641_cast_fp16_0)[name = tensor<string, []>("squeeze_514_cast_fp16")];
            tensor<int32, [1]> squeeze_515_axes_0 = const()[name = tensor<string, []>("squeeze_515_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_515_cast_fp16 = squeeze(axes = squeeze_515_axes_0, x = var_15641_cast_fp16_1)[name = tensor<string, []>("squeeze_515_cast_fp16")];
            tensor<fp16, []> const_2103_promoted_to_fp16 = const()[name = tensor<string, []>("const_2103_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15644_cast_fp16 = mul(x = squeeze_515_cast_fp16, y = const_2103_promoted_to_fp16)[name = tensor<string, []>("op_15644_cast_fp16")];
            tensor<int32, []> x_1755_axis_0 = const()[name = tensor<string, []>("x_1755_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1755_cast_fp16 = stack(axis = x_1755_axis_0, values = (var_15644_cast_fp16, squeeze_514_cast_fp16))[name = tensor<string, []>("x_1755_cast_fp16")];
            tensor<int32, [4]> var_15650 = const()[name = tensor<string, []>("op_15650"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15651_cast_fp16 = reshape(shape = var_15650, x = x_1755_cast_fp16)[name = tensor<string, []>("op_15651_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15652_cast_fp16 = mul(x = var_15651_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15652_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_257_cast_fp16 = add(x = var_15631_cast_fp16, y = var_15652_cast_fp16)[name = tensor<string, []>("k_patches_257_cast_fp16")];
            tensor<bool, []> var_15655_interleave_0 = const()[name = tensor<string, []>("op_15655_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15655_cast_fp16 = concat(axis = var_20, interleave = var_15655_interleave_0, values = (var_15594_cast_fp16, q_patches_257_cast_fp16))[name = tensor<string, []>("op_15655_cast_fp16")];
            tensor<bool, []> var_15658_interleave_0 = const()[name = tensor<string, []>("op_15658_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15658_cast_fp16 = concat(axis = var_20, interleave = var_15658_interleave_0, values = (var_15598_cast_fp16, k_patches_257_cast_fp16))[name = tensor<string, []>("op_15658_cast_fp16")];
            tensor<fp16, []> var_15660_to_fp16 = const()[name = tensor<string, []>("op_15660_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1289_cast_fp16 = mul(x = var_15655_cast_fp16, y = var_15660_to_fp16)[name = tensor<string, []>("q_1289_cast_fp16")];
            tensor<bool, []> attn_513_transpose_x_1 = const()[name = tensor<string, []>("attn_513_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_513_transpose_y_1 = const()[name = tensor<string, []>("attn_513_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_513_cast_fp16 = matmul(transpose_x = attn_513_transpose_x_1, transpose_y = attn_513_transpose_y_1, x = q_1289_cast_fp16, y = var_15658_cast_fp16)[name = tensor<string, []>("attn_513_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_515_cast_fp16 = softmax(axis = var_21, x = attn_513_cast_fp16)[name = tensor<string, []>("attn_515_cast_fp16")];
            tensor<bool, []> var_15665_transpose_x_0 = const()[name = tensor<string, []>("op_15665_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15665_transpose_y_0 = const()[name = tensor<string, []>("op_15665_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15665_cast_fp16 = matmul(transpose_x = var_15665_transpose_x_0, transpose_y = var_15665_transpose_y_0, x = attn_515_cast_fp16, y = v_773_cast_fp16)[name = tensor<string, []>("op_15665_cast_fp16")];
            tensor<int32, [1]> var_15666_axes_0 = const()[name = tensor<string, []>("op_15666_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15666_cast_fp16 = squeeze(axes = var_15666_axes_0, x = var_15665_cast_fp16)[name = tensor<string, []>("op_15666_cast_fp16")];
            tensor<string, []> var_15672_pad_type_0 = const()[name = tensor<string, []>("op_15672_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15672_strides_0 = const()[name = tensor<string, []>("op_15672_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15672_pad_0 = const()[name = tensor<string, []>("op_15672_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15672_dilations_0 = const()[name = tensor<string, []>("op_15672_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15672_groups_0 = const()[name = tensor<string, []>("op_15672_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146072064)))];
            tensor<fp16, [1, 64, 1, 197]> var_15672_cast_fp16 = conv(dilations = var_15672_dilations_0, groups = var_15672_groups_0, pad = var_15672_pad_0, pad_type = var_15672_pad_type_0, strides = var_15672_strides_0, weight = model_blocks_10_attn_q_projs_9_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15672_cast_fp16")];
            tensor<int32, [1]> var_15673_axes_0 = const()[name = tensor<string, []>("op_15673_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15673_cast_fp16 = squeeze(axes = var_15673_axes_0, x = var_15672_cast_fp16)[name = tensor<string, []>("op_15673_cast_fp16")];
            tensor<int32, [3]> q_1291_perm_0 = const()[name = tensor<string, []>("q_1291_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15680_pad_type_0 = const()[name = tensor<string, []>("op_15680_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15680_strides_0 = const()[name = tensor<string, []>("op_15680_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15680_pad_0 = const()[name = tensor<string, []>("op_15680_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15680_dilations_0 = const()[name = tensor<string, []>("op_15680_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15680_groups_0 = const()[name = tensor<string, []>("op_15680_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146170432)))];
            tensor<fp16, [1, 64, 1, 197]> var_15680_cast_fp16 = conv(dilations = var_15680_dilations_0, groups = var_15680_groups_0, pad = var_15680_pad_0, pad_type = var_15680_pad_type_0, strides = var_15680_strides_0, weight = model_blocks_10_attn_k_projs_9_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15680_cast_fp16")];
            tensor<int32, [1]> var_15681_axes_0 = const()[name = tensor<string, []>("op_15681_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15681_cast_fp16 = squeeze(axes = var_15681_axes_0, x = var_15680_cast_fp16)[name = tensor<string, []>("op_15681_cast_fp16")];
            tensor<int32, [3]> k_775_perm_0 = const()[name = tensor<string, []>("k_775_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15688_pad_type_0 = const()[name = tensor<string, []>("op_15688_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15688_strides_0 = const()[name = tensor<string, []>("op_15688_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15688_pad_0 = const()[name = tensor<string, []>("op_15688_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15688_dilations_0 = const()[name = tensor<string, []>("op_15688_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15688_groups_0 = const()[name = tensor<string, []>("op_15688_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146268800)))];
            tensor<fp16, [1, 64, 1, 197]> var_15688_cast_fp16 = conv(dilations = var_15688_dilations_0, groups = var_15688_groups_0, pad = var_15688_pad_0, pad_type = var_15688_pad_type_0, strides = var_15688_strides_0, weight = model_blocks_10_attn_v_projs_9_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15688_cast_fp16")];
            tensor<int32, [1]> var_15689_axes_0 = const()[name = tensor<string, []>("op_15689_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15689_cast_fp16 = squeeze(axes = var_15689_axes_0, x = var_15688_cast_fp16)[name = tensor<string, []>("op_15689_cast_fp16")];
            tensor<int32, [3]> v_775_perm_0 = const()[name = tensor<string, []>("v_775_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146367168)))];
            tensor<fp16, [1, 197, 64]> q_1291_cast_fp16 = transpose(perm = q_1291_perm_0, x = var_15673_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [1, 197, 64]> q_1293_cast_fp16 = add(x = q_1291_cast_fp16, y = model_blocks_10_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_1293_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146367360)))];
            tensor<fp16, [1, 197, 64]> v_775_cast_fp16 = transpose(perm = v_775_perm_0, x = var_15689_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 197, 64]> v_777_cast_fp16 = add(x = v_775_cast_fp16, y = model_blocks_10_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_777_cast_fp16")];
            tensor<int32, [1]> q_1295_axes_0 = const()[name = tensor<string, []>("q_1295_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1295_cast_fp16 = expand_dims(axes = q_1295_axes_0, x = q_1293_cast_fp16)[name = tensor<string, []>("q_1295_cast_fp16")];
            tensor<int32, [1]> k_777_axes_0 = const()[name = tensor<string, []>("k_777_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_775_cast_fp16 = transpose(perm = k_775_perm_0, x = var_15681_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [1, 1, 197, 64]> k_777_cast_fp16 = expand_dims(axes = k_777_axes_0, x = k_775_cast_fp16)[name = tensor<string, []>("k_777_cast_fp16")];
            tensor<int32, [1]> v_779_axes_0 = const()[name = tensor<string, []>("v_779_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_779_cast_fp16 = expand_dims(axes = v_779_axes_0, x = v_777_cast_fp16)[name = tensor<string, []>("v_779_cast_fp16")];
            tensor<int32, [4]> var_15698_begin_0 = const()[name = tensor<string, []>("op_15698_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15698_end_0 = const()[name = tensor<string, []>("op_15698_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15698_end_mask_0 = const()[name = tensor<string, []>("op_15698_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15698_cast_fp16 = slice_by_index(begin = var_15698_begin_0, end = var_15698_end_0, end_mask = var_15698_end_mask_0, x = q_1295_cast_fp16)[name = tensor<string, []>("op_15698_cast_fp16")];
            tensor<int32, [4]> var_15702_begin_0 = const()[name = tensor<string, []>("op_15702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15702_end_0 = const()[name = tensor<string, []>("op_15702_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15702_end_mask_0 = const()[name = tensor<string, []>("op_15702_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15702_cast_fp16 = slice_by_index(begin = var_15702_begin_0, end = var_15702_end_0, end_mask = var_15702_end_mask_0, x = k_777_cast_fp16)[name = tensor<string, []>("op_15702_cast_fp16")];
            tensor<int32, [4]> var_15706_begin_0 = const()[name = tensor<string, []>("op_15706_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15706_end_0 = const()[name = tensor<string, []>("op_15706_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15706_end_mask_0 = const()[name = tensor<string, []>("op_15706_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15706_cast_fp16 = slice_by_index(begin = var_15706_begin_0, end = var_15706_end_0, end_mask = var_15706_end_mask_0, x = q_1295_cast_fp16)[name = tensor<string, []>("op_15706_cast_fp16")];
            tensor<int32, [4]> var_15710_begin_0 = const()[name = tensor<string, []>("op_15710_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15710_end_0 = const()[name = tensor<string, []>("op_15710_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15710_end_mask_0 = const()[name = tensor<string, []>("op_15710_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15710_cast_fp16 = slice_by_index(begin = var_15710_begin_0, end = var_15710_end_0, end_mask = var_15710_end_mask_0, x = k_777_cast_fp16)[name = tensor<string, []>("op_15710_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15712_cast_fp16 = mul(x = var_15706_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15712_cast_fp16")];
            tensor<int32, [5]> var_15720 = const()[name = tensor<string, []>("op_15720"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1759_cast_fp16 = reshape(shape = var_15720, x = var_15706_cast_fp16)[name = tensor<string, []>("x_1759_cast_fp16")];
            tensor<int32, [2]> var_15722_split_sizes_0 = const()[name = tensor<string, []>("op_15722_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15722_axis_0 = const()[name = tensor<string, []>("op_15722_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15722_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15722_cast_fp16_1 = split(axis = var_15722_axis_0, split_sizes = var_15722_split_sizes_0, x = x_1759_cast_fp16)[name = tensor<string, []>("op_15722_cast_fp16")];
            tensor<int32, [1]> squeeze_516_axes_0 = const()[name = tensor<string, []>("squeeze_516_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_516_cast_fp16 = squeeze(axes = squeeze_516_axes_0, x = var_15722_cast_fp16_0)[name = tensor<string, []>("squeeze_516_cast_fp16")];
            tensor<int32, [1]> squeeze_517_axes_0 = const()[name = tensor<string, []>("squeeze_517_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_517_cast_fp16 = squeeze(axes = squeeze_517_axes_0, x = var_15722_cast_fp16_1)[name = tensor<string, []>("squeeze_517_cast_fp16")];
            tensor<fp16, []> const_2111_promoted_to_fp16 = const()[name = tensor<string, []>("const_2111_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15725_cast_fp16 = mul(x = squeeze_517_cast_fp16, y = const_2111_promoted_to_fp16)[name = tensor<string, []>("op_15725_cast_fp16")];
            tensor<int32, []> x_1761_axis_0 = const()[name = tensor<string, []>("x_1761_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1761_cast_fp16 = stack(axis = x_1761_axis_0, values = (var_15725_cast_fp16, squeeze_516_cast_fp16))[name = tensor<string, []>("x_1761_cast_fp16")];
            tensor<int32, [4]> var_15731 = const()[name = tensor<string, []>("op_15731"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15732_cast_fp16 = reshape(shape = var_15731, x = x_1761_cast_fp16)[name = tensor<string, []>("op_15732_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15733_cast_fp16 = mul(x = var_15732_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15733_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_259_cast_fp16 = add(x = var_15712_cast_fp16, y = var_15733_cast_fp16)[name = tensor<string, []>("q_patches_259_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15735_cast_fp16 = mul(x = var_15710_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15735_cast_fp16")];
            tensor<int32, [5]> var_15743 = const()[name = tensor<string, []>("op_15743"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1765_cast_fp16 = reshape(shape = var_15743, x = var_15710_cast_fp16)[name = tensor<string, []>("x_1765_cast_fp16")];
            tensor<int32, [2]> var_15745_split_sizes_0 = const()[name = tensor<string, []>("op_15745_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15745_axis_0 = const()[name = tensor<string, []>("op_15745_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15745_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15745_cast_fp16_1 = split(axis = var_15745_axis_0, split_sizes = var_15745_split_sizes_0, x = x_1765_cast_fp16)[name = tensor<string, []>("op_15745_cast_fp16")];
            tensor<int32, [1]> squeeze_518_axes_0 = const()[name = tensor<string, []>("squeeze_518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_518_cast_fp16 = squeeze(axes = squeeze_518_axes_0, x = var_15745_cast_fp16_0)[name = tensor<string, []>("squeeze_518_cast_fp16")];
            tensor<int32, [1]> squeeze_519_axes_0 = const()[name = tensor<string, []>("squeeze_519_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_519_cast_fp16 = squeeze(axes = squeeze_519_axes_0, x = var_15745_cast_fp16_1)[name = tensor<string, []>("squeeze_519_cast_fp16")];
            tensor<fp16, []> const_2119_promoted_to_fp16 = const()[name = tensor<string, []>("const_2119_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15748_cast_fp16 = mul(x = squeeze_519_cast_fp16, y = const_2119_promoted_to_fp16)[name = tensor<string, []>("op_15748_cast_fp16")];
            tensor<int32, []> x_1767_axis_0 = const()[name = tensor<string, []>("x_1767_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1767_cast_fp16 = stack(axis = x_1767_axis_0, values = (var_15748_cast_fp16, squeeze_518_cast_fp16))[name = tensor<string, []>("x_1767_cast_fp16")];
            tensor<int32, [4]> var_15754 = const()[name = tensor<string, []>("op_15754"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15755_cast_fp16 = reshape(shape = var_15754, x = x_1767_cast_fp16)[name = tensor<string, []>("op_15755_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15756_cast_fp16 = mul(x = var_15755_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15756_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_259_cast_fp16 = add(x = var_15735_cast_fp16, y = var_15756_cast_fp16)[name = tensor<string, []>("k_patches_259_cast_fp16")];
            tensor<bool, []> var_15759_interleave_0 = const()[name = tensor<string, []>("op_15759_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15759_cast_fp16 = concat(axis = var_20, interleave = var_15759_interleave_0, values = (var_15698_cast_fp16, q_patches_259_cast_fp16))[name = tensor<string, []>("op_15759_cast_fp16")];
            tensor<bool, []> var_15762_interleave_0 = const()[name = tensor<string, []>("op_15762_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15762_cast_fp16 = concat(axis = var_20, interleave = var_15762_interleave_0, values = (var_15702_cast_fp16, k_patches_259_cast_fp16))[name = tensor<string, []>("op_15762_cast_fp16")];
            tensor<fp16, []> var_15764_to_fp16 = const()[name = tensor<string, []>("op_15764_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1299_cast_fp16 = mul(x = var_15759_cast_fp16, y = var_15764_to_fp16)[name = tensor<string, []>("q_1299_cast_fp16")];
            tensor<bool, []> attn_517_transpose_x_1 = const()[name = tensor<string, []>("attn_517_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_517_transpose_y_1 = const()[name = tensor<string, []>("attn_517_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_517_cast_fp16 = matmul(transpose_x = attn_517_transpose_x_1, transpose_y = attn_517_transpose_y_1, x = q_1299_cast_fp16, y = var_15762_cast_fp16)[name = tensor<string, []>("attn_517_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_519_cast_fp16 = softmax(axis = var_21, x = attn_517_cast_fp16)[name = tensor<string, []>("attn_519_cast_fp16")];
            tensor<bool, []> var_15769_transpose_x_0 = const()[name = tensor<string, []>("op_15769_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15769_transpose_y_0 = const()[name = tensor<string, []>("op_15769_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15769_cast_fp16 = matmul(transpose_x = var_15769_transpose_x_0, transpose_y = var_15769_transpose_y_0, x = attn_519_cast_fp16, y = v_779_cast_fp16)[name = tensor<string, []>("op_15769_cast_fp16")];
            tensor<int32, [1]> var_15770_axes_0 = const()[name = tensor<string, []>("op_15770_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15770_cast_fp16 = squeeze(axes = var_15770_axes_0, x = var_15769_cast_fp16)[name = tensor<string, []>("op_15770_cast_fp16")];
            tensor<string, []> var_15776_pad_type_0 = const()[name = tensor<string, []>("op_15776_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15776_strides_0 = const()[name = tensor<string, []>("op_15776_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15776_pad_0 = const()[name = tensor<string, []>("op_15776_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15776_dilations_0 = const()[name = tensor<string, []>("op_15776_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15776_groups_0 = const()[name = tensor<string, []>("op_15776_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146367552)))];
            tensor<fp16, [1, 64, 1, 197]> var_15776_cast_fp16 = conv(dilations = var_15776_dilations_0, groups = var_15776_groups_0, pad = var_15776_pad_0, pad_type = var_15776_pad_type_0, strides = var_15776_strides_0, weight = model_blocks_10_attn_q_projs_10_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15776_cast_fp16")];
            tensor<int32, [1]> var_15777_axes_0 = const()[name = tensor<string, []>("op_15777_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15777_cast_fp16 = squeeze(axes = var_15777_axes_0, x = var_15776_cast_fp16)[name = tensor<string, []>("op_15777_cast_fp16")];
            tensor<int32, [3]> q_1301_perm_0 = const()[name = tensor<string, []>("q_1301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15784_pad_type_0 = const()[name = tensor<string, []>("op_15784_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15784_strides_0 = const()[name = tensor<string, []>("op_15784_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15784_pad_0 = const()[name = tensor<string, []>("op_15784_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15784_dilations_0 = const()[name = tensor<string, []>("op_15784_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15784_groups_0 = const()[name = tensor<string, []>("op_15784_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146465920)))];
            tensor<fp16, [1, 64, 1, 197]> var_15784_cast_fp16 = conv(dilations = var_15784_dilations_0, groups = var_15784_groups_0, pad = var_15784_pad_0, pad_type = var_15784_pad_type_0, strides = var_15784_strides_0, weight = model_blocks_10_attn_k_projs_10_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15784_cast_fp16")];
            tensor<int32, [1]> var_15785_axes_0 = const()[name = tensor<string, []>("op_15785_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15785_cast_fp16 = squeeze(axes = var_15785_axes_0, x = var_15784_cast_fp16)[name = tensor<string, []>("op_15785_cast_fp16")];
            tensor<int32, [3]> k_781_perm_0 = const()[name = tensor<string, []>("k_781_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15792_pad_type_0 = const()[name = tensor<string, []>("op_15792_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15792_strides_0 = const()[name = tensor<string, []>("op_15792_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15792_pad_0 = const()[name = tensor<string, []>("op_15792_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15792_dilations_0 = const()[name = tensor<string, []>("op_15792_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15792_groups_0 = const()[name = tensor<string, []>("op_15792_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146564288)))];
            tensor<fp16, [1, 64, 1, 197]> var_15792_cast_fp16 = conv(dilations = var_15792_dilations_0, groups = var_15792_groups_0, pad = var_15792_pad_0, pad_type = var_15792_pad_type_0, strides = var_15792_strides_0, weight = model_blocks_10_attn_v_projs_10_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15792_cast_fp16")];
            tensor<int32, [1]> var_15793_axes_0 = const()[name = tensor<string, []>("op_15793_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15793_cast_fp16 = squeeze(axes = var_15793_axes_0, x = var_15792_cast_fp16)[name = tensor<string, []>("op_15793_cast_fp16")];
            tensor<int32, [3]> v_781_perm_0 = const()[name = tensor<string, []>("v_781_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146662656)))];
            tensor<fp16, [1, 197, 64]> q_1301_cast_fp16 = transpose(perm = q_1301_perm_0, x = var_15777_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [1, 197, 64]> q_1303_cast_fp16 = add(x = q_1301_cast_fp16, y = model_blocks_10_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_1303_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146662848)))];
            tensor<fp16, [1, 197, 64]> v_781_cast_fp16 = transpose(perm = v_781_perm_0, x = var_15793_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 197, 64]> v_783_cast_fp16 = add(x = v_781_cast_fp16, y = model_blocks_10_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_783_cast_fp16")];
            tensor<int32, [1]> q_1305_axes_0 = const()[name = tensor<string, []>("q_1305_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1305_cast_fp16 = expand_dims(axes = q_1305_axes_0, x = q_1303_cast_fp16)[name = tensor<string, []>("q_1305_cast_fp16")];
            tensor<int32, [1]> k_783_axes_0 = const()[name = tensor<string, []>("k_783_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_781_cast_fp16 = transpose(perm = k_781_perm_0, x = var_15785_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 1, 197, 64]> k_783_cast_fp16 = expand_dims(axes = k_783_axes_0, x = k_781_cast_fp16)[name = tensor<string, []>("k_783_cast_fp16")];
            tensor<int32, [1]> v_785_axes_0 = const()[name = tensor<string, []>("v_785_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_785_cast_fp16 = expand_dims(axes = v_785_axes_0, x = v_783_cast_fp16)[name = tensor<string, []>("v_785_cast_fp16")];
            tensor<int32, [4]> var_15802_begin_0 = const()[name = tensor<string, []>("op_15802_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15802_end_0 = const()[name = tensor<string, []>("op_15802_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15802_end_mask_0 = const()[name = tensor<string, []>("op_15802_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15802_cast_fp16 = slice_by_index(begin = var_15802_begin_0, end = var_15802_end_0, end_mask = var_15802_end_mask_0, x = q_1305_cast_fp16)[name = tensor<string, []>("op_15802_cast_fp16")];
            tensor<int32, [4]> var_15806_begin_0 = const()[name = tensor<string, []>("op_15806_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15806_end_0 = const()[name = tensor<string, []>("op_15806_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15806_end_mask_0 = const()[name = tensor<string, []>("op_15806_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15806_cast_fp16 = slice_by_index(begin = var_15806_begin_0, end = var_15806_end_0, end_mask = var_15806_end_mask_0, x = k_783_cast_fp16)[name = tensor<string, []>("op_15806_cast_fp16")];
            tensor<int32, [4]> var_15810_begin_0 = const()[name = tensor<string, []>("op_15810_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15810_end_0 = const()[name = tensor<string, []>("op_15810_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15810_end_mask_0 = const()[name = tensor<string, []>("op_15810_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15810_cast_fp16 = slice_by_index(begin = var_15810_begin_0, end = var_15810_end_0, end_mask = var_15810_end_mask_0, x = q_1305_cast_fp16)[name = tensor<string, []>("op_15810_cast_fp16")];
            tensor<int32, [4]> var_15814_begin_0 = const()[name = tensor<string, []>("op_15814_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15814_end_0 = const()[name = tensor<string, []>("op_15814_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15814_end_mask_0 = const()[name = tensor<string, []>("op_15814_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15814_cast_fp16 = slice_by_index(begin = var_15814_begin_0, end = var_15814_end_0, end_mask = var_15814_end_mask_0, x = k_783_cast_fp16)[name = tensor<string, []>("op_15814_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15816_cast_fp16 = mul(x = var_15810_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15816_cast_fp16")];
            tensor<int32, [5]> var_15824 = const()[name = tensor<string, []>("op_15824"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1771_cast_fp16 = reshape(shape = var_15824, x = var_15810_cast_fp16)[name = tensor<string, []>("x_1771_cast_fp16")];
            tensor<int32, [2]> var_15826_split_sizes_0 = const()[name = tensor<string, []>("op_15826_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15826_axis_0 = const()[name = tensor<string, []>("op_15826_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15826_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15826_cast_fp16_1 = split(axis = var_15826_axis_0, split_sizes = var_15826_split_sizes_0, x = x_1771_cast_fp16)[name = tensor<string, []>("op_15826_cast_fp16")];
            tensor<int32, [1]> squeeze_520_axes_0 = const()[name = tensor<string, []>("squeeze_520_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_520_cast_fp16 = squeeze(axes = squeeze_520_axes_0, x = var_15826_cast_fp16_0)[name = tensor<string, []>("squeeze_520_cast_fp16")];
            tensor<int32, [1]> squeeze_521_axes_0 = const()[name = tensor<string, []>("squeeze_521_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_521_cast_fp16 = squeeze(axes = squeeze_521_axes_0, x = var_15826_cast_fp16_1)[name = tensor<string, []>("squeeze_521_cast_fp16")];
            tensor<fp16, []> const_2127_promoted_to_fp16 = const()[name = tensor<string, []>("const_2127_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15829_cast_fp16 = mul(x = squeeze_521_cast_fp16, y = const_2127_promoted_to_fp16)[name = tensor<string, []>("op_15829_cast_fp16")];
            tensor<int32, []> x_1773_axis_0 = const()[name = tensor<string, []>("x_1773_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1773_cast_fp16 = stack(axis = x_1773_axis_0, values = (var_15829_cast_fp16, squeeze_520_cast_fp16))[name = tensor<string, []>("x_1773_cast_fp16")];
            tensor<int32, [4]> var_15835 = const()[name = tensor<string, []>("op_15835"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15836_cast_fp16 = reshape(shape = var_15835, x = x_1773_cast_fp16)[name = tensor<string, []>("op_15836_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15837_cast_fp16 = mul(x = var_15836_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15837_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_261_cast_fp16 = add(x = var_15816_cast_fp16, y = var_15837_cast_fp16)[name = tensor<string, []>("q_patches_261_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15839_cast_fp16 = mul(x = var_15814_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15839_cast_fp16")];
            tensor<int32, [5]> var_15847 = const()[name = tensor<string, []>("op_15847"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1777_cast_fp16 = reshape(shape = var_15847, x = var_15814_cast_fp16)[name = tensor<string, []>("x_1777_cast_fp16")];
            tensor<int32, [2]> var_15849_split_sizes_0 = const()[name = tensor<string, []>("op_15849_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15849_axis_0 = const()[name = tensor<string, []>("op_15849_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15849_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15849_cast_fp16_1 = split(axis = var_15849_axis_0, split_sizes = var_15849_split_sizes_0, x = x_1777_cast_fp16)[name = tensor<string, []>("op_15849_cast_fp16")];
            tensor<int32, [1]> squeeze_522_axes_0 = const()[name = tensor<string, []>("squeeze_522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_522_cast_fp16 = squeeze(axes = squeeze_522_axes_0, x = var_15849_cast_fp16_0)[name = tensor<string, []>("squeeze_522_cast_fp16")];
            tensor<int32, [1]> squeeze_523_axes_0 = const()[name = tensor<string, []>("squeeze_523_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_523_cast_fp16 = squeeze(axes = squeeze_523_axes_0, x = var_15849_cast_fp16_1)[name = tensor<string, []>("squeeze_523_cast_fp16")];
            tensor<fp16, []> const_2135_promoted_to_fp16 = const()[name = tensor<string, []>("const_2135_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15852_cast_fp16 = mul(x = squeeze_523_cast_fp16, y = const_2135_promoted_to_fp16)[name = tensor<string, []>("op_15852_cast_fp16")];
            tensor<int32, []> x_1779_axis_0 = const()[name = tensor<string, []>("x_1779_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1779_cast_fp16 = stack(axis = x_1779_axis_0, values = (var_15852_cast_fp16, squeeze_522_cast_fp16))[name = tensor<string, []>("x_1779_cast_fp16")];
            tensor<int32, [4]> var_15858 = const()[name = tensor<string, []>("op_15858"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15859_cast_fp16 = reshape(shape = var_15858, x = x_1779_cast_fp16)[name = tensor<string, []>("op_15859_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15860_cast_fp16 = mul(x = var_15859_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15860_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_261_cast_fp16 = add(x = var_15839_cast_fp16, y = var_15860_cast_fp16)[name = tensor<string, []>("k_patches_261_cast_fp16")];
            tensor<bool, []> var_15863_interleave_0 = const()[name = tensor<string, []>("op_15863_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15863_cast_fp16 = concat(axis = var_20, interleave = var_15863_interleave_0, values = (var_15802_cast_fp16, q_patches_261_cast_fp16))[name = tensor<string, []>("op_15863_cast_fp16")];
            tensor<bool, []> var_15866_interleave_0 = const()[name = tensor<string, []>("op_15866_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15866_cast_fp16 = concat(axis = var_20, interleave = var_15866_interleave_0, values = (var_15806_cast_fp16, k_patches_261_cast_fp16))[name = tensor<string, []>("op_15866_cast_fp16")];
            tensor<fp16, []> var_15868_to_fp16 = const()[name = tensor<string, []>("op_15868_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1309_cast_fp16 = mul(x = var_15863_cast_fp16, y = var_15868_to_fp16)[name = tensor<string, []>("q_1309_cast_fp16")];
            tensor<bool, []> attn_521_transpose_x_1 = const()[name = tensor<string, []>("attn_521_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_521_transpose_y_1 = const()[name = tensor<string, []>("attn_521_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_521_cast_fp16 = matmul(transpose_x = attn_521_transpose_x_1, transpose_y = attn_521_transpose_y_1, x = q_1309_cast_fp16, y = var_15866_cast_fp16)[name = tensor<string, []>("attn_521_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_523_cast_fp16 = softmax(axis = var_21, x = attn_521_cast_fp16)[name = tensor<string, []>("attn_523_cast_fp16")];
            tensor<bool, []> var_15873_transpose_x_0 = const()[name = tensor<string, []>("op_15873_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15873_transpose_y_0 = const()[name = tensor<string, []>("op_15873_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15873_cast_fp16 = matmul(transpose_x = var_15873_transpose_x_0, transpose_y = var_15873_transpose_y_0, x = attn_523_cast_fp16, y = v_785_cast_fp16)[name = tensor<string, []>("op_15873_cast_fp16")];
            tensor<int32, [1]> var_15874_axes_0 = const()[name = tensor<string, []>("op_15874_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_15874_cast_fp16 = squeeze(axes = var_15874_axes_0, x = var_15873_cast_fp16)[name = tensor<string, []>("op_15874_cast_fp16")];
            tensor<string, []> var_15880_pad_type_0 = const()[name = tensor<string, []>("op_15880_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15880_strides_0 = const()[name = tensor<string, []>("op_15880_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15880_pad_0 = const()[name = tensor<string, []>("op_15880_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15880_dilations_0 = const()[name = tensor<string, []>("op_15880_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15880_groups_0 = const()[name = tensor<string, []>("op_15880_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146663040)))];
            tensor<fp16, [1, 64, 1, 197]> var_15880_cast_fp16 = conv(dilations = var_15880_dilations_0, groups = var_15880_groups_0, pad = var_15880_pad_0, pad_type = var_15880_pad_type_0, strides = var_15880_strides_0, weight = model_blocks_10_attn_q_projs_11_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15880_cast_fp16")];
            tensor<int32, [1]> var_15881_axes_0 = const()[name = tensor<string, []>("op_15881_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15881_cast_fp16 = squeeze(axes = var_15881_axes_0, x = var_15880_cast_fp16)[name = tensor<string, []>("op_15881_cast_fp16")];
            tensor<int32, [3]> q_1311_perm_0 = const()[name = tensor<string, []>("q_1311_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15888_pad_type_0 = const()[name = tensor<string, []>("op_15888_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15888_strides_0 = const()[name = tensor<string, []>("op_15888_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15888_pad_0 = const()[name = tensor<string, []>("op_15888_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15888_dilations_0 = const()[name = tensor<string, []>("op_15888_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15888_groups_0 = const()[name = tensor<string, []>("op_15888_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146761408)))];
            tensor<fp16, [1, 64, 1, 197]> var_15888_cast_fp16 = conv(dilations = var_15888_dilations_0, groups = var_15888_groups_0, pad = var_15888_pad_0, pad_type = var_15888_pad_type_0, strides = var_15888_strides_0, weight = model_blocks_10_attn_k_projs_11_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15888_cast_fp16")];
            tensor<int32, [1]> var_15889_axes_0 = const()[name = tensor<string, []>("op_15889_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15889_cast_fp16 = squeeze(axes = var_15889_axes_0, x = var_15888_cast_fp16)[name = tensor<string, []>("op_15889_cast_fp16")];
            tensor<int32, [3]> k_787_perm_0 = const()[name = tensor<string, []>("k_787_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_15896_pad_type_0 = const()[name = tensor<string, []>("op_15896_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_15896_strides_0 = const()[name = tensor<string, []>("op_15896_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_15896_pad_0 = const()[name = tensor<string, []>("op_15896_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_15896_dilations_0 = const()[name = tensor<string, []>("op_15896_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15896_groups_0 = const()[name = tensor<string, []>("op_15896_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_10_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146859776)))];
            tensor<fp16, [1, 64, 1, 197]> var_15896_cast_fp16 = conv(dilations = var_15896_dilations_0, groups = var_15896_groups_0, pad = var_15896_pad_0, pad_type = var_15896_pad_type_0, strides = var_15896_strides_0, weight = model_blocks_10_attn_v_projs_11_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("op_15896_cast_fp16")];
            tensor<int32, [1]> var_15897_axes_0 = const()[name = tensor<string, []>("op_15897_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_15897_cast_fp16 = squeeze(axes = var_15897_axes_0, x = var_15896_cast_fp16)[name = tensor<string, []>("op_15897_cast_fp16")];
            tensor<int32, [3]> v_787_perm_0 = const()[name = tensor<string, []>("v_787_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_10_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146958144)))];
            tensor<fp16, [1, 197, 64]> q_1311_cast_fp16 = transpose(perm = q_1311_perm_0, x = var_15881_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 197, 64]> q_1313_cast_fp16 = add(x = q_1311_cast_fp16, y = model_blocks_10_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_1313_cast_fp16")];
            tensor<fp16, [64]> model_blocks_10_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146958336)))];
            tensor<fp16, [1, 197, 64]> v_787_cast_fp16 = transpose(perm = v_787_perm_0, x = var_15897_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 197, 64]> v_789_cast_fp16 = add(x = v_787_cast_fp16, y = model_blocks_10_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_789_cast_fp16")];
            tensor<int32, [1]> q_1315_axes_0 = const()[name = tensor<string, []>("q_1315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1315_cast_fp16 = expand_dims(axes = q_1315_axes_0, x = q_1313_cast_fp16)[name = tensor<string, []>("q_1315_cast_fp16")];
            tensor<int32, [1]> k_789_axes_0 = const()[name = tensor<string, []>("k_789_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_787_cast_fp16 = transpose(perm = k_787_perm_0, x = var_15889_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 1, 197, 64]> k_789_cast_fp16 = expand_dims(axes = k_789_axes_0, x = k_787_cast_fp16)[name = tensor<string, []>("k_789_cast_fp16")];
            tensor<int32, [1]> v_791_axes_0 = const()[name = tensor<string, []>("v_791_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_791_cast_fp16 = expand_dims(axes = v_791_axes_0, x = v_789_cast_fp16)[name = tensor<string, []>("v_791_cast_fp16")];
            tensor<int32, [4]> var_15906_begin_0 = const()[name = tensor<string, []>("op_15906_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15906_end_0 = const()[name = tensor<string, []>("op_15906_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15906_end_mask_0 = const()[name = tensor<string, []>("op_15906_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15906_cast_fp16 = slice_by_index(begin = var_15906_begin_0, end = var_15906_end_0, end_mask = var_15906_end_mask_0, x = q_1315_cast_fp16)[name = tensor<string, []>("op_15906_cast_fp16")];
            tensor<int32, [4]> var_15910_begin_0 = const()[name = tensor<string, []>("op_15910_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15910_end_0 = const()[name = tensor<string, []>("op_15910_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_15910_end_mask_0 = const()[name = tensor<string, []>("op_15910_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_15910_cast_fp16 = slice_by_index(begin = var_15910_begin_0, end = var_15910_end_0, end_mask = var_15910_end_mask_0, x = k_789_cast_fp16)[name = tensor<string, []>("op_15910_cast_fp16")];
            tensor<int32, [4]> var_15914_begin_0 = const()[name = tensor<string, []>("op_15914_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15914_end_0 = const()[name = tensor<string, []>("op_15914_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15914_end_mask_0 = const()[name = tensor<string, []>("op_15914_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15914_cast_fp16 = slice_by_index(begin = var_15914_begin_0, end = var_15914_end_0, end_mask = var_15914_end_mask_0, x = q_1315_cast_fp16)[name = tensor<string, []>("op_15914_cast_fp16")];
            tensor<int32, [4]> var_15918_begin_0 = const()[name = tensor<string, []>("op_15918_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15918_end_0 = const()[name = tensor<string, []>("op_15918_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_15918_end_mask_0 = const()[name = tensor<string, []>("op_15918_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_15918_cast_fp16 = slice_by_index(begin = var_15918_begin_0, end = var_15918_end_0, end_mask = var_15918_end_mask_0, x = k_789_cast_fp16)[name = tensor<string, []>("op_15918_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15920_cast_fp16 = mul(x = var_15914_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15920_cast_fp16")];
            tensor<int32, [5]> var_15928 = const()[name = tensor<string, []>("op_15928"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1783_cast_fp16 = reshape(shape = var_15928, x = var_15914_cast_fp16)[name = tensor<string, []>("x_1783_cast_fp16")];
            tensor<int32, [2]> var_15930_split_sizes_0 = const()[name = tensor<string, []>("op_15930_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15930_axis_0 = const()[name = tensor<string, []>("op_15930_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15930_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15930_cast_fp16_1 = split(axis = var_15930_axis_0, split_sizes = var_15930_split_sizes_0, x = x_1783_cast_fp16)[name = tensor<string, []>("op_15930_cast_fp16")];
            tensor<int32, [1]> squeeze_524_axes_0 = const()[name = tensor<string, []>("squeeze_524_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_524_cast_fp16 = squeeze(axes = squeeze_524_axes_0, x = var_15930_cast_fp16_0)[name = tensor<string, []>("squeeze_524_cast_fp16")];
            tensor<int32, [1]> squeeze_525_axes_0 = const()[name = tensor<string, []>("squeeze_525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_525_cast_fp16 = squeeze(axes = squeeze_525_axes_0, x = var_15930_cast_fp16_1)[name = tensor<string, []>("squeeze_525_cast_fp16")];
            tensor<fp16, []> const_2143_promoted_to_fp16 = const()[name = tensor<string, []>("const_2143_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15933_cast_fp16 = mul(x = squeeze_525_cast_fp16, y = const_2143_promoted_to_fp16)[name = tensor<string, []>("op_15933_cast_fp16")];
            tensor<int32, []> x_1785_axis_0 = const()[name = tensor<string, []>("x_1785_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1785_cast_fp16 = stack(axis = x_1785_axis_0, values = (var_15933_cast_fp16, squeeze_524_cast_fp16))[name = tensor<string, []>("x_1785_cast_fp16")];
            tensor<int32, [4]> var_15939 = const()[name = tensor<string, []>("op_15939"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15940_cast_fp16 = reshape(shape = var_15939, x = x_1785_cast_fp16)[name = tensor<string, []>("op_15940_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15941_cast_fp16 = mul(x = var_15940_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15941_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_263_cast_fp16 = add(x = var_15920_cast_fp16, y = var_15941_cast_fp16)[name = tensor<string, []>("q_patches_263_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15943_cast_fp16 = mul(x = var_15918_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_15943_cast_fp16")];
            tensor<int32, [5]> var_15951 = const()[name = tensor<string, []>("op_15951"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1789_cast_fp16 = reshape(shape = var_15951, x = var_15918_cast_fp16)[name = tensor<string, []>("x_1789_cast_fp16")];
            tensor<int32, [2]> var_15953_split_sizes_0 = const()[name = tensor<string, []>("op_15953_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_15953_axis_0 = const()[name = tensor<string, []>("op_15953_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_15953_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_15953_cast_fp16_1 = split(axis = var_15953_axis_0, split_sizes = var_15953_split_sizes_0, x = x_1789_cast_fp16)[name = tensor<string, []>("op_15953_cast_fp16")];
            tensor<int32, [1]> squeeze_526_axes_0 = const()[name = tensor<string, []>("squeeze_526_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_526_cast_fp16 = squeeze(axes = squeeze_526_axes_0, x = var_15953_cast_fp16_0)[name = tensor<string, []>("squeeze_526_cast_fp16")];
            tensor<int32, [1]> squeeze_527_axes_0 = const()[name = tensor<string, []>("squeeze_527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_527_cast_fp16 = squeeze(axes = squeeze_527_axes_0, x = var_15953_cast_fp16_1)[name = tensor<string, []>("squeeze_527_cast_fp16")];
            tensor<fp16, []> const_2151_promoted_to_fp16 = const()[name = tensor<string, []>("const_2151_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_15956_cast_fp16 = mul(x = squeeze_527_cast_fp16, y = const_2151_promoted_to_fp16)[name = tensor<string, []>("op_15956_cast_fp16")];
            tensor<int32, []> x_1791_axis_0 = const()[name = tensor<string, []>("x_1791_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1791_cast_fp16 = stack(axis = x_1791_axis_0, values = (var_15956_cast_fp16, squeeze_526_cast_fp16))[name = tensor<string, []>("x_1791_cast_fp16")];
            tensor<int32, [4]> var_15962 = const()[name = tensor<string, []>("op_15962"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_15963_cast_fp16 = reshape(shape = var_15962, x = x_1791_cast_fp16)[name = tensor<string, []>("op_15963_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_15964_cast_fp16 = mul(x = var_15963_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_15964_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_263_cast_fp16 = add(x = var_15943_cast_fp16, y = var_15964_cast_fp16)[name = tensor<string, []>("k_patches_263_cast_fp16")];
            tensor<bool, []> var_15967_interleave_0 = const()[name = tensor<string, []>("op_15967_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15967_cast_fp16 = concat(axis = var_20, interleave = var_15967_interleave_0, values = (var_15906_cast_fp16, q_patches_263_cast_fp16))[name = tensor<string, []>("op_15967_cast_fp16")];
            tensor<bool, []> var_15970_interleave_0 = const()[name = tensor<string, []>("op_15970_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15970_cast_fp16 = concat(axis = var_20, interleave = var_15970_interleave_0, values = (var_15910_cast_fp16, k_patches_263_cast_fp16))[name = tensor<string, []>("op_15970_cast_fp16")];
            tensor<fp16, []> var_15972_to_fp16 = const()[name = tensor<string, []>("op_15972_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1319_cast_fp16 = mul(x = var_15967_cast_fp16, y = var_15972_to_fp16)[name = tensor<string, []>("q_1319_cast_fp16")];
            tensor<bool, []> attn_525_transpose_x_1 = const()[name = tensor<string, []>("attn_525_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_525_transpose_y_1 = const()[name = tensor<string, []>("attn_525_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_525_cast_fp16 = matmul(transpose_x = attn_525_transpose_x_1, transpose_y = attn_525_transpose_y_1, x = q_1319_cast_fp16, y = var_15970_cast_fp16)[name = tensor<string, []>("attn_525_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_527_cast_fp16 = softmax(axis = var_21, x = attn_525_cast_fp16)[name = tensor<string, []>("attn_527_cast_fp16")];
            tensor<bool, []> var_15977_transpose_x_0 = const()[name = tensor<string, []>("op_15977_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_15977_transpose_y_0 = const()[name = tensor<string, []>("op_15977_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_15977_cast_fp16 = matmul(transpose_x = var_15977_transpose_x_0, transpose_y = var_15977_transpose_y_0, x = attn_527_cast_fp16, y = v_791_cast_fp16)[name = tensor<string, []>("op_15977_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_21_cast_fp16 = squeeze(axes = out_21_axes_0, x = var_15977_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<bool, []> input_185_interleave_0 = const()[name = tensor<string, []>("input_185_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_185_cast_fp16 = concat(axis = var_21, interleave = input_185_interleave_0, values = (var_14834_cast_fp16, var_14938_cast_fp16, var_15042_cast_fp16, var_15146_cast_fp16, var_15250_cast_fp16, var_15354_cast_fp16, var_15458_cast_fp16, var_15562_cast_fp16, var_15666_cast_fp16, var_15770_cast_fp16, var_15874_cast_fp16, out_21_cast_fp16))[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [1]> x_1793_axes_0 = const()[name = tensor<string, []>("x_1793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_10_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146958528)))];
            tensor<fp16, [768]> model_blocks_10_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146960128)))];
            tensor<fp16, [1, 197, 768]> x_1793_cast_fp16 = layer_norm(axes = x_1793_axes_0, beta = model_blocks_10_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_10_attn_inner_attn_ln_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("x_1793_cast_fp16")];
            tensor<int32, [4]> var_15985 = const()[name = tensor<string, []>("op_15985"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1795_cast_fp16 = reshape(shape = var_15985, x = x_1793_cast_fp16)[name = tensor<string, []>("x_1795_cast_fp16")];
            tensor<int32, [4]> input_187_perm_0 = const()[name = tensor<string, []>("input_187_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1797_pad_type_0 = const()[name = tensor<string, []>("x_1797_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1797_strides_0 = const()[name = tensor<string, []>("x_1797_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1797_pad_0 = const()[name = tensor<string, []>("x_1797_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1797_dilations_0 = const()[name = tensor<string, []>("x_1797_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1797_groups_0 = const()[name = tensor<string, []>("x_1797_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_10_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146961728)))];
            tensor<fp16, [768]> model_blocks_10_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148141440)))];
            tensor<fp16, [1, 768, 1, 197]> input_187_cast_fp16 = transpose(perm = input_187_perm_0, x = x_1795_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 768, 1, 197]> x_1797_cast_fp16 = conv(bias = model_blocks_10_attn_proj_bias_to_fp16, dilations = x_1797_dilations_0, groups = x_1797_groups_0, pad = x_1797_pad_0, pad_type = x_1797_pad_type_0, strides = x_1797_strides_0, weight = model_blocks_10_attn_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("x_1797_cast_fp16")];
            tensor<int32, [4]> x_1799_perm_0 = const()[name = tensor<string, []>("x_1799_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_15996 = const()[name = tensor<string, []>("op_15996"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1799_cast_fp16 = transpose(perm = x_1799_perm_0, x = x_1797_cast_fp16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [1, 197, 768]> var_15997_cast_fp16 = reshape(shape = var_15996, x = x_1799_cast_fp16)[name = tensor<string, []>("op_15997_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_189_cast_fp16 = add(x = input_181_cast_fp16, y = var_15997_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<int32, [1]> x_1801_axes_0 = const()[name = tensor<string, []>("x_1801_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_10_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148143040)))];
            tensor<fp16, [768]> model_blocks_10_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148144640)))];
            tensor<fp16, [1, 197, 768]> x_1801_cast_fp16 = layer_norm(axes = x_1801_axes_0, beta = model_blocks_10_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_10_norm2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("x_1801_cast_fp16")];
            tensor<int32, [4]> var_16009 = const()[name = tensor<string, []>("op_16009"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1803_cast_fp16 = reshape(shape = var_16009, x = x_1801_cast_fp16)[name = tensor<string, []>("x_1803_cast_fp16")];
            tensor<int32, [4]> input_191_perm_0 = const()[name = tensor<string, []>("input_191_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_193_pad_type_0 = const()[name = tensor<string, []>("input_193_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_193_strides_0 = const()[name = tensor<string, []>("input_193_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_193_pad_0 = const()[name = tensor<string, []>("input_193_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_193_dilations_0 = const()[name = tensor<string, []>("input_193_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_193_groups_0 = const()[name = tensor<string, []>("input_193_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_10_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148146240)))];
            tensor<fp16, [2048]> model_blocks_10_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151292032)))];
            tensor<fp16, [1, 768, 1, 197]> input_191_cast_fp16 = transpose(perm = input_191_perm_0, x = x_1803_cast_fp16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 2048, 1, 197]> input_193_cast_fp16 = conv(bias = model_blocks_10_mlp_w1_bias_to_fp16, dilations = input_193_dilations_0, groups = input_193_groups_0, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = input_193_strides_0, weight = model_blocks_10_mlp_w1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<string, []> x2_549_pad_type_0 = const()[name = tensor<string, []>("x2_549_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_549_strides_0 = const()[name = tensor<string, []>("x2_549_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_549_pad_0 = const()[name = tensor<string, []>("x2_549_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_549_dilations_0 = const()[name = tensor<string, []>("x2_549_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_549_groups_0 = const()[name = tensor<string, []>("x2_549_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_10_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151296192)))];
            tensor<fp16, [2048]> model_blocks_10_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154441984)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_549_cast_fp16 = conv(bias = model_blocks_10_mlp_w2_bias_to_fp16, dilations = x2_549_dilations_0, groups = x2_549_groups_0, pad = x2_549_pad_0, pad_type = x2_549_pad_type_0, strides = x2_549_strides_0, weight = model_blocks_10_mlp_w2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("x2_549_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_16026_cast_fp16 = silu(x = input_193_cast_fp16)[name = tensor<string, []>("op_16026_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_81_cast_fp16 = mul(x = var_16026_cast_fp16, y = x2_549_cast_fp16)[name = tensor<string, []>("hidden_81_cast_fp16")];
            tensor<int32, [4]> hidden_83_perm_0 = const()[name = tensor<string, []>("hidden_83_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_16029 = const()[name = tensor<string, []>("op_16029"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_83_cast_fp16 = transpose(perm = hidden_83_perm_0, x = hidden_81_cast_fp16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [1, 197, 2048]> input_195_cast_fp16 = reshape(shape = var_16029, x = hidden_83_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<int32, [1]> hidden_85_axes_0 = const()[name = tensor<string, []>("hidden_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_10_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154446144)))];
            tensor<fp16, [2048]> model_blocks_10_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154450304)))];
            tensor<fp16, [1, 197, 2048]> hidden_85_cast_fp16 = layer_norm(axes = hidden_85_axes_0, beta = model_blocks_10_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_10_mlp_ffn_ln_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("hidden_85_cast_fp16")];
            tensor<int32, [4]> var_16035 = const()[name = tensor<string, []>("op_16035"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_87_cast_fp16 = reshape(shape = var_16035, x = hidden_85_cast_fp16)[name = tensor<string, []>("hidden_87_cast_fp16")];
            tensor<int32, [4]> input_197_perm_0 = const()[name = tensor<string, []>("input_197_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1805_pad_type_0 = const()[name = tensor<string, []>("x_1805_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1805_strides_0 = const()[name = tensor<string, []>("x_1805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1805_pad_0 = const()[name = tensor<string, []>("x_1805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1805_dilations_0 = const()[name = tensor<string, []>("x_1805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1805_groups_0 = const()[name = tensor<string, []>("x_1805_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_10_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154454464)))];
            tensor<fp16, [768]> model_blocks_10_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_10_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157600256)))];
            tensor<fp16, [1, 2048, 1, 197]> input_197_cast_fp16 = transpose(perm = input_197_perm_0, x = hidden_87_cast_fp16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [1, 768, 1, 197]> x_1805_cast_fp16 = conv(bias = model_blocks_10_mlp_w3_bias_to_fp16, dilations = x_1805_dilations_0, groups = x_1805_groups_0, pad = x_1805_pad_0, pad_type = x_1805_pad_type_0, strides = x_1805_strides_0, weight = model_blocks_10_mlp_w3_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("x_1805_cast_fp16")];
            tensor<int32, [4]> x_1807_perm_0 = const()[name = tensor<string, []>("x_1807_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_16046 = const()[name = tensor<string, []>("op_16046"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1807_cast_fp16 = transpose(perm = x_1807_perm_0, x = x_1805_cast_fp16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [1, 197, 768]> var_16047_cast_fp16 = reshape(shape = var_16046, x = x_1807_cast_fp16)[name = tensor<string, []>("op_16047_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_199_cast_fp16 = add(x = input_189_cast_fp16, y = var_16047_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<int32, [1]> x_1809_axes_0 = const()[name = tensor<string, []>("x_1809_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_11_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157601856)))];
            tensor<fp16, [768]> model_blocks_11_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157603456)))];
            tensor<fp16, [1, 197, 768]> x_1809_cast_fp16 = layer_norm(axes = x_1809_axes_0, beta = model_blocks_11_norm1_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_11_norm1_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("x_1809_cast_fp16")];
            tensor<int32, [4]> var_16181 = const()[name = tensor<string, []>("op_16181"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1811_cast_fp16 = reshape(shape = var_16181, x = x_1809_cast_fp16)[name = tensor<string, []>("x_1811_cast_fp16")];
            tensor<int32, [4]> input_201_perm_0 = const()[name = tensor<string, []>("input_201_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> var_16189_pad_type_0 = const()[name = tensor<string, []>("op_16189_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16189_strides_0 = const()[name = tensor<string, []>("op_16189_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16189_pad_0 = const()[name = tensor<string, []>("op_16189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16189_dilations_0 = const()[name = tensor<string, []>("op_16189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16189_groups_0 = const()[name = tensor<string, []>("op_16189_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157605056)))];
            tensor<fp16, [1, 768, 1, 197]> input_201_cast_fp16 = transpose(perm = input_201_perm_0, x = x_1811_cast_fp16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 64, 1, 197]> var_16189_cast_fp16 = conv(dilations = var_16189_dilations_0, groups = var_16189_groups_0, pad = var_16189_pad_0, pad_type = var_16189_pad_type_0, strides = var_16189_strides_0, weight = model_blocks_11_attn_q_projs_0_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16189_cast_fp16")];
            tensor<int32, [1]> var_16190_axes_0 = const()[name = tensor<string, []>("op_16190_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16190_cast_fp16 = squeeze(axes = var_16190_axes_0, x = var_16189_cast_fp16)[name = tensor<string, []>("op_16190_cast_fp16")];
            tensor<int32, [3]> q_1321_perm_0 = const()[name = tensor<string, []>("q_1321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16197_pad_type_0 = const()[name = tensor<string, []>("op_16197_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16197_strides_0 = const()[name = tensor<string, []>("op_16197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16197_pad_0 = const()[name = tensor<string, []>("op_16197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16197_dilations_0 = const()[name = tensor<string, []>("op_16197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16197_groups_0 = const()[name = tensor<string, []>("op_16197_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157703424)))];
            tensor<fp16, [1, 64, 1, 197]> var_16197_cast_fp16 = conv(dilations = var_16197_dilations_0, groups = var_16197_groups_0, pad = var_16197_pad_0, pad_type = var_16197_pad_type_0, strides = var_16197_strides_0, weight = model_blocks_11_attn_k_projs_0_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16197_cast_fp16")];
            tensor<int32, [1]> var_16198_axes_0 = const()[name = tensor<string, []>("op_16198_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16198_cast_fp16 = squeeze(axes = var_16198_axes_0, x = var_16197_cast_fp16)[name = tensor<string, []>("op_16198_cast_fp16")];
            tensor<int32, [3]> k_793_perm_0 = const()[name = tensor<string, []>("k_793_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16205_pad_type_0 = const()[name = tensor<string, []>("op_16205_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16205_strides_0 = const()[name = tensor<string, []>("op_16205_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16205_pad_0 = const()[name = tensor<string, []>("op_16205_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16205_dilations_0 = const()[name = tensor<string, []>("op_16205_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16205_groups_0 = const()[name = tensor<string, []>("op_16205_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_0_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_0_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157801792)))];
            tensor<fp16, [1, 64, 1, 197]> var_16205_cast_fp16 = conv(dilations = var_16205_dilations_0, groups = var_16205_groups_0, pad = var_16205_pad_0, pad_type = var_16205_pad_type_0, strides = var_16205_strides_0, weight = model_blocks_11_attn_v_projs_0_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16205_cast_fp16")];
            tensor<int32, [1]> var_16206_axes_0 = const()[name = tensor<string, []>("op_16206_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16206_cast_fp16 = squeeze(axes = var_16206_axes_0, x = var_16205_cast_fp16)[name = tensor<string, []>("op_16206_cast_fp16")];
            tensor<int32, [3]> v_793_perm_0 = const()[name = tensor<string, []>("v_793_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157900160)))];
            tensor<fp16, [1, 197, 64]> q_1321_cast_fp16 = transpose(perm = q_1321_perm_0, x = var_16190_cast_fp16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [1, 197, 64]> q_1323_cast_fp16 = add(x = q_1321_cast_fp16, y = model_blocks_11_attn_q_biases_0_to_fp16)[name = tensor<string, []>("q_1323_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_0_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_0_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157900352)))];
            tensor<fp16, [1, 197, 64]> v_793_cast_fp16 = transpose(perm = v_793_perm_0, x = var_16206_cast_fp16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [1, 197, 64]> v_795_cast_fp16 = add(x = v_793_cast_fp16, y = model_blocks_11_attn_v_biases_0_to_fp16)[name = tensor<string, []>("v_795_cast_fp16")];
            tensor<int32, [1]> q_1325_axes_0 = const()[name = tensor<string, []>("q_1325_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1325_cast_fp16 = expand_dims(axes = q_1325_axes_0, x = q_1323_cast_fp16)[name = tensor<string, []>("q_1325_cast_fp16")];
            tensor<int32, [1]> k_795_axes_0 = const()[name = tensor<string, []>("k_795_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_793_cast_fp16 = transpose(perm = k_793_perm_0, x = var_16198_cast_fp16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [1, 1, 197, 64]> k_795_cast_fp16 = expand_dims(axes = k_795_axes_0, x = k_793_cast_fp16)[name = tensor<string, []>("k_795_cast_fp16")];
            tensor<int32, [1]> v_797_axes_0 = const()[name = tensor<string, []>("v_797_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_797_cast_fp16 = expand_dims(axes = v_797_axes_0, x = v_795_cast_fp16)[name = tensor<string, []>("v_797_cast_fp16")];
            tensor<int32, [4]> var_16215_begin_0 = const()[name = tensor<string, []>("op_16215_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16215_end_0 = const()[name = tensor<string, []>("op_16215_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16215_end_mask_0 = const()[name = tensor<string, []>("op_16215_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16215_cast_fp16 = slice_by_index(begin = var_16215_begin_0, end = var_16215_end_0, end_mask = var_16215_end_mask_0, x = q_1325_cast_fp16)[name = tensor<string, []>("op_16215_cast_fp16")];
            tensor<int32, [4]> var_16219_begin_0 = const()[name = tensor<string, []>("op_16219_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16219_end_0 = const()[name = tensor<string, []>("op_16219_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16219_end_mask_0 = const()[name = tensor<string, []>("op_16219_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16219_cast_fp16 = slice_by_index(begin = var_16219_begin_0, end = var_16219_end_0, end_mask = var_16219_end_mask_0, x = k_795_cast_fp16)[name = tensor<string, []>("op_16219_cast_fp16")];
            tensor<int32, [4]> var_16223_begin_0 = const()[name = tensor<string, []>("op_16223_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16223_end_0 = const()[name = tensor<string, []>("op_16223_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16223_end_mask_0 = const()[name = tensor<string, []>("op_16223_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16223_cast_fp16 = slice_by_index(begin = var_16223_begin_0, end = var_16223_end_0, end_mask = var_16223_end_mask_0, x = q_1325_cast_fp16)[name = tensor<string, []>("op_16223_cast_fp16")];
            tensor<int32, [4]> var_16227_begin_0 = const()[name = tensor<string, []>("op_16227_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16227_end_0 = const()[name = tensor<string, []>("op_16227_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16227_end_mask_0 = const()[name = tensor<string, []>("op_16227_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16227_cast_fp16 = slice_by_index(begin = var_16227_begin_0, end = var_16227_end_0, end_mask = var_16227_end_mask_0, x = k_795_cast_fp16)[name = tensor<string, []>("op_16227_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16229_cast_fp16 = mul(x = var_16223_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16229_cast_fp16")];
            tensor<int32, [5]> var_16237 = const()[name = tensor<string, []>("op_16237"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1815_cast_fp16 = reshape(shape = var_16237, x = var_16223_cast_fp16)[name = tensor<string, []>("x_1815_cast_fp16")];
            tensor<int32, [2]> var_16239_split_sizes_0 = const()[name = tensor<string, []>("op_16239_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16239_axis_0 = const()[name = tensor<string, []>("op_16239_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16239_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16239_cast_fp16_1 = split(axis = var_16239_axis_0, split_sizes = var_16239_split_sizes_0, x = x_1815_cast_fp16)[name = tensor<string, []>("op_16239_cast_fp16")];
            tensor<int32, [1]> squeeze_528_axes_0 = const()[name = tensor<string, []>("squeeze_528_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_528_cast_fp16 = squeeze(axes = squeeze_528_axes_0, x = var_16239_cast_fp16_0)[name = tensor<string, []>("squeeze_528_cast_fp16")];
            tensor<int32, [1]> squeeze_529_axes_0 = const()[name = tensor<string, []>("squeeze_529_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_529_cast_fp16 = squeeze(axes = squeeze_529_axes_0, x = var_16239_cast_fp16_1)[name = tensor<string, []>("squeeze_529_cast_fp16")];
            tensor<fp16, []> const_2163_promoted_to_fp16 = const()[name = tensor<string, []>("const_2163_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16242_cast_fp16 = mul(x = squeeze_529_cast_fp16, y = const_2163_promoted_to_fp16)[name = tensor<string, []>("op_16242_cast_fp16")];
            tensor<int32, []> x_1817_axis_0 = const()[name = tensor<string, []>("x_1817_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1817_cast_fp16 = stack(axis = x_1817_axis_0, values = (var_16242_cast_fp16, squeeze_528_cast_fp16))[name = tensor<string, []>("x_1817_cast_fp16")];
            tensor<int32, [4]> var_16248 = const()[name = tensor<string, []>("op_16248"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16249_cast_fp16 = reshape(shape = var_16248, x = x_1817_cast_fp16)[name = tensor<string, []>("op_16249_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16250_cast_fp16 = mul(x = var_16249_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16250_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_265_cast_fp16 = add(x = var_16229_cast_fp16, y = var_16250_cast_fp16)[name = tensor<string, []>("q_patches_265_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16252_cast_fp16 = mul(x = var_16227_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16252_cast_fp16")];
            tensor<int32, [5]> var_16260 = const()[name = tensor<string, []>("op_16260"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1821_cast_fp16 = reshape(shape = var_16260, x = var_16227_cast_fp16)[name = tensor<string, []>("x_1821_cast_fp16")];
            tensor<int32, [2]> var_16262_split_sizes_0 = const()[name = tensor<string, []>("op_16262_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16262_axis_0 = const()[name = tensor<string, []>("op_16262_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16262_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16262_cast_fp16_1 = split(axis = var_16262_axis_0, split_sizes = var_16262_split_sizes_0, x = x_1821_cast_fp16)[name = tensor<string, []>("op_16262_cast_fp16")];
            tensor<int32, [1]> squeeze_530_axes_0 = const()[name = tensor<string, []>("squeeze_530_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_530_cast_fp16 = squeeze(axes = squeeze_530_axes_0, x = var_16262_cast_fp16_0)[name = tensor<string, []>("squeeze_530_cast_fp16")];
            tensor<int32, [1]> squeeze_531_axes_0 = const()[name = tensor<string, []>("squeeze_531_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_531_cast_fp16 = squeeze(axes = squeeze_531_axes_0, x = var_16262_cast_fp16_1)[name = tensor<string, []>("squeeze_531_cast_fp16")];
            tensor<fp16, []> const_2171_promoted_to_fp16 = const()[name = tensor<string, []>("const_2171_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16265_cast_fp16 = mul(x = squeeze_531_cast_fp16, y = const_2171_promoted_to_fp16)[name = tensor<string, []>("op_16265_cast_fp16")];
            tensor<int32, []> x_1823_axis_0 = const()[name = tensor<string, []>("x_1823_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1823_cast_fp16 = stack(axis = x_1823_axis_0, values = (var_16265_cast_fp16, squeeze_530_cast_fp16))[name = tensor<string, []>("x_1823_cast_fp16")];
            tensor<int32, [4]> var_16271 = const()[name = tensor<string, []>("op_16271"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16272_cast_fp16 = reshape(shape = var_16271, x = x_1823_cast_fp16)[name = tensor<string, []>("op_16272_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16273_cast_fp16 = mul(x = var_16272_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16273_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_265_cast_fp16 = add(x = var_16252_cast_fp16, y = var_16273_cast_fp16)[name = tensor<string, []>("k_patches_265_cast_fp16")];
            tensor<bool, []> var_16276_interleave_0 = const()[name = tensor<string, []>("op_16276_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16276_cast_fp16 = concat(axis = var_20, interleave = var_16276_interleave_0, values = (var_16215_cast_fp16, q_patches_265_cast_fp16))[name = tensor<string, []>("op_16276_cast_fp16")];
            tensor<bool, []> var_16279_interleave_0 = const()[name = tensor<string, []>("op_16279_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16279_cast_fp16 = concat(axis = var_20, interleave = var_16279_interleave_0, values = (var_16219_cast_fp16, k_patches_265_cast_fp16))[name = tensor<string, []>("op_16279_cast_fp16")];
            tensor<fp16, []> var_16281_to_fp16 = const()[name = tensor<string, []>("op_16281_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1329_cast_fp16 = mul(x = var_16276_cast_fp16, y = var_16281_to_fp16)[name = tensor<string, []>("q_1329_cast_fp16")];
            tensor<bool, []> attn_529_transpose_x_1 = const()[name = tensor<string, []>("attn_529_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_529_transpose_y_1 = const()[name = tensor<string, []>("attn_529_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_529_cast_fp16 = matmul(transpose_x = attn_529_transpose_x_1, transpose_y = attn_529_transpose_y_1, x = q_1329_cast_fp16, y = var_16279_cast_fp16)[name = tensor<string, []>("attn_529_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_531_cast_fp16 = softmax(axis = var_21, x = attn_529_cast_fp16)[name = tensor<string, []>("attn_531_cast_fp16")];
            tensor<bool, []> var_16286_transpose_x_0 = const()[name = tensor<string, []>("op_16286_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16286_transpose_y_0 = const()[name = tensor<string, []>("op_16286_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16286_cast_fp16 = matmul(transpose_x = var_16286_transpose_x_0, transpose_y = var_16286_transpose_y_0, x = attn_531_cast_fp16, y = v_797_cast_fp16)[name = tensor<string, []>("op_16286_cast_fp16")];
            tensor<int32, [1]> var_16287_axes_0 = const()[name = tensor<string, []>("op_16287_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16287_cast_fp16 = squeeze(axes = var_16287_axes_0, x = var_16286_cast_fp16)[name = tensor<string, []>("op_16287_cast_fp16")];
            tensor<string, []> var_16293_pad_type_0 = const()[name = tensor<string, []>("op_16293_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16293_strides_0 = const()[name = tensor<string, []>("op_16293_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16293_pad_0 = const()[name = tensor<string, []>("op_16293_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16293_dilations_0 = const()[name = tensor<string, []>("op_16293_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16293_groups_0 = const()[name = tensor<string, []>("op_16293_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157900544)))];
            tensor<fp16, [1, 64, 1, 197]> var_16293_cast_fp16 = conv(dilations = var_16293_dilations_0, groups = var_16293_groups_0, pad = var_16293_pad_0, pad_type = var_16293_pad_type_0, strides = var_16293_strides_0, weight = model_blocks_11_attn_q_projs_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16293_cast_fp16")];
            tensor<int32, [1]> var_16294_axes_0 = const()[name = tensor<string, []>("op_16294_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16294_cast_fp16 = squeeze(axes = var_16294_axes_0, x = var_16293_cast_fp16)[name = tensor<string, []>("op_16294_cast_fp16")];
            tensor<int32, [3]> q_1331_perm_0 = const()[name = tensor<string, []>("q_1331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16301_pad_type_0 = const()[name = tensor<string, []>("op_16301_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16301_strides_0 = const()[name = tensor<string, []>("op_16301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16301_pad_0 = const()[name = tensor<string, []>("op_16301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16301_dilations_0 = const()[name = tensor<string, []>("op_16301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16301_groups_0 = const()[name = tensor<string, []>("op_16301_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157998912)))];
            tensor<fp16, [1, 64, 1, 197]> var_16301_cast_fp16 = conv(dilations = var_16301_dilations_0, groups = var_16301_groups_0, pad = var_16301_pad_0, pad_type = var_16301_pad_type_0, strides = var_16301_strides_0, weight = model_blocks_11_attn_k_projs_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16301_cast_fp16")];
            tensor<int32, [1]> var_16302_axes_0 = const()[name = tensor<string, []>("op_16302_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16302_cast_fp16 = squeeze(axes = var_16302_axes_0, x = var_16301_cast_fp16)[name = tensor<string, []>("op_16302_cast_fp16")];
            tensor<int32, [3]> k_799_perm_0 = const()[name = tensor<string, []>("k_799_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16309_pad_type_0 = const()[name = tensor<string, []>("op_16309_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16309_strides_0 = const()[name = tensor<string, []>("op_16309_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16309_pad_0 = const()[name = tensor<string, []>("op_16309_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16309_dilations_0 = const()[name = tensor<string, []>("op_16309_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16309_groups_0 = const()[name = tensor<string, []>("op_16309_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_1_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158097280)))];
            tensor<fp16, [1, 64, 1, 197]> var_16309_cast_fp16 = conv(dilations = var_16309_dilations_0, groups = var_16309_groups_0, pad = var_16309_pad_0, pad_type = var_16309_pad_type_0, strides = var_16309_strides_0, weight = model_blocks_11_attn_v_projs_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16309_cast_fp16")];
            tensor<int32, [1]> var_16310_axes_0 = const()[name = tensor<string, []>("op_16310_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16310_cast_fp16 = squeeze(axes = var_16310_axes_0, x = var_16309_cast_fp16)[name = tensor<string, []>("op_16310_cast_fp16")];
            tensor<int32, [3]> v_799_perm_0 = const()[name = tensor<string, []>("v_799_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158195648)))];
            tensor<fp16, [1, 197, 64]> q_1331_cast_fp16 = transpose(perm = q_1331_perm_0, x = var_16294_cast_fp16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 197, 64]> q_1333_cast_fp16 = add(x = q_1331_cast_fp16, y = model_blocks_11_attn_q_biases_1_to_fp16)[name = tensor<string, []>("q_1333_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_1_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158195840)))];
            tensor<fp16, [1, 197, 64]> v_799_cast_fp16 = transpose(perm = v_799_perm_0, x = var_16310_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, 197, 64]> v_801_cast_fp16 = add(x = v_799_cast_fp16, y = model_blocks_11_attn_v_biases_1_to_fp16)[name = tensor<string, []>("v_801_cast_fp16")];
            tensor<int32, [1]> q_1335_axes_0 = const()[name = tensor<string, []>("q_1335_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1335_cast_fp16 = expand_dims(axes = q_1335_axes_0, x = q_1333_cast_fp16)[name = tensor<string, []>("q_1335_cast_fp16")];
            tensor<int32, [1]> k_801_axes_0 = const()[name = tensor<string, []>("k_801_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_799_cast_fp16 = transpose(perm = k_799_perm_0, x = var_16302_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [1, 1, 197, 64]> k_801_cast_fp16 = expand_dims(axes = k_801_axes_0, x = k_799_cast_fp16)[name = tensor<string, []>("k_801_cast_fp16")];
            tensor<int32, [1]> v_803_axes_0 = const()[name = tensor<string, []>("v_803_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_803_cast_fp16 = expand_dims(axes = v_803_axes_0, x = v_801_cast_fp16)[name = tensor<string, []>("v_803_cast_fp16")];
            tensor<int32, [4]> var_16319_begin_0 = const()[name = tensor<string, []>("op_16319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16319_end_0 = const()[name = tensor<string, []>("op_16319_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16319_end_mask_0 = const()[name = tensor<string, []>("op_16319_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16319_cast_fp16 = slice_by_index(begin = var_16319_begin_0, end = var_16319_end_0, end_mask = var_16319_end_mask_0, x = q_1335_cast_fp16)[name = tensor<string, []>("op_16319_cast_fp16")];
            tensor<int32, [4]> var_16323_begin_0 = const()[name = tensor<string, []>("op_16323_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16323_end_0 = const()[name = tensor<string, []>("op_16323_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16323_end_mask_0 = const()[name = tensor<string, []>("op_16323_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16323_cast_fp16 = slice_by_index(begin = var_16323_begin_0, end = var_16323_end_0, end_mask = var_16323_end_mask_0, x = k_801_cast_fp16)[name = tensor<string, []>("op_16323_cast_fp16")];
            tensor<int32, [4]> var_16327_begin_0 = const()[name = tensor<string, []>("op_16327_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16327_end_0 = const()[name = tensor<string, []>("op_16327_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16327_end_mask_0 = const()[name = tensor<string, []>("op_16327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16327_cast_fp16 = slice_by_index(begin = var_16327_begin_0, end = var_16327_end_0, end_mask = var_16327_end_mask_0, x = q_1335_cast_fp16)[name = tensor<string, []>("op_16327_cast_fp16")];
            tensor<int32, [4]> var_16331_begin_0 = const()[name = tensor<string, []>("op_16331_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16331_end_0 = const()[name = tensor<string, []>("op_16331_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16331_end_mask_0 = const()[name = tensor<string, []>("op_16331_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16331_cast_fp16 = slice_by_index(begin = var_16331_begin_0, end = var_16331_end_0, end_mask = var_16331_end_mask_0, x = k_801_cast_fp16)[name = tensor<string, []>("op_16331_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16333_cast_fp16 = mul(x = var_16327_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16333_cast_fp16")];
            tensor<int32, [5]> var_16341 = const()[name = tensor<string, []>("op_16341"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1827_cast_fp16 = reshape(shape = var_16341, x = var_16327_cast_fp16)[name = tensor<string, []>("x_1827_cast_fp16")];
            tensor<int32, [2]> var_16343_split_sizes_0 = const()[name = tensor<string, []>("op_16343_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16343_axis_0 = const()[name = tensor<string, []>("op_16343_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16343_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16343_cast_fp16_1 = split(axis = var_16343_axis_0, split_sizes = var_16343_split_sizes_0, x = x_1827_cast_fp16)[name = tensor<string, []>("op_16343_cast_fp16")];
            tensor<int32, [1]> squeeze_532_axes_0 = const()[name = tensor<string, []>("squeeze_532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_532_cast_fp16 = squeeze(axes = squeeze_532_axes_0, x = var_16343_cast_fp16_0)[name = tensor<string, []>("squeeze_532_cast_fp16")];
            tensor<int32, [1]> squeeze_533_axes_0 = const()[name = tensor<string, []>("squeeze_533_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_533_cast_fp16 = squeeze(axes = squeeze_533_axes_0, x = var_16343_cast_fp16_1)[name = tensor<string, []>("squeeze_533_cast_fp16")];
            tensor<fp16, []> const_2179_promoted_to_fp16 = const()[name = tensor<string, []>("const_2179_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16346_cast_fp16 = mul(x = squeeze_533_cast_fp16, y = const_2179_promoted_to_fp16)[name = tensor<string, []>("op_16346_cast_fp16")];
            tensor<int32, []> x_1829_axis_0 = const()[name = tensor<string, []>("x_1829_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1829_cast_fp16 = stack(axis = x_1829_axis_0, values = (var_16346_cast_fp16, squeeze_532_cast_fp16))[name = tensor<string, []>("x_1829_cast_fp16")];
            tensor<int32, [4]> var_16352 = const()[name = tensor<string, []>("op_16352"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16353_cast_fp16 = reshape(shape = var_16352, x = x_1829_cast_fp16)[name = tensor<string, []>("op_16353_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16354_cast_fp16 = mul(x = var_16353_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16354_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_267_cast_fp16 = add(x = var_16333_cast_fp16, y = var_16354_cast_fp16)[name = tensor<string, []>("q_patches_267_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16356_cast_fp16 = mul(x = var_16331_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16356_cast_fp16")];
            tensor<int32, [5]> var_16364 = const()[name = tensor<string, []>("op_16364"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1833_cast_fp16 = reshape(shape = var_16364, x = var_16331_cast_fp16)[name = tensor<string, []>("x_1833_cast_fp16")];
            tensor<int32, [2]> var_16366_split_sizes_0 = const()[name = tensor<string, []>("op_16366_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16366_axis_0 = const()[name = tensor<string, []>("op_16366_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16366_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16366_cast_fp16_1 = split(axis = var_16366_axis_0, split_sizes = var_16366_split_sizes_0, x = x_1833_cast_fp16)[name = tensor<string, []>("op_16366_cast_fp16")];
            tensor<int32, [1]> squeeze_534_axes_0 = const()[name = tensor<string, []>("squeeze_534_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_534_cast_fp16 = squeeze(axes = squeeze_534_axes_0, x = var_16366_cast_fp16_0)[name = tensor<string, []>("squeeze_534_cast_fp16")];
            tensor<int32, [1]> squeeze_535_axes_0 = const()[name = tensor<string, []>("squeeze_535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_535_cast_fp16 = squeeze(axes = squeeze_535_axes_0, x = var_16366_cast_fp16_1)[name = tensor<string, []>("squeeze_535_cast_fp16")];
            tensor<fp16, []> const_2187_promoted_to_fp16 = const()[name = tensor<string, []>("const_2187_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16369_cast_fp16 = mul(x = squeeze_535_cast_fp16, y = const_2187_promoted_to_fp16)[name = tensor<string, []>("op_16369_cast_fp16")];
            tensor<int32, []> x_1835_axis_0 = const()[name = tensor<string, []>("x_1835_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1835_cast_fp16 = stack(axis = x_1835_axis_0, values = (var_16369_cast_fp16, squeeze_534_cast_fp16))[name = tensor<string, []>("x_1835_cast_fp16")];
            tensor<int32, [4]> var_16375 = const()[name = tensor<string, []>("op_16375"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16376_cast_fp16 = reshape(shape = var_16375, x = x_1835_cast_fp16)[name = tensor<string, []>("op_16376_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16377_cast_fp16 = mul(x = var_16376_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16377_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_267_cast_fp16 = add(x = var_16356_cast_fp16, y = var_16377_cast_fp16)[name = tensor<string, []>("k_patches_267_cast_fp16")];
            tensor<bool, []> var_16380_interleave_0 = const()[name = tensor<string, []>("op_16380_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16380_cast_fp16 = concat(axis = var_20, interleave = var_16380_interleave_0, values = (var_16319_cast_fp16, q_patches_267_cast_fp16))[name = tensor<string, []>("op_16380_cast_fp16")];
            tensor<bool, []> var_16383_interleave_0 = const()[name = tensor<string, []>("op_16383_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16383_cast_fp16 = concat(axis = var_20, interleave = var_16383_interleave_0, values = (var_16323_cast_fp16, k_patches_267_cast_fp16))[name = tensor<string, []>("op_16383_cast_fp16")];
            tensor<fp16, []> var_16385_to_fp16 = const()[name = tensor<string, []>("op_16385_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1339_cast_fp16 = mul(x = var_16380_cast_fp16, y = var_16385_to_fp16)[name = tensor<string, []>("q_1339_cast_fp16")];
            tensor<bool, []> attn_533_transpose_x_1 = const()[name = tensor<string, []>("attn_533_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_533_transpose_y_1 = const()[name = tensor<string, []>("attn_533_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_533_cast_fp16 = matmul(transpose_x = attn_533_transpose_x_1, transpose_y = attn_533_transpose_y_1, x = q_1339_cast_fp16, y = var_16383_cast_fp16)[name = tensor<string, []>("attn_533_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_535_cast_fp16 = softmax(axis = var_21, x = attn_533_cast_fp16)[name = tensor<string, []>("attn_535_cast_fp16")];
            tensor<bool, []> var_16390_transpose_x_0 = const()[name = tensor<string, []>("op_16390_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16390_transpose_y_0 = const()[name = tensor<string, []>("op_16390_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16390_cast_fp16 = matmul(transpose_x = var_16390_transpose_x_0, transpose_y = var_16390_transpose_y_0, x = attn_535_cast_fp16, y = v_803_cast_fp16)[name = tensor<string, []>("op_16390_cast_fp16")];
            tensor<int32, [1]> var_16391_axes_0 = const()[name = tensor<string, []>("op_16391_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16391_cast_fp16 = squeeze(axes = var_16391_axes_0, x = var_16390_cast_fp16)[name = tensor<string, []>("op_16391_cast_fp16")];
            tensor<string, []> var_16397_pad_type_0 = const()[name = tensor<string, []>("op_16397_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16397_strides_0 = const()[name = tensor<string, []>("op_16397_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16397_pad_0 = const()[name = tensor<string, []>("op_16397_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16397_dilations_0 = const()[name = tensor<string, []>("op_16397_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16397_groups_0 = const()[name = tensor<string, []>("op_16397_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158196032)))];
            tensor<fp16, [1, 64, 1, 197]> var_16397_cast_fp16 = conv(dilations = var_16397_dilations_0, groups = var_16397_groups_0, pad = var_16397_pad_0, pad_type = var_16397_pad_type_0, strides = var_16397_strides_0, weight = model_blocks_11_attn_q_projs_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16397_cast_fp16")];
            tensor<int32, [1]> var_16398_axes_0 = const()[name = tensor<string, []>("op_16398_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16398_cast_fp16 = squeeze(axes = var_16398_axes_0, x = var_16397_cast_fp16)[name = tensor<string, []>("op_16398_cast_fp16")];
            tensor<int32, [3]> q_1341_perm_0 = const()[name = tensor<string, []>("q_1341_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16405_pad_type_0 = const()[name = tensor<string, []>("op_16405_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16405_strides_0 = const()[name = tensor<string, []>("op_16405_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16405_pad_0 = const()[name = tensor<string, []>("op_16405_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16405_dilations_0 = const()[name = tensor<string, []>("op_16405_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16405_groups_0 = const()[name = tensor<string, []>("op_16405_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158294400)))];
            tensor<fp16, [1, 64, 1, 197]> var_16405_cast_fp16 = conv(dilations = var_16405_dilations_0, groups = var_16405_groups_0, pad = var_16405_pad_0, pad_type = var_16405_pad_type_0, strides = var_16405_strides_0, weight = model_blocks_11_attn_k_projs_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16405_cast_fp16")];
            tensor<int32, [1]> var_16406_axes_0 = const()[name = tensor<string, []>("op_16406_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16406_cast_fp16 = squeeze(axes = var_16406_axes_0, x = var_16405_cast_fp16)[name = tensor<string, []>("op_16406_cast_fp16")];
            tensor<int32, [3]> k_805_perm_0 = const()[name = tensor<string, []>("k_805_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16413_pad_type_0 = const()[name = tensor<string, []>("op_16413_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16413_strides_0 = const()[name = tensor<string, []>("op_16413_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16413_pad_0 = const()[name = tensor<string, []>("op_16413_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16413_dilations_0 = const()[name = tensor<string, []>("op_16413_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16413_groups_0 = const()[name = tensor<string, []>("op_16413_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_2_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158392768)))];
            tensor<fp16, [1, 64, 1, 197]> var_16413_cast_fp16 = conv(dilations = var_16413_dilations_0, groups = var_16413_groups_0, pad = var_16413_pad_0, pad_type = var_16413_pad_type_0, strides = var_16413_strides_0, weight = model_blocks_11_attn_v_projs_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16413_cast_fp16")];
            tensor<int32, [1]> var_16414_axes_0 = const()[name = tensor<string, []>("op_16414_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16414_cast_fp16 = squeeze(axes = var_16414_axes_0, x = var_16413_cast_fp16)[name = tensor<string, []>("op_16414_cast_fp16")];
            tensor<int32, [3]> v_805_perm_0 = const()[name = tensor<string, []>("v_805_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158491136)))];
            tensor<fp16, [1, 197, 64]> q_1341_cast_fp16 = transpose(perm = q_1341_perm_0, x = var_16398_cast_fp16)[name = tensor<string, []>("transpose_35")];
            tensor<fp16, [1, 197, 64]> q_1343_cast_fp16 = add(x = q_1341_cast_fp16, y = model_blocks_11_attn_q_biases_2_to_fp16)[name = tensor<string, []>("q_1343_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_2_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_2_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158491328)))];
            tensor<fp16, [1, 197, 64]> v_805_cast_fp16 = transpose(perm = v_805_perm_0, x = var_16414_cast_fp16)[name = tensor<string, []>("transpose_33")];
            tensor<fp16, [1, 197, 64]> v_807_cast_fp16 = add(x = v_805_cast_fp16, y = model_blocks_11_attn_v_biases_2_to_fp16)[name = tensor<string, []>("v_807_cast_fp16")];
            tensor<int32, [1]> q_1345_axes_0 = const()[name = tensor<string, []>("q_1345_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1345_cast_fp16 = expand_dims(axes = q_1345_axes_0, x = q_1343_cast_fp16)[name = tensor<string, []>("q_1345_cast_fp16")];
            tensor<int32, [1]> k_807_axes_0 = const()[name = tensor<string, []>("k_807_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_805_cast_fp16 = transpose(perm = k_805_perm_0, x = var_16406_cast_fp16)[name = tensor<string, []>("transpose_34")];
            tensor<fp16, [1, 1, 197, 64]> k_807_cast_fp16 = expand_dims(axes = k_807_axes_0, x = k_805_cast_fp16)[name = tensor<string, []>("k_807_cast_fp16")];
            tensor<int32, [1]> v_809_axes_0 = const()[name = tensor<string, []>("v_809_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_809_cast_fp16 = expand_dims(axes = v_809_axes_0, x = v_807_cast_fp16)[name = tensor<string, []>("v_809_cast_fp16")];
            tensor<int32, [4]> var_16423_begin_0 = const()[name = tensor<string, []>("op_16423_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16423_end_0 = const()[name = tensor<string, []>("op_16423_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16423_end_mask_0 = const()[name = tensor<string, []>("op_16423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16423_cast_fp16 = slice_by_index(begin = var_16423_begin_0, end = var_16423_end_0, end_mask = var_16423_end_mask_0, x = q_1345_cast_fp16)[name = tensor<string, []>("op_16423_cast_fp16")];
            tensor<int32, [4]> var_16427_begin_0 = const()[name = tensor<string, []>("op_16427_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16427_end_0 = const()[name = tensor<string, []>("op_16427_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16427_end_mask_0 = const()[name = tensor<string, []>("op_16427_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16427_cast_fp16 = slice_by_index(begin = var_16427_begin_0, end = var_16427_end_0, end_mask = var_16427_end_mask_0, x = k_807_cast_fp16)[name = tensor<string, []>("op_16427_cast_fp16")];
            tensor<int32, [4]> var_16431_begin_0 = const()[name = tensor<string, []>("op_16431_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16431_end_0 = const()[name = tensor<string, []>("op_16431_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16431_end_mask_0 = const()[name = tensor<string, []>("op_16431_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16431_cast_fp16 = slice_by_index(begin = var_16431_begin_0, end = var_16431_end_0, end_mask = var_16431_end_mask_0, x = q_1345_cast_fp16)[name = tensor<string, []>("op_16431_cast_fp16")];
            tensor<int32, [4]> var_16435_begin_0 = const()[name = tensor<string, []>("op_16435_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16435_end_0 = const()[name = tensor<string, []>("op_16435_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16435_end_mask_0 = const()[name = tensor<string, []>("op_16435_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16435_cast_fp16 = slice_by_index(begin = var_16435_begin_0, end = var_16435_end_0, end_mask = var_16435_end_mask_0, x = k_807_cast_fp16)[name = tensor<string, []>("op_16435_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16437_cast_fp16 = mul(x = var_16431_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16437_cast_fp16")];
            tensor<int32, [5]> var_16445 = const()[name = tensor<string, []>("op_16445"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1839_cast_fp16 = reshape(shape = var_16445, x = var_16431_cast_fp16)[name = tensor<string, []>("x_1839_cast_fp16")];
            tensor<int32, [2]> var_16447_split_sizes_0 = const()[name = tensor<string, []>("op_16447_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16447_axis_0 = const()[name = tensor<string, []>("op_16447_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16447_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16447_cast_fp16_1 = split(axis = var_16447_axis_0, split_sizes = var_16447_split_sizes_0, x = x_1839_cast_fp16)[name = tensor<string, []>("op_16447_cast_fp16")];
            tensor<int32, [1]> squeeze_536_axes_0 = const()[name = tensor<string, []>("squeeze_536_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_536_cast_fp16 = squeeze(axes = squeeze_536_axes_0, x = var_16447_cast_fp16_0)[name = tensor<string, []>("squeeze_536_cast_fp16")];
            tensor<int32, [1]> squeeze_537_axes_0 = const()[name = tensor<string, []>("squeeze_537_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_537_cast_fp16 = squeeze(axes = squeeze_537_axes_0, x = var_16447_cast_fp16_1)[name = tensor<string, []>("squeeze_537_cast_fp16")];
            tensor<fp16, []> const_2195_promoted_to_fp16 = const()[name = tensor<string, []>("const_2195_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16450_cast_fp16 = mul(x = squeeze_537_cast_fp16, y = const_2195_promoted_to_fp16)[name = tensor<string, []>("op_16450_cast_fp16")];
            tensor<int32, []> x_1841_axis_0 = const()[name = tensor<string, []>("x_1841_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1841_cast_fp16 = stack(axis = x_1841_axis_0, values = (var_16450_cast_fp16, squeeze_536_cast_fp16))[name = tensor<string, []>("x_1841_cast_fp16")];
            tensor<int32, [4]> var_16456 = const()[name = tensor<string, []>("op_16456"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16457_cast_fp16 = reshape(shape = var_16456, x = x_1841_cast_fp16)[name = tensor<string, []>("op_16457_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16458_cast_fp16 = mul(x = var_16457_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16458_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_269_cast_fp16 = add(x = var_16437_cast_fp16, y = var_16458_cast_fp16)[name = tensor<string, []>("q_patches_269_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16460_cast_fp16 = mul(x = var_16435_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16460_cast_fp16")];
            tensor<int32, [5]> var_16468 = const()[name = tensor<string, []>("op_16468"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1845_cast_fp16 = reshape(shape = var_16468, x = var_16435_cast_fp16)[name = tensor<string, []>("x_1845_cast_fp16")];
            tensor<int32, [2]> var_16470_split_sizes_0 = const()[name = tensor<string, []>("op_16470_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16470_axis_0 = const()[name = tensor<string, []>("op_16470_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16470_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16470_cast_fp16_1 = split(axis = var_16470_axis_0, split_sizes = var_16470_split_sizes_0, x = x_1845_cast_fp16)[name = tensor<string, []>("op_16470_cast_fp16")];
            tensor<int32, [1]> squeeze_538_axes_0 = const()[name = tensor<string, []>("squeeze_538_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_538_cast_fp16 = squeeze(axes = squeeze_538_axes_0, x = var_16470_cast_fp16_0)[name = tensor<string, []>("squeeze_538_cast_fp16")];
            tensor<int32, [1]> squeeze_539_axes_0 = const()[name = tensor<string, []>("squeeze_539_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_539_cast_fp16 = squeeze(axes = squeeze_539_axes_0, x = var_16470_cast_fp16_1)[name = tensor<string, []>("squeeze_539_cast_fp16")];
            tensor<fp16, []> const_2203_promoted_to_fp16 = const()[name = tensor<string, []>("const_2203_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16473_cast_fp16 = mul(x = squeeze_539_cast_fp16, y = const_2203_promoted_to_fp16)[name = tensor<string, []>("op_16473_cast_fp16")];
            tensor<int32, []> x_1847_axis_0 = const()[name = tensor<string, []>("x_1847_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1847_cast_fp16 = stack(axis = x_1847_axis_0, values = (var_16473_cast_fp16, squeeze_538_cast_fp16))[name = tensor<string, []>("x_1847_cast_fp16")];
            tensor<int32, [4]> var_16479 = const()[name = tensor<string, []>("op_16479"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16480_cast_fp16 = reshape(shape = var_16479, x = x_1847_cast_fp16)[name = tensor<string, []>("op_16480_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16481_cast_fp16 = mul(x = var_16480_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16481_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_269_cast_fp16 = add(x = var_16460_cast_fp16, y = var_16481_cast_fp16)[name = tensor<string, []>("k_patches_269_cast_fp16")];
            tensor<bool, []> var_16484_interleave_0 = const()[name = tensor<string, []>("op_16484_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16484_cast_fp16 = concat(axis = var_20, interleave = var_16484_interleave_0, values = (var_16423_cast_fp16, q_patches_269_cast_fp16))[name = tensor<string, []>("op_16484_cast_fp16")];
            tensor<bool, []> var_16487_interleave_0 = const()[name = tensor<string, []>("op_16487_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16487_cast_fp16 = concat(axis = var_20, interleave = var_16487_interleave_0, values = (var_16427_cast_fp16, k_patches_269_cast_fp16))[name = tensor<string, []>("op_16487_cast_fp16")];
            tensor<fp16, []> var_16489_to_fp16 = const()[name = tensor<string, []>("op_16489_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1349_cast_fp16 = mul(x = var_16484_cast_fp16, y = var_16489_to_fp16)[name = tensor<string, []>("q_1349_cast_fp16")];
            tensor<bool, []> attn_537_transpose_x_1 = const()[name = tensor<string, []>("attn_537_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_537_transpose_y_1 = const()[name = tensor<string, []>("attn_537_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_537_cast_fp16 = matmul(transpose_x = attn_537_transpose_x_1, transpose_y = attn_537_transpose_y_1, x = q_1349_cast_fp16, y = var_16487_cast_fp16)[name = tensor<string, []>("attn_537_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_539_cast_fp16 = softmax(axis = var_21, x = attn_537_cast_fp16)[name = tensor<string, []>("attn_539_cast_fp16")];
            tensor<bool, []> var_16494_transpose_x_0 = const()[name = tensor<string, []>("op_16494_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16494_transpose_y_0 = const()[name = tensor<string, []>("op_16494_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16494_cast_fp16 = matmul(transpose_x = var_16494_transpose_x_0, transpose_y = var_16494_transpose_y_0, x = attn_539_cast_fp16, y = v_809_cast_fp16)[name = tensor<string, []>("op_16494_cast_fp16")];
            tensor<int32, [1]> var_16495_axes_0 = const()[name = tensor<string, []>("op_16495_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16495_cast_fp16 = squeeze(axes = var_16495_axes_0, x = var_16494_cast_fp16)[name = tensor<string, []>("op_16495_cast_fp16")];
            tensor<string, []> var_16501_pad_type_0 = const()[name = tensor<string, []>("op_16501_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16501_strides_0 = const()[name = tensor<string, []>("op_16501_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16501_pad_0 = const()[name = tensor<string, []>("op_16501_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16501_dilations_0 = const()[name = tensor<string, []>("op_16501_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16501_groups_0 = const()[name = tensor<string, []>("op_16501_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158491520)))];
            tensor<fp16, [1, 64, 1, 197]> var_16501_cast_fp16 = conv(dilations = var_16501_dilations_0, groups = var_16501_groups_0, pad = var_16501_pad_0, pad_type = var_16501_pad_type_0, strides = var_16501_strides_0, weight = model_blocks_11_attn_q_projs_3_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16501_cast_fp16")];
            tensor<int32, [1]> var_16502_axes_0 = const()[name = tensor<string, []>("op_16502_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16502_cast_fp16 = squeeze(axes = var_16502_axes_0, x = var_16501_cast_fp16)[name = tensor<string, []>("op_16502_cast_fp16")];
            tensor<int32, [3]> q_1351_perm_0 = const()[name = tensor<string, []>("q_1351_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16509_pad_type_0 = const()[name = tensor<string, []>("op_16509_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16509_strides_0 = const()[name = tensor<string, []>("op_16509_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16509_pad_0 = const()[name = tensor<string, []>("op_16509_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16509_dilations_0 = const()[name = tensor<string, []>("op_16509_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16509_groups_0 = const()[name = tensor<string, []>("op_16509_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158589888)))];
            tensor<fp16, [1, 64, 1, 197]> var_16509_cast_fp16 = conv(dilations = var_16509_dilations_0, groups = var_16509_groups_0, pad = var_16509_pad_0, pad_type = var_16509_pad_type_0, strides = var_16509_strides_0, weight = model_blocks_11_attn_k_projs_3_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16509_cast_fp16")];
            tensor<int32, [1]> var_16510_axes_0 = const()[name = tensor<string, []>("op_16510_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16510_cast_fp16 = squeeze(axes = var_16510_axes_0, x = var_16509_cast_fp16)[name = tensor<string, []>("op_16510_cast_fp16")];
            tensor<int32, [3]> k_811_perm_0 = const()[name = tensor<string, []>("k_811_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16517_pad_type_0 = const()[name = tensor<string, []>("op_16517_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16517_strides_0 = const()[name = tensor<string, []>("op_16517_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16517_pad_0 = const()[name = tensor<string, []>("op_16517_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16517_dilations_0 = const()[name = tensor<string, []>("op_16517_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16517_groups_0 = const()[name = tensor<string, []>("op_16517_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_3_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158688256)))];
            tensor<fp16, [1, 64, 1, 197]> var_16517_cast_fp16 = conv(dilations = var_16517_dilations_0, groups = var_16517_groups_0, pad = var_16517_pad_0, pad_type = var_16517_pad_type_0, strides = var_16517_strides_0, weight = model_blocks_11_attn_v_projs_3_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16517_cast_fp16")];
            tensor<int32, [1]> var_16518_axes_0 = const()[name = tensor<string, []>("op_16518_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16518_cast_fp16 = squeeze(axes = var_16518_axes_0, x = var_16517_cast_fp16)[name = tensor<string, []>("op_16518_cast_fp16")];
            tensor<int32, [3]> v_811_perm_0 = const()[name = tensor<string, []>("v_811_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158786624)))];
            tensor<fp16, [1, 197, 64]> q_1351_cast_fp16 = transpose(perm = q_1351_perm_0, x = var_16502_cast_fp16)[name = tensor<string, []>("transpose_32")];
            tensor<fp16, [1, 197, 64]> q_1353_cast_fp16 = add(x = q_1351_cast_fp16, y = model_blocks_11_attn_q_biases_3_to_fp16)[name = tensor<string, []>("q_1353_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_3_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_3_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158786816)))];
            tensor<fp16, [1, 197, 64]> v_811_cast_fp16 = transpose(perm = v_811_perm_0, x = var_16518_cast_fp16)[name = tensor<string, []>("transpose_30")];
            tensor<fp16, [1, 197, 64]> v_813_cast_fp16 = add(x = v_811_cast_fp16, y = model_blocks_11_attn_v_biases_3_to_fp16)[name = tensor<string, []>("v_813_cast_fp16")];
            tensor<int32, [1]> q_1355_axes_0 = const()[name = tensor<string, []>("q_1355_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1355_cast_fp16 = expand_dims(axes = q_1355_axes_0, x = q_1353_cast_fp16)[name = tensor<string, []>("q_1355_cast_fp16")];
            tensor<int32, [1]> k_813_axes_0 = const()[name = tensor<string, []>("k_813_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_811_cast_fp16 = transpose(perm = k_811_perm_0, x = var_16510_cast_fp16)[name = tensor<string, []>("transpose_31")];
            tensor<fp16, [1, 1, 197, 64]> k_813_cast_fp16 = expand_dims(axes = k_813_axes_0, x = k_811_cast_fp16)[name = tensor<string, []>("k_813_cast_fp16")];
            tensor<int32, [1]> v_815_axes_0 = const()[name = tensor<string, []>("v_815_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_815_cast_fp16 = expand_dims(axes = v_815_axes_0, x = v_813_cast_fp16)[name = tensor<string, []>("v_815_cast_fp16")];
            tensor<int32, [4]> var_16527_begin_0 = const()[name = tensor<string, []>("op_16527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16527_end_0 = const()[name = tensor<string, []>("op_16527_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16527_end_mask_0 = const()[name = tensor<string, []>("op_16527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16527_cast_fp16 = slice_by_index(begin = var_16527_begin_0, end = var_16527_end_0, end_mask = var_16527_end_mask_0, x = q_1355_cast_fp16)[name = tensor<string, []>("op_16527_cast_fp16")];
            tensor<int32, [4]> var_16531_begin_0 = const()[name = tensor<string, []>("op_16531_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16531_end_0 = const()[name = tensor<string, []>("op_16531_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16531_end_mask_0 = const()[name = tensor<string, []>("op_16531_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16531_cast_fp16 = slice_by_index(begin = var_16531_begin_0, end = var_16531_end_0, end_mask = var_16531_end_mask_0, x = k_813_cast_fp16)[name = tensor<string, []>("op_16531_cast_fp16")];
            tensor<int32, [4]> var_16535_begin_0 = const()[name = tensor<string, []>("op_16535_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16535_end_0 = const()[name = tensor<string, []>("op_16535_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16535_end_mask_0 = const()[name = tensor<string, []>("op_16535_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16535_cast_fp16 = slice_by_index(begin = var_16535_begin_0, end = var_16535_end_0, end_mask = var_16535_end_mask_0, x = q_1355_cast_fp16)[name = tensor<string, []>("op_16535_cast_fp16")];
            tensor<int32, [4]> var_16539_begin_0 = const()[name = tensor<string, []>("op_16539_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16539_end_0 = const()[name = tensor<string, []>("op_16539_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16539_end_mask_0 = const()[name = tensor<string, []>("op_16539_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16539_cast_fp16 = slice_by_index(begin = var_16539_begin_0, end = var_16539_end_0, end_mask = var_16539_end_mask_0, x = k_813_cast_fp16)[name = tensor<string, []>("op_16539_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16541_cast_fp16 = mul(x = var_16535_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16541_cast_fp16")];
            tensor<int32, [5]> var_16549 = const()[name = tensor<string, []>("op_16549"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1851_cast_fp16 = reshape(shape = var_16549, x = var_16535_cast_fp16)[name = tensor<string, []>("x_1851_cast_fp16")];
            tensor<int32, [2]> var_16551_split_sizes_0 = const()[name = tensor<string, []>("op_16551_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16551_axis_0 = const()[name = tensor<string, []>("op_16551_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16551_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16551_cast_fp16_1 = split(axis = var_16551_axis_0, split_sizes = var_16551_split_sizes_0, x = x_1851_cast_fp16)[name = tensor<string, []>("op_16551_cast_fp16")];
            tensor<int32, [1]> squeeze_540_axes_0 = const()[name = tensor<string, []>("squeeze_540_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_540_cast_fp16 = squeeze(axes = squeeze_540_axes_0, x = var_16551_cast_fp16_0)[name = tensor<string, []>("squeeze_540_cast_fp16")];
            tensor<int32, [1]> squeeze_541_axes_0 = const()[name = tensor<string, []>("squeeze_541_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_541_cast_fp16 = squeeze(axes = squeeze_541_axes_0, x = var_16551_cast_fp16_1)[name = tensor<string, []>("squeeze_541_cast_fp16")];
            tensor<fp16, []> const_2211_promoted_to_fp16 = const()[name = tensor<string, []>("const_2211_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16554_cast_fp16 = mul(x = squeeze_541_cast_fp16, y = const_2211_promoted_to_fp16)[name = tensor<string, []>("op_16554_cast_fp16")];
            tensor<int32, []> x_1853_axis_0 = const()[name = tensor<string, []>("x_1853_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1853_cast_fp16 = stack(axis = x_1853_axis_0, values = (var_16554_cast_fp16, squeeze_540_cast_fp16))[name = tensor<string, []>("x_1853_cast_fp16")];
            tensor<int32, [4]> var_16560 = const()[name = tensor<string, []>("op_16560"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16561_cast_fp16 = reshape(shape = var_16560, x = x_1853_cast_fp16)[name = tensor<string, []>("op_16561_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16562_cast_fp16 = mul(x = var_16561_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16562_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_271_cast_fp16 = add(x = var_16541_cast_fp16, y = var_16562_cast_fp16)[name = tensor<string, []>("q_patches_271_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16564_cast_fp16 = mul(x = var_16539_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16564_cast_fp16")];
            tensor<int32, [5]> var_16572 = const()[name = tensor<string, []>("op_16572"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1857_cast_fp16 = reshape(shape = var_16572, x = var_16539_cast_fp16)[name = tensor<string, []>("x_1857_cast_fp16")];
            tensor<int32, [2]> var_16574_split_sizes_0 = const()[name = tensor<string, []>("op_16574_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16574_axis_0 = const()[name = tensor<string, []>("op_16574_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16574_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16574_cast_fp16_1 = split(axis = var_16574_axis_0, split_sizes = var_16574_split_sizes_0, x = x_1857_cast_fp16)[name = tensor<string, []>("op_16574_cast_fp16")];
            tensor<int32, [1]> squeeze_542_axes_0 = const()[name = tensor<string, []>("squeeze_542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_542_cast_fp16 = squeeze(axes = squeeze_542_axes_0, x = var_16574_cast_fp16_0)[name = tensor<string, []>("squeeze_542_cast_fp16")];
            tensor<int32, [1]> squeeze_543_axes_0 = const()[name = tensor<string, []>("squeeze_543_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_543_cast_fp16 = squeeze(axes = squeeze_543_axes_0, x = var_16574_cast_fp16_1)[name = tensor<string, []>("squeeze_543_cast_fp16")];
            tensor<fp16, []> const_2219_promoted_to_fp16 = const()[name = tensor<string, []>("const_2219_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16577_cast_fp16 = mul(x = squeeze_543_cast_fp16, y = const_2219_promoted_to_fp16)[name = tensor<string, []>("op_16577_cast_fp16")];
            tensor<int32, []> x_1859_axis_0 = const()[name = tensor<string, []>("x_1859_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1859_cast_fp16 = stack(axis = x_1859_axis_0, values = (var_16577_cast_fp16, squeeze_542_cast_fp16))[name = tensor<string, []>("x_1859_cast_fp16")];
            tensor<int32, [4]> var_16583 = const()[name = tensor<string, []>("op_16583"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16584_cast_fp16 = reshape(shape = var_16583, x = x_1859_cast_fp16)[name = tensor<string, []>("op_16584_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16585_cast_fp16 = mul(x = var_16584_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16585_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_271_cast_fp16 = add(x = var_16564_cast_fp16, y = var_16585_cast_fp16)[name = tensor<string, []>("k_patches_271_cast_fp16")];
            tensor<bool, []> var_16588_interleave_0 = const()[name = tensor<string, []>("op_16588_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16588_cast_fp16 = concat(axis = var_20, interleave = var_16588_interleave_0, values = (var_16527_cast_fp16, q_patches_271_cast_fp16))[name = tensor<string, []>("op_16588_cast_fp16")];
            tensor<bool, []> var_16591_interleave_0 = const()[name = tensor<string, []>("op_16591_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16591_cast_fp16 = concat(axis = var_20, interleave = var_16591_interleave_0, values = (var_16531_cast_fp16, k_patches_271_cast_fp16))[name = tensor<string, []>("op_16591_cast_fp16")];
            tensor<fp16, []> var_16593_to_fp16 = const()[name = tensor<string, []>("op_16593_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1359_cast_fp16 = mul(x = var_16588_cast_fp16, y = var_16593_to_fp16)[name = tensor<string, []>("q_1359_cast_fp16")];
            tensor<bool, []> attn_541_transpose_x_1 = const()[name = tensor<string, []>("attn_541_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_541_transpose_y_1 = const()[name = tensor<string, []>("attn_541_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_541_cast_fp16 = matmul(transpose_x = attn_541_transpose_x_1, transpose_y = attn_541_transpose_y_1, x = q_1359_cast_fp16, y = var_16591_cast_fp16)[name = tensor<string, []>("attn_541_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_543_cast_fp16 = softmax(axis = var_21, x = attn_541_cast_fp16)[name = tensor<string, []>("attn_543_cast_fp16")];
            tensor<bool, []> var_16598_transpose_x_0 = const()[name = tensor<string, []>("op_16598_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16598_transpose_y_0 = const()[name = tensor<string, []>("op_16598_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16598_cast_fp16 = matmul(transpose_x = var_16598_transpose_x_0, transpose_y = var_16598_transpose_y_0, x = attn_543_cast_fp16, y = v_815_cast_fp16)[name = tensor<string, []>("op_16598_cast_fp16")];
            tensor<int32, [1]> var_16599_axes_0 = const()[name = tensor<string, []>("op_16599_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16599_cast_fp16 = squeeze(axes = var_16599_axes_0, x = var_16598_cast_fp16)[name = tensor<string, []>("op_16599_cast_fp16")];
            tensor<string, []> var_16605_pad_type_0 = const()[name = tensor<string, []>("op_16605_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16605_strides_0 = const()[name = tensor<string, []>("op_16605_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16605_pad_0 = const()[name = tensor<string, []>("op_16605_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16605_dilations_0 = const()[name = tensor<string, []>("op_16605_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16605_groups_0 = const()[name = tensor<string, []>("op_16605_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158787008)))];
            tensor<fp16, [1, 64, 1, 197]> var_16605_cast_fp16 = conv(dilations = var_16605_dilations_0, groups = var_16605_groups_0, pad = var_16605_pad_0, pad_type = var_16605_pad_type_0, strides = var_16605_strides_0, weight = model_blocks_11_attn_q_projs_4_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16605_cast_fp16")];
            tensor<int32, [1]> var_16606_axes_0 = const()[name = tensor<string, []>("op_16606_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16606_cast_fp16 = squeeze(axes = var_16606_axes_0, x = var_16605_cast_fp16)[name = tensor<string, []>("op_16606_cast_fp16")];
            tensor<int32, [3]> q_1361_perm_0 = const()[name = tensor<string, []>("q_1361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16613_pad_type_0 = const()[name = tensor<string, []>("op_16613_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16613_strides_0 = const()[name = tensor<string, []>("op_16613_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16613_pad_0 = const()[name = tensor<string, []>("op_16613_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16613_dilations_0 = const()[name = tensor<string, []>("op_16613_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16613_groups_0 = const()[name = tensor<string, []>("op_16613_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158885376)))];
            tensor<fp16, [1, 64, 1, 197]> var_16613_cast_fp16 = conv(dilations = var_16613_dilations_0, groups = var_16613_groups_0, pad = var_16613_pad_0, pad_type = var_16613_pad_type_0, strides = var_16613_strides_0, weight = model_blocks_11_attn_k_projs_4_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16613_cast_fp16")];
            tensor<int32, [1]> var_16614_axes_0 = const()[name = tensor<string, []>("op_16614_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16614_cast_fp16 = squeeze(axes = var_16614_axes_0, x = var_16613_cast_fp16)[name = tensor<string, []>("op_16614_cast_fp16")];
            tensor<int32, [3]> k_817_perm_0 = const()[name = tensor<string, []>("k_817_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16621_pad_type_0 = const()[name = tensor<string, []>("op_16621_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16621_strides_0 = const()[name = tensor<string, []>("op_16621_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16621_pad_0 = const()[name = tensor<string, []>("op_16621_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16621_dilations_0 = const()[name = tensor<string, []>("op_16621_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16621_groups_0 = const()[name = tensor<string, []>("op_16621_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_4_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_4_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158983744)))];
            tensor<fp16, [1, 64, 1, 197]> var_16621_cast_fp16 = conv(dilations = var_16621_dilations_0, groups = var_16621_groups_0, pad = var_16621_pad_0, pad_type = var_16621_pad_type_0, strides = var_16621_strides_0, weight = model_blocks_11_attn_v_projs_4_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16621_cast_fp16")];
            tensor<int32, [1]> var_16622_axes_0 = const()[name = tensor<string, []>("op_16622_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16622_cast_fp16 = squeeze(axes = var_16622_axes_0, x = var_16621_cast_fp16)[name = tensor<string, []>("op_16622_cast_fp16")];
            tensor<int32, [3]> v_817_perm_0 = const()[name = tensor<string, []>("v_817_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159082112)))];
            tensor<fp16, [1, 197, 64]> q_1361_cast_fp16 = transpose(perm = q_1361_perm_0, x = var_16606_cast_fp16)[name = tensor<string, []>("transpose_29")];
            tensor<fp16, [1, 197, 64]> q_1363_cast_fp16 = add(x = q_1361_cast_fp16, y = model_blocks_11_attn_q_biases_4_to_fp16)[name = tensor<string, []>("q_1363_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_4_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_4_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159082304)))];
            tensor<fp16, [1, 197, 64]> v_817_cast_fp16 = transpose(perm = v_817_perm_0, x = var_16622_cast_fp16)[name = tensor<string, []>("transpose_27")];
            tensor<fp16, [1, 197, 64]> v_819_cast_fp16 = add(x = v_817_cast_fp16, y = model_blocks_11_attn_v_biases_4_to_fp16)[name = tensor<string, []>("v_819_cast_fp16")];
            tensor<int32, [1]> q_1365_axes_0 = const()[name = tensor<string, []>("q_1365_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1365_cast_fp16 = expand_dims(axes = q_1365_axes_0, x = q_1363_cast_fp16)[name = tensor<string, []>("q_1365_cast_fp16")];
            tensor<int32, [1]> k_819_axes_0 = const()[name = tensor<string, []>("k_819_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_817_cast_fp16 = transpose(perm = k_817_perm_0, x = var_16614_cast_fp16)[name = tensor<string, []>("transpose_28")];
            tensor<fp16, [1, 1, 197, 64]> k_819_cast_fp16 = expand_dims(axes = k_819_axes_0, x = k_817_cast_fp16)[name = tensor<string, []>("k_819_cast_fp16")];
            tensor<int32, [1]> v_821_axes_0 = const()[name = tensor<string, []>("v_821_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_821_cast_fp16 = expand_dims(axes = v_821_axes_0, x = v_819_cast_fp16)[name = tensor<string, []>("v_821_cast_fp16")];
            tensor<int32, [4]> var_16631_begin_0 = const()[name = tensor<string, []>("op_16631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16631_end_0 = const()[name = tensor<string, []>("op_16631_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16631_end_mask_0 = const()[name = tensor<string, []>("op_16631_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16631_cast_fp16 = slice_by_index(begin = var_16631_begin_0, end = var_16631_end_0, end_mask = var_16631_end_mask_0, x = q_1365_cast_fp16)[name = tensor<string, []>("op_16631_cast_fp16")];
            tensor<int32, [4]> var_16635_begin_0 = const()[name = tensor<string, []>("op_16635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16635_end_0 = const()[name = tensor<string, []>("op_16635_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16635_end_mask_0 = const()[name = tensor<string, []>("op_16635_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16635_cast_fp16 = slice_by_index(begin = var_16635_begin_0, end = var_16635_end_0, end_mask = var_16635_end_mask_0, x = k_819_cast_fp16)[name = tensor<string, []>("op_16635_cast_fp16")];
            tensor<int32, [4]> var_16639_begin_0 = const()[name = tensor<string, []>("op_16639_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16639_end_0 = const()[name = tensor<string, []>("op_16639_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16639_end_mask_0 = const()[name = tensor<string, []>("op_16639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16639_cast_fp16 = slice_by_index(begin = var_16639_begin_0, end = var_16639_end_0, end_mask = var_16639_end_mask_0, x = q_1365_cast_fp16)[name = tensor<string, []>("op_16639_cast_fp16")];
            tensor<int32, [4]> var_16643_begin_0 = const()[name = tensor<string, []>("op_16643_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16643_end_0 = const()[name = tensor<string, []>("op_16643_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16643_end_mask_0 = const()[name = tensor<string, []>("op_16643_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16643_cast_fp16 = slice_by_index(begin = var_16643_begin_0, end = var_16643_end_0, end_mask = var_16643_end_mask_0, x = k_819_cast_fp16)[name = tensor<string, []>("op_16643_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16645_cast_fp16 = mul(x = var_16639_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16645_cast_fp16")];
            tensor<int32, [5]> var_16653 = const()[name = tensor<string, []>("op_16653"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1863_cast_fp16 = reshape(shape = var_16653, x = var_16639_cast_fp16)[name = tensor<string, []>("x_1863_cast_fp16")];
            tensor<int32, [2]> var_16655_split_sizes_0 = const()[name = tensor<string, []>("op_16655_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16655_axis_0 = const()[name = tensor<string, []>("op_16655_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16655_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16655_cast_fp16_1 = split(axis = var_16655_axis_0, split_sizes = var_16655_split_sizes_0, x = x_1863_cast_fp16)[name = tensor<string, []>("op_16655_cast_fp16")];
            tensor<int32, [1]> squeeze_544_axes_0 = const()[name = tensor<string, []>("squeeze_544_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_544_cast_fp16 = squeeze(axes = squeeze_544_axes_0, x = var_16655_cast_fp16_0)[name = tensor<string, []>("squeeze_544_cast_fp16")];
            tensor<int32, [1]> squeeze_545_axes_0 = const()[name = tensor<string, []>("squeeze_545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_545_cast_fp16 = squeeze(axes = squeeze_545_axes_0, x = var_16655_cast_fp16_1)[name = tensor<string, []>("squeeze_545_cast_fp16")];
            tensor<fp16, []> const_2227_promoted_to_fp16 = const()[name = tensor<string, []>("const_2227_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16658_cast_fp16 = mul(x = squeeze_545_cast_fp16, y = const_2227_promoted_to_fp16)[name = tensor<string, []>("op_16658_cast_fp16")];
            tensor<int32, []> x_1865_axis_0 = const()[name = tensor<string, []>("x_1865_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1865_cast_fp16 = stack(axis = x_1865_axis_0, values = (var_16658_cast_fp16, squeeze_544_cast_fp16))[name = tensor<string, []>("x_1865_cast_fp16")];
            tensor<int32, [4]> var_16664 = const()[name = tensor<string, []>("op_16664"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16665_cast_fp16 = reshape(shape = var_16664, x = x_1865_cast_fp16)[name = tensor<string, []>("op_16665_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16666_cast_fp16 = mul(x = var_16665_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16666_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_273_cast_fp16 = add(x = var_16645_cast_fp16, y = var_16666_cast_fp16)[name = tensor<string, []>("q_patches_273_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16668_cast_fp16 = mul(x = var_16643_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16668_cast_fp16")];
            tensor<int32, [5]> var_16676 = const()[name = tensor<string, []>("op_16676"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1869_cast_fp16 = reshape(shape = var_16676, x = var_16643_cast_fp16)[name = tensor<string, []>("x_1869_cast_fp16")];
            tensor<int32, [2]> var_16678_split_sizes_0 = const()[name = tensor<string, []>("op_16678_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16678_axis_0 = const()[name = tensor<string, []>("op_16678_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16678_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16678_cast_fp16_1 = split(axis = var_16678_axis_0, split_sizes = var_16678_split_sizes_0, x = x_1869_cast_fp16)[name = tensor<string, []>("op_16678_cast_fp16")];
            tensor<int32, [1]> squeeze_546_axes_0 = const()[name = tensor<string, []>("squeeze_546_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_546_cast_fp16 = squeeze(axes = squeeze_546_axes_0, x = var_16678_cast_fp16_0)[name = tensor<string, []>("squeeze_546_cast_fp16")];
            tensor<int32, [1]> squeeze_547_axes_0 = const()[name = tensor<string, []>("squeeze_547_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_547_cast_fp16 = squeeze(axes = squeeze_547_axes_0, x = var_16678_cast_fp16_1)[name = tensor<string, []>("squeeze_547_cast_fp16")];
            tensor<fp16, []> const_2235_promoted_to_fp16 = const()[name = tensor<string, []>("const_2235_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16681_cast_fp16 = mul(x = squeeze_547_cast_fp16, y = const_2235_promoted_to_fp16)[name = tensor<string, []>("op_16681_cast_fp16")];
            tensor<int32, []> x_1871_axis_0 = const()[name = tensor<string, []>("x_1871_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1871_cast_fp16 = stack(axis = x_1871_axis_0, values = (var_16681_cast_fp16, squeeze_546_cast_fp16))[name = tensor<string, []>("x_1871_cast_fp16")];
            tensor<int32, [4]> var_16687 = const()[name = tensor<string, []>("op_16687"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16688_cast_fp16 = reshape(shape = var_16687, x = x_1871_cast_fp16)[name = tensor<string, []>("op_16688_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16689_cast_fp16 = mul(x = var_16688_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16689_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_273_cast_fp16 = add(x = var_16668_cast_fp16, y = var_16689_cast_fp16)[name = tensor<string, []>("k_patches_273_cast_fp16")];
            tensor<bool, []> var_16692_interleave_0 = const()[name = tensor<string, []>("op_16692_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16692_cast_fp16 = concat(axis = var_20, interleave = var_16692_interleave_0, values = (var_16631_cast_fp16, q_patches_273_cast_fp16))[name = tensor<string, []>("op_16692_cast_fp16")];
            tensor<bool, []> var_16695_interleave_0 = const()[name = tensor<string, []>("op_16695_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16695_cast_fp16 = concat(axis = var_20, interleave = var_16695_interleave_0, values = (var_16635_cast_fp16, k_patches_273_cast_fp16))[name = tensor<string, []>("op_16695_cast_fp16")];
            tensor<fp16, []> var_16697_to_fp16 = const()[name = tensor<string, []>("op_16697_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1369_cast_fp16 = mul(x = var_16692_cast_fp16, y = var_16697_to_fp16)[name = tensor<string, []>("q_1369_cast_fp16")];
            tensor<bool, []> attn_545_transpose_x_1 = const()[name = tensor<string, []>("attn_545_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_545_transpose_y_1 = const()[name = tensor<string, []>("attn_545_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_545_cast_fp16 = matmul(transpose_x = attn_545_transpose_x_1, transpose_y = attn_545_transpose_y_1, x = q_1369_cast_fp16, y = var_16695_cast_fp16)[name = tensor<string, []>("attn_545_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_547_cast_fp16 = softmax(axis = var_21, x = attn_545_cast_fp16)[name = tensor<string, []>("attn_547_cast_fp16")];
            tensor<bool, []> var_16702_transpose_x_0 = const()[name = tensor<string, []>("op_16702_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16702_transpose_y_0 = const()[name = tensor<string, []>("op_16702_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16702_cast_fp16 = matmul(transpose_x = var_16702_transpose_x_0, transpose_y = var_16702_transpose_y_0, x = attn_547_cast_fp16, y = v_821_cast_fp16)[name = tensor<string, []>("op_16702_cast_fp16")];
            tensor<int32, [1]> var_16703_axes_0 = const()[name = tensor<string, []>("op_16703_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16703_cast_fp16 = squeeze(axes = var_16703_axes_0, x = var_16702_cast_fp16)[name = tensor<string, []>("op_16703_cast_fp16")];
            tensor<string, []> var_16709_pad_type_0 = const()[name = tensor<string, []>("op_16709_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16709_strides_0 = const()[name = tensor<string, []>("op_16709_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16709_pad_0 = const()[name = tensor<string, []>("op_16709_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16709_dilations_0 = const()[name = tensor<string, []>("op_16709_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16709_groups_0 = const()[name = tensor<string, []>("op_16709_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159082496)))];
            tensor<fp16, [1, 64, 1, 197]> var_16709_cast_fp16 = conv(dilations = var_16709_dilations_0, groups = var_16709_groups_0, pad = var_16709_pad_0, pad_type = var_16709_pad_type_0, strides = var_16709_strides_0, weight = model_blocks_11_attn_q_projs_5_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16709_cast_fp16")];
            tensor<int32, [1]> var_16710_axes_0 = const()[name = tensor<string, []>("op_16710_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16710_cast_fp16 = squeeze(axes = var_16710_axes_0, x = var_16709_cast_fp16)[name = tensor<string, []>("op_16710_cast_fp16")];
            tensor<int32, [3]> q_1371_perm_0 = const()[name = tensor<string, []>("q_1371_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16717_pad_type_0 = const()[name = tensor<string, []>("op_16717_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16717_strides_0 = const()[name = tensor<string, []>("op_16717_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16717_pad_0 = const()[name = tensor<string, []>("op_16717_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16717_dilations_0 = const()[name = tensor<string, []>("op_16717_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16717_groups_0 = const()[name = tensor<string, []>("op_16717_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159180864)))];
            tensor<fp16, [1, 64, 1, 197]> var_16717_cast_fp16 = conv(dilations = var_16717_dilations_0, groups = var_16717_groups_0, pad = var_16717_pad_0, pad_type = var_16717_pad_type_0, strides = var_16717_strides_0, weight = model_blocks_11_attn_k_projs_5_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16717_cast_fp16")];
            tensor<int32, [1]> var_16718_axes_0 = const()[name = tensor<string, []>("op_16718_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16718_cast_fp16 = squeeze(axes = var_16718_axes_0, x = var_16717_cast_fp16)[name = tensor<string, []>("op_16718_cast_fp16")];
            tensor<int32, [3]> k_823_perm_0 = const()[name = tensor<string, []>("k_823_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16725_pad_type_0 = const()[name = tensor<string, []>("op_16725_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16725_strides_0 = const()[name = tensor<string, []>("op_16725_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16725_pad_0 = const()[name = tensor<string, []>("op_16725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16725_dilations_0 = const()[name = tensor<string, []>("op_16725_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16725_groups_0 = const()[name = tensor<string, []>("op_16725_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_5_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_5_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159279232)))];
            tensor<fp16, [1, 64, 1, 197]> var_16725_cast_fp16 = conv(dilations = var_16725_dilations_0, groups = var_16725_groups_0, pad = var_16725_pad_0, pad_type = var_16725_pad_type_0, strides = var_16725_strides_0, weight = model_blocks_11_attn_v_projs_5_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16725_cast_fp16")];
            tensor<int32, [1]> var_16726_axes_0 = const()[name = tensor<string, []>("op_16726_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16726_cast_fp16 = squeeze(axes = var_16726_axes_0, x = var_16725_cast_fp16)[name = tensor<string, []>("op_16726_cast_fp16")];
            tensor<int32, [3]> v_823_perm_0 = const()[name = tensor<string, []>("v_823_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159377600)))];
            tensor<fp16, [1, 197, 64]> q_1371_cast_fp16 = transpose(perm = q_1371_perm_0, x = var_16710_cast_fp16)[name = tensor<string, []>("transpose_26")];
            tensor<fp16, [1, 197, 64]> q_1373_cast_fp16 = add(x = q_1371_cast_fp16, y = model_blocks_11_attn_q_biases_5_to_fp16)[name = tensor<string, []>("q_1373_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_5_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_5_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159377792)))];
            tensor<fp16, [1, 197, 64]> v_823_cast_fp16 = transpose(perm = v_823_perm_0, x = var_16726_cast_fp16)[name = tensor<string, []>("transpose_24")];
            tensor<fp16, [1, 197, 64]> v_825_cast_fp16 = add(x = v_823_cast_fp16, y = model_blocks_11_attn_v_biases_5_to_fp16)[name = tensor<string, []>("v_825_cast_fp16")];
            tensor<int32, [1]> q_1375_axes_0 = const()[name = tensor<string, []>("q_1375_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1375_cast_fp16 = expand_dims(axes = q_1375_axes_0, x = q_1373_cast_fp16)[name = tensor<string, []>("q_1375_cast_fp16")];
            tensor<int32, [1]> k_825_axes_0 = const()[name = tensor<string, []>("k_825_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_823_cast_fp16 = transpose(perm = k_823_perm_0, x = var_16718_cast_fp16)[name = tensor<string, []>("transpose_25")];
            tensor<fp16, [1, 1, 197, 64]> k_825_cast_fp16 = expand_dims(axes = k_825_axes_0, x = k_823_cast_fp16)[name = tensor<string, []>("k_825_cast_fp16")];
            tensor<int32, [1]> v_827_axes_0 = const()[name = tensor<string, []>("v_827_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_827_cast_fp16 = expand_dims(axes = v_827_axes_0, x = v_825_cast_fp16)[name = tensor<string, []>("v_827_cast_fp16")];
            tensor<int32, [4]> var_16735_begin_0 = const()[name = tensor<string, []>("op_16735_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16735_end_0 = const()[name = tensor<string, []>("op_16735_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16735_end_mask_0 = const()[name = tensor<string, []>("op_16735_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16735_cast_fp16 = slice_by_index(begin = var_16735_begin_0, end = var_16735_end_0, end_mask = var_16735_end_mask_0, x = q_1375_cast_fp16)[name = tensor<string, []>("op_16735_cast_fp16")];
            tensor<int32, [4]> var_16739_begin_0 = const()[name = tensor<string, []>("op_16739_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16739_end_0 = const()[name = tensor<string, []>("op_16739_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16739_end_mask_0 = const()[name = tensor<string, []>("op_16739_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16739_cast_fp16 = slice_by_index(begin = var_16739_begin_0, end = var_16739_end_0, end_mask = var_16739_end_mask_0, x = k_825_cast_fp16)[name = tensor<string, []>("op_16739_cast_fp16")];
            tensor<int32, [4]> var_16743_begin_0 = const()[name = tensor<string, []>("op_16743_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16743_end_0 = const()[name = tensor<string, []>("op_16743_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16743_end_mask_0 = const()[name = tensor<string, []>("op_16743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16743_cast_fp16 = slice_by_index(begin = var_16743_begin_0, end = var_16743_end_0, end_mask = var_16743_end_mask_0, x = q_1375_cast_fp16)[name = tensor<string, []>("op_16743_cast_fp16")];
            tensor<int32, [4]> var_16747_begin_0 = const()[name = tensor<string, []>("op_16747_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16747_end_0 = const()[name = tensor<string, []>("op_16747_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16747_end_mask_0 = const()[name = tensor<string, []>("op_16747_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16747_cast_fp16 = slice_by_index(begin = var_16747_begin_0, end = var_16747_end_0, end_mask = var_16747_end_mask_0, x = k_825_cast_fp16)[name = tensor<string, []>("op_16747_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16749_cast_fp16 = mul(x = var_16743_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16749_cast_fp16")];
            tensor<int32, [5]> var_16757 = const()[name = tensor<string, []>("op_16757"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1875_cast_fp16 = reshape(shape = var_16757, x = var_16743_cast_fp16)[name = tensor<string, []>("x_1875_cast_fp16")];
            tensor<int32, [2]> var_16759_split_sizes_0 = const()[name = tensor<string, []>("op_16759_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16759_axis_0 = const()[name = tensor<string, []>("op_16759_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16759_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16759_cast_fp16_1 = split(axis = var_16759_axis_0, split_sizes = var_16759_split_sizes_0, x = x_1875_cast_fp16)[name = tensor<string, []>("op_16759_cast_fp16")];
            tensor<int32, [1]> squeeze_548_axes_0 = const()[name = tensor<string, []>("squeeze_548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_548_cast_fp16 = squeeze(axes = squeeze_548_axes_0, x = var_16759_cast_fp16_0)[name = tensor<string, []>("squeeze_548_cast_fp16")];
            tensor<int32, [1]> squeeze_549_axes_0 = const()[name = tensor<string, []>("squeeze_549_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_549_cast_fp16 = squeeze(axes = squeeze_549_axes_0, x = var_16759_cast_fp16_1)[name = tensor<string, []>("squeeze_549_cast_fp16")];
            tensor<fp16, []> const_2243_promoted_to_fp16 = const()[name = tensor<string, []>("const_2243_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16762_cast_fp16 = mul(x = squeeze_549_cast_fp16, y = const_2243_promoted_to_fp16)[name = tensor<string, []>("op_16762_cast_fp16")];
            tensor<int32, []> x_1877_axis_0 = const()[name = tensor<string, []>("x_1877_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1877_cast_fp16 = stack(axis = x_1877_axis_0, values = (var_16762_cast_fp16, squeeze_548_cast_fp16))[name = tensor<string, []>("x_1877_cast_fp16")];
            tensor<int32, [4]> var_16768 = const()[name = tensor<string, []>("op_16768"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16769_cast_fp16 = reshape(shape = var_16768, x = x_1877_cast_fp16)[name = tensor<string, []>("op_16769_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16770_cast_fp16 = mul(x = var_16769_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16770_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_275_cast_fp16 = add(x = var_16749_cast_fp16, y = var_16770_cast_fp16)[name = tensor<string, []>("q_patches_275_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16772_cast_fp16 = mul(x = var_16747_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16772_cast_fp16")];
            tensor<int32, [5]> var_16780 = const()[name = tensor<string, []>("op_16780"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1881_cast_fp16 = reshape(shape = var_16780, x = var_16747_cast_fp16)[name = tensor<string, []>("x_1881_cast_fp16")];
            tensor<int32, [2]> var_16782_split_sizes_0 = const()[name = tensor<string, []>("op_16782_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16782_axis_0 = const()[name = tensor<string, []>("op_16782_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16782_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16782_cast_fp16_1 = split(axis = var_16782_axis_0, split_sizes = var_16782_split_sizes_0, x = x_1881_cast_fp16)[name = tensor<string, []>("op_16782_cast_fp16")];
            tensor<int32, [1]> squeeze_550_axes_0 = const()[name = tensor<string, []>("squeeze_550_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_550_cast_fp16 = squeeze(axes = squeeze_550_axes_0, x = var_16782_cast_fp16_0)[name = tensor<string, []>("squeeze_550_cast_fp16")];
            tensor<int32, [1]> squeeze_551_axes_0 = const()[name = tensor<string, []>("squeeze_551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_551_cast_fp16 = squeeze(axes = squeeze_551_axes_0, x = var_16782_cast_fp16_1)[name = tensor<string, []>("squeeze_551_cast_fp16")];
            tensor<fp16, []> const_2251_promoted_to_fp16 = const()[name = tensor<string, []>("const_2251_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16785_cast_fp16 = mul(x = squeeze_551_cast_fp16, y = const_2251_promoted_to_fp16)[name = tensor<string, []>("op_16785_cast_fp16")];
            tensor<int32, []> x_1883_axis_0 = const()[name = tensor<string, []>("x_1883_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1883_cast_fp16 = stack(axis = x_1883_axis_0, values = (var_16785_cast_fp16, squeeze_550_cast_fp16))[name = tensor<string, []>("x_1883_cast_fp16")];
            tensor<int32, [4]> var_16791 = const()[name = tensor<string, []>("op_16791"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16792_cast_fp16 = reshape(shape = var_16791, x = x_1883_cast_fp16)[name = tensor<string, []>("op_16792_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16793_cast_fp16 = mul(x = var_16792_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16793_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_275_cast_fp16 = add(x = var_16772_cast_fp16, y = var_16793_cast_fp16)[name = tensor<string, []>("k_patches_275_cast_fp16")];
            tensor<bool, []> var_16796_interleave_0 = const()[name = tensor<string, []>("op_16796_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16796_cast_fp16 = concat(axis = var_20, interleave = var_16796_interleave_0, values = (var_16735_cast_fp16, q_patches_275_cast_fp16))[name = tensor<string, []>("op_16796_cast_fp16")];
            tensor<bool, []> var_16799_interleave_0 = const()[name = tensor<string, []>("op_16799_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16799_cast_fp16 = concat(axis = var_20, interleave = var_16799_interleave_0, values = (var_16739_cast_fp16, k_patches_275_cast_fp16))[name = tensor<string, []>("op_16799_cast_fp16")];
            tensor<fp16, []> var_16801_to_fp16 = const()[name = tensor<string, []>("op_16801_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1379_cast_fp16 = mul(x = var_16796_cast_fp16, y = var_16801_to_fp16)[name = tensor<string, []>("q_1379_cast_fp16")];
            tensor<bool, []> attn_549_transpose_x_1 = const()[name = tensor<string, []>("attn_549_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_549_transpose_y_1 = const()[name = tensor<string, []>("attn_549_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_549_cast_fp16 = matmul(transpose_x = attn_549_transpose_x_1, transpose_y = attn_549_transpose_y_1, x = q_1379_cast_fp16, y = var_16799_cast_fp16)[name = tensor<string, []>("attn_549_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_551_cast_fp16 = softmax(axis = var_21, x = attn_549_cast_fp16)[name = tensor<string, []>("attn_551_cast_fp16")];
            tensor<bool, []> var_16806_transpose_x_0 = const()[name = tensor<string, []>("op_16806_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16806_transpose_y_0 = const()[name = tensor<string, []>("op_16806_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16806_cast_fp16 = matmul(transpose_x = var_16806_transpose_x_0, transpose_y = var_16806_transpose_y_0, x = attn_551_cast_fp16, y = v_827_cast_fp16)[name = tensor<string, []>("op_16806_cast_fp16")];
            tensor<int32, [1]> var_16807_axes_0 = const()[name = tensor<string, []>("op_16807_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16807_cast_fp16 = squeeze(axes = var_16807_axes_0, x = var_16806_cast_fp16)[name = tensor<string, []>("op_16807_cast_fp16")];
            tensor<string, []> var_16813_pad_type_0 = const()[name = tensor<string, []>("op_16813_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16813_strides_0 = const()[name = tensor<string, []>("op_16813_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16813_pad_0 = const()[name = tensor<string, []>("op_16813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16813_dilations_0 = const()[name = tensor<string, []>("op_16813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16813_groups_0 = const()[name = tensor<string, []>("op_16813_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159377984)))];
            tensor<fp16, [1, 64, 1, 197]> var_16813_cast_fp16 = conv(dilations = var_16813_dilations_0, groups = var_16813_groups_0, pad = var_16813_pad_0, pad_type = var_16813_pad_type_0, strides = var_16813_strides_0, weight = model_blocks_11_attn_q_projs_6_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16813_cast_fp16")];
            tensor<int32, [1]> var_16814_axes_0 = const()[name = tensor<string, []>("op_16814_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16814_cast_fp16 = squeeze(axes = var_16814_axes_0, x = var_16813_cast_fp16)[name = tensor<string, []>("op_16814_cast_fp16")];
            tensor<int32, [3]> q_1381_perm_0 = const()[name = tensor<string, []>("q_1381_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16821_pad_type_0 = const()[name = tensor<string, []>("op_16821_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16821_strides_0 = const()[name = tensor<string, []>("op_16821_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16821_pad_0 = const()[name = tensor<string, []>("op_16821_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16821_dilations_0 = const()[name = tensor<string, []>("op_16821_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16821_groups_0 = const()[name = tensor<string, []>("op_16821_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159476352)))];
            tensor<fp16, [1, 64, 1, 197]> var_16821_cast_fp16 = conv(dilations = var_16821_dilations_0, groups = var_16821_groups_0, pad = var_16821_pad_0, pad_type = var_16821_pad_type_0, strides = var_16821_strides_0, weight = model_blocks_11_attn_k_projs_6_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16821_cast_fp16")];
            tensor<int32, [1]> var_16822_axes_0 = const()[name = tensor<string, []>("op_16822_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16822_cast_fp16 = squeeze(axes = var_16822_axes_0, x = var_16821_cast_fp16)[name = tensor<string, []>("op_16822_cast_fp16")];
            tensor<int32, [3]> k_829_perm_0 = const()[name = tensor<string, []>("k_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16829_pad_type_0 = const()[name = tensor<string, []>("op_16829_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16829_strides_0 = const()[name = tensor<string, []>("op_16829_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16829_pad_0 = const()[name = tensor<string, []>("op_16829_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16829_dilations_0 = const()[name = tensor<string, []>("op_16829_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16829_groups_0 = const()[name = tensor<string, []>("op_16829_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_6_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_6_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159574720)))];
            tensor<fp16, [1, 64, 1, 197]> var_16829_cast_fp16 = conv(dilations = var_16829_dilations_0, groups = var_16829_groups_0, pad = var_16829_pad_0, pad_type = var_16829_pad_type_0, strides = var_16829_strides_0, weight = model_blocks_11_attn_v_projs_6_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16829_cast_fp16")];
            tensor<int32, [1]> var_16830_axes_0 = const()[name = tensor<string, []>("op_16830_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16830_cast_fp16 = squeeze(axes = var_16830_axes_0, x = var_16829_cast_fp16)[name = tensor<string, []>("op_16830_cast_fp16")];
            tensor<int32, [3]> v_829_perm_0 = const()[name = tensor<string, []>("v_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159673088)))];
            tensor<fp16, [1, 197, 64]> q_1381_cast_fp16 = transpose(perm = q_1381_perm_0, x = var_16814_cast_fp16)[name = tensor<string, []>("transpose_23")];
            tensor<fp16, [1, 197, 64]> q_1383_cast_fp16 = add(x = q_1381_cast_fp16, y = model_blocks_11_attn_q_biases_6_to_fp16)[name = tensor<string, []>("q_1383_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_6_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_6_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159673280)))];
            tensor<fp16, [1, 197, 64]> v_829_cast_fp16 = transpose(perm = v_829_perm_0, x = var_16830_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, 197, 64]> v_831_cast_fp16 = add(x = v_829_cast_fp16, y = model_blocks_11_attn_v_biases_6_to_fp16)[name = tensor<string, []>("v_831_cast_fp16")];
            tensor<int32, [1]> q_1385_axes_0 = const()[name = tensor<string, []>("q_1385_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1385_cast_fp16 = expand_dims(axes = q_1385_axes_0, x = q_1383_cast_fp16)[name = tensor<string, []>("q_1385_cast_fp16")];
            tensor<int32, [1]> k_831_axes_0 = const()[name = tensor<string, []>("k_831_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_829_cast_fp16 = transpose(perm = k_829_perm_0, x = var_16822_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 1, 197, 64]> k_831_cast_fp16 = expand_dims(axes = k_831_axes_0, x = k_829_cast_fp16)[name = tensor<string, []>("k_831_cast_fp16")];
            tensor<int32, [1]> v_833_axes_0 = const()[name = tensor<string, []>("v_833_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_833_cast_fp16 = expand_dims(axes = v_833_axes_0, x = v_831_cast_fp16)[name = tensor<string, []>("v_833_cast_fp16")];
            tensor<int32, [4]> var_16839_begin_0 = const()[name = tensor<string, []>("op_16839_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16839_end_0 = const()[name = tensor<string, []>("op_16839_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16839_end_mask_0 = const()[name = tensor<string, []>("op_16839_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16839_cast_fp16 = slice_by_index(begin = var_16839_begin_0, end = var_16839_end_0, end_mask = var_16839_end_mask_0, x = q_1385_cast_fp16)[name = tensor<string, []>("op_16839_cast_fp16")];
            tensor<int32, [4]> var_16843_begin_0 = const()[name = tensor<string, []>("op_16843_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16843_end_0 = const()[name = tensor<string, []>("op_16843_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16843_end_mask_0 = const()[name = tensor<string, []>("op_16843_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16843_cast_fp16 = slice_by_index(begin = var_16843_begin_0, end = var_16843_end_0, end_mask = var_16843_end_mask_0, x = k_831_cast_fp16)[name = tensor<string, []>("op_16843_cast_fp16")];
            tensor<int32, [4]> var_16847_begin_0 = const()[name = tensor<string, []>("op_16847_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16847_end_0 = const()[name = tensor<string, []>("op_16847_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16847_end_mask_0 = const()[name = tensor<string, []>("op_16847_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16847_cast_fp16 = slice_by_index(begin = var_16847_begin_0, end = var_16847_end_0, end_mask = var_16847_end_mask_0, x = q_1385_cast_fp16)[name = tensor<string, []>("op_16847_cast_fp16")];
            tensor<int32, [4]> var_16851_begin_0 = const()[name = tensor<string, []>("op_16851_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16851_end_0 = const()[name = tensor<string, []>("op_16851_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16851_end_mask_0 = const()[name = tensor<string, []>("op_16851_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16851_cast_fp16 = slice_by_index(begin = var_16851_begin_0, end = var_16851_end_0, end_mask = var_16851_end_mask_0, x = k_831_cast_fp16)[name = tensor<string, []>("op_16851_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16853_cast_fp16 = mul(x = var_16847_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16853_cast_fp16")];
            tensor<int32, [5]> var_16861 = const()[name = tensor<string, []>("op_16861"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1887_cast_fp16 = reshape(shape = var_16861, x = var_16847_cast_fp16)[name = tensor<string, []>("x_1887_cast_fp16")];
            tensor<int32, [2]> var_16863_split_sizes_0 = const()[name = tensor<string, []>("op_16863_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16863_axis_0 = const()[name = tensor<string, []>("op_16863_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16863_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16863_cast_fp16_1 = split(axis = var_16863_axis_0, split_sizes = var_16863_split_sizes_0, x = x_1887_cast_fp16)[name = tensor<string, []>("op_16863_cast_fp16")];
            tensor<int32, [1]> squeeze_552_axes_0 = const()[name = tensor<string, []>("squeeze_552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_552_cast_fp16 = squeeze(axes = squeeze_552_axes_0, x = var_16863_cast_fp16_0)[name = tensor<string, []>("squeeze_552_cast_fp16")];
            tensor<int32, [1]> squeeze_553_axes_0 = const()[name = tensor<string, []>("squeeze_553_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_553_cast_fp16 = squeeze(axes = squeeze_553_axes_0, x = var_16863_cast_fp16_1)[name = tensor<string, []>("squeeze_553_cast_fp16")];
            tensor<fp16, []> const_2259_promoted_to_fp16 = const()[name = tensor<string, []>("const_2259_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16866_cast_fp16 = mul(x = squeeze_553_cast_fp16, y = const_2259_promoted_to_fp16)[name = tensor<string, []>("op_16866_cast_fp16")];
            tensor<int32, []> x_1889_axis_0 = const()[name = tensor<string, []>("x_1889_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1889_cast_fp16 = stack(axis = x_1889_axis_0, values = (var_16866_cast_fp16, squeeze_552_cast_fp16))[name = tensor<string, []>("x_1889_cast_fp16")];
            tensor<int32, [4]> var_16872 = const()[name = tensor<string, []>("op_16872"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16873_cast_fp16 = reshape(shape = var_16872, x = x_1889_cast_fp16)[name = tensor<string, []>("op_16873_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16874_cast_fp16 = mul(x = var_16873_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16874_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_277_cast_fp16 = add(x = var_16853_cast_fp16, y = var_16874_cast_fp16)[name = tensor<string, []>("q_patches_277_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16876_cast_fp16 = mul(x = var_16851_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16876_cast_fp16")];
            tensor<int32, [5]> var_16884 = const()[name = tensor<string, []>("op_16884"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1893_cast_fp16 = reshape(shape = var_16884, x = var_16851_cast_fp16)[name = tensor<string, []>("x_1893_cast_fp16")];
            tensor<int32, [2]> var_16886_split_sizes_0 = const()[name = tensor<string, []>("op_16886_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16886_axis_0 = const()[name = tensor<string, []>("op_16886_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16886_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16886_cast_fp16_1 = split(axis = var_16886_axis_0, split_sizes = var_16886_split_sizes_0, x = x_1893_cast_fp16)[name = tensor<string, []>("op_16886_cast_fp16")];
            tensor<int32, [1]> squeeze_554_axes_0 = const()[name = tensor<string, []>("squeeze_554_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_554_cast_fp16 = squeeze(axes = squeeze_554_axes_0, x = var_16886_cast_fp16_0)[name = tensor<string, []>("squeeze_554_cast_fp16")];
            tensor<int32, [1]> squeeze_555_axes_0 = const()[name = tensor<string, []>("squeeze_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_555_cast_fp16 = squeeze(axes = squeeze_555_axes_0, x = var_16886_cast_fp16_1)[name = tensor<string, []>("squeeze_555_cast_fp16")];
            tensor<fp16, []> const_2267_promoted_to_fp16 = const()[name = tensor<string, []>("const_2267_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16889_cast_fp16 = mul(x = squeeze_555_cast_fp16, y = const_2267_promoted_to_fp16)[name = tensor<string, []>("op_16889_cast_fp16")];
            tensor<int32, []> x_1895_axis_0 = const()[name = tensor<string, []>("x_1895_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1895_cast_fp16 = stack(axis = x_1895_axis_0, values = (var_16889_cast_fp16, squeeze_554_cast_fp16))[name = tensor<string, []>("x_1895_cast_fp16")];
            tensor<int32, [4]> var_16895 = const()[name = tensor<string, []>("op_16895"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16896_cast_fp16 = reshape(shape = var_16895, x = x_1895_cast_fp16)[name = tensor<string, []>("op_16896_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16897_cast_fp16 = mul(x = var_16896_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16897_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_277_cast_fp16 = add(x = var_16876_cast_fp16, y = var_16897_cast_fp16)[name = tensor<string, []>("k_patches_277_cast_fp16")];
            tensor<bool, []> var_16900_interleave_0 = const()[name = tensor<string, []>("op_16900_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16900_cast_fp16 = concat(axis = var_20, interleave = var_16900_interleave_0, values = (var_16839_cast_fp16, q_patches_277_cast_fp16))[name = tensor<string, []>("op_16900_cast_fp16")];
            tensor<bool, []> var_16903_interleave_0 = const()[name = tensor<string, []>("op_16903_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16903_cast_fp16 = concat(axis = var_20, interleave = var_16903_interleave_0, values = (var_16843_cast_fp16, k_patches_277_cast_fp16))[name = tensor<string, []>("op_16903_cast_fp16")];
            tensor<fp16, []> var_16905_to_fp16 = const()[name = tensor<string, []>("op_16905_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1389_cast_fp16 = mul(x = var_16900_cast_fp16, y = var_16905_to_fp16)[name = tensor<string, []>("q_1389_cast_fp16")];
            tensor<bool, []> attn_553_transpose_x_1 = const()[name = tensor<string, []>("attn_553_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_553_transpose_y_1 = const()[name = tensor<string, []>("attn_553_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_553_cast_fp16 = matmul(transpose_x = attn_553_transpose_x_1, transpose_y = attn_553_transpose_y_1, x = q_1389_cast_fp16, y = var_16903_cast_fp16)[name = tensor<string, []>("attn_553_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_555_cast_fp16 = softmax(axis = var_21, x = attn_553_cast_fp16)[name = tensor<string, []>("attn_555_cast_fp16")];
            tensor<bool, []> var_16910_transpose_x_0 = const()[name = tensor<string, []>("op_16910_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_16910_transpose_y_0 = const()[name = tensor<string, []>("op_16910_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_16910_cast_fp16 = matmul(transpose_x = var_16910_transpose_x_0, transpose_y = var_16910_transpose_y_0, x = attn_555_cast_fp16, y = v_833_cast_fp16)[name = tensor<string, []>("op_16910_cast_fp16")];
            tensor<int32, [1]> var_16911_axes_0 = const()[name = tensor<string, []>("op_16911_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_16911_cast_fp16 = squeeze(axes = var_16911_axes_0, x = var_16910_cast_fp16)[name = tensor<string, []>("op_16911_cast_fp16")];
            tensor<string, []> var_16917_pad_type_0 = const()[name = tensor<string, []>("op_16917_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16917_strides_0 = const()[name = tensor<string, []>("op_16917_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16917_pad_0 = const()[name = tensor<string, []>("op_16917_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16917_dilations_0 = const()[name = tensor<string, []>("op_16917_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16917_groups_0 = const()[name = tensor<string, []>("op_16917_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159673472)))];
            tensor<fp16, [1, 64, 1, 197]> var_16917_cast_fp16 = conv(dilations = var_16917_dilations_0, groups = var_16917_groups_0, pad = var_16917_pad_0, pad_type = var_16917_pad_type_0, strides = var_16917_strides_0, weight = model_blocks_11_attn_q_projs_7_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16917_cast_fp16")];
            tensor<int32, [1]> var_16918_axes_0 = const()[name = tensor<string, []>("op_16918_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16918_cast_fp16 = squeeze(axes = var_16918_axes_0, x = var_16917_cast_fp16)[name = tensor<string, []>("op_16918_cast_fp16")];
            tensor<int32, [3]> q_1391_perm_0 = const()[name = tensor<string, []>("q_1391_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16925_pad_type_0 = const()[name = tensor<string, []>("op_16925_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16925_strides_0 = const()[name = tensor<string, []>("op_16925_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16925_pad_0 = const()[name = tensor<string, []>("op_16925_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16925_dilations_0 = const()[name = tensor<string, []>("op_16925_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16925_groups_0 = const()[name = tensor<string, []>("op_16925_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159771840)))];
            tensor<fp16, [1, 64, 1, 197]> var_16925_cast_fp16 = conv(dilations = var_16925_dilations_0, groups = var_16925_groups_0, pad = var_16925_pad_0, pad_type = var_16925_pad_type_0, strides = var_16925_strides_0, weight = model_blocks_11_attn_k_projs_7_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16925_cast_fp16")];
            tensor<int32, [1]> var_16926_axes_0 = const()[name = tensor<string, []>("op_16926_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16926_cast_fp16 = squeeze(axes = var_16926_axes_0, x = var_16925_cast_fp16)[name = tensor<string, []>("op_16926_cast_fp16")];
            tensor<int32, [3]> k_835_perm_0 = const()[name = tensor<string, []>("k_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_16933_pad_type_0 = const()[name = tensor<string, []>("op_16933_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_16933_strides_0 = const()[name = tensor<string, []>("op_16933_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_16933_pad_0 = const()[name = tensor<string, []>("op_16933_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_16933_dilations_0 = const()[name = tensor<string, []>("op_16933_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16933_groups_0 = const()[name = tensor<string, []>("op_16933_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_7_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_7_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159870208)))];
            tensor<fp16, [1, 64, 1, 197]> var_16933_cast_fp16 = conv(dilations = var_16933_dilations_0, groups = var_16933_groups_0, pad = var_16933_pad_0, pad_type = var_16933_pad_type_0, strides = var_16933_strides_0, weight = model_blocks_11_attn_v_projs_7_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_16933_cast_fp16")];
            tensor<int32, [1]> var_16934_axes_0 = const()[name = tensor<string, []>("op_16934_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_16934_cast_fp16 = squeeze(axes = var_16934_axes_0, x = var_16933_cast_fp16)[name = tensor<string, []>("op_16934_cast_fp16")];
            tensor<int32, [3]> v_835_perm_0 = const()[name = tensor<string, []>("v_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159968576)))];
            tensor<fp16, [1, 197, 64]> q_1391_cast_fp16 = transpose(perm = q_1391_perm_0, x = var_16918_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 197, 64]> q_1393_cast_fp16 = add(x = q_1391_cast_fp16, y = model_blocks_11_attn_q_biases_7_to_fp16)[name = tensor<string, []>("q_1393_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_7_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_7_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159968768)))];
            tensor<fp16, [1, 197, 64]> v_835_cast_fp16 = transpose(perm = v_835_perm_0, x = var_16934_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, 197, 64]> v_837_cast_fp16 = add(x = v_835_cast_fp16, y = model_blocks_11_attn_v_biases_7_to_fp16)[name = tensor<string, []>("v_837_cast_fp16")];
            tensor<int32, [1]> q_1395_axes_0 = const()[name = tensor<string, []>("q_1395_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1395_cast_fp16 = expand_dims(axes = q_1395_axes_0, x = q_1393_cast_fp16)[name = tensor<string, []>("q_1395_cast_fp16")];
            tensor<int32, [1]> k_837_axes_0 = const()[name = tensor<string, []>("k_837_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_835_cast_fp16 = transpose(perm = k_835_perm_0, x = var_16926_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, 1, 197, 64]> k_837_cast_fp16 = expand_dims(axes = k_837_axes_0, x = k_835_cast_fp16)[name = tensor<string, []>("k_837_cast_fp16")];
            tensor<int32, [1]> v_839_axes_0 = const()[name = tensor<string, []>("v_839_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_839_cast_fp16 = expand_dims(axes = v_839_axes_0, x = v_837_cast_fp16)[name = tensor<string, []>("v_839_cast_fp16")];
            tensor<int32, [4]> var_16943_begin_0 = const()[name = tensor<string, []>("op_16943_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16943_end_0 = const()[name = tensor<string, []>("op_16943_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16943_end_mask_0 = const()[name = tensor<string, []>("op_16943_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16943_cast_fp16 = slice_by_index(begin = var_16943_begin_0, end = var_16943_end_0, end_mask = var_16943_end_mask_0, x = q_1395_cast_fp16)[name = tensor<string, []>("op_16943_cast_fp16")];
            tensor<int32, [4]> var_16947_begin_0 = const()[name = tensor<string, []>("op_16947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16947_end_0 = const()[name = tensor<string, []>("op_16947_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_16947_end_mask_0 = const()[name = tensor<string, []>("op_16947_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_16947_cast_fp16 = slice_by_index(begin = var_16947_begin_0, end = var_16947_end_0, end_mask = var_16947_end_mask_0, x = k_837_cast_fp16)[name = tensor<string, []>("op_16947_cast_fp16")];
            tensor<int32, [4]> var_16951_begin_0 = const()[name = tensor<string, []>("op_16951_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16951_end_0 = const()[name = tensor<string, []>("op_16951_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16951_end_mask_0 = const()[name = tensor<string, []>("op_16951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16951_cast_fp16 = slice_by_index(begin = var_16951_begin_0, end = var_16951_end_0, end_mask = var_16951_end_mask_0, x = q_1395_cast_fp16)[name = tensor<string, []>("op_16951_cast_fp16")];
            tensor<int32, [4]> var_16955_begin_0 = const()[name = tensor<string, []>("op_16955_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16955_end_0 = const()[name = tensor<string, []>("op_16955_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_16955_end_mask_0 = const()[name = tensor<string, []>("op_16955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_16955_cast_fp16 = slice_by_index(begin = var_16955_begin_0, end = var_16955_end_0, end_mask = var_16955_end_mask_0, x = k_837_cast_fp16)[name = tensor<string, []>("op_16955_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16957_cast_fp16 = mul(x = var_16951_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16957_cast_fp16")];
            tensor<int32, [5]> var_16965 = const()[name = tensor<string, []>("op_16965"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1899_cast_fp16 = reshape(shape = var_16965, x = var_16951_cast_fp16)[name = tensor<string, []>("x_1899_cast_fp16")];
            tensor<int32, [2]> var_16967_split_sizes_0 = const()[name = tensor<string, []>("op_16967_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16967_axis_0 = const()[name = tensor<string, []>("op_16967_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16967_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16967_cast_fp16_1 = split(axis = var_16967_axis_0, split_sizes = var_16967_split_sizes_0, x = x_1899_cast_fp16)[name = tensor<string, []>("op_16967_cast_fp16")];
            tensor<int32, [1]> squeeze_556_axes_0 = const()[name = tensor<string, []>("squeeze_556_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_556_cast_fp16 = squeeze(axes = squeeze_556_axes_0, x = var_16967_cast_fp16_0)[name = tensor<string, []>("squeeze_556_cast_fp16")];
            tensor<int32, [1]> squeeze_557_axes_0 = const()[name = tensor<string, []>("squeeze_557_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_557_cast_fp16 = squeeze(axes = squeeze_557_axes_0, x = var_16967_cast_fp16_1)[name = tensor<string, []>("squeeze_557_cast_fp16")];
            tensor<fp16, []> const_2275_promoted_to_fp16 = const()[name = tensor<string, []>("const_2275_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16970_cast_fp16 = mul(x = squeeze_557_cast_fp16, y = const_2275_promoted_to_fp16)[name = tensor<string, []>("op_16970_cast_fp16")];
            tensor<int32, []> x_1901_axis_0 = const()[name = tensor<string, []>("x_1901_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1901_cast_fp16 = stack(axis = x_1901_axis_0, values = (var_16970_cast_fp16, squeeze_556_cast_fp16))[name = tensor<string, []>("x_1901_cast_fp16")];
            tensor<int32, [4]> var_16976 = const()[name = tensor<string, []>("op_16976"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_16977_cast_fp16 = reshape(shape = var_16976, x = x_1901_cast_fp16)[name = tensor<string, []>("op_16977_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16978_cast_fp16 = mul(x = var_16977_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_16978_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_279_cast_fp16 = add(x = var_16957_cast_fp16, y = var_16978_cast_fp16)[name = tensor<string, []>("q_patches_279_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_16980_cast_fp16 = mul(x = var_16955_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_16980_cast_fp16")];
            tensor<int32, [5]> var_16988 = const()[name = tensor<string, []>("op_16988"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1905_cast_fp16 = reshape(shape = var_16988, x = var_16955_cast_fp16)[name = tensor<string, []>("x_1905_cast_fp16")];
            tensor<int32, [2]> var_16990_split_sizes_0 = const()[name = tensor<string, []>("op_16990_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_16990_axis_0 = const()[name = tensor<string, []>("op_16990_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_16990_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_16990_cast_fp16_1 = split(axis = var_16990_axis_0, split_sizes = var_16990_split_sizes_0, x = x_1905_cast_fp16)[name = tensor<string, []>("op_16990_cast_fp16")];
            tensor<int32, [1]> squeeze_558_axes_0 = const()[name = tensor<string, []>("squeeze_558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_558_cast_fp16 = squeeze(axes = squeeze_558_axes_0, x = var_16990_cast_fp16_0)[name = tensor<string, []>("squeeze_558_cast_fp16")];
            tensor<int32, [1]> squeeze_559_axes_0 = const()[name = tensor<string, []>("squeeze_559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_559_cast_fp16 = squeeze(axes = squeeze_559_axes_0, x = var_16990_cast_fp16_1)[name = tensor<string, []>("squeeze_559_cast_fp16")];
            tensor<fp16, []> const_2283_promoted_to_fp16 = const()[name = tensor<string, []>("const_2283_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_16993_cast_fp16 = mul(x = squeeze_559_cast_fp16, y = const_2283_promoted_to_fp16)[name = tensor<string, []>("op_16993_cast_fp16")];
            tensor<int32, []> x_1907_axis_0 = const()[name = tensor<string, []>("x_1907_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1907_cast_fp16 = stack(axis = x_1907_axis_0, values = (var_16993_cast_fp16, squeeze_558_cast_fp16))[name = tensor<string, []>("x_1907_cast_fp16")];
            tensor<int32, [4]> var_16999 = const()[name = tensor<string, []>("op_16999"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17000_cast_fp16 = reshape(shape = var_16999, x = x_1907_cast_fp16)[name = tensor<string, []>("op_17000_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17001_cast_fp16 = mul(x = var_17000_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17001_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_279_cast_fp16 = add(x = var_16980_cast_fp16, y = var_17001_cast_fp16)[name = tensor<string, []>("k_patches_279_cast_fp16")];
            tensor<bool, []> var_17004_interleave_0 = const()[name = tensor<string, []>("op_17004_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17004_cast_fp16 = concat(axis = var_20, interleave = var_17004_interleave_0, values = (var_16943_cast_fp16, q_patches_279_cast_fp16))[name = tensor<string, []>("op_17004_cast_fp16")];
            tensor<bool, []> var_17007_interleave_0 = const()[name = tensor<string, []>("op_17007_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17007_cast_fp16 = concat(axis = var_20, interleave = var_17007_interleave_0, values = (var_16947_cast_fp16, k_patches_279_cast_fp16))[name = tensor<string, []>("op_17007_cast_fp16")];
            tensor<fp16, []> var_17009_to_fp16 = const()[name = tensor<string, []>("op_17009_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1399_cast_fp16 = mul(x = var_17004_cast_fp16, y = var_17009_to_fp16)[name = tensor<string, []>("q_1399_cast_fp16")];
            tensor<bool, []> attn_557_transpose_x_1 = const()[name = tensor<string, []>("attn_557_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_557_transpose_y_1 = const()[name = tensor<string, []>("attn_557_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_557_cast_fp16 = matmul(transpose_x = attn_557_transpose_x_1, transpose_y = attn_557_transpose_y_1, x = q_1399_cast_fp16, y = var_17007_cast_fp16)[name = tensor<string, []>("attn_557_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_559_cast_fp16 = softmax(axis = var_21, x = attn_557_cast_fp16)[name = tensor<string, []>("attn_559_cast_fp16")];
            tensor<bool, []> var_17014_transpose_x_0 = const()[name = tensor<string, []>("op_17014_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_17014_transpose_y_0 = const()[name = tensor<string, []>("op_17014_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17014_cast_fp16 = matmul(transpose_x = var_17014_transpose_x_0, transpose_y = var_17014_transpose_y_0, x = attn_559_cast_fp16, y = v_839_cast_fp16)[name = tensor<string, []>("op_17014_cast_fp16")];
            tensor<int32, [1]> var_17015_axes_0 = const()[name = tensor<string, []>("op_17015_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_17015_cast_fp16 = squeeze(axes = var_17015_axes_0, x = var_17014_cast_fp16)[name = tensor<string, []>("op_17015_cast_fp16")];
            tensor<string, []> var_17021_pad_type_0 = const()[name = tensor<string, []>("op_17021_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17021_strides_0 = const()[name = tensor<string, []>("op_17021_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17021_pad_0 = const()[name = tensor<string, []>("op_17021_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17021_dilations_0 = const()[name = tensor<string, []>("op_17021_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17021_groups_0 = const()[name = tensor<string, []>("op_17021_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159968960)))];
            tensor<fp16, [1, 64, 1, 197]> var_17021_cast_fp16 = conv(dilations = var_17021_dilations_0, groups = var_17021_groups_0, pad = var_17021_pad_0, pad_type = var_17021_pad_type_0, strides = var_17021_strides_0, weight = model_blocks_11_attn_q_projs_8_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17021_cast_fp16")];
            tensor<int32, [1]> var_17022_axes_0 = const()[name = tensor<string, []>("op_17022_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17022_cast_fp16 = squeeze(axes = var_17022_axes_0, x = var_17021_cast_fp16)[name = tensor<string, []>("op_17022_cast_fp16")];
            tensor<int32, [3]> q_1401_perm_0 = const()[name = tensor<string, []>("q_1401_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17029_pad_type_0 = const()[name = tensor<string, []>("op_17029_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17029_strides_0 = const()[name = tensor<string, []>("op_17029_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17029_pad_0 = const()[name = tensor<string, []>("op_17029_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17029_dilations_0 = const()[name = tensor<string, []>("op_17029_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17029_groups_0 = const()[name = tensor<string, []>("op_17029_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160067328)))];
            tensor<fp16, [1, 64, 1, 197]> var_17029_cast_fp16 = conv(dilations = var_17029_dilations_0, groups = var_17029_groups_0, pad = var_17029_pad_0, pad_type = var_17029_pad_type_0, strides = var_17029_strides_0, weight = model_blocks_11_attn_k_projs_8_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17029_cast_fp16")];
            tensor<int32, [1]> var_17030_axes_0 = const()[name = tensor<string, []>("op_17030_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17030_cast_fp16 = squeeze(axes = var_17030_axes_0, x = var_17029_cast_fp16)[name = tensor<string, []>("op_17030_cast_fp16")];
            tensor<int32, [3]> k_841_perm_0 = const()[name = tensor<string, []>("k_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17037_pad_type_0 = const()[name = tensor<string, []>("op_17037_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17037_strides_0 = const()[name = tensor<string, []>("op_17037_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17037_pad_0 = const()[name = tensor<string, []>("op_17037_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17037_dilations_0 = const()[name = tensor<string, []>("op_17037_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17037_groups_0 = const()[name = tensor<string, []>("op_17037_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_8_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_8_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160165696)))];
            tensor<fp16, [1, 64, 1, 197]> var_17037_cast_fp16 = conv(dilations = var_17037_dilations_0, groups = var_17037_groups_0, pad = var_17037_pad_0, pad_type = var_17037_pad_type_0, strides = var_17037_strides_0, weight = model_blocks_11_attn_v_projs_8_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17037_cast_fp16")];
            tensor<int32, [1]> var_17038_axes_0 = const()[name = tensor<string, []>("op_17038_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17038_cast_fp16 = squeeze(axes = var_17038_axes_0, x = var_17037_cast_fp16)[name = tensor<string, []>("op_17038_cast_fp16")];
            tensor<int32, [3]> v_841_perm_0 = const()[name = tensor<string, []>("v_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160264064)))];
            tensor<fp16, [1, 197, 64]> q_1401_cast_fp16 = transpose(perm = q_1401_perm_0, x = var_17022_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 197, 64]> q_1403_cast_fp16 = add(x = q_1401_cast_fp16, y = model_blocks_11_attn_q_biases_8_to_fp16)[name = tensor<string, []>("q_1403_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_8_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_8_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160264256)))];
            tensor<fp16, [1, 197, 64]> v_841_cast_fp16 = transpose(perm = v_841_perm_0, x = var_17038_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, 197, 64]> v_843_cast_fp16 = add(x = v_841_cast_fp16, y = model_blocks_11_attn_v_biases_8_to_fp16)[name = tensor<string, []>("v_843_cast_fp16")];
            tensor<int32, [1]> q_1405_axes_0 = const()[name = tensor<string, []>("q_1405_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1405_cast_fp16 = expand_dims(axes = q_1405_axes_0, x = q_1403_cast_fp16)[name = tensor<string, []>("q_1405_cast_fp16")];
            tensor<int32, [1]> k_843_axes_0 = const()[name = tensor<string, []>("k_843_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_841_cast_fp16 = transpose(perm = k_841_perm_0, x = var_17030_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, 1, 197, 64]> k_843_cast_fp16 = expand_dims(axes = k_843_axes_0, x = k_841_cast_fp16)[name = tensor<string, []>("k_843_cast_fp16")];
            tensor<int32, [1]> v_845_axes_0 = const()[name = tensor<string, []>("v_845_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_845_cast_fp16 = expand_dims(axes = v_845_axes_0, x = v_843_cast_fp16)[name = tensor<string, []>("v_845_cast_fp16")];
            tensor<int32, [4]> var_17047_begin_0 = const()[name = tensor<string, []>("op_17047_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17047_end_0 = const()[name = tensor<string, []>("op_17047_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17047_end_mask_0 = const()[name = tensor<string, []>("op_17047_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17047_cast_fp16 = slice_by_index(begin = var_17047_begin_0, end = var_17047_end_0, end_mask = var_17047_end_mask_0, x = q_1405_cast_fp16)[name = tensor<string, []>("op_17047_cast_fp16")];
            tensor<int32, [4]> var_17051_begin_0 = const()[name = tensor<string, []>("op_17051_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17051_end_0 = const()[name = tensor<string, []>("op_17051_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17051_end_mask_0 = const()[name = tensor<string, []>("op_17051_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17051_cast_fp16 = slice_by_index(begin = var_17051_begin_0, end = var_17051_end_0, end_mask = var_17051_end_mask_0, x = k_843_cast_fp16)[name = tensor<string, []>("op_17051_cast_fp16")];
            tensor<int32, [4]> var_17055_begin_0 = const()[name = tensor<string, []>("op_17055_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17055_end_0 = const()[name = tensor<string, []>("op_17055_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17055_end_mask_0 = const()[name = tensor<string, []>("op_17055_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17055_cast_fp16 = slice_by_index(begin = var_17055_begin_0, end = var_17055_end_0, end_mask = var_17055_end_mask_0, x = q_1405_cast_fp16)[name = tensor<string, []>("op_17055_cast_fp16")];
            tensor<int32, [4]> var_17059_begin_0 = const()[name = tensor<string, []>("op_17059_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17059_end_0 = const()[name = tensor<string, []>("op_17059_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17059_end_mask_0 = const()[name = tensor<string, []>("op_17059_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17059_cast_fp16 = slice_by_index(begin = var_17059_begin_0, end = var_17059_end_0, end_mask = var_17059_end_mask_0, x = k_843_cast_fp16)[name = tensor<string, []>("op_17059_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17061_cast_fp16 = mul(x = var_17055_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17061_cast_fp16")];
            tensor<int32, [5]> var_17069 = const()[name = tensor<string, []>("op_17069"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1911_cast_fp16 = reshape(shape = var_17069, x = var_17055_cast_fp16)[name = tensor<string, []>("x_1911_cast_fp16")];
            tensor<int32, [2]> var_17071_split_sizes_0 = const()[name = tensor<string, []>("op_17071_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17071_axis_0 = const()[name = tensor<string, []>("op_17071_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17071_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17071_cast_fp16_1 = split(axis = var_17071_axis_0, split_sizes = var_17071_split_sizes_0, x = x_1911_cast_fp16)[name = tensor<string, []>("op_17071_cast_fp16")];
            tensor<int32, [1]> squeeze_560_axes_0 = const()[name = tensor<string, []>("squeeze_560_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_560_cast_fp16 = squeeze(axes = squeeze_560_axes_0, x = var_17071_cast_fp16_0)[name = tensor<string, []>("squeeze_560_cast_fp16")];
            tensor<int32, [1]> squeeze_561_axes_0 = const()[name = tensor<string, []>("squeeze_561_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_561_cast_fp16 = squeeze(axes = squeeze_561_axes_0, x = var_17071_cast_fp16_1)[name = tensor<string, []>("squeeze_561_cast_fp16")];
            tensor<fp16, []> const_2291_promoted_to_fp16 = const()[name = tensor<string, []>("const_2291_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17074_cast_fp16 = mul(x = squeeze_561_cast_fp16, y = const_2291_promoted_to_fp16)[name = tensor<string, []>("op_17074_cast_fp16")];
            tensor<int32, []> x_1913_axis_0 = const()[name = tensor<string, []>("x_1913_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1913_cast_fp16 = stack(axis = x_1913_axis_0, values = (var_17074_cast_fp16, squeeze_560_cast_fp16))[name = tensor<string, []>("x_1913_cast_fp16")];
            tensor<int32, [4]> var_17080 = const()[name = tensor<string, []>("op_17080"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17081_cast_fp16 = reshape(shape = var_17080, x = x_1913_cast_fp16)[name = tensor<string, []>("op_17081_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17082_cast_fp16 = mul(x = var_17081_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17082_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_281_cast_fp16 = add(x = var_17061_cast_fp16, y = var_17082_cast_fp16)[name = tensor<string, []>("q_patches_281_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17084_cast_fp16 = mul(x = var_17059_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17084_cast_fp16")];
            tensor<int32, [5]> var_17092 = const()[name = tensor<string, []>("op_17092"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1917_cast_fp16 = reshape(shape = var_17092, x = var_17059_cast_fp16)[name = tensor<string, []>("x_1917_cast_fp16")];
            tensor<int32, [2]> var_17094_split_sizes_0 = const()[name = tensor<string, []>("op_17094_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17094_axis_0 = const()[name = tensor<string, []>("op_17094_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17094_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17094_cast_fp16_1 = split(axis = var_17094_axis_0, split_sizes = var_17094_split_sizes_0, x = x_1917_cast_fp16)[name = tensor<string, []>("op_17094_cast_fp16")];
            tensor<int32, [1]> squeeze_562_axes_0 = const()[name = tensor<string, []>("squeeze_562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_562_cast_fp16 = squeeze(axes = squeeze_562_axes_0, x = var_17094_cast_fp16_0)[name = tensor<string, []>("squeeze_562_cast_fp16")];
            tensor<int32, [1]> squeeze_563_axes_0 = const()[name = tensor<string, []>("squeeze_563_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_563_cast_fp16 = squeeze(axes = squeeze_563_axes_0, x = var_17094_cast_fp16_1)[name = tensor<string, []>("squeeze_563_cast_fp16")];
            tensor<fp16, []> const_2299_promoted_to_fp16 = const()[name = tensor<string, []>("const_2299_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17097_cast_fp16 = mul(x = squeeze_563_cast_fp16, y = const_2299_promoted_to_fp16)[name = tensor<string, []>("op_17097_cast_fp16")];
            tensor<int32, []> x_1919_axis_0 = const()[name = tensor<string, []>("x_1919_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1919_cast_fp16 = stack(axis = x_1919_axis_0, values = (var_17097_cast_fp16, squeeze_562_cast_fp16))[name = tensor<string, []>("x_1919_cast_fp16")];
            tensor<int32, [4]> var_17103 = const()[name = tensor<string, []>("op_17103"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17104_cast_fp16 = reshape(shape = var_17103, x = x_1919_cast_fp16)[name = tensor<string, []>("op_17104_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17105_cast_fp16 = mul(x = var_17104_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17105_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_281_cast_fp16 = add(x = var_17084_cast_fp16, y = var_17105_cast_fp16)[name = tensor<string, []>("k_patches_281_cast_fp16")];
            tensor<bool, []> var_17108_interleave_0 = const()[name = tensor<string, []>("op_17108_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17108_cast_fp16 = concat(axis = var_20, interleave = var_17108_interleave_0, values = (var_17047_cast_fp16, q_patches_281_cast_fp16))[name = tensor<string, []>("op_17108_cast_fp16")];
            tensor<bool, []> var_17111_interleave_0 = const()[name = tensor<string, []>("op_17111_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17111_cast_fp16 = concat(axis = var_20, interleave = var_17111_interleave_0, values = (var_17051_cast_fp16, k_patches_281_cast_fp16))[name = tensor<string, []>("op_17111_cast_fp16")];
            tensor<fp16, []> var_17113_to_fp16 = const()[name = tensor<string, []>("op_17113_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1409_cast_fp16 = mul(x = var_17108_cast_fp16, y = var_17113_to_fp16)[name = tensor<string, []>("q_1409_cast_fp16")];
            tensor<bool, []> attn_561_transpose_x_1 = const()[name = tensor<string, []>("attn_561_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_561_transpose_y_1 = const()[name = tensor<string, []>("attn_561_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_561_cast_fp16 = matmul(transpose_x = attn_561_transpose_x_1, transpose_y = attn_561_transpose_y_1, x = q_1409_cast_fp16, y = var_17111_cast_fp16)[name = tensor<string, []>("attn_561_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_563_cast_fp16 = softmax(axis = var_21, x = attn_561_cast_fp16)[name = tensor<string, []>("attn_563_cast_fp16")];
            tensor<bool, []> var_17118_transpose_x_0 = const()[name = tensor<string, []>("op_17118_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_17118_transpose_y_0 = const()[name = tensor<string, []>("op_17118_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17118_cast_fp16 = matmul(transpose_x = var_17118_transpose_x_0, transpose_y = var_17118_transpose_y_0, x = attn_563_cast_fp16, y = v_845_cast_fp16)[name = tensor<string, []>("op_17118_cast_fp16")];
            tensor<int32, [1]> var_17119_axes_0 = const()[name = tensor<string, []>("op_17119_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_17119_cast_fp16 = squeeze(axes = var_17119_axes_0, x = var_17118_cast_fp16)[name = tensor<string, []>("op_17119_cast_fp16")];
            tensor<string, []> var_17125_pad_type_0 = const()[name = tensor<string, []>("op_17125_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17125_strides_0 = const()[name = tensor<string, []>("op_17125_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17125_pad_0 = const()[name = tensor<string, []>("op_17125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17125_dilations_0 = const()[name = tensor<string, []>("op_17125_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17125_groups_0 = const()[name = tensor<string, []>("op_17125_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160264448)))];
            tensor<fp16, [1, 64, 1, 197]> var_17125_cast_fp16 = conv(dilations = var_17125_dilations_0, groups = var_17125_groups_0, pad = var_17125_pad_0, pad_type = var_17125_pad_type_0, strides = var_17125_strides_0, weight = model_blocks_11_attn_q_projs_9_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17125_cast_fp16")];
            tensor<int32, [1]> var_17126_axes_0 = const()[name = tensor<string, []>("op_17126_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17126_cast_fp16 = squeeze(axes = var_17126_axes_0, x = var_17125_cast_fp16)[name = tensor<string, []>("op_17126_cast_fp16")];
            tensor<int32, [3]> q_1411_perm_0 = const()[name = tensor<string, []>("q_1411_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17133_pad_type_0 = const()[name = tensor<string, []>("op_17133_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17133_strides_0 = const()[name = tensor<string, []>("op_17133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17133_pad_0 = const()[name = tensor<string, []>("op_17133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17133_dilations_0 = const()[name = tensor<string, []>("op_17133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17133_groups_0 = const()[name = tensor<string, []>("op_17133_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160362816)))];
            tensor<fp16, [1, 64, 1, 197]> var_17133_cast_fp16 = conv(dilations = var_17133_dilations_0, groups = var_17133_groups_0, pad = var_17133_pad_0, pad_type = var_17133_pad_type_0, strides = var_17133_strides_0, weight = model_blocks_11_attn_k_projs_9_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17133_cast_fp16")];
            tensor<int32, [1]> var_17134_axes_0 = const()[name = tensor<string, []>("op_17134_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17134_cast_fp16 = squeeze(axes = var_17134_axes_0, x = var_17133_cast_fp16)[name = tensor<string, []>("op_17134_cast_fp16")];
            tensor<int32, [3]> k_847_perm_0 = const()[name = tensor<string, []>("k_847_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17141_pad_type_0 = const()[name = tensor<string, []>("op_17141_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17141_strides_0 = const()[name = tensor<string, []>("op_17141_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17141_pad_0 = const()[name = tensor<string, []>("op_17141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17141_dilations_0 = const()[name = tensor<string, []>("op_17141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17141_groups_0 = const()[name = tensor<string, []>("op_17141_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_9_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_9_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160461184)))];
            tensor<fp16, [1, 64, 1, 197]> var_17141_cast_fp16 = conv(dilations = var_17141_dilations_0, groups = var_17141_groups_0, pad = var_17141_pad_0, pad_type = var_17141_pad_type_0, strides = var_17141_strides_0, weight = model_blocks_11_attn_v_projs_9_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17141_cast_fp16")];
            tensor<int32, [1]> var_17142_axes_0 = const()[name = tensor<string, []>("op_17142_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17142_cast_fp16 = squeeze(axes = var_17142_axes_0, x = var_17141_cast_fp16)[name = tensor<string, []>("op_17142_cast_fp16")];
            tensor<int32, [3]> v_847_perm_0 = const()[name = tensor<string, []>("v_847_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160559552)))];
            tensor<fp16, [1, 197, 64]> q_1411_cast_fp16 = transpose(perm = q_1411_perm_0, x = var_17126_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, 197, 64]> q_1413_cast_fp16 = add(x = q_1411_cast_fp16, y = model_blocks_11_attn_q_biases_9_to_fp16)[name = tensor<string, []>("q_1413_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_9_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_9_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160559744)))];
            tensor<fp16, [1, 197, 64]> v_847_cast_fp16 = transpose(perm = v_847_perm_0, x = var_17142_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 197, 64]> v_849_cast_fp16 = add(x = v_847_cast_fp16, y = model_blocks_11_attn_v_biases_9_to_fp16)[name = tensor<string, []>("v_849_cast_fp16")];
            tensor<int32, [1]> q_1415_axes_0 = const()[name = tensor<string, []>("q_1415_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1415_cast_fp16 = expand_dims(axes = q_1415_axes_0, x = q_1413_cast_fp16)[name = tensor<string, []>("q_1415_cast_fp16")];
            tensor<int32, [1]> k_849_axes_0 = const()[name = tensor<string, []>("k_849_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_847_cast_fp16 = transpose(perm = k_847_perm_0, x = var_17134_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 1, 197, 64]> k_849_cast_fp16 = expand_dims(axes = k_849_axes_0, x = k_847_cast_fp16)[name = tensor<string, []>("k_849_cast_fp16")];
            tensor<int32, [1]> v_851_axes_0 = const()[name = tensor<string, []>("v_851_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_851_cast_fp16 = expand_dims(axes = v_851_axes_0, x = v_849_cast_fp16)[name = tensor<string, []>("v_851_cast_fp16")];
            tensor<int32, [4]> var_17151_begin_0 = const()[name = tensor<string, []>("op_17151_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17151_end_0 = const()[name = tensor<string, []>("op_17151_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17151_end_mask_0 = const()[name = tensor<string, []>("op_17151_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17151_cast_fp16 = slice_by_index(begin = var_17151_begin_0, end = var_17151_end_0, end_mask = var_17151_end_mask_0, x = q_1415_cast_fp16)[name = tensor<string, []>("op_17151_cast_fp16")];
            tensor<int32, [4]> var_17155_begin_0 = const()[name = tensor<string, []>("op_17155_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17155_end_0 = const()[name = tensor<string, []>("op_17155_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17155_end_mask_0 = const()[name = tensor<string, []>("op_17155_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17155_cast_fp16 = slice_by_index(begin = var_17155_begin_0, end = var_17155_end_0, end_mask = var_17155_end_mask_0, x = k_849_cast_fp16)[name = tensor<string, []>("op_17155_cast_fp16")];
            tensor<int32, [4]> var_17159_begin_0 = const()[name = tensor<string, []>("op_17159_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17159_end_0 = const()[name = tensor<string, []>("op_17159_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17159_end_mask_0 = const()[name = tensor<string, []>("op_17159_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17159_cast_fp16 = slice_by_index(begin = var_17159_begin_0, end = var_17159_end_0, end_mask = var_17159_end_mask_0, x = q_1415_cast_fp16)[name = tensor<string, []>("op_17159_cast_fp16")];
            tensor<int32, [4]> var_17163_begin_0 = const()[name = tensor<string, []>("op_17163_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17163_end_0 = const()[name = tensor<string, []>("op_17163_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17163_end_mask_0 = const()[name = tensor<string, []>("op_17163_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17163_cast_fp16 = slice_by_index(begin = var_17163_begin_0, end = var_17163_end_0, end_mask = var_17163_end_mask_0, x = k_849_cast_fp16)[name = tensor<string, []>("op_17163_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17165_cast_fp16 = mul(x = var_17159_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17165_cast_fp16")];
            tensor<int32, [5]> var_17173 = const()[name = tensor<string, []>("op_17173"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1923_cast_fp16 = reshape(shape = var_17173, x = var_17159_cast_fp16)[name = tensor<string, []>("x_1923_cast_fp16")];
            tensor<int32, [2]> var_17175_split_sizes_0 = const()[name = tensor<string, []>("op_17175_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17175_axis_0 = const()[name = tensor<string, []>("op_17175_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17175_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17175_cast_fp16_1 = split(axis = var_17175_axis_0, split_sizes = var_17175_split_sizes_0, x = x_1923_cast_fp16)[name = tensor<string, []>("op_17175_cast_fp16")];
            tensor<int32, [1]> squeeze_564_axes_0 = const()[name = tensor<string, []>("squeeze_564_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_564_cast_fp16 = squeeze(axes = squeeze_564_axes_0, x = var_17175_cast_fp16_0)[name = tensor<string, []>("squeeze_564_cast_fp16")];
            tensor<int32, [1]> squeeze_565_axes_0 = const()[name = tensor<string, []>("squeeze_565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_565_cast_fp16 = squeeze(axes = squeeze_565_axes_0, x = var_17175_cast_fp16_1)[name = tensor<string, []>("squeeze_565_cast_fp16")];
            tensor<fp16, []> const_2307_promoted_to_fp16 = const()[name = tensor<string, []>("const_2307_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17178_cast_fp16 = mul(x = squeeze_565_cast_fp16, y = const_2307_promoted_to_fp16)[name = tensor<string, []>("op_17178_cast_fp16")];
            tensor<int32, []> x_1925_axis_0 = const()[name = tensor<string, []>("x_1925_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1925_cast_fp16 = stack(axis = x_1925_axis_0, values = (var_17178_cast_fp16, squeeze_564_cast_fp16))[name = tensor<string, []>("x_1925_cast_fp16")];
            tensor<int32, [4]> var_17184 = const()[name = tensor<string, []>("op_17184"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17185_cast_fp16 = reshape(shape = var_17184, x = x_1925_cast_fp16)[name = tensor<string, []>("op_17185_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17186_cast_fp16 = mul(x = var_17185_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17186_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_283_cast_fp16 = add(x = var_17165_cast_fp16, y = var_17186_cast_fp16)[name = tensor<string, []>("q_patches_283_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17188_cast_fp16 = mul(x = var_17163_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17188_cast_fp16")];
            tensor<int32, [5]> var_17196 = const()[name = tensor<string, []>("op_17196"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1929_cast_fp16 = reshape(shape = var_17196, x = var_17163_cast_fp16)[name = tensor<string, []>("x_1929_cast_fp16")];
            tensor<int32, [2]> var_17198_split_sizes_0 = const()[name = tensor<string, []>("op_17198_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17198_axis_0 = const()[name = tensor<string, []>("op_17198_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17198_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17198_cast_fp16_1 = split(axis = var_17198_axis_0, split_sizes = var_17198_split_sizes_0, x = x_1929_cast_fp16)[name = tensor<string, []>("op_17198_cast_fp16")];
            tensor<int32, [1]> squeeze_566_axes_0 = const()[name = tensor<string, []>("squeeze_566_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_566_cast_fp16 = squeeze(axes = squeeze_566_axes_0, x = var_17198_cast_fp16_0)[name = tensor<string, []>("squeeze_566_cast_fp16")];
            tensor<int32, [1]> squeeze_567_axes_0 = const()[name = tensor<string, []>("squeeze_567_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_567_cast_fp16 = squeeze(axes = squeeze_567_axes_0, x = var_17198_cast_fp16_1)[name = tensor<string, []>("squeeze_567_cast_fp16")];
            tensor<fp16, []> const_2315_promoted_to_fp16 = const()[name = tensor<string, []>("const_2315_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17201_cast_fp16 = mul(x = squeeze_567_cast_fp16, y = const_2315_promoted_to_fp16)[name = tensor<string, []>("op_17201_cast_fp16")];
            tensor<int32, []> x_1931_axis_0 = const()[name = tensor<string, []>("x_1931_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1931_cast_fp16 = stack(axis = x_1931_axis_0, values = (var_17201_cast_fp16, squeeze_566_cast_fp16))[name = tensor<string, []>("x_1931_cast_fp16")];
            tensor<int32, [4]> var_17207 = const()[name = tensor<string, []>("op_17207"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17208_cast_fp16 = reshape(shape = var_17207, x = x_1931_cast_fp16)[name = tensor<string, []>("op_17208_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17209_cast_fp16 = mul(x = var_17208_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17209_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_283_cast_fp16 = add(x = var_17188_cast_fp16, y = var_17209_cast_fp16)[name = tensor<string, []>("k_patches_283_cast_fp16")];
            tensor<bool, []> var_17212_interleave_0 = const()[name = tensor<string, []>("op_17212_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17212_cast_fp16 = concat(axis = var_20, interleave = var_17212_interleave_0, values = (var_17151_cast_fp16, q_patches_283_cast_fp16))[name = tensor<string, []>("op_17212_cast_fp16")];
            tensor<bool, []> var_17215_interleave_0 = const()[name = tensor<string, []>("op_17215_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17215_cast_fp16 = concat(axis = var_20, interleave = var_17215_interleave_0, values = (var_17155_cast_fp16, k_patches_283_cast_fp16))[name = tensor<string, []>("op_17215_cast_fp16")];
            tensor<fp16, []> var_17217_to_fp16 = const()[name = tensor<string, []>("op_17217_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1419_cast_fp16 = mul(x = var_17212_cast_fp16, y = var_17217_to_fp16)[name = tensor<string, []>("q_1419_cast_fp16")];
            tensor<bool, []> attn_565_transpose_x_1 = const()[name = tensor<string, []>("attn_565_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_565_transpose_y_1 = const()[name = tensor<string, []>("attn_565_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_565_cast_fp16 = matmul(transpose_x = attn_565_transpose_x_1, transpose_y = attn_565_transpose_y_1, x = q_1419_cast_fp16, y = var_17215_cast_fp16)[name = tensor<string, []>("attn_565_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_567_cast_fp16 = softmax(axis = var_21, x = attn_565_cast_fp16)[name = tensor<string, []>("attn_567_cast_fp16")];
            tensor<bool, []> var_17222_transpose_x_0 = const()[name = tensor<string, []>("op_17222_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_17222_transpose_y_0 = const()[name = tensor<string, []>("op_17222_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17222_cast_fp16 = matmul(transpose_x = var_17222_transpose_x_0, transpose_y = var_17222_transpose_y_0, x = attn_567_cast_fp16, y = v_851_cast_fp16)[name = tensor<string, []>("op_17222_cast_fp16")];
            tensor<int32, [1]> var_17223_axes_0 = const()[name = tensor<string, []>("op_17223_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_17223_cast_fp16 = squeeze(axes = var_17223_axes_0, x = var_17222_cast_fp16)[name = tensor<string, []>("op_17223_cast_fp16")];
            tensor<string, []> var_17229_pad_type_0 = const()[name = tensor<string, []>("op_17229_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17229_strides_0 = const()[name = tensor<string, []>("op_17229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17229_pad_0 = const()[name = tensor<string, []>("op_17229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17229_dilations_0 = const()[name = tensor<string, []>("op_17229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17229_groups_0 = const()[name = tensor<string, []>("op_17229_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160559936)))];
            tensor<fp16, [1, 64, 1, 197]> var_17229_cast_fp16 = conv(dilations = var_17229_dilations_0, groups = var_17229_groups_0, pad = var_17229_pad_0, pad_type = var_17229_pad_type_0, strides = var_17229_strides_0, weight = model_blocks_11_attn_q_projs_10_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17229_cast_fp16")];
            tensor<int32, [1]> var_17230_axes_0 = const()[name = tensor<string, []>("op_17230_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17230_cast_fp16 = squeeze(axes = var_17230_axes_0, x = var_17229_cast_fp16)[name = tensor<string, []>("op_17230_cast_fp16")];
            tensor<int32, [3]> q_1421_perm_0 = const()[name = tensor<string, []>("q_1421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17237_pad_type_0 = const()[name = tensor<string, []>("op_17237_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17237_strides_0 = const()[name = tensor<string, []>("op_17237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17237_pad_0 = const()[name = tensor<string, []>("op_17237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17237_dilations_0 = const()[name = tensor<string, []>("op_17237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17237_groups_0 = const()[name = tensor<string, []>("op_17237_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160658304)))];
            tensor<fp16, [1, 64, 1, 197]> var_17237_cast_fp16 = conv(dilations = var_17237_dilations_0, groups = var_17237_groups_0, pad = var_17237_pad_0, pad_type = var_17237_pad_type_0, strides = var_17237_strides_0, weight = model_blocks_11_attn_k_projs_10_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17237_cast_fp16")];
            tensor<int32, [1]> var_17238_axes_0 = const()[name = tensor<string, []>("op_17238_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17238_cast_fp16 = squeeze(axes = var_17238_axes_0, x = var_17237_cast_fp16)[name = tensor<string, []>("op_17238_cast_fp16")];
            tensor<int32, [3]> k_853_perm_0 = const()[name = tensor<string, []>("k_853_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17245_pad_type_0 = const()[name = tensor<string, []>("op_17245_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17245_strides_0 = const()[name = tensor<string, []>("op_17245_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17245_pad_0 = const()[name = tensor<string, []>("op_17245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17245_dilations_0 = const()[name = tensor<string, []>("op_17245_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17245_groups_0 = const()[name = tensor<string, []>("op_17245_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_10_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_10_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160756672)))];
            tensor<fp16, [1, 64, 1, 197]> var_17245_cast_fp16 = conv(dilations = var_17245_dilations_0, groups = var_17245_groups_0, pad = var_17245_pad_0, pad_type = var_17245_pad_type_0, strides = var_17245_strides_0, weight = model_blocks_11_attn_v_projs_10_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17245_cast_fp16")];
            tensor<int32, [1]> var_17246_axes_0 = const()[name = tensor<string, []>("op_17246_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17246_cast_fp16 = squeeze(axes = var_17246_axes_0, x = var_17245_cast_fp16)[name = tensor<string, []>("op_17246_cast_fp16")];
            tensor<int32, [3]> v_853_perm_0 = const()[name = tensor<string, []>("v_853_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160855040)))];
            tensor<fp16, [1, 197, 64]> q_1421_cast_fp16 = transpose(perm = q_1421_perm_0, x = var_17230_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, 197, 64]> q_1423_cast_fp16 = add(x = q_1421_cast_fp16, y = model_blocks_11_attn_q_biases_10_to_fp16)[name = tensor<string, []>("q_1423_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_10_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_10_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160855232)))];
            tensor<fp16, [1, 197, 64]> v_853_cast_fp16 = transpose(perm = v_853_perm_0, x = var_17246_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, 197, 64]> v_855_cast_fp16 = add(x = v_853_cast_fp16, y = model_blocks_11_attn_v_biases_10_to_fp16)[name = tensor<string, []>("v_855_cast_fp16")];
            tensor<int32, [1]> q_1425_axes_0 = const()[name = tensor<string, []>("q_1425_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1425_cast_fp16 = expand_dims(axes = q_1425_axes_0, x = q_1423_cast_fp16)[name = tensor<string, []>("q_1425_cast_fp16")];
            tensor<int32, [1]> k_855_axes_0 = const()[name = tensor<string, []>("k_855_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_853_cast_fp16 = transpose(perm = k_853_perm_0, x = var_17238_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 1, 197, 64]> k_855_cast_fp16 = expand_dims(axes = k_855_axes_0, x = k_853_cast_fp16)[name = tensor<string, []>("k_855_cast_fp16")];
            tensor<int32, [1]> v_857_axes_0 = const()[name = tensor<string, []>("v_857_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_857_cast_fp16 = expand_dims(axes = v_857_axes_0, x = v_855_cast_fp16)[name = tensor<string, []>("v_857_cast_fp16")];
            tensor<int32, [4]> var_17255_begin_0 = const()[name = tensor<string, []>("op_17255_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17255_end_0 = const()[name = tensor<string, []>("op_17255_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17255_end_mask_0 = const()[name = tensor<string, []>("op_17255_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17255_cast_fp16 = slice_by_index(begin = var_17255_begin_0, end = var_17255_end_0, end_mask = var_17255_end_mask_0, x = q_1425_cast_fp16)[name = tensor<string, []>("op_17255_cast_fp16")];
            tensor<int32, [4]> var_17259_begin_0 = const()[name = tensor<string, []>("op_17259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17259_end_0 = const()[name = tensor<string, []>("op_17259_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17259_end_mask_0 = const()[name = tensor<string, []>("op_17259_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17259_cast_fp16 = slice_by_index(begin = var_17259_begin_0, end = var_17259_end_0, end_mask = var_17259_end_mask_0, x = k_855_cast_fp16)[name = tensor<string, []>("op_17259_cast_fp16")];
            tensor<int32, [4]> var_17263_begin_0 = const()[name = tensor<string, []>("op_17263_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17263_end_0 = const()[name = tensor<string, []>("op_17263_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17263_end_mask_0 = const()[name = tensor<string, []>("op_17263_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17263_cast_fp16 = slice_by_index(begin = var_17263_begin_0, end = var_17263_end_0, end_mask = var_17263_end_mask_0, x = q_1425_cast_fp16)[name = tensor<string, []>("op_17263_cast_fp16")];
            tensor<int32, [4]> var_17267_begin_0 = const()[name = tensor<string, []>("op_17267_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17267_end_0 = const()[name = tensor<string, []>("op_17267_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17267_end_mask_0 = const()[name = tensor<string, []>("op_17267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17267_cast_fp16 = slice_by_index(begin = var_17267_begin_0, end = var_17267_end_0, end_mask = var_17267_end_mask_0, x = k_855_cast_fp16)[name = tensor<string, []>("op_17267_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17269_cast_fp16 = mul(x = var_17263_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17269_cast_fp16")];
            tensor<int32, [5]> var_17277 = const()[name = tensor<string, []>("op_17277"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1935_cast_fp16 = reshape(shape = var_17277, x = var_17263_cast_fp16)[name = tensor<string, []>("x_1935_cast_fp16")];
            tensor<int32, [2]> var_17279_split_sizes_0 = const()[name = tensor<string, []>("op_17279_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17279_axis_0 = const()[name = tensor<string, []>("op_17279_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17279_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17279_cast_fp16_1 = split(axis = var_17279_axis_0, split_sizes = var_17279_split_sizes_0, x = x_1935_cast_fp16)[name = tensor<string, []>("op_17279_cast_fp16")];
            tensor<int32, [1]> squeeze_568_axes_0 = const()[name = tensor<string, []>("squeeze_568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_568_cast_fp16 = squeeze(axes = squeeze_568_axes_0, x = var_17279_cast_fp16_0)[name = tensor<string, []>("squeeze_568_cast_fp16")];
            tensor<int32, [1]> squeeze_569_axes_0 = const()[name = tensor<string, []>("squeeze_569_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_569_cast_fp16 = squeeze(axes = squeeze_569_axes_0, x = var_17279_cast_fp16_1)[name = tensor<string, []>("squeeze_569_cast_fp16")];
            tensor<fp16, []> const_2323_promoted_to_fp16 = const()[name = tensor<string, []>("const_2323_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17282_cast_fp16 = mul(x = squeeze_569_cast_fp16, y = const_2323_promoted_to_fp16)[name = tensor<string, []>("op_17282_cast_fp16")];
            tensor<int32, []> x_1937_axis_0 = const()[name = tensor<string, []>("x_1937_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1937_cast_fp16 = stack(axis = x_1937_axis_0, values = (var_17282_cast_fp16, squeeze_568_cast_fp16))[name = tensor<string, []>("x_1937_cast_fp16")];
            tensor<int32, [4]> var_17288 = const()[name = tensor<string, []>("op_17288"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17289_cast_fp16 = reshape(shape = var_17288, x = x_1937_cast_fp16)[name = tensor<string, []>("op_17289_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17290_cast_fp16 = mul(x = var_17289_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17290_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_285_cast_fp16 = add(x = var_17269_cast_fp16, y = var_17290_cast_fp16)[name = tensor<string, []>("q_patches_285_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17292_cast_fp16 = mul(x = var_17267_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17292_cast_fp16")];
            tensor<int32, [5]> var_17300 = const()[name = tensor<string, []>("op_17300"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1941_cast_fp16 = reshape(shape = var_17300, x = var_17267_cast_fp16)[name = tensor<string, []>("x_1941_cast_fp16")];
            tensor<int32, [2]> var_17302_split_sizes_0 = const()[name = tensor<string, []>("op_17302_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17302_axis_0 = const()[name = tensor<string, []>("op_17302_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17302_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17302_cast_fp16_1 = split(axis = var_17302_axis_0, split_sizes = var_17302_split_sizes_0, x = x_1941_cast_fp16)[name = tensor<string, []>("op_17302_cast_fp16")];
            tensor<int32, [1]> squeeze_570_axes_0 = const()[name = tensor<string, []>("squeeze_570_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_570_cast_fp16 = squeeze(axes = squeeze_570_axes_0, x = var_17302_cast_fp16_0)[name = tensor<string, []>("squeeze_570_cast_fp16")];
            tensor<int32, [1]> squeeze_571_axes_0 = const()[name = tensor<string, []>("squeeze_571_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_571_cast_fp16 = squeeze(axes = squeeze_571_axes_0, x = var_17302_cast_fp16_1)[name = tensor<string, []>("squeeze_571_cast_fp16")];
            tensor<fp16, []> const_2331_promoted_to_fp16 = const()[name = tensor<string, []>("const_2331_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17305_cast_fp16 = mul(x = squeeze_571_cast_fp16, y = const_2331_promoted_to_fp16)[name = tensor<string, []>("op_17305_cast_fp16")];
            tensor<int32, []> x_1943_axis_0 = const()[name = tensor<string, []>("x_1943_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1943_cast_fp16 = stack(axis = x_1943_axis_0, values = (var_17305_cast_fp16, squeeze_570_cast_fp16))[name = tensor<string, []>("x_1943_cast_fp16")];
            tensor<int32, [4]> var_17311 = const()[name = tensor<string, []>("op_17311"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17312_cast_fp16 = reshape(shape = var_17311, x = x_1943_cast_fp16)[name = tensor<string, []>("op_17312_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17313_cast_fp16 = mul(x = var_17312_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17313_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_285_cast_fp16 = add(x = var_17292_cast_fp16, y = var_17313_cast_fp16)[name = tensor<string, []>("k_patches_285_cast_fp16")];
            tensor<bool, []> var_17316_interleave_0 = const()[name = tensor<string, []>("op_17316_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17316_cast_fp16 = concat(axis = var_20, interleave = var_17316_interleave_0, values = (var_17255_cast_fp16, q_patches_285_cast_fp16))[name = tensor<string, []>("op_17316_cast_fp16")];
            tensor<bool, []> var_17319_interleave_0 = const()[name = tensor<string, []>("op_17319_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17319_cast_fp16 = concat(axis = var_20, interleave = var_17319_interleave_0, values = (var_17259_cast_fp16, k_patches_285_cast_fp16))[name = tensor<string, []>("op_17319_cast_fp16")];
            tensor<fp16, []> var_17321_to_fp16 = const()[name = tensor<string, []>("op_17321_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_1429_cast_fp16 = mul(x = var_17316_cast_fp16, y = var_17321_to_fp16)[name = tensor<string, []>("q_1429_cast_fp16")];
            tensor<bool, []> attn_569_transpose_x_1 = const()[name = tensor<string, []>("attn_569_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_569_transpose_y_1 = const()[name = tensor<string, []>("attn_569_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_569_cast_fp16 = matmul(transpose_x = attn_569_transpose_x_1, transpose_y = attn_569_transpose_y_1, x = q_1429_cast_fp16, y = var_17319_cast_fp16)[name = tensor<string, []>("attn_569_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_571_cast_fp16 = softmax(axis = var_21, x = attn_569_cast_fp16)[name = tensor<string, []>("attn_571_cast_fp16")];
            tensor<bool, []> var_17326_transpose_x_0 = const()[name = tensor<string, []>("op_17326_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_17326_transpose_y_0 = const()[name = tensor<string, []>("op_17326_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17326_cast_fp16 = matmul(transpose_x = var_17326_transpose_x_0, transpose_y = var_17326_transpose_y_0, x = attn_571_cast_fp16, y = v_857_cast_fp16)[name = tensor<string, []>("op_17326_cast_fp16")];
            tensor<int32, [1]> var_17327_axes_0 = const()[name = tensor<string, []>("op_17327_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> var_17327_cast_fp16 = squeeze(axes = var_17327_axes_0, x = var_17326_cast_fp16)[name = tensor<string, []>("op_17327_cast_fp16")];
            tensor<string, []> var_17333_pad_type_0 = const()[name = tensor<string, []>("op_17333_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17333_strides_0 = const()[name = tensor<string, []>("op_17333_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17333_pad_0 = const()[name = tensor<string, []>("op_17333_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17333_dilations_0 = const()[name = tensor<string, []>("op_17333_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17333_groups_0 = const()[name = tensor<string, []>("op_17333_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_q_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160855424)))];
            tensor<fp16, [1, 64, 1, 197]> var_17333_cast_fp16 = conv(dilations = var_17333_dilations_0, groups = var_17333_groups_0, pad = var_17333_pad_0, pad_type = var_17333_pad_type_0, strides = var_17333_strides_0, weight = model_blocks_11_attn_q_projs_11_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17333_cast_fp16")];
            tensor<int32, [1]> var_17334_axes_0 = const()[name = tensor<string, []>("op_17334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17334_cast_fp16 = squeeze(axes = var_17334_axes_0, x = var_17333_cast_fp16)[name = tensor<string, []>("op_17334_cast_fp16")];
            tensor<int32, [3]> q_1431_perm_0 = const()[name = tensor<string, []>("q_1431_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17341_pad_type_0 = const()[name = tensor<string, []>("op_17341_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17341_strides_0 = const()[name = tensor<string, []>("op_17341_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17341_pad_0 = const()[name = tensor<string, []>("op_17341_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17341_dilations_0 = const()[name = tensor<string, []>("op_17341_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17341_groups_0 = const()[name = tensor<string, []>("op_17341_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_k_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_k_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160953792)))];
            tensor<fp16, [1, 64, 1, 197]> var_17341_cast_fp16 = conv(dilations = var_17341_dilations_0, groups = var_17341_groups_0, pad = var_17341_pad_0, pad_type = var_17341_pad_type_0, strides = var_17341_strides_0, weight = model_blocks_11_attn_k_projs_11_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17341_cast_fp16")];
            tensor<int32, [1]> var_17342_axes_0 = const()[name = tensor<string, []>("op_17342_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17342_cast_fp16 = squeeze(axes = var_17342_axes_0, x = var_17341_cast_fp16)[name = tensor<string, []>("op_17342_cast_fp16")];
            tensor<int32, [3]> k_859_perm_0 = const()[name = tensor<string, []>("k_859_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_17349_pad_type_0 = const()[name = tensor<string, []>("op_17349_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_17349_strides_0 = const()[name = tensor<string, []>("op_17349_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_17349_pad_0 = const()[name = tensor<string, []>("op_17349_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_17349_dilations_0 = const()[name = tensor<string, []>("op_17349_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17349_groups_0 = const()[name = tensor<string, []>("op_17349_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [64, 768, 1, 1]> model_blocks_11_attn_v_projs_11_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_projs_11_weight_to_fp16"), val = tensor<fp16, [64, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161052160)))];
            tensor<fp16, [1, 64, 1, 197]> var_17349_cast_fp16 = conv(dilations = var_17349_dilations_0, groups = var_17349_groups_0, pad = var_17349_pad_0, pad_type = var_17349_pad_type_0, strides = var_17349_strides_0, weight = model_blocks_11_attn_v_projs_11_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_17349_cast_fp16")];
            tensor<int32, [1]> var_17350_axes_0 = const()[name = tensor<string, []>("op_17350_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 64, 197]> var_17350_cast_fp16 = squeeze(axes = var_17350_axes_0, x = var_17349_cast_fp16)[name = tensor<string, []>("op_17350_cast_fp16")];
            tensor<int32, [3]> v_859_perm_0 = const()[name = tensor<string, []>("v_859_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [64]> model_blocks_11_attn_q_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_q_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161150528)))];
            tensor<fp16, [1, 197, 64]> q_1431_cast_fp16 = transpose(perm = q_1431_perm_0, x = var_17334_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 197, 64]> q_1433_cast_fp16 = add(x = q_1431_cast_fp16, y = model_blocks_11_attn_q_biases_11_to_fp16)[name = tensor<string, []>("q_1433_cast_fp16")];
            tensor<fp16, [64]> model_blocks_11_attn_v_biases_11_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_v_biases_11_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161150720)))];
            tensor<fp16, [1, 197, 64]> v_859_cast_fp16 = transpose(perm = v_859_perm_0, x = var_17350_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 197, 64]> v_861_cast_fp16 = add(x = v_859_cast_fp16, y = model_blocks_11_attn_v_biases_11_to_fp16)[name = tensor<string, []>("v_861_cast_fp16")];
            tensor<int32, [1]> q_1435_axes_0 = const()[name = tensor<string, []>("q_1435_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> q_1435_cast_fp16 = expand_dims(axes = q_1435_axes_0, x = q_1433_cast_fp16)[name = tensor<string, []>("q_1435_cast_fp16")];
            tensor<int32, [1]> k_861_axes_0 = const()[name = tensor<string, []>("k_861_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> k_859_cast_fp16 = transpose(perm = k_859_perm_0, x = var_17342_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, 1, 197, 64]> k_861_cast_fp16 = expand_dims(axes = k_861_axes_0, x = k_859_cast_fp16)[name = tensor<string, []>("k_861_cast_fp16")];
            tensor<int32, [1]> v_axes_0 = const()[name = tensor<string, []>("v_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 197, 64]> v_cast_fp16 = expand_dims(axes = v_axes_0, x = v_861_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
            tensor<int32, [4]> var_17359_begin_0 = const()[name = tensor<string, []>("op_17359_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17359_end_0 = const()[name = tensor<string, []>("op_17359_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17359_end_mask_0 = const()[name = tensor<string, []>("op_17359_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17359_cast_fp16 = slice_by_index(begin = var_17359_begin_0, end = var_17359_end_0, end_mask = var_17359_end_mask_0, x = q_1435_cast_fp16)[name = tensor<string, []>("op_17359_cast_fp16")];
            tensor<int32, [4]> var_17363_begin_0 = const()[name = tensor<string, []>("op_17363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17363_end_0 = const()[name = tensor<string, []>("op_17363_end_0"), val = tensor<int32, [4]>([1, 1, 1, 64])];
            tensor<bool, [4]> var_17363_end_mask_0 = const()[name = tensor<string, []>("op_17363_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 1, 64]> var_17363_cast_fp16 = slice_by_index(begin = var_17363_begin_0, end = var_17363_end_0, end_mask = var_17363_end_mask_0, x = k_861_cast_fp16)[name = tensor<string, []>("op_17363_cast_fp16")];
            tensor<int32, [4]> var_17367_begin_0 = const()[name = tensor<string, []>("op_17367_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17367_end_0 = const()[name = tensor<string, []>("op_17367_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17367_end_mask_0 = const()[name = tensor<string, []>("op_17367_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17367_cast_fp16 = slice_by_index(begin = var_17367_begin_0, end = var_17367_end_0, end_mask = var_17367_end_mask_0, x = q_1435_cast_fp16)[name = tensor<string, []>("op_17367_cast_fp16")];
            tensor<int32, [4]> var_17371_begin_0 = const()[name = tensor<string, []>("op_17371_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17371_end_0 = const()[name = tensor<string, []>("op_17371_end_0"), val = tensor<int32, [4]>([1, 1, 197, 64])];
            tensor<bool, [4]> var_17371_end_mask_0 = const()[name = tensor<string, []>("op_17371_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 196, 64]> var_17371_cast_fp16 = slice_by_index(begin = var_17371_begin_0, end = var_17371_end_0, end_mask = var_17371_end_mask_0, x = k_861_cast_fp16)[name = tensor<string, []>("op_17371_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17373_cast_fp16 = mul(x = var_17367_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17373_cast_fp16")];
            tensor<int32, [5]> var_17381 = const()[name = tensor<string, []>("op_17381"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1947_cast_fp16 = reshape(shape = var_17381, x = var_17367_cast_fp16)[name = tensor<string, []>("x_1947_cast_fp16")];
            tensor<int32, [2]> var_17383_split_sizes_0 = const()[name = tensor<string, []>("op_17383_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17383_axis_0 = const()[name = tensor<string, []>("op_17383_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17383_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17383_cast_fp16_1 = split(axis = var_17383_axis_0, split_sizes = var_17383_split_sizes_0, x = x_1947_cast_fp16)[name = tensor<string, []>("op_17383_cast_fp16")];
            tensor<int32, [1]> squeeze_572_axes_0 = const()[name = tensor<string, []>("squeeze_572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_572_cast_fp16 = squeeze(axes = squeeze_572_axes_0, x = var_17383_cast_fp16_0)[name = tensor<string, []>("squeeze_572_cast_fp16")];
            tensor<int32, [1]> squeeze_573_axes_0 = const()[name = tensor<string, []>("squeeze_573_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_573_cast_fp16 = squeeze(axes = squeeze_573_axes_0, x = var_17383_cast_fp16_1)[name = tensor<string, []>("squeeze_573_cast_fp16")];
            tensor<fp16, []> const_2339_promoted_to_fp16 = const()[name = tensor<string, []>("const_2339_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17386_cast_fp16 = mul(x = squeeze_573_cast_fp16, y = const_2339_promoted_to_fp16)[name = tensor<string, []>("op_17386_cast_fp16")];
            tensor<int32, []> x_1949_axis_0 = const()[name = tensor<string, []>("x_1949_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1949_cast_fp16 = stack(axis = x_1949_axis_0, values = (var_17386_cast_fp16, squeeze_572_cast_fp16))[name = tensor<string, []>("x_1949_cast_fp16")];
            tensor<int32, [4]> var_17392 = const()[name = tensor<string, []>("op_17392"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17393_cast_fp16 = reshape(shape = var_17392, x = x_1949_cast_fp16)[name = tensor<string, []>("op_17393_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17394_cast_fp16 = mul(x = var_17393_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17394_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> q_patches_cast_fp16 = add(x = var_17373_cast_fp16, y = var_17394_cast_fp16)[name = tensor<string, []>("q_patches_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17396_cast_fp16 = mul(x = var_17371_cast_fp16, y = rope_cos)[name = tensor<string, []>("op_17396_cast_fp16")];
            tensor<int32, [5]> var_17404 = const()[name = tensor<string, []>("op_17404"), val = tensor<int32, [5]>([1, 1, 196, 32, 2])];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1953_cast_fp16 = reshape(shape = var_17404, x = var_17371_cast_fp16)[name = tensor<string, []>("x_1953_cast_fp16")];
            tensor<int32, [2]> var_17406_split_sizes_0 = const()[name = tensor<string, []>("op_17406_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_17406_axis_0 = const()[name = tensor<string, []>("op_17406_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 1]> var_17406_cast_fp16_0, tensor<fp16, [1, 1, 196, 32, 1]> var_17406_cast_fp16_1 = split(axis = var_17406_axis_0, split_sizes = var_17406_split_sizes_0, x = x_1953_cast_fp16)[name = tensor<string, []>("op_17406_cast_fp16")];
            tensor<int32, [1]> squeeze_574_axes_0 = const()[name = tensor<string, []>("squeeze_574_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_574_cast_fp16 = squeeze(axes = squeeze_574_axes_0, x = var_17406_cast_fp16_0)[name = tensor<string, []>("squeeze_574_cast_fp16")];
            tensor<int32, [1]> squeeze_575_axes_0 = const()[name = tensor<string, []>("squeeze_575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 196, 32]> squeeze_575_cast_fp16 = squeeze(axes = squeeze_575_axes_0, x = var_17406_cast_fp16_1)[name = tensor<string, []>("squeeze_575_cast_fp16")];
            tensor<fp16, []> const_2347_promoted_to_fp16 = const()[name = tensor<string, []>("const_2347_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1, 196, 32]> var_17409_cast_fp16 = mul(x = squeeze_575_cast_fp16, y = const_2347_promoted_to_fp16)[name = tensor<string, []>("op_17409_cast_fp16")];
            tensor<int32, []> x_1955_axis_0 = const()[name = tensor<string, []>("x_1955_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 1, 196, 32, 2]> x_1955_cast_fp16 = stack(axis = x_1955_axis_0, values = (var_17409_cast_fp16, squeeze_574_cast_fp16))[name = tensor<string, []>("x_1955_cast_fp16")];
            tensor<int32, [4]> var_17415 = const()[name = tensor<string, []>("op_17415"), val = tensor<int32, [4]>([1, 1, 196, -1])];
            tensor<fp16, [1, 1, 196, 64]> var_17416_cast_fp16 = reshape(shape = var_17415, x = x_1955_cast_fp16)[name = tensor<string, []>("op_17416_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> var_17417_cast_fp16 = mul(x = var_17416_cast_fp16, y = rope_sin)[name = tensor<string, []>("op_17417_cast_fp16")];
            tensor<fp16, [1, 1, 196, 64]> k_patches_cast_fp16 = add(x = var_17396_cast_fp16, y = var_17417_cast_fp16)[name = tensor<string, []>("k_patches_cast_fp16")];
            tensor<bool, []> var_17420_interleave_0 = const()[name = tensor<string, []>("op_17420_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17420_cast_fp16 = concat(axis = var_20, interleave = var_17420_interleave_0, values = (var_17359_cast_fp16, q_patches_cast_fp16))[name = tensor<string, []>("op_17420_cast_fp16")];
            tensor<bool, []> var_17423_interleave_0 = const()[name = tensor<string, []>("op_17423_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17423_cast_fp16 = concat(axis = var_20, interleave = var_17423_interleave_0, values = (var_17363_cast_fp16, k_patches_cast_fp16))[name = tensor<string, []>("op_17423_cast_fp16")];
            tensor<fp16, []> var_17425_to_fp16 = const()[name = tensor<string, []>("op_17425_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 1, 197, 64]> q_cast_fp16 = mul(x = var_17420_cast_fp16, y = var_17425_to_fp16)[name = tensor<string, []>("q_cast_fp16")];
            tensor<bool, []> attn_573_transpose_x_1 = const()[name = tensor<string, []>("attn_573_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_573_transpose_y_1 = const()[name = tensor<string, []>("attn_573_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 197, 197]> attn_573_cast_fp16 = matmul(transpose_x = attn_573_transpose_x_1, transpose_y = attn_573_transpose_y_1, x = q_cast_fp16, y = var_17423_cast_fp16)[name = tensor<string, []>("attn_573_cast_fp16")];
            tensor<fp16, [1, 1, 197, 197]> attn_575_cast_fp16 = softmax(axis = var_21, x = attn_573_cast_fp16)[name = tensor<string, []>("attn_575_cast_fp16")];
            tensor<bool, []> var_17430_transpose_x_0 = const()[name = tensor<string, []>("op_17430_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_17430_transpose_y_0 = const()[name = tensor<string, []>("op_17430_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 197, 64]> var_17430_cast_fp16 = matmul(transpose_x = var_17430_transpose_x_0, transpose_y = var_17430_transpose_y_0, x = attn_575_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("op_17430_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 197, 64]> out_cast_fp16 = squeeze(axes = out_axes_0, x = var_17430_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<bool, []> input_203_interleave_0 = const()[name = tensor<string, []>("input_203_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 197, 768]> input_203_cast_fp16 = concat(axis = var_21, interleave = input_203_interleave_0, values = (var_16287_cast_fp16, var_16391_cast_fp16, var_16495_cast_fp16, var_16599_cast_fp16, var_16703_cast_fp16, var_16807_cast_fp16, var_16911_cast_fp16, var_17015_cast_fp16, var_17119_cast_fp16, var_17223_cast_fp16, var_17327_cast_fp16, out_cast_fp16))[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<int32, [1]> x_1957_axes_0 = const()[name = tensor<string, []>("x_1957_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_11_attn_inner_attn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_inner_attn_ln_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161150912)))];
            tensor<fp16, [768]> model_blocks_11_attn_inner_attn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_inner_attn_ln_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161152512)))];
            tensor<fp16, [1, 197, 768]> x_1957_cast_fp16 = layer_norm(axes = x_1957_axes_0, beta = model_blocks_11_attn_inner_attn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_11_attn_inner_attn_ln_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_1957_cast_fp16")];
            tensor<int32, [4]> var_17438 = const()[name = tensor<string, []>("op_17438"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1959_cast_fp16 = reshape(shape = var_17438, x = x_1957_cast_fp16)[name = tensor<string, []>("x_1959_cast_fp16")];
            tensor<int32, [4]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1961_pad_type_0 = const()[name = tensor<string, []>("x_1961_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1961_strides_0 = const()[name = tensor<string, []>("x_1961_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1961_pad_0 = const()[name = tensor<string, []>("x_1961_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1961_dilations_0 = const()[name = tensor<string, []>("x_1961_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1961_groups_0 = const()[name = tensor<string, []>("x_1961_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 768, 1, 1]> model_blocks_11_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161154112)))];
            tensor<fp16, [768]> model_blocks_11_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162333824)))];
            tensor<fp16, [1, 768, 1, 197]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_1959_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, 768, 1, 197]> x_1961_cast_fp16 = conv(bias = model_blocks_11_attn_proj_bias_to_fp16, dilations = x_1961_dilations_0, groups = x_1961_groups_0, pad = x_1961_pad_0, pad_type = x_1961_pad_type_0, strides = x_1961_strides_0, weight = model_blocks_11_attn_proj_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("x_1961_cast_fp16")];
            tensor<int32, [4]> x_1963_perm_0 = const()[name = tensor<string, []>("x_1963_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_17449 = const()[name = tensor<string, []>("op_17449"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1963_cast_fp16 = transpose(perm = x_1963_perm_0, x = x_1961_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 197, 768]> var_17450_cast_fp16 = reshape(shape = var_17449, x = x_1963_cast_fp16)[name = tensor<string, []>("op_17450_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_207_cast_fp16 = add(x = input_199_cast_fp16, y = var_17450_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<int32, [1]> x_1965_axes_0 = const()[name = tensor<string, []>("x_1965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_blocks_11_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162335424)))];
            tensor<fp16, [768]> model_blocks_11_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162337024)))];
            tensor<fp16, [1, 197, 768]> x_1965_cast_fp16 = layer_norm(axes = x_1965_axes_0, beta = model_blocks_11_norm2_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_11_norm2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("x_1965_cast_fp16")];
            tensor<int32, [4]> var_17462 = const()[name = tensor<string, []>("op_17462"), val = tensor<int32, [4]>([1, 197, 1, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1967_cast_fp16 = reshape(shape = var_17462, x = x_1965_cast_fp16)[name = tensor<string, []>("x_1967_cast_fp16")];
            tensor<int32, [4]> input_209_perm_0 = const()[name = tensor<string, []>("input_209_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> input_211_pad_type_0 = const()[name = tensor<string, []>("input_211_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_211_strides_0 = const()[name = tensor<string, []>("input_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_211_pad_0 = const()[name = tensor<string, []>("input_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_211_dilations_0 = const()[name = tensor<string, []>("input_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_211_groups_0 = const()[name = tensor<string, []>("input_211_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_11_mlp_w1_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w1_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162338624)))];
            tensor<fp16, [2048]> model_blocks_11_mlp_w1_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165484416)))];
            tensor<fp16, [1, 768, 1, 197]> input_209_cast_fp16 = transpose(perm = input_209_perm_0, x = x_1967_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, 2048, 1, 197]> input_211_cast_fp16 = conv(bias = model_blocks_11_mlp_w1_bias_to_fp16, dilations = input_211_dilations_0, groups = input_211_groups_0, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = input_211_strides_0, weight = model_blocks_11_mlp_w1_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<string, []> x2_pad_type_0 = const()[name = tensor<string, []>("x2_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x2_strides_0 = const()[name = tensor<string, []>("x2_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x2_pad_0 = const()[name = tensor<string, []>("x2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x2_dilations_0 = const()[name = tensor<string, []>("x2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x2_groups_0 = const()[name = tensor<string, []>("x2_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [2048, 768, 1, 1]> model_blocks_11_mlp_w2_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w2_weight_to_fp16"), val = tensor<fp16, [2048, 768, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165488576)))];
            tensor<fp16, [2048]> model_blocks_11_mlp_w2_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168634368)))];
            tensor<fp16, [1, 2048, 1, 197]> x2_cast_fp16 = conv(bias = model_blocks_11_mlp_w2_bias_to_fp16, dilations = x2_dilations_0, groups = x2_groups_0, pad = x2_pad_0, pad_type = x2_pad_type_0, strides = x2_strides_0, weight = model_blocks_11_mlp_w2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("x2_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> var_17479_cast_fp16 = silu(x = input_211_cast_fp16)[name = tensor<string, []>("op_17479_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 197]> hidden_89_cast_fp16 = mul(x = var_17479_cast_fp16, y = x2_cast_fp16)[name = tensor<string, []>("hidden_89_cast_fp16")];
            tensor<int32, [4]> hidden_91_perm_0 = const()[name = tensor<string, []>("hidden_91_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_17482 = const()[name = tensor<string, []>("op_17482"), val = tensor<int32, [3]>([1, 197, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_91_cast_fp16 = transpose(perm = hidden_91_perm_0, x = hidden_89_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<fp16, [1, 197, 2048]> input_213_cast_fp16 = reshape(shape = var_17482, x = hidden_91_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<int32, [1]> hidden_93_axes_0 = const()[name = tensor<string, []>("hidden_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2048]> model_blocks_11_mlp_ffn_ln_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_ffn_ln_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168638528)))];
            tensor<fp16, [2048]> model_blocks_11_mlp_ffn_ln_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_ffn_ln_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168642688)))];
            tensor<fp16, [1, 197, 2048]> hidden_93_cast_fp16 = layer_norm(axes = hidden_93_axes_0, beta = model_blocks_11_mlp_ffn_ln_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_blocks_11_mlp_ffn_ln_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("hidden_93_cast_fp16")];
            tensor<int32, [4]> var_17488 = const()[name = tensor<string, []>("op_17488"), val = tensor<int32, [4]>([1, 197, 1, 2048])];
            tensor<fp16, [1, 197, 1, 2048]> hidden_cast_fp16 = reshape(shape = var_17488, x = hidden_93_cast_fp16)[name = tensor<string, []>("hidden_cast_fp16")];
            tensor<int32, [4]> input_215_perm_0 = const()[name = tensor<string, []>("input_215_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<string, []> x_1969_pad_type_0 = const()[name = tensor<string, []>("x_1969_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_1969_strides_0 = const()[name = tensor<string, []>("x_1969_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_1969_pad_0 = const()[name = tensor<string, []>("x_1969_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_1969_dilations_0 = const()[name = tensor<string, []>("x_1969_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_1969_groups_0 = const()[name = tensor<string, []>("x_1969_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 2048, 1, 1]> model_blocks_11_mlp_w3_weight_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w3_weight_to_fp16"), val = tensor<fp16, [768, 2048, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168646848)))];
            tensor<fp16, [768]> model_blocks_11_mlp_w3_bias_to_fp16 = const()[name = tensor<string, []>("model_blocks_11_mlp_w3_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171792640)))];
            tensor<fp16, [1, 2048, 1, 197]> input_215_cast_fp16 = transpose(perm = input_215_perm_0, x = hidden_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp16, [1, 768, 1, 197]> x_1969_cast_fp16 = conv(bias = model_blocks_11_mlp_w3_bias_to_fp16, dilations = x_1969_dilations_0, groups = x_1969_groups_0, pad = x_1969_pad_0, pad_type = x_1969_pad_type_0, strides = x_1969_strides_0, weight = model_blocks_11_mlp_w3_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_1969_cast_fp16")];
            tensor<int32, [4]> x_1971_perm_0 = const()[name = tensor<string, []>("x_1971_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_17499 = const()[name = tensor<string, []>("op_17499"), val = tensor<int32, [3]>([1, 197, 768])];
            tensor<fp16, [1, 197, 1, 768]> x_1971_cast_fp16 = transpose(perm = x_1971_perm_0, x = x_1969_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 197, 768]> var_17500_cast_fp16 = reshape(shape = var_17499, x = x_1971_cast_fp16)[name = tensor<string, []>("op_17500_cast_fp16")];
            tensor<fp16, [1, 197, 768]> input_217_cast_fp16 = add(x = input_207_cast_fp16, y = var_17500_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> model_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_norm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171794240)))];
            tensor<fp16, [768]> model_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_norm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171795840)))];
            tensor<fp16, [1, 197, 768]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = model_norm_bias_to_fp16, epsilon = var_6_to_fp16, gamma = model_norm_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<int32, [3]> input_begin_0 = const()[name = tensor<string, []>("input_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_end_0 = const()[name = tensor<string, []>("input_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> input_end_mask_0 = const()[name = tensor<string, []>("input_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> input_squeeze_mask_0 = const()[name = tensor<string, []>("input_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> input_cast_fp16 = slice_by_index(begin = input_begin_0, end = input_end_0, end_mask = input_end_mask_0, squeeze_mask = input_squeeze_mask_0, x = x_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<int32, [1]> var_17508 = const()[name = tensor<string, []>("op_17508"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> var_17509_cast_fp16 = reduce_l2_norm(axes = var_17508, keep_dims = var_15, x = input_cast_fp16)[name = tensor<string, []>("op_17509_cast_fp16")];
            tensor<fp16, []> var_5_to_fp16 = const()[name = tensor<string, []>("op_5_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
            tensor<fp16, [1, 1]> var_17510_cast_fp16 = maximum(x = var_17509_cast_fp16, y = var_5_to_fp16)[name = tensor<string, []>("op_17510_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = tensor<string, []>("denom_reps_0"), val = tensor<int32, [2]>([1, 768])];
            tensor<fp16, [1, 768]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_17510_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
            tensor<fp16, [1, 768]> image_embeddings = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_17512_cast_fp16")];
        } -> (image_embeddings);
}