diff --git "a/openai_whisper-large-v2_1050MB/AudioEncoder.mlmodelc/model.mil" "b/openai_whisper-large-v2_1050MB/AudioEncoder.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/openai_whisper-large-v2_1050MB/AudioEncoder.mlmodelc/model.mil"
@@ -0,0 +1,3086 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})]
+{
+    func main<ios17>(tensor<fp16, [1, 80, 1, 3000]> melspectrogram_features) {
+            tensor<int32, [2]> var_90 = const()[name = tensor<string, []>("op_90"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_96 = const()[name = tensor<string, []>("op_96"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, []>(1)];
+            tensor<string, []> var_106_pad_type_0 = const()[name = tensor<string, []>("op_106_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> var_106_pad_0 = const()[name = tensor<string, []>("op_106_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
+            tensor<fp16, [1280, 80, 1, 3]> var_81_to_fp16 = const()[name = tensor<string, []>("op_81_to_fp16"), val = tensor<fp16, [1280, 80, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1280]> var_87_to_fp16 = const()[name = tensor<string, []>("op_87_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(614528)))];
+            tensor<fp16, [1, 1280, 1, 3000]> var_106_cast_fp16 = conv(bias = var_87_to_fp16, dilations = var_96, groups = var_101, pad = var_106_pad_0, pad_type = var_106_pad_type_0, strides = var_90, weight = var_81_to_fp16, x = melspectrogram_features)[name = tensor<string, []>("op_106_cast_fp16")];
+            tensor<string, []> hidden_states_1_mode_0 = const()[name = tensor<string, []>("hidden_states_1_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1280, 1, 3000]> hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_106_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
+            tensor<int32, [2]> var_130 = const()[name = tensor<string, []>("op_130"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [2]> var_136 = const()[name = tensor<string, []>("op_136"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> var_141 = const()[name = tensor<string, []>("op_141"), val = tensor<int32, []>(1)];
+            tensor<string, []> var_146_pad_type_0 = const()[name = tensor<string, []>("op_146_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> var_146_pad_0 = const()[name = tensor<string, []>("op_146_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
+            tensor<fp16, [1280, 1280, 1, 3]> var_121_to_fp16 = const()[name = tensor<string, []>("op_121_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617152)))];
+            tensor<fp16, [1280]> var_127_to_fp16 = const()[name = tensor<string, []>("op_127_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10447616)))];
+            tensor<fp16, [1, 1280, 1, 1500]> var_146_cast_fp16 = conv(bias = var_127_to_fp16, dilations = var_136, groups = var_141, pad = var_146_pad_0, pad_type = var_146_pad_type_0, strides = var_130, weight = var_121_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_146_cast_fp16")];
+            tensor<string, []> hidden_states_3_mode_0 = const()[name = tensor<string, []>("hidden_states_3_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_146_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> var_164_to_fp16 = const()[name = tensor<string, []>("op_164_to_fp16"), val = tensor<fp16, [1, 1280, 1, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10450240)))];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_164_to_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
+            tensor<int32, []> var_178 = const()[name = tensor<string, []>("op_178"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_180 = const()[name = tensor<string, []>("op_180"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_181 = const()[name = tensor<string, []>("op_181"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_191 = const()[name = tensor<string, []>("op_191"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_191, keep_dims = var_181, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")];
+            tensor<int32, [1]> var_195 = const()[name = tensor<string, []>("op_195"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_196_cast_fp16 = reduce_mean(axes = var_195, keep_dims = var_181, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
+            tensor<fp16, []> var_197_to_fp16 = const()[name = tensor<string, []>("op_197_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_198_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_to_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
+            tensor<fp32, []> denom_1_epsilon_0 = const()[name = tensor<string, []>("denom_1_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0, x = var_198_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14290304)))];
+            tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14292928)))];
+            tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14295552)))];
+            tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14298176)))];
+            tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
+            tensor<int32, [2]> var_213 = const()[name = tensor<string, []>("op_213"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_215 = const()[name = tensor<string, []>("op_215"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14300800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15120064))), name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15120192)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_215, groups = var_180, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_213, weight = layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
+            tensor<int32, [2]> var_219 = const()[name = tensor<string, []>("op_219"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_1_pad_type_0 = const()[name = tensor<string, []>("key_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_1_pad_0 = const()[name = tensor<string, []>("key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15122816))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15942080))), name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_1_cast_fp16 = conv(dilations = var_221, groups = var_180, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = var_219, weight = layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<int32, [2]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_228 = const()[name = tensor<string, []>("op_228"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_1_pad_type_0 = const()[name = tensor<string, []>("value_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_1_pad_0 = const()[name = tensor<string, []>("value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15942208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16761472))), name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16761600)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_228, groups = var_180, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = var_226, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
+            tensor<int32, [4]> var_232 = const()[name = tensor<string, []>("op_232"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_233_cast_fp16 = reshape(shape = var_232, x = query_1_cast_fp16)[name = tensor<string, []>("op_233_cast_fp16")];
+            tensor<fp16, []> var_234_to_fp16 = const()[name = tensor<string, []>("op_234_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_235_cast_fp16 = mul(x = var_233_cast_fp16, y = var_234_to_fp16)[name = tensor<string, []>("op_235_cast_fp16")];
+            tensor<int32, [4]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_237_cast_fp16 = reshape(shape = var_236, x = key_1_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
+            tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_235_cast_fp16, y = var_237_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_240_cast_fp16 = softmax(axis = var_178, x = mh_w_1_cast_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
+            tensor<int32, [4]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_242_cast_fp16 = reshape(shape = var_241, x = value_1_cast_fp16)[name = tensor<string, []>("op_242_cast_fp16")];
+            tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_242_cast_fp16, y = var_240_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
+            tensor<int32, [4]> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_1_cast_fp16 = reshape(shape = var_245, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [2]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_251 = const()[name = tensor<string, []>("op_251"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_3_pad_type_0 = const()[name = tensor<string, []>("obj_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_3_pad_0 = const()[name = tensor<string, []>("obj_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16764224))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17583488))), name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17583616)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_251, groups = var_180, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = var_249, weight = layers_0_self_attn_o_proj_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
+            tensor<int32, [1]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_257, keep_dims = var_181, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")];
+            tensor<int32, [1]> var_261 = const()[name = tensor<string, []>("op_261"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_262_cast_fp16 = reduce_mean(axes = var_261, keep_dims = var_181, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_262_cast_fp16")];
+            tensor<fp16, []> var_263_to_fp16 = const()[name = tensor<string, []>("op_263_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_264_cast_fp16 = add(x = var_262_cast_fp16, y = var_263_to_fp16)[name = tensor<string, []>("op_264_cast_fp16")];
+            tensor<fp32, []> denom_3_epsilon_0 = const()[name = tensor<string, []>("denom_3_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0, x = var_264_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<fp16, [1280]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17586240)))];
+            tensor<fp16, [1280]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17588864)))];
+            tensor<fp16, []> input_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [2]> var_275 = const()[name = tensor<string, []>("op_275"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17591488)))];
+            tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30698752)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_5_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_277, groups = var_180, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_275, weight = layers_0_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [2]> var_283 = const()[name = tensor<string, []>("op_283"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_285 = const()[name = tensor<string, []>("op_285"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30709056)))];
+            tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43816320)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_285, groups = var_180, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_283, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
+            tensor<int32, []> var_296 = const()[name = tensor<string, []>("op_296"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_298 = const()[name = tensor<string, []>("op_298"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_299 = const()[name = tensor<string, []>("op_299"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_309 = const()[name = tensor<string, []>("op_309"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_309, keep_dims = var_299, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")];
+            tensor<int32, [1]> var_313 = const()[name = tensor<string, []>("op_313"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_314_cast_fp16 = reduce_mean(axes = var_313, keep_dims = var_299, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_314_cast_fp16")];
+            tensor<fp16, []> var_315_to_fp16 = const()[name = tensor<string, []>("op_315_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_316_cast_fp16 = add(x = var_314_cast_fp16, y = var_315_to_fp16)[name = tensor<string, []>("op_316_cast_fp16")];
+            tensor<fp32, []> denom_5_epsilon_0 = const()[name = tensor<string, []>("denom_5_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0, x = var_316_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1280]> obj_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43818944)))];
+            tensor<fp16, [1280]> obj_5_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43821568)))];
+            tensor<fp16, []> obj_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, [2]> var_331 = const()[name = tensor<string, []>("op_331"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43824192)))];
+            tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47101056)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_333, groups = var_298, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_331, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
+            tensor<int32, [2]> var_337 = const()[name = tensor<string, []>("op_337"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47103680))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47922944))), name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = var_339, groups = var_298, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_337, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<int32, [2]> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_346 = const()[name = tensor<string, []>("op_346"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47923072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48742336))), name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48742464)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_346, groups = var_298, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_344, weight = layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_351_cast_fp16 = reshape(shape = var_350, x = query_3_cast_fp16)[name = tensor<string, []>("op_351_cast_fp16")];
+            tensor<fp16, []> var_352_to_fp16 = const()[name = tensor<string, []>("op_352_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_353_cast_fp16 = mul(x = var_351_cast_fp16, y = var_352_to_fp16)[name = tensor<string, []>("op_353_cast_fp16")];
+            tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_355_cast_fp16 = reshape(shape = var_354, x = key_3_cast_fp16)[name = tensor<string, []>("op_355_cast_fp16")];
+            tensor<bool, []> mh_w_3_transpose_x_0 = const()[name = tensor<string, []>("mh_w_3_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_3_transpose_y_0 = const()[name = tensor<string, []>("mh_w_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_353_cast_fp16, y = var_355_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_358_cast_fp16 = softmax(axis = var_296, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
+            tensor<int32, [4]> var_359 = const()[name = tensor<string, []>("op_359"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_360_cast_fp16 = reshape(shape = var_359, x = value_3_cast_fp16)[name = tensor<string, []>("op_360_cast_fp16")];
+            tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_360_cast_fp16, y = var_358_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
+            tensor<int32, [4]> var_363 = const()[name = tensor<string, []>("op_363"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_9_cast_fp16 = reshape(shape = var_363, x = attn_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<int32, [2]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48745088)))];
+            tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52021952)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_369, groups = var_298, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_367, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
+            tensor<int32, [1]> var_375 = const()[name = tensor<string, []>("op_375"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_375, keep_dims = var_299, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")];
+            tensor<int32, [1]> var_379 = const()[name = tensor<string, []>("op_379"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_380_cast_fp16 = reduce_mean(axes = var_379, keep_dims = var_299, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_380_cast_fp16")];
+            tensor<fp16, []> var_381_to_fp16 = const()[name = tensor<string, []>("op_381_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_382_cast_fp16 = add(x = var_380_cast_fp16, y = var_381_to_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
+            tensor<fp32, []> denom_7_epsilon_0 = const()[name = tensor<string, []>("denom_7_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_382_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1280]> input_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_11_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52024576)))];
+            tensor<fp16, [1280]> input_11_beta_0_to_fp16 = const()[name = tensor<string, []>("input_11_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52027200)))];
+            tensor<fp16, []> input_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<int32, [2]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_395 = const()[name = tensor<string, []>("op_395"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52029824))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55306688))), name = tensor<string, []>("layers_1_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55306816)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_13_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_395, groups = var_298, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_393, weight = layers_1_fc1_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<string, []> input_15_mode_0 = const()[name = tensor<string, []>("input_15_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [2]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_403 = const()[name = tensor<string, []>("op_403"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55317120))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58593984))), name = tensor<string, []>("layers_1_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58594112)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_403, groups = var_298, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_401, weight = layers_1_fc2_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
+            tensor<int32, []> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_416 = const()[name = tensor<string, []>("op_416"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_427 = const()[name = tensor<string, []>("op_427"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_427, keep_dims = var_417, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")];
+            tensor<int32, [1]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_432_cast_fp16 = reduce_mean(axes = var_431, keep_dims = var_417, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_432_cast_fp16")];
+            tensor<fp16, []> var_433_to_fp16 = const()[name = tensor<string, []>("op_433_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_434_cast_fp16 = add(x = var_432_cast_fp16, y = var_433_to_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<fp32, []> denom_9_epsilon_0 = const()[name = tensor<string, []>("denom_9_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_434_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58596736)))];
+            tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58599360)))];
+            tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
+            tensor<int32, [2]> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58601984))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59421248))), name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59421376)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_451, groups = var_416, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_449, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
+            tensor<int32, [2]> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_457 = const()[name = tensor<string, []>("op_457"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_5_pad_type_0 = const()[name = tensor<string, []>("key_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_5_pad_0 = const()[name = tensor<string, []>("key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59424000))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60243264))), name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_5_cast_fp16 = conv(dilations = var_457, groups = var_416, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = var_455, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<int32, [2]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_464 = const()[name = tensor<string, []>("op_464"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_5_pad_type_0 = const()[name = tensor<string, []>("value_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_5_pad_0 = const()[name = tensor<string, []>("value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60243392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61062656))), name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61062784)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_464, groups = var_416, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = var_462, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
+            tensor<int32, [4]> var_468 = const()[name = tensor<string, []>("op_468"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_469_cast_fp16 = reshape(shape = var_468, x = query_5_cast_fp16)[name = tensor<string, []>("op_469_cast_fp16")];
+            tensor<fp16, []> var_470_to_fp16 = const()[name = tensor<string, []>("op_470_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_471_cast_fp16 = mul(x = var_469_cast_fp16, y = var_470_to_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
+            tensor<int32, [4]> var_472 = const()[name = tensor<string, []>("op_472"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_473_cast_fp16 = reshape(shape = var_472, x = key_5_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
+            tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_471_cast_fp16, y = var_473_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_476_cast_fp16 = softmax(axis = var_414, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
+            tensor<int32, [4]> var_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
+            tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_478_cast_fp16, y = var_476_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
+            tensor<int32, [4]> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_17_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<int32, [2]> var_485 = const()[name = tensor<string, []>("op_485"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_487 = const()[name = tensor<string, []>("op_487"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61065408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61884672))), name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61884800)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_487, groups = var_416, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_485, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
+            tensor<int32, [1]> var_493 = const()[name = tensor<string, []>("op_493"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_493, keep_dims = var_417, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
+            tensor<int32, [1]> var_497 = const()[name = tensor<string, []>("op_497"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_498_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_417, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_498_cast_fp16")];
+            tensor<fp16, []> var_499_to_fp16 = const()[name = tensor<string, []>("op_499_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_500_cast_fp16 = add(x = var_498_cast_fp16, y = var_499_to_fp16)[name = tensor<string, []>("op_500_cast_fp16")];
+            tensor<fp32, []> denom_11_epsilon_0 = const()[name = tensor<string, []>("denom_11_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_500_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
+            tensor<fp16, [1280]> input_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_19_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61887424)))];
+            tensor<fp16, [1280]> input_19_beta_0_to_fp16 = const()[name = tensor<string, []>("input_19_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61890048)))];
+            tensor<fp16, []> input_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<int32, [2]> var_511 = const()[name = tensor<string, []>("op_511"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_513 = const()[name = tensor<string, []>("op_513"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61892672))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65169536))), name = tensor<string, []>("layers_2_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65169664)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_513, groups = var_416, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_511, weight = layers_2_fc1_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<int32, [2]> var_519 = const()[name = tensor<string, []>("op_519"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_521 = const()[name = tensor<string, []>("op_521"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65179968))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68456832))), name = tensor<string, []>("layers_2_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68456960)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_521, groups = var_416, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_519, weight = layers_2_fc2_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
+            tensor<int32, []> var_532 = const()[name = tensor<string, []>("op_532"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_534 = const()[name = tensor<string, []>("op_534"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_535 = const()[name = tensor<string, []>("op_535"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_545 = const()[name = tensor<string, []>("op_545"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_545, keep_dims = var_535, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_13_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_sq_13_cast_fp16")];
+            tensor<int32, [1]> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_550_cast_fp16 = reduce_mean(axes = var_549, keep_dims = var_535, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_550_cast_fp16")];
+            tensor<fp16, []> var_551_to_fp16 = const()[name = tensor<string, []>("op_551_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_552_cast_fp16 = add(x = var_550_cast_fp16, y = var_551_to_fp16)[name = tensor<string, []>("op_552_cast_fp16")];
+            tensor<fp32, []> denom_13_epsilon_0 = const()[name = tensor<string, []>("denom_13_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_552_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
+            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68459584)))];
+            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68462208)))];
+            tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
+            tensor<int32, [2]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_569 = const()[name = tensor<string, []>("op_569"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68464832))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69284096))), name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69284224)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_7_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_569, groups = var_534, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_567, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
+            tensor<int32, [2]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69286848)))];
+            tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = conv(dilations = var_575, groups = var_534, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_573, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<int32, [2]> var_580 = const()[name = tensor<string, []>("op_580"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_582 = const()[name = tensor<string, []>("op_582"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72563712))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73382976))), name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73383104)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_582, groups = var_534, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_580, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
+            tensor<int32, [4]> var_586 = const()[name = tensor<string, []>("op_586"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_587_cast_fp16 = reshape(shape = var_586, x = query_7_cast_fp16)[name = tensor<string, []>("op_587_cast_fp16")];
+            tensor<fp16, []> var_588_to_fp16 = const()[name = tensor<string, []>("op_588_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_589_cast_fp16 = mul(x = var_587_cast_fp16, y = var_588_to_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
+            tensor<int32, [4]> var_590 = const()[name = tensor<string, []>("op_590"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_591_cast_fp16 = reshape(shape = var_590, x = key_7_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
+            tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_589_cast_fp16, y = var_591_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_594_cast_fp16 = softmax(axis = var_532, x = mh_w_7_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
+            tensor<int32, [4]> var_595 = const()[name = tensor<string, []>("op_595"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_596_cast_fp16 = reshape(shape = var_595, x = value_7_cast_fp16)[name = tensor<string, []>("op_596_cast_fp16")];
+            tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_596_cast_fp16, y = var_594_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
+            tensor<int32, [4]> var_599 = const()[name = tensor<string, []>("op_599"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_25_cast_fp16 = reshape(shape = var_599, x = attn_7_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<int32, [2]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_605 = const()[name = tensor<string, []>("op_605"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_15_pad_type_0 = const()[name = tensor<string, []>("obj_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_15_pad_0 = const()[name = tensor<string, []>("obj_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73385728))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74204992))), name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74205120)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_15_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_605, groups = var_534, pad = obj_15_pad_0, pad_type = obj_15_pad_type_0, strides = var_603, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
+            tensor<int32, [1]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_611, keep_dims = var_535, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_15_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_sq_15_cast_fp16")];
+            tensor<int32, [1]> var_615 = const()[name = tensor<string, []>("op_615"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_616_cast_fp16 = reduce_mean(axes = var_615, keep_dims = var_535, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_616_cast_fp16")];
+            tensor<fp16, []> var_617_to_fp16 = const()[name = tensor<string, []>("op_617_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_618_cast_fp16 = add(x = var_616_cast_fp16, y = var_617_to_fp16)[name = tensor<string, []>("op_618_cast_fp16")];
+            tensor<fp32, []> denom_15_epsilon_0 = const()[name = tensor<string, []>("denom_15_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_618_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
+            tensor<fp16, [1280]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74207744)))];
+            tensor<fp16, [1280]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74210368)))];
+            tensor<fp16, []> input_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<int32, [2]> var_629 = const()[name = tensor<string, []>("op_629"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_631 = const()[name = tensor<string, []>("op_631"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74212992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77489856))), name = tensor<string, []>("layers_3_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77489984)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_631, groups = var_534, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_629, weight = layers_3_fc1_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<int32, [2]> var_637 = const()[name = tensor<string, []>("op_637"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_639 = const()[name = tensor<string, []>("op_639"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77500288)))];
+            tensor<fp16, [1280]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90607552)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_11_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_639, groups = var_534, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_637, weight = layers_3_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
+            tensor<int32, []> var_650 = const()[name = tensor<string, []>("op_650"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_652 = const()[name = tensor<string, []>("op_652"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_663 = const()[name = tensor<string, []>("op_663"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_17_cast_fp16 = reduce_mean(axes = var_663, keep_dims = var_653, x = inputs_17_cast_fp16)[name = tensor<string, []>("channels_mean_17_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_17_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_sq_17_cast_fp16")];
+            tensor<int32, [1]> var_667 = const()[name = tensor<string, []>("op_667"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_668_cast_fp16 = reduce_mean(axes = var_667, keep_dims = var_653, x = zero_mean_sq_17_cast_fp16)[name = tensor<string, []>("op_668_cast_fp16")];
+            tensor<fp16, []> var_669_to_fp16 = const()[name = tensor<string, []>("op_669_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_670_cast_fp16 = add(x = var_668_cast_fp16, y = var_669_to_fp16)[name = tensor<string, []>("op_670_cast_fp16")];
+            tensor<fp32, []> denom_17_epsilon_0 = const()[name = tensor<string, []>("denom_17_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_670_cast_fp16)[name = tensor<string, []>("denom_17_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
+            tensor<fp16, [1280]> obj_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_17_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90610176)))];
+            tensor<fp16, [1280]> obj_17_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_17_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90612800)))];
+            tensor<fp16, []> obj_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("obj_17_cast_fp16")];
+            tensor<int32, [2]> var_685 = const()[name = tensor<string, []>("op_685"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_687 = const()[name = tensor<string, []>("op_687"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90615424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91434688))), name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91434816)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_9_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_687, groups = var_652, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_685, weight = layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
+            tensor<int32, [2]> var_691 = const()[name = tensor<string, []>("op_691"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_693 = const()[name = tensor<string, []>("op_693"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_9_pad_type_0 = const()[name = tensor<string, []>("key_9_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_9_pad_0 = const()[name = tensor<string, []>("key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91437440))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92256704))), name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_9_cast_fp16 = conv(dilations = var_693, groups = var_652, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = var_691, weight = layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<int32, [2]> var_698 = const()[name = tensor<string, []>("op_698"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_700 = const()[name = tensor<string, []>("op_700"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_9_pad_type_0 = const()[name = tensor<string, []>("value_9_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_9_pad_0 = const()[name = tensor<string, []>("value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92256832))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93076096))), name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93076224)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_700, groups = var_652, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = var_698, weight = layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
+            tensor<int32, [4]> var_704 = const()[name = tensor<string, []>("op_704"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_705_cast_fp16 = reshape(shape = var_704, x = query_9_cast_fp16)[name = tensor<string, []>("op_705_cast_fp16")];
+            tensor<fp16, []> var_706_to_fp16 = const()[name = tensor<string, []>("op_706_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_707_cast_fp16 = mul(x = var_705_cast_fp16, y = var_706_to_fp16)[name = tensor<string, []>("op_707_cast_fp16")];
+            tensor<int32, [4]> var_708 = const()[name = tensor<string, []>("op_708"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_709_cast_fp16 = reshape(shape = var_708, x = key_9_cast_fp16)[name = tensor<string, []>("op_709_cast_fp16")];
+            tensor<bool, []> mh_w_9_transpose_x_0 = const()[name = tensor<string, []>("mh_w_9_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_9_transpose_y_0 = const()[name = tensor<string, []>("mh_w_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_707_cast_fp16, y = var_709_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_712_cast_fp16 = softmax(axis = var_650, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
+            tensor<int32, [4]> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_714_cast_fp16 = reshape(shape = var_713, x = value_9_cast_fp16)[name = tensor<string, []>("op_714_cast_fp16")];
+            tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_714_cast_fp16, y = var_712_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
+            tensor<int32, [4]> var_717 = const()[name = tensor<string, []>("op_717"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_33_cast_fp16 = reshape(shape = var_717, x = attn_9_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<int32, [2]> var_721 = const()[name = tensor<string, []>("op_721"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_723 = const()[name = tensor<string, []>("op_723"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93078848))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93898112))), name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93898240)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_19_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_723, groups = var_652, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_721, weight = layers_4_self_attn_o_proj_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
+            tensor<int32, [1]> var_729 = const()[name = tensor<string, []>("op_729"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_19_cast_fp16 = reduce_mean(axes = var_729, keep_dims = var_653, x = inputs_19_cast_fp16)[name = tensor<string, []>("channels_mean_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_sq_19_cast_fp16")];
+            tensor<int32, [1]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_734_cast_fp16 = reduce_mean(axes = var_733, keep_dims = var_653, x = zero_mean_sq_19_cast_fp16)[name = tensor<string, []>("op_734_cast_fp16")];
+            tensor<fp16, []> var_735_to_fp16 = const()[name = tensor<string, []>("op_735_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_736_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_to_fp16)[name = tensor<string, []>("op_736_cast_fp16")];
+            tensor<fp32, []> denom_19_epsilon_0 = const()[name = tensor<string, []>("denom_19_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_736_cast_fp16)[name = tensor<string, []>("denom_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
+            tensor<fp16, [1280]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93900864)))];
+            tensor<fp16, [1280]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93903488)))];
+            tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
+            tensor<int32, [2]> var_747 = const()[name = tensor<string, []>("op_747"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_749 = const()[name = tensor<string, []>("op_749"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_4_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93906112))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97182976))), name = tensor<string, []>("layers_4_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_4_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97183104)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_37_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_749, groups = var_652, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_747, weight = layers_4_fc1_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<int32, [2]> var_755 = const()[name = tensor<string, []>("op_755"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_757 = const()[name = tensor<string, []>("op_757"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_4_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97193408)))];
+            tensor<fp16, [1280]> layers_4_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110300672)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_13_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_757, groups = var_652, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_755, weight = layers_4_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
+            tensor<int32, []> var_768 = const()[name = tensor<string, []>("op_768"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_770 = const()[name = tensor<string, []>("op_770"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_781 = const()[name = tensor<string, []>("op_781"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_21_cast_fp16 = reduce_mean(axes = var_781, keep_dims = var_771, x = inputs_21_cast_fp16)[name = tensor<string, []>("channels_mean_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_sq_21_cast_fp16")];
+            tensor<int32, [1]> var_785 = const()[name = tensor<string, []>("op_785"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_786_cast_fp16 = reduce_mean(axes = var_785, keep_dims = var_771, x = zero_mean_sq_21_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")];
+            tensor<fp16, []> var_787_to_fp16 = const()[name = tensor<string, []>("op_787_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_788_cast_fp16 = add(x = var_786_cast_fp16, y = var_787_to_fp16)[name = tensor<string, []>("op_788_cast_fp16")];
+            tensor<fp32, []> denom_21_epsilon_0 = const()[name = tensor<string, []>("denom_21_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_788_cast_fp16)[name = tensor<string, []>("denom_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
+            tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110303296)))];
+            tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110305920)))];
+            tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
+            tensor<int32, [2]> var_803 = const()[name = tensor<string, []>("op_803"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110308544))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111127808))), name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111127936)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_11_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_805, groups = var_770, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_803, weight = layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
+            tensor<int32, [2]> var_809 = const()[name = tensor<string, []>("op_809"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_811 = const()[name = tensor<string, []>("op_811"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111130560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111949824))), name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = conv(dilations = var_811, groups = var_770, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_809, weight = layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<int32, [2]> var_816 = const()[name = tensor<string, []>("op_816"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111949952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112769216))), name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112769344)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_818, groups = var_770, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_816, weight = layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
+            tensor<int32, [4]> var_822 = const()[name = tensor<string, []>("op_822"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_823_cast_fp16 = reshape(shape = var_822, x = query_11_cast_fp16)[name = tensor<string, []>("op_823_cast_fp16")];
+            tensor<fp16, []> var_824_to_fp16 = const()[name = tensor<string, []>("op_824_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_825_cast_fp16 = mul(x = var_823_cast_fp16, y = var_824_to_fp16)[name = tensor<string, []>("op_825_cast_fp16")];
+            tensor<int32, [4]> var_826 = const()[name = tensor<string, []>("op_826"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_827_cast_fp16 = reshape(shape = var_826, x = key_11_cast_fp16)[name = tensor<string, []>("op_827_cast_fp16")];
+            tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_825_cast_fp16, y = var_827_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_830_cast_fp16 = softmax(axis = var_768, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_830_cast_fp16")];
+            tensor<int32, [4]> var_831 = const()[name = tensor<string, []>("op_831"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_832_cast_fp16 = reshape(shape = var_831, x = value_11_cast_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
+            tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_832_cast_fp16, y = var_830_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
+            tensor<int32, [4]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_41_cast_fp16 = reshape(shape = var_835, x = attn_11_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<int32, [2]> var_839 = const()[name = tensor<string, []>("op_839"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_841 = const()[name = tensor<string, []>("op_841"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112771968))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113591232))), name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113591360)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_23_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_841, groups = var_770, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_839, weight = layers_5_self_attn_o_proj_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
+            tensor<int32, [1]> var_847 = const()[name = tensor<string, []>("op_847"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_23_cast_fp16 = reduce_mean(axes = var_847, keep_dims = var_771, x = inputs_23_cast_fp16)[name = tensor<string, []>("channels_mean_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_sq_23_cast_fp16")];
+            tensor<int32, [1]> var_851 = const()[name = tensor<string, []>("op_851"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_852_cast_fp16 = reduce_mean(axes = var_851, keep_dims = var_771, x = zero_mean_sq_23_cast_fp16)[name = tensor<string, []>("op_852_cast_fp16")];
+            tensor<fp16, []> var_853_to_fp16 = const()[name = tensor<string, []>("op_853_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_854_cast_fp16 = add(x = var_852_cast_fp16, y = var_853_to_fp16)[name = tensor<string, []>("op_854_cast_fp16")];
+            tensor<fp32, []> denom_23_epsilon_0 = const()[name = tensor<string, []>("denom_23_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_854_cast_fp16)[name = tensor<string, []>("denom_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
+            tensor<fp16, [1280]> input_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113593984)))];
+            tensor<fp16, [1280]> input_43_beta_0_to_fp16 = const()[name = tensor<string, []>("input_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113596608)))];
+            tensor<fp16, []> input_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<int32, [2]> var_865 = const()[name = tensor<string, []>("op_865"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_867 = const()[name = tensor<string, []>("op_867"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_45_pad_type_0 = const()[name = tensor<string, []>("input_45_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_45_pad_0 = const()[name = tensor<string, []>("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_5_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113599232))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116876096))), name = tensor<string, []>("layers_5_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_5_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116876224)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_45_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_867, groups = var_770, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = var_865, weight = layers_5_fc1_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
+            tensor<string, []> input_47_mode_0 = const()[name = tensor<string, []>("input_47_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [2]> var_873 = const()[name = tensor<string, []>("op_873"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_15_pad_type_0 = const()[name = tensor<string, []>("hidden_states_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = tensor<string, []>("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_5_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116886528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120163392))), name = tensor<string, []>("layers_5_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_5_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120163520)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_15_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_875, groups = var_770, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = var_873, weight = layers_5_fc2_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
+            tensor<int32, []> var_886 = const()[name = tensor<string, []>("op_886"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_888 = const()[name = tensor<string, []>("op_888"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_889 = const()[name = tensor<string, []>("op_889"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_899 = const()[name = tensor<string, []>("op_899"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_25_cast_fp16 = reduce_mean(axes = var_899, keep_dims = var_889, x = inputs_25_cast_fp16)[name = tensor<string, []>("channels_mean_25_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor<string, []>("zero_mean_25_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor<string, []>("zero_mean_sq_25_cast_fp16")];
+            tensor<int32, [1]> var_903 = const()[name = tensor<string, []>("op_903"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_904_cast_fp16 = reduce_mean(axes = var_903, keep_dims = var_889, x = zero_mean_sq_25_cast_fp16)[name = tensor<string, []>("op_904_cast_fp16")];
+            tensor<fp16, []> var_905_to_fp16 = const()[name = tensor<string, []>("op_905_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_906_cast_fp16 = add(x = var_904_cast_fp16, y = var_905_to_fp16)[name = tensor<string, []>("op_906_cast_fp16")];
+            tensor<fp32, []> denom_25_epsilon_0 = const()[name = tensor<string, []>("denom_25_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0, x = var_906_cast_fp16)[name = tensor<string, []>("denom_25_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
+            tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120166144)))];
+            tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120168768)))];
+            tensor<fp16, []> obj_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
+            tensor<int32, [2]> var_921 = const()[name = tensor<string, []>("op_921"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_923 = const()[name = tensor<string, []>("op_923"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120171392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120990656))), name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120990784)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_13_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = var_923, groups = var_888, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_921, weight = layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
+            tensor<int32, [2]> var_927 = const()[name = tensor<string, []>("op_927"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_929 = const()[name = tensor<string, []>("op_929"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_13_pad_type_0 = const()[name = tensor<string, []>("key_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_13_pad_0 = const()[name = tensor<string, []>("key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120993408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121812672))), name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_13_cast_fp16 = conv(dilations = var_929, groups = var_888, pad = key_13_pad_0, pad_type = key_13_pad_type_0, strides = var_927, weight = layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<int32, [2]> var_934 = const()[name = tensor<string, []>("op_934"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_936 = const()[name = tensor<string, []>("op_936"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_13_pad_type_0 = const()[name = tensor<string, []>("value_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_13_pad_0 = const()[name = tensor<string, []>("value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121812800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122632064))), name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122632192)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = var_936, groups = var_888, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = var_934, weight = layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
+            tensor<int32, [4]> var_940 = const()[name = tensor<string, []>("op_940"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_941_cast_fp16 = reshape(shape = var_940, x = query_13_cast_fp16)[name = tensor<string, []>("op_941_cast_fp16")];
+            tensor<fp16, []> var_942_to_fp16 = const()[name = tensor<string, []>("op_942_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_943_cast_fp16 = mul(x = var_941_cast_fp16, y = var_942_to_fp16)[name = tensor<string, []>("op_943_cast_fp16")];
+            tensor<int32, [4]> var_944 = const()[name = tensor<string, []>("op_944"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_945_cast_fp16 = reshape(shape = var_944, x = key_13_cast_fp16)[name = tensor<string, []>("op_945_cast_fp16")];
+            tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_943_cast_fp16, y = var_945_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_948_cast_fp16 = softmax(axis = var_886, x = mh_w_13_cast_fp16)[name = tensor<string, []>("op_948_cast_fp16")];
+            tensor<int32, [4]> var_949 = const()[name = tensor<string, []>("op_949"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_950_cast_fp16 = reshape(shape = var_949, x = value_13_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
+            tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_950_cast_fp16, y = var_948_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
+            tensor<int32, [4]> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_49_cast_fp16 = reshape(shape = var_953, x = attn_13_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
+            tensor<int32, [2]> var_957 = const()[name = tensor<string, []>("op_957"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_959 = const()[name = tensor<string, []>("op_959"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_27_pad_type_0 = const()[name = tensor<string, []>("obj_27_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_27_pad_0 = const()[name = tensor<string, []>("obj_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122634816))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123454080))), name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123454208)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_27_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = var_959, groups = var_888, pad = obj_27_pad_0, pad_type = obj_27_pad_type_0, strides = var_957, weight = layers_6_self_attn_o_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
+            tensor<int32, [1]> var_965 = const()[name = tensor<string, []>("op_965"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_27_cast_fp16 = reduce_mean(axes = var_965, keep_dims = var_889, x = inputs_27_cast_fp16)[name = tensor<string, []>("channels_mean_27_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor<string, []>("zero_mean_27_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor<string, []>("zero_mean_sq_27_cast_fp16")];
+            tensor<int32, [1]> var_969 = const()[name = tensor<string, []>("op_969"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_970_cast_fp16 = reduce_mean(axes = var_969, keep_dims = var_889, x = zero_mean_sq_27_cast_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
+            tensor<fp16, []> var_971_to_fp16 = const()[name = tensor<string, []>("op_971_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_972_cast_fp16 = add(x = var_970_cast_fp16, y = var_971_to_fp16)[name = tensor<string, []>("op_972_cast_fp16")];
+            tensor<fp32, []> denom_27_epsilon_0 = const()[name = tensor<string, []>("denom_27_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0, x = var_972_cast_fp16)[name = tensor<string, []>("denom_27_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
+            tensor<fp16, [1280]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123456832)))];
+            tensor<fp16, [1280]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123459456)))];
+            tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, [2]> var_983 = const()[name = tensor<string, []>("op_983"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_985 = const()[name = tensor<string, []>("op_985"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_53_pad_type_0 = const()[name = tensor<string, []>("input_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_53_pad_0 = const()[name = tensor<string, []>("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_6_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123462080))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126738944))), name = tensor<string, []>("layers_6_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_6_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126739072)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_53_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = var_985, groups = var_888, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = var_983, weight = layers_6_fc1_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<string, []> input_55_mode_0 = const()[name = tensor<string, []>("input_55_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<int32, [2]> var_991 = const()[name = tensor<string, []>("op_991"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_993 = const()[name = tensor<string, []>("op_993"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_17_pad_type_0 = const()[name = tensor<string, []>("hidden_states_17_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = tensor<string, []>("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_6_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126749376))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130026240))), name = tensor<string, []>("layers_6_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_6_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130026368)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_17_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = var_993, groups = var_888, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = var_991, weight = layers_6_fc2_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
+            tensor<int32, []> var_1004 = const()[name = tensor<string, []>("op_1004"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1006 = const()[name = tensor<string, []>("op_1006"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1007 = const()[name = tensor<string, []>("op_1007"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1017 = const()[name = tensor<string, []>("op_1017"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_29_cast_fp16 = reduce_mean(axes = var_1017, keep_dims = var_1007, x = inputs_29_cast_fp16)[name = tensor<string, []>("channels_mean_29_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor<string, []>("zero_mean_29_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor<string, []>("zero_mean_sq_29_cast_fp16")];
+            tensor<int32, [1]> var_1021 = const()[name = tensor<string, []>("op_1021"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1022_cast_fp16 = reduce_mean(axes = var_1021, keep_dims = var_1007, x = zero_mean_sq_29_cast_fp16)[name = tensor<string, []>("op_1022_cast_fp16")];
+            tensor<fp16, []> var_1023_to_fp16 = const()[name = tensor<string, []>("op_1023_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1024_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_to_fp16)[name = tensor<string, []>("op_1024_cast_fp16")];
+            tensor<fp32, []> denom_29_epsilon_0 = const()[name = tensor<string, []>("denom_29_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0, x = var_1024_cast_fp16)[name = tensor<string, []>("denom_29_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
+            tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130028992)))];
+            tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130031616)))];
+            tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
+            tensor<int32, [2]> var_1039 = const()[name = tensor<string, []>("op_1039"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1041 = const()[name = tensor<string, []>("op_1041"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_15_pad_type_0 = const()[name = tensor<string, []>("query_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_15_pad_0 = const()[name = tensor<string, []>("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130034240))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130853504))), name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130853632)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_15_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = var_1041, groups = var_1006, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_1039, weight = layers_7_self_attn_q_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
+            tensor<int32, [2]> var_1045 = const()[name = tensor<string, []>("op_1045"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1047 = const()[name = tensor<string, []>("op_1047"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130856256))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131675520))), name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_15_cast_fp16 = conv(dilations = var_1047, groups = var_1006, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_1045, weight = layers_7_self_attn_k_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<int32, [2]> var_1052 = const()[name = tensor<string, []>("op_1052"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1054 = const()[name = tensor<string, []>("op_1054"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_15_pad_type_0 = const()[name = tensor<string, []>("value_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_15_pad_0 = const()[name = tensor<string, []>("value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131675648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132494912))), name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132495040)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = var_1054, groups = var_1006, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_1052, weight = layers_7_self_attn_v_proj_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")];
+            tensor<int32, [4]> var_1058 = const()[name = tensor<string, []>("op_1058"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1059_cast_fp16 = reshape(shape = var_1058, x = query_15_cast_fp16)[name = tensor<string, []>("op_1059_cast_fp16")];
+            tensor<fp16, []> var_1060_to_fp16 = const()[name = tensor<string, []>("op_1060_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1061_cast_fp16 = mul(x = var_1059_cast_fp16, y = var_1060_to_fp16)[name = tensor<string, []>("op_1061_cast_fp16")];
+            tensor<int32, [4]> var_1062 = const()[name = tensor<string, []>("op_1062"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1063_cast_fp16 = reshape(shape = var_1062, x = key_15_cast_fp16)[name = tensor<string, []>("op_1063_cast_fp16")];
+            tensor<bool, []> mh_w_15_transpose_x_0 = const()[name = tensor<string, []>("mh_w_15_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_15_transpose_y_0 = const()[name = tensor<string, []>("mh_w_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1061_cast_fp16, y = var_1063_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1066_cast_fp16 = softmax(axis = var_1004, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_1066_cast_fp16")];
+            tensor<int32, [4]> var_1067 = const()[name = tensor<string, []>("op_1067"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1068_cast_fp16 = reshape(shape = var_1067, x = value_15_cast_fp16)[name = tensor<string, []>("op_1068_cast_fp16")];
+            tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1068_cast_fp16, y = var_1066_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
+            tensor<int32, [4]> var_1071 = const()[name = tensor<string, []>("op_1071"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_57_cast_fp16 = reshape(shape = var_1071, x = attn_15_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
+            tensor<int32, [2]> var_1075 = const()[name = tensor<string, []>("op_1075"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1077 = const()[name = tensor<string, []>("op_1077"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_31_pad_type_0 = const()[name = tensor<string, []>("obj_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_31_pad_0 = const()[name = tensor<string, []>("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132497664))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133316928))), name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133317056)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_31_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = var_1077, groups = var_1006, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = var_1075, weight = layers_7_self_attn_o_proj_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
+            tensor<int32, [1]> var_1083 = const()[name = tensor<string, []>("op_1083"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_31_cast_fp16 = reduce_mean(axes = var_1083, keep_dims = var_1007, x = inputs_31_cast_fp16)[name = tensor<string, []>("channels_mean_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor<string, []>("zero_mean_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor<string, []>("zero_mean_sq_31_cast_fp16")];
+            tensor<int32, [1]> var_1087 = const()[name = tensor<string, []>("op_1087"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1088_cast_fp16 = reduce_mean(axes = var_1087, keep_dims = var_1007, x = zero_mean_sq_31_cast_fp16)[name = tensor<string, []>("op_1088_cast_fp16")];
+            tensor<fp16, []> var_1089_to_fp16 = const()[name = tensor<string, []>("op_1089_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1090_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_to_fp16)[name = tensor<string, []>("op_1090_cast_fp16")];
+            tensor<fp32, []> denom_31_epsilon_0 = const()[name = tensor<string, []>("denom_31_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0, x = var_1090_cast_fp16)[name = tensor<string, []>("denom_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
+            tensor<fp16, [1280]> input_59_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_59_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133319680)))];
+            tensor<fp16, [1280]> input_59_beta_0_to_fp16 = const()[name = tensor<string, []>("input_59_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133322304)))];
+            tensor<fp16, []> input_59_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_59_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_59_cast_fp16 = batch_norm(beta = input_59_beta_0_to_fp16, epsilon = input_59_epsilon_0_to_fp16, gamma = input_59_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<int32, [2]> var_1101 = const()[name = tensor<string, []>("op_1101"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1103 = const()[name = tensor<string, []>("op_1103"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_61_pad_type_0 = const()[name = tensor<string, []>("input_61_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_61_pad_0 = const()[name = tensor<string, []>("input_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_7_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133324928))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136601792))), name = tensor<string, []>("layers_7_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_7_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136601920)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_61_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = var_1103, groups = var_1006, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = var_1101, weight = layers_7_fc1_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<string, []> input_63_mode_0 = const()[name = tensor<string, []>("input_63_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [2]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1111 = const()[name = tensor<string, []>("op_1111"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_19_pad_type_0 = const()[name = tensor<string, []>("hidden_states_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = tensor<string, []>("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_7_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136612224))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139889088))), name = tensor<string, []>("layers_7_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_7_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139889216)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_19_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = var_1111, groups = var_1006, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = var_1109, weight = layers_7_fc2_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
+            tensor<int32, []> var_1122 = const()[name = tensor<string, []>("op_1122"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1124 = const()[name = tensor<string, []>("op_1124"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1125 = const()[name = tensor<string, []>("op_1125"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1135 = const()[name = tensor<string, []>("op_1135"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_33_cast_fp16 = reduce_mean(axes = var_1135, keep_dims = var_1125, x = inputs_33_cast_fp16)[name = tensor<string, []>("channels_mean_33_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor<string, []>("zero_mean_33_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor<string, []>("zero_mean_sq_33_cast_fp16")];
+            tensor<int32, [1]> var_1139 = const()[name = tensor<string, []>("op_1139"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1140_cast_fp16 = reduce_mean(axes = var_1139, keep_dims = var_1125, x = zero_mean_sq_33_cast_fp16)[name = tensor<string, []>("op_1140_cast_fp16")];
+            tensor<fp16, []> var_1141_to_fp16 = const()[name = tensor<string, []>("op_1141_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1142_cast_fp16 = add(x = var_1140_cast_fp16, y = var_1141_to_fp16)[name = tensor<string, []>("op_1142_cast_fp16")];
+            tensor<fp32, []> denom_33_epsilon_0 = const()[name = tensor<string, []>("denom_33_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0, x = var_1142_cast_fp16)[name = tensor<string, []>("denom_33_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
+            tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139891840)))];
+            tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139894464)))];
+            tensor<fp16, []> obj_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
+            tensor<int32, [2]> var_1157 = const()[name = tensor<string, []>("op_1157"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1159 = const()[name = tensor<string, []>("op_1159"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139897088))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140716352))), name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140716480)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_17_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = var_1159, groups = var_1124, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_1157, weight = layers_8_self_attn_q_proj_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
+            tensor<int32, [2]> var_1163 = const()[name = tensor<string, []>("op_1163"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1165 = const()[name = tensor<string, []>("op_1165"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_17_pad_type_0 = const()[name = tensor<string, []>("key_17_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_17_pad_0 = const()[name = tensor<string, []>("key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140719104))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141538368))), name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_17_cast_fp16 = conv(dilations = var_1165, groups = var_1124, pad = key_17_pad_0, pad_type = key_17_pad_type_0, strides = var_1163, weight = layers_8_self_attn_k_proj_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<int32, [2]> var_1170 = const()[name = tensor<string, []>("op_1170"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1172 = const()[name = tensor<string, []>("op_1172"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_17_pad_type_0 = const()[name = tensor<string, []>("value_17_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_17_pad_0 = const()[name = tensor<string, []>("value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141538496))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142357760))), name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142357888)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = var_1172, groups = var_1124, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = var_1170, weight = layers_8_self_attn_v_proj_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
+            tensor<int32, [4]> var_1176 = const()[name = tensor<string, []>("op_1176"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1177_cast_fp16 = reshape(shape = var_1176, x = query_17_cast_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
+            tensor<fp16, []> var_1178_to_fp16 = const()[name = tensor<string, []>("op_1178_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1179_cast_fp16 = mul(x = var_1177_cast_fp16, y = var_1178_to_fp16)[name = tensor<string, []>("op_1179_cast_fp16")];
+            tensor<int32, [4]> var_1180 = const()[name = tensor<string, []>("op_1180"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1181_cast_fp16 = reshape(shape = var_1180, x = key_17_cast_fp16)[name = tensor<string, []>("op_1181_cast_fp16")];
+            tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1179_cast_fp16, y = var_1181_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1184_cast_fp16 = softmax(axis = var_1122, x = mh_w_17_cast_fp16)[name = tensor<string, []>("op_1184_cast_fp16")];
+            tensor<int32, [4]> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1186_cast_fp16 = reshape(shape = var_1185, x = value_17_cast_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
+            tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1186_cast_fp16, y = var_1184_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
+            tensor<int32, [4]> var_1189 = const()[name = tensor<string, []>("op_1189"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_65_cast_fp16 = reshape(shape = var_1189, x = attn_17_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<int32, [2]> var_1193 = const()[name = tensor<string, []>("op_1193"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1195 = const()[name = tensor<string, []>("op_1195"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142360512))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143179776))), name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143179904)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_35_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = var_1195, groups = var_1124, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_1193, weight = layers_8_self_attn_o_proj_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
+            tensor<int32, [1]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_35_cast_fp16 = reduce_mean(axes = var_1201, keep_dims = var_1125, x = inputs_35_cast_fp16)[name = tensor<string, []>("channels_mean_35_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor<string, []>("zero_mean_35_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor<string, []>("zero_mean_sq_35_cast_fp16")];
+            tensor<int32, [1]> var_1205 = const()[name = tensor<string, []>("op_1205"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1206_cast_fp16 = reduce_mean(axes = var_1205, keep_dims = var_1125, x = zero_mean_sq_35_cast_fp16)[name = tensor<string, []>("op_1206_cast_fp16")];
+            tensor<fp16, []> var_1207_to_fp16 = const()[name = tensor<string, []>("op_1207_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1208_cast_fp16 = add(x = var_1206_cast_fp16, y = var_1207_to_fp16)[name = tensor<string, []>("op_1208_cast_fp16")];
+            tensor<fp32, []> denom_35_epsilon_0 = const()[name = tensor<string, []>("denom_35_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0, x = var_1208_cast_fp16)[name = tensor<string, []>("denom_35_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
+            tensor<fp16, [1280]> input_67_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_67_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143182528)))];
+            tensor<fp16, [1280]> input_67_beta_0_to_fp16 = const()[name = tensor<string, []>("input_67_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143185152)))];
+            tensor<fp16, []> input_67_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_67_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_67_cast_fp16 = batch_norm(beta = input_67_beta_0_to_fp16, epsilon = input_67_epsilon_0_to_fp16, gamma = input_67_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
+            tensor<int32, [2]> var_1219 = const()[name = tensor<string, []>("op_1219"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1221 = const()[name = tensor<string, []>("op_1221"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_8_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143187776))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146464640))), name = tensor<string, []>("layers_8_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_8_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146464768)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_69_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = var_1221, groups = var_1124, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = var_1219, weight = layers_8_fc1_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<int32, [2]> var_1227 = const()[name = tensor<string, []>("op_1227"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1229 = const()[name = tensor<string, []>("op_1229"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_21_pad_type_0 = const()[name = tensor<string, []>("hidden_states_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = tensor<string, []>("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_8_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146475072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149751936))), name = tensor<string, []>("layers_8_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_8_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149752064)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_21_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = var_1229, groups = var_1124, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = var_1227, weight = layers_8_fc2_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
+            tensor<int32, []> var_1240 = const()[name = tensor<string, []>("op_1240"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1242 = const()[name = tensor<string, []>("op_1242"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1243 = const()[name = tensor<string, []>("op_1243"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1253 = const()[name = tensor<string, []>("op_1253"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_37_cast_fp16 = reduce_mean(axes = var_1253, keep_dims = var_1243, x = inputs_37_cast_fp16)[name = tensor<string, []>("channels_mean_37_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor<string, []>("zero_mean_37_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor<string, []>("zero_mean_sq_37_cast_fp16")];
+            tensor<int32, [1]> var_1257 = const()[name = tensor<string, []>("op_1257"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1258_cast_fp16 = reduce_mean(axes = var_1257, keep_dims = var_1243, x = zero_mean_sq_37_cast_fp16)[name = tensor<string, []>("op_1258_cast_fp16")];
+            tensor<fp16, []> var_1259_to_fp16 = const()[name = tensor<string, []>("op_1259_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1260_cast_fp16 = add(x = var_1258_cast_fp16, y = var_1259_to_fp16)[name = tensor<string, []>("op_1260_cast_fp16")];
+            tensor<fp32, []> denom_37_epsilon_0 = const()[name = tensor<string, []>("denom_37_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0, x = var_1260_cast_fp16)[name = tensor<string, []>("denom_37_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
+            tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149754688)))];
+            tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149757312)))];
+            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
+            tensor<int32, [2]> var_1275 = const()[name = tensor<string, []>("op_1275"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1277 = const()[name = tensor<string, []>("op_1277"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_19_pad_type_0 = const()[name = tensor<string, []>("query_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_19_pad_0 = const()[name = tensor<string, []>("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149759936))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150579200))), name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150579328)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_19_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = var_1277, groups = var_1242, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1275, weight = layers_9_self_attn_q_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
+            tensor<int32, [2]> var_1281 = const()[name = tensor<string, []>("op_1281"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1283 = const()[name = tensor<string, []>("op_1283"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150581952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151401216))), name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_19_cast_fp16 = conv(dilations = var_1283, groups = var_1242, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1281, weight = layers_9_self_attn_k_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<int32, [2]> var_1288 = const()[name = tensor<string, []>("op_1288"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1290 = const()[name = tensor<string, []>("op_1290"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_19_pad_type_0 = const()[name = tensor<string, []>("value_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_19_pad_0 = const()[name = tensor<string, []>("value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151401344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152220608))), name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152220736)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = var_1290, groups = var_1242, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1288, weight = layers_9_self_attn_v_proj_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")];
+            tensor<int32, [4]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1295_cast_fp16 = reshape(shape = var_1294, x = query_19_cast_fp16)[name = tensor<string, []>("op_1295_cast_fp16")];
+            tensor<fp16, []> var_1296_to_fp16 = const()[name = tensor<string, []>("op_1296_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor<string, []>("op_1297_cast_fp16")];
+            tensor<int32, [4]> var_1298 = const()[name = tensor<string, []>("op_1298"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1299_cast_fp16 = reshape(shape = var_1298, x = key_19_cast_fp16)[name = tensor<string, []>("op_1299_cast_fp16")];
+            tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1297_cast_fp16, y = var_1299_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1302_cast_fp16 = softmax(axis = var_1240, x = mh_w_19_cast_fp16)[name = tensor<string, []>("op_1302_cast_fp16")];
+            tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1304_cast_fp16 = reshape(shape = var_1303, x = value_19_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
+            tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1304_cast_fp16, y = var_1302_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
+            tensor<int32, [4]> var_1307 = const()[name = tensor<string, []>("op_1307"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_73_cast_fp16 = reshape(shape = var_1307, x = attn_19_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
+            tensor<int32, [2]> var_1311 = const()[name = tensor<string, []>("op_1311"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1313 = const()[name = tensor<string, []>("op_1313"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152223360))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153042624))), name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153042752)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_39_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = var_1313, groups = var_1242, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_1311, weight = layers_9_self_attn_o_proj_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
+            tensor<int32, [1]> var_1319 = const()[name = tensor<string, []>("op_1319"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_39_cast_fp16 = reduce_mean(axes = var_1319, keep_dims = var_1243, x = inputs_39_cast_fp16)[name = tensor<string, []>("channels_mean_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor<string, []>("zero_mean_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor<string, []>("zero_mean_sq_39_cast_fp16")];
+            tensor<int32, [1]> var_1323 = const()[name = tensor<string, []>("op_1323"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1324_cast_fp16 = reduce_mean(axes = var_1323, keep_dims = var_1243, x = zero_mean_sq_39_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")];
+            tensor<fp16, []> var_1325_to_fp16 = const()[name = tensor<string, []>("op_1325_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1326_cast_fp16 = add(x = var_1324_cast_fp16, y = var_1325_to_fp16)[name = tensor<string, []>("op_1326_cast_fp16")];
+            tensor<fp32, []> denom_39_epsilon_0 = const()[name = tensor<string, []>("denom_39_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0, x = var_1326_cast_fp16)[name = tensor<string, []>("denom_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
+            tensor<fp16, [1280]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153045376)))];
+            tensor<fp16, [1280]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153048000)))];
+            tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [2]> var_1337 = const()[name = tensor<string, []>("op_1337"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1339 = const()[name = tensor<string, []>("op_1339"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_9_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153050624))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156327488))), name = tensor<string, []>("layers_9_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_9_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156327616)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_77_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = var_1339, groups = var_1242, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = var_1337, weight = layers_9_fc1_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<int32, [2]> var_1345 = const()[name = tensor<string, []>("op_1345"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1347 = const()[name = tensor<string, []>("op_1347"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_23_pad_type_0 = const()[name = tensor<string, []>("hidden_states_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = tensor<string, []>("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_9_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156337920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159614784))), name = tensor<string, []>("layers_9_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_9_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159614912)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_23_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = var_1347, groups = var_1242, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = var_1345, weight = layers_9_fc2_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
+            tensor<int32, []> var_1358 = const()[name = tensor<string, []>("op_1358"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1360 = const()[name = tensor<string, []>("op_1360"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1361 = const()[name = tensor<string, []>("op_1361"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1371 = const()[name = tensor<string, []>("op_1371"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_41_cast_fp16 = reduce_mean(axes = var_1371, keep_dims = var_1361, x = inputs_41_cast_fp16)[name = tensor<string, []>("channels_mean_41_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor<string, []>("zero_mean_41_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor<string, []>("zero_mean_sq_41_cast_fp16")];
+            tensor<int32, [1]> var_1375 = const()[name = tensor<string, []>("op_1375"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1376_cast_fp16 = reduce_mean(axes = var_1375, keep_dims = var_1361, x = zero_mean_sq_41_cast_fp16)[name = tensor<string, []>("op_1376_cast_fp16")];
+            tensor<fp16, []> var_1377_to_fp16 = const()[name = tensor<string, []>("op_1377_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1378_cast_fp16 = add(x = var_1376_cast_fp16, y = var_1377_to_fp16)[name = tensor<string, []>("op_1378_cast_fp16")];
+            tensor<fp32, []> denom_41_epsilon_0 = const()[name = tensor<string, []>("denom_41_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0, x = var_1378_cast_fp16)[name = tensor<string, []>("denom_41_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
+            tensor<fp16, [1280]> obj_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_41_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159617536)))];
+            tensor<fp16, [1280]> obj_41_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_41_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159620160)))];
+            tensor<fp16, []> obj_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
+            tensor<int32, [2]> var_1393 = const()[name = tensor<string, []>("op_1393"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1395 = const()[name = tensor<string, []>("op_1395"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159622784))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160442048))), name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160442176)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_21_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = var_1395, groups = var_1360, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1393, weight = layers_10_self_attn_q_proj_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
+            tensor<int32, [2]> var_1399 = const()[name = tensor<string, []>("op_1399"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1401 = const()[name = tensor<string, []>("op_1401"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_21_pad_type_0 = const()[name = tensor<string, []>("key_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_21_pad_0 = const()[name = tensor<string, []>("key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160444800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161264064))), name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_21_cast_fp16 = conv(dilations = var_1401, groups = var_1360, pad = key_21_pad_0, pad_type = key_21_pad_type_0, strides = var_1399, weight = layers_10_self_attn_k_proj_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<int32, [2]> var_1406 = const()[name = tensor<string, []>("op_1406"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1408 = const()[name = tensor<string, []>("op_1408"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_21_pad_type_0 = const()[name = tensor<string, []>("value_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_21_pad_0 = const()[name = tensor<string, []>("value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161264192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162083456))), name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162083584)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = var_1408, groups = var_1360, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = var_1406, weight = layers_10_self_attn_v_proj_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
+            tensor<int32, [4]> var_1412 = const()[name = tensor<string, []>("op_1412"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1413_cast_fp16 = reshape(shape = var_1412, x = query_21_cast_fp16)[name = tensor<string, []>("op_1413_cast_fp16")];
+            tensor<fp16, []> var_1414_to_fp16 = const()[name = tensor<string, []>("op_1414_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1415_cast_fp16 = mul(x = var_1413_cast_fp16, y = var_1414_to_fp16)[name = tensor<string, []>("op_1415_cast_fp16")];
+            tensor<int32, [4]> var_1416 = const()[name = tensor<string, []>("op_1416"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1417_cast_fp16 = reshape(shape = var_1416, x = key_21_cast_fp16)[name = tensor<string, []>("op_1417_cast_fp16")];
+            tensor<bool, []> mh_w_21_transpose_x_0 = const()[name = tensor<string, []>("mh_w_21_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_21_transpose_y_0 = const()[name = tensor<string, []>("mh_w_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1415_cast_fp16, y = var_1417_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1420_cast_fp16 = softmax(axis = var_1358, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
+            tensor<int32, [4]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1422_cast_fp16 = reshape(shape = var_1421, x = value_21_cast_fp16)[name = tensor<string, []>("op_1422_cast_fp16")];
+            tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1422_cast_fp16, y = var_1420_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
+            tensor<int32, [4]> var_1425 = const()[name = tensor<string, []>("op_1425"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_81_cast_fp16 = reshape(shape = var_1425, x = attn_21_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
+            tensor<int32, [2]> var_1429 = const()[name = tensor<string, []>("op_1429"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1431 = const()[name = tensor<string, []>("op_1431"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_43_pad_type_0 = const()[name = tensor<string, []>("obj_43_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_43_pad_0 = const()[name = tensor<string, []>("obj_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162086208))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162905472))), name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162905600)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_43_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = var_1431, groups = var_1360, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = var_1429, weight = layers_10_self_attn_o_proj_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
+            tensor<int32, [1]> var_1437 = const()[name = tensor<string, []>("op_1437"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_43_cast_fp16 = reduce_mean(axes = var_1437, keep_dims = var_1361, x = inputs_43_cast_fp16)[name = tensor<string, []>("channels_mean_43_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor<string, []>("zero_mean_43_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor<string, []>("zero_mean_sq_43_cast_fp16")];
+            tensor<int32, [1]> var_1441 = const()[name = tensor<string, []>("op_1441"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1442_cast_fp16 = reduce_mean(axes = var_1441, keep_dims = var_1361, x = zero_mean_sq_43_cast_fp16)[name = tensor<string, []>("op_1442_cast_fp16")];
+            tensor<fp16, []> var_1443_to_fp16 = const()[name = tensor<string, []>("op_1443_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1444_cast_fp16 = add(x = var_1442_cast_fp16, y = var_1443_to_fp16)[name = tensor<string, []>("op_1444_cast_fp16")];
+            tensor<fp32, []> denom_43_epsilon_0 = const()[name = tensor<string, []>("denom_43_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0, x = var_1444_cast_fp16)[name = tensor<string, []>("denom_43_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
+            tensor<fp16, [1280]> input_83_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_83_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162908224)))];
+            tensor<fp16, [1280]> input_83_beta_0_to_fp16 = const()[name = tensor<string, []>("input_83_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162910848)))];
+            tensor<fp16, []> input_83_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_83_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_83_cast_fp16 = batch_norm(beta = input_83_beta_0_to_fp16, epsilon = input_83_epsilon_0_to_fp16, gamma = input_83_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<int32, [2]> var_1455 = const()[name = tensor<string, []>("op_1455"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1457 = const()[name = tensor<string, []>("op_1457"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_85_pad_type_0 = const()[name = tensor<string, []>("input_85_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_85_pad_0 = const()[name = tensor<string, []>("input_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_10_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162913472))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166190336))), name = tensor<string, []>("layers_10_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_10_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166190464)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_85_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = var_1457, groups = var_1360, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = var_1455, weight = layers_10_fc1_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
+            tensor<int32, [2]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1465 = const()[name = tensor<string, []>("op_1465"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_25_pad_type_0 = const()[name = tensor<string, []>("hidden_states_25_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = tensor<string, []>("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_10_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166200768))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169477632))), name = tensor<string, []>("layers_10_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_10_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169477760)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_25_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = var_1465, groups = var_1360, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = var_1463, weight = layers_10_fc2_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
+            tensor<int32, []> var_1476 = const()[name = tensor<string, []>("op_1476"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1478 = const()[name = tensor<string, []>("op_1478"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1479 = const()[name = tensor<string, []>("op_1479"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1489 = const()[name = tensor<string, []>("op_1489"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_45_cast_fp16 = reduce_mean(axes = var_1489, keep_dims = var_1479, x = inputs_45_cast_fp16)[name = tensor<string, []>("channels_mean_45_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor<string, []>("zero_mean_45_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor<string, []>("zero_mean_sq_45_cast_fp16")];
+            tensor<int32, [1]> var_1493 = const()[name = tensor<string, []>("op_1493"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1494_cast_fp16 = reduce_mean(axes = var_1493, keep_dims = var_1479, x = zero_mean_sq_45_cast_fp16)[name = tensor<string, []>("op_1494_cast_fp16")];
+            tensor<fp16, []> var_1495_to_fp16 = const()[name = tensor<string, []>("op_1495_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1496_cast_fp16 = add(x = var_1494_cast_fp16, y = var_1495_to_fp16)[name = tensor<string, []>("op_1496_cast_fp16")];
+            tensor<fp32, []> denom_45_epsilon_0 = const()[name = tensor<string, []>("denom_45_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0, x = var_1496_cast_fp16)[name = tensor<string, []>("denom_45_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
+            tensor<fp16, [1280]> obj_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169480384)))];
+            tensor<fp16, [1280]> obj_45_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_45_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169483008)))];
+            tensor<fp16, []> obj_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
+            tensor<int32, [2]> var_1511 = const()[name = tensor<string, []>("op_1511"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1513 = const()[name = tensor<string, []>("op_1513"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_23_pad_type_0 = const()[name = tensor<string, []>("query_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_23_pad_0 = const()[name = tensor<string, []>("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169485632))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170304896))), name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170305024)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_23_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = var_1513, groups = var_1478, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = var_1511, weight = layers_11_self_attn_q_proj_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
+            tensor<int32, [2]> var_1517 = const()[name = tensor<string, []>("op_1517"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1519 = const()[name = tensor<string, []>("op_1519"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170307648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171126912))), name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_23_cast_fp16 = conv(dilations = var_1519, groups = var_1478, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = var_1517, weight = layers_11_self_attn_k_proj_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<int32, [2]> var_1524 = const()[name = tensor<string, []>("op_1524"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1526 = const()[name = tensor<string, []>("op_1526"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_23_pad_type_0 = const()[name = tensor<string, []>("value_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_23_pad_0 = const()[name = tensor<string, []>("value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171127040))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171946304))), name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171946432)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = var_1526, groups = var_1478, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = var_1524, weight = layers_11_self_attn_v_proj_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor<string, []>("value_23_cast_fp16")];
+            tensor<int32, [4]> var_1530 = const()[name = tensor<string, []>("op_1530"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1531_cast_fp16 = reshape(shape = var_1530, x = query_23_cast_fp16)[name = tensor<string, []>("op_1531_cast_fp16")];
+            tensor<fp16, []> var_1532_to_fp16 = const()[name = tensor<string, []>("op_1532_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1533_cast_fp16 = mul(x = var_1531_cast_fp16, y = var_1532_to_fp16)[name = tensor<string, []>("op_1533_cast_fp16")];
+            tensor<int32, [4]> var_1534 = const()[name = tensor<string, []>("op_1534"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1535_cast_fp16 = reshape(shape = var_1534, x = key_23_cast_fp16)[name = tensor<string, []>("op_1535_cast_fp16")];
+            tensor<bool, []> mh_w_23_transpose_x_0 = const()[name = tensor<string, []>("mh_w_23_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_23_transpose_y_0 = const()[name = tensor<string, []>("mh_w_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1533_cast_fp16, y = var_1535_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1538_cast_fp16 = softmax(axis = var_1476, x = mh_w_23_cast_fp16)[name = tensor<string, []>("op_1538_cast_fp16")];
+            tensor<int32, [4]> var_1539 = const()[name = tensor<string, []>("op_1539"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1540_cast_fp16 = reshape(shape = var_1539, x = value_23_cast_fp16)[name = tensor<string, []>("op_1540_cast_fp16")];
+            tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1540_cast_fp16, y = var_1538_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
+            tensor<int32, [4]> var_1543 = const()[name = tensor<string, []>("op_1543"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_89_cast_fp16 = reshape(shape = var_1543, x = attn_23_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [2]> var_1547 = const()[name = tensor<string, []>("op_1547"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1549 = const()[name = tensor<string, []>("op_1549"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_47_pad_type_0 = const()[name = tensor<string, []>("obj_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_47_pad_0 = const()[name = tensor<string, []>("obj_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171949056))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172768320))), name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172768448)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_47_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = var_1549, groups = var_1478, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = var_1547, weight = layers_11_self_attn_o_proj_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
+            tensor<int32, [1]> var_1555 = const()[name = tensor<string, []>("op_1555"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_47_cast_fp16 = reduce_mean(axes = var_1555, keep_dims = var_1479, x = inputs_47_cast_fp16)[name = tensor<string, []>("channels_mean_47_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor<string, []>("zero_mean_47_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor<string, []>("zero_mean_sq_47_cast_fp16")];
+            tensor<int32, [1]> var_1559 = const()[name = tensor<string, []>("op_1559"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1560_cast_fp16 = reduce_mean(axes = var_1559, keep_dims = var_1479, x = zero_mean_sq_47_cast_fp16)[name = tensor<string, []>("op_1560_cast_fp16")];
+            tensor<fp16, []> var_1561_to_fp16 = const()[name = tensor<string, []>("op_1561_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1562_cast_fp16 = add(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor<string, []>("op_1562_cast_fp16")];
+            tensor<fp32, []> denom_47_epsilon_0 = const()[name = tensor<string, []>("denom_47_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0, x = var_1562_cast_fp16)[name = tensor<string, []>("denom_47_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
+            tensor<fp16, [1280]> input_91_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_91_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172771072)))];
+            tensor<fp16, [1280]> input_91_beta_0_to_fp16 = const()[name = tensor<string, []>("input_91_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172773696)))];
+            tensor<fp16, []> input_91_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_91_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [2]> var_1573 = const()[name = tensor<string, []>("op_1573"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1575 = const()[name = tensor<string, []>("op_1575"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_93_pad_type_0 = const()[name = tensor<string, []>("input_93_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_93_pad_0 = const()[name = tensor<string, []>("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_11_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172776320))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176053184))), name = tensor<string, []>("layers_11_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_11_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176053312)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_93_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = var_1575, groups = var_1478, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = var_1573, weight = layers_11_fc1_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<string, []> input_95_mode_0 = const()[name = tensor<string, []>("input_95_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<int32, [2]> var_1581 = const()[name = tensor<string, []>("op_1581"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1583 = const()[name = tensor<string, []>("op_1583"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_27_pad_type_0 = const()[name = tensor<string, []>("hidden_states_27_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = tensor<string, []>("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_11_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176063616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179340480))), name = tensor<string, []>("layers_11_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_11_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179340608)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_27_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = var_1583, groups = var_1478, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = var_1581, weight = layers_11_fc2_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = tensor<string, []>("hidden_states_27_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
+            tensor<int32, []> var_1594 = const()[name = tensor<string, []>("op_1594"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1596 = const()[name = tensor<string, []>("op_1596"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1597 = const()[name = tensor<string, []>("op_1597"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1607 = const()[name = tensor<string, []>("op_1607"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_49_cast_fp16 = reduce_mean(axes = var_1607, keep_dims = var_1597, x = inputs_49_cast_fp16)[name = tensor<string, []>("channels_mean_49_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor<string, []>("zero_mean_49_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor<string, []>("zero_mean_sq_49_cast_fp16")];
+            tensor<int32, [1]> var_1611 = const()[name = tensor<string, []>("op_1611"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1612_cast_fp16 = reduce_mean(axes = var_1611, keep_dims = var_1597, x = zero_mean_sq_49_cast_fp16)[name = tensor<string, []>("op_1612_cast_fp16")];
+            tensor<fp16, []> var_1613_to_fp16 = const()[name = tensor<string, []>("op_1613_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1614_cast_fp16 = add(x = var_1612_cast_fp16, y = var_1613_to_fp16)[name = tensor<string, []>("op_1614_cast_fp16")];
+            tensor<fp32, []> denom_49_epsilon_0 = const()[name = tensor<string, []>("denom_49_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0, x = var_1614_cast_fp16)[name = tensor<string, []>("denom_49_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
+            tensor<fp16, [1280]> obj_49_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_49_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179343232)))];
+            tensor<fp16, [1280]> obj_49_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_49_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179345856)))];
+            tensor<fp16, []> obj_49_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_49_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
+            tensor<int32, [2]> var_1629 = const()[name = tensor<string, []>("op_1629"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1631 = const()[name = tensor<string, []>("op_1631"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179348480))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180167744))), name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180167872)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_25_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = var_1631, groups = var_1596, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = var_1629, weight = layers_12_self_attn_q_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
+            tensor<int32, [2]> var_1635 = const()[name = tensor<string, []>("op_1635"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1637 = const()[name = tensor<string, []>("op_1637"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_25_pad_type_0 = const()[name = tensor<string, []>("key_25_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_25_pad_0 = const()[name = tensor<string, []>("key_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180170496))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180989760))), name = tensor<string, []>("layers_12_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_25_cast_fp16 = conv(dilations = var_1637, groups = var_1596, pad = key_25_pad_0, pad_type = key_25_pad_type_0, strides = var_1635, weight = layers_12_self_attn_k_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<int32, [2]> var_1642 = const()[name = tensor<string, []>("op_1642"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1644 = const()[name = tensor<string, []>("op_1644"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_25_pad_type_0 = const()[name = tensor<string, []>("value_25_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_25_pad_0 = const()[name = tensor<string, []>("value_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180989888))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181809152))), name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181809280)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = var_1644, groups = var_1596, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = var_1642, weight = layers_12_self_attn_v_proj_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
+            tensor<int32, [4]> var_1648 = const()[name = tensor<string, []>("op_1648"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1649_cast_fp16 = reshape(shape = var_1648, x = query_25_cast_fp16)[name = tensor<string, []>("op_1649_cast_fp16")];
+            tensor<fp16, []> var_1650_to_fp16 = const()[name = tensor<string, []>("op_1650_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1651_cast_fp16 = mul(x = var_1649_cast_fp16, y = var_1650_to_fp16)[name = tensor<string, []>("op_1651_cast_fp16")];
+            tensor<int32, [4]> var_1652 = const()[name = tensor<string, []>("op_1652"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1653_cast_fp16 = reshape(shape = var_1652, x = key_25_cast_fp16)[name = tensor<string, []>("op_1653_cast_fp16")];
+            tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1651_cast_fp16, y = var_1653_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1656_cast_fp16 = softmax(axis = var_1594, x = mh_w_25_cast_fp16)[name = tensor<string, []>("op_1656_cast_fp16")];
+            tensor<int32, [4]> var_1657 = const()[name = tensor<string, []>("op_1657"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1658_cast_fp16 = reshape(shape = var_1657, x = value_25_cast_fp16)[name = tensor<string, []>("op_1658_cast_fp16")];
+            tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1658_cast_fp16, y = var_1656_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
+            tensor<int32, [4]> var_1661 = const()[name = tensor<string, []>("op_1661"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_97_cast_fp16 = reshape(shape = var_1661, x = attn_25_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
+            tensor<int32, [2]> var_1665 = const()[name = tensor<string, []>("op_1665"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1667 = const()[name = tensor<string, []>("op_1667"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_51_pad_type_0 = const()[name = tensor<string, []>("obj_51_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_51_pad_0 = const()[name = tensor<string, []>("obj_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_12_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181811904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182631168))), name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182631296)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_51_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = var_1667, groups = var_1596, pad = obj_51_pad_0, pad_type = obj_51_pad_type_0, strides = var_1665, weight = layers_12_self_attn_o_proj_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
+            tensor<int32, [1]> var_1673 = const()[name = tensor<string, []>("op_1673"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_51_cast_fp16 = reduce_mean(axes = var_1673, keep_dims = var_1597, x = inputs_51_cast_fp16)[name = tensor<string, []>("channels_mean_51_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor<string, []>("zero_mean_51_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor<string, []>("zero_mean_sq_51_cast_fp16")];
+            tensor<int32, [1]> var_1677 = const()[name = tensor<string, []>("op_1677"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1678_cast_fp16 = reduce_mean(axes = var_1677, keep_dims = var_1597, x = zero_mean_sq_51_cast_fp16)[name = tensor<string, []>("op_1678_cast_fp16")];
+            tensor<fp16, []> var_1679_to_fp16 = const()[name = tensor<string, []>("op_1679_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1680_cast_fp16 = add(x = var_1678_cast_fp16, y = var_1679_to_fp16)[name = tensor<string, []>("op_1680_cast_fp16")];
+            tensor<fp32, []> denom_51_epsilon_0 = const()[name = tensor<string, []>("denom_51_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0, x = var_1680_cast_fp16)[name = tensor<string, []>("denom_51_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
+            tensor<fp16, [1280]> input_99_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_99_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182633920)))];
+            tensor<fp16, [1280]> input_99_beta_0_to_fp16 = const()[name = tensor<string, []>("input_99_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182636544)))];
+            tensor<fp16, []> input_99_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_99_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_99_cast_fp16 = batch_norm(beta = input_99_beta_0_to_fp16, epsilon = input_99_epsilon_0_to_fp16, gamma = input_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [2]> var_1691 = const()[name = tensor<string, []>("op_1691"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1693 = const()[name = tensor<string, []>("op_1693"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_101_pad_type_0 = const()[name = tensor<string, []>("input_101_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_12_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182639168))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185916032))), name = tensor<string, []>("layers_12_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_12_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185916160)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_101_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = var_1693, groups = var_1596, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = var_1691, weight = layers_12_fc1_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
+            tensor<string, []> input_103_mode_0 = const()[name = tensor<string, []>("input_103_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, [2]> var_1699 = const()[name = tensor<string, []>("op_1699"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1701 = const()[name = tensor<string, []>("op_1701"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_29_pad_type_0 = const()[name = tensor<string, []>("hidden_states_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = tensor<string, []>("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_12_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185926464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189203328))), name = tensor<string, []>("layers_12_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_12_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189203456)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_29_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = var_1701, groups = var_1596, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = var_1699, weight = layers_12_fc2_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = tensor<string, []>("hidden_states_29_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
+            tensor<int32, []> var_1712 = const()[name = tensor<string, []>("op_1712"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1714 = const()[name = tensor<string, []>("op_1714"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1725 = const()[name = tensor<string, []>("op_1725"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_53_cast_fp16 = reduce_mean(axes = var_1725, keep_dims = var_1715, x = inputs_53_cast_fp16)[name = tensor<string, []>("channels_mean_53_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor<string, []>("zero_mean_53_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor<string, []>("zero_mean_sq_53_cast_fp16")];
+            tensor<int32, [1]> var_1729 = const()[name = tensor<string, []>("op_1729"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1730_cast_fp16 = reduce_mean(axes = var_1729, keep_dims = var_1715, x = zero_mean_sq_53_cast_fp16)[name = tensor<string, []>("op_1730_cast_fp16")];
+            tensor<fp16, []> var_1731_to_fp16 = const()[name = tensor<string, []>("op_1731_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1732_cast_fp16 = add(x = var_1730_cast_fp16, y = var_1731_to_fp16)[name = tensor<string, []>("op_1732_cast_fp16")];
+            tensor<fp32, []> denom_53_epsilon_0 = const()[name = tensor<string, []>("denom_53_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0, x = var_1732_cast_fp16)[name = tensor<string, []>("denom_53_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
+            tensor<fp16, [1280]> obj_53_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_53_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189206080)))];
+            tensor<fp16, [1280]> obj_53_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_53_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189208704)))];
+            tensor<fp16, []> obj_53_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_53_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
+            tensor<int32, [2]> var_1747 = const()[name = tensor<string, []>("op_1747"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1749 = const()[name = tensor<string, []>("op_1749"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_27_pad_type_0 = const()[name = tensor<string, []>("query_27_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_27_pad_0 = const()[name = tensor<string, []>("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189211328))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190030592))), name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190030720)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_27_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = var_1749, groups = var_1714, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = var_1747, weight = layers_13_self_attn_q_proj_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
+            tensor<int32, [2]> var_1753 = const()[name = tensor<string, []>("op_1753"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1755 = const()[name = tensor<string, []>("op_1755"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190033344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190852608))), name = tensor<string, []>("layers_13_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_27_cast_fp16 = conv(dilations = var_1755, groups = var_1714, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = var_1753, weight = layers_13_self_attn_k_proj_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<int32, [2]> var_1760 = const()[name = tensor<string, []>("op_1760"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1762 = const()[name = tensor<string, []>("op_1762"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_27_pad_type_0 = const()[name = tensor<string, []>("value_27_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_27_pad_0 = const()[name = tensor<string, []>("value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190852736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191672000))), name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191672128)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = var_1762, groups = var_1714, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = var_1760, weight = layers_13_self_attn_v_proj_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor<string, []>("value_27_cast_fp16")];
+            tensor<int32, [4]> var_1766 = const()[name = tensor<string, []>("op_1766"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1767_cast_fp16 = reshape(shape = var_1766, x = query_27_cast_fp16)[name = tensor<string, []>("op_1767_cast_fp16")];
+            tensor<fp16, []> var_1768_to_fp16 = const()[name = tensor<string, []>("op_1768_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1769_cast_fp16 = mul(x = var_1767_cast_fp16, y = var_1768_to_fp16)[name = tensor<string, []>("op_1769_cast_fp16")];
+            tensor<int32, [4]> var_1770 = const()[name = tensor<string, []>("op_1770"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1771_cast_fp16 = reshape(shape = var_1770, x = key_27_cast_fp16)[name = tensor<string, []>("op_1771_cast_fp16")];
+            tensor<bool, []> mh_w_27_transpose_x_0 = const()[name = tensor<string, []>("mh_w_27_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_27_transpose_y_0 = const()[name = tensor<string, []>("mh_w_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_1769_cast_fp16, y = var_1771_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1774_cast_fp16 = softmax(axis = var_1712, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1774_cast_fp16")];
+            tensor<int32, [4]> var_1775 = const()[name = tensor<string, []>("op_1775"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1776_cast_fp16 = reshape(shape = var_1775, x = value_27_cast_fp16)[name = tensor<string, []>("op_1776_cast_fp16")];
+            tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1776_cast_fp16, y = var_1774_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
+            tensor<int32, [4]> var_1779 = const()[name = tensor<string, []>("op_1779"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_105_cast_fp16 = reshape(shape = var_1779, x = attn_27_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<int32, [2]> var_1783 = const()[name = tensor<string, []>("op_1783"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1785 = const()[name = tensor<string, []>("op_1785"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_55_pad_type_0 = const()[name = tensor<string, []>("obj_55_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_55_pad_0 = const()[name = tensor<string, []>("obj_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191674752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192494016))), name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192494144)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_55_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = var_1785, groups = var_1714, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = var_1783, weight = layers_13_self_attn_o_proj_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")];
+            tensor<int32, [1]> var_1791 = const()[name = tensor<string, []>("op_1791"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_55_cast_fp16 = reduce_mean(axes = var_1791, keep_dims = var_1715, x = inputs_55_cast_fp16)[name = tensor<string, []>("channels_mean_55_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor<string, []>("zero_mean_55_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor<string, []>("zero_mean_sq_55_cast_fp16")];
+            tensor<int32, [1]> var_1795 = const()[name = tensor<string, []>("op_1795"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1796_cast_fp16 = reduce_mean(axes = var_1795, keep_dims = var_1715, x = zero_mean_sq_55_cast_fp16)[name = tensor<string, []>("op_1796_cast_fp16")];
+            tensor<fp16, []> var_1797_to_fp16 = const()[name = tensor<string, []>("op_1797_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1798_cast_fp16 = add(x = var_1796_cast_fp16, y = var_1797_to_fp16)[name = tensor<string, []>("op_1798_cast_fp16")];
+            tensor<fp32, []> denom_55_epsilon_0 = const()[name = tensor<string, []>("denom_55_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0, x = var_1798_cast_fp16)[name = tensor<string, []>("denom_55_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
+            tensor<fp16, [1280]> input_107_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_107_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192496768)))];
+            tensor<fp16, [1280]> input_107_beta_0_to_fp16 = const()[name = tensor<string, []>("input_107_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192499392)))];
+            tensor<fp16, []> input_107_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_107_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_107_cast_fp16 = batch_norm(beta = input_107_beta_0_to_fp16, epsilon = input_107_epsilon_0_to_fp16, gamma = input_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
+            tensor<int32, [2]> var_1809 = const()[name = tensor<string, []>("op_1809"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1811 = const()[name = tensor<string, []>("op_1811"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_13_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192502016))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195778880))), name = tensor<string, []>("layers_13_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_13_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195779008)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_109_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = var_1811, groups = var_1714, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = var_1809, weight = layers_13_fc1_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
+            tensor<string, []> input_111_mode_0 = const()[name = tensor<string, []>("input_111_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<int32, [2]> var_1817 = const()[name = tensor<string, []>("op_1817"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1819 = const()[name = tensor<string, []>("op_1819"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_31_pad_type_0 = const()[name = tensor<string, []>("hidden_states_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_31_pad_0 = const()[name = tensor<string, []>("hidden_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_13_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195789312))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199066176))), name = tensor<string, []>("layers_13_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_13_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199066304)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_31_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = var_1819, groups = var_1714, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = var_1817, weight = layers_13_fc2_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
+            tensor<int32, []> var_1830 = const()[name = tensor<string, []>("op_1830"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1832 = const()[name = tensor<string, []>("op_1832"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1833 = const()[name = tensor<string, []>("op_1833"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1843 = const()[name = tensor<string, []>("op_1843"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_57_cast_fp16 = reduce_mean(axes = var_1843, keep_dims = var_1833, x = inputs_57_cast_fp16)[name = tensor<string, []>("channels_mean_57_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor<string, []>("zero_mean_57_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor<string, []>("zero_mean_sq_57_cast_fp16")];
+            tensor<int32, [1]> var_1847 = const()[name = tensor<string, []>("op_1847"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1848_cast_fp16 = reduce_mean(axes = var_1847, keep_dims = var_1833, x = zero_mean_sq_57_cast_fp16)[name = tensor<string, []>("op_1848_cast_fp16")];
+            tensor<fp16, []> var_1849_to_fp16 = const()[name = tensor<string, []>("op_1849_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1850_cast_fp16 = add(x = var_1848_cast_fp16, y = var_1849_to_fp16)[name = tensor<string, []>("op_1850_cast_fp16")];
+            tensor<fp32, []> denom_57_epsilon_0 = const()[name = tensor<string, []>("denom_57_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0, x = var_1850_cast_fp16)[name = tensor<string, []>("denom_57_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
+            tensor<fp16, [1280]> obj_57_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199068928)))];
+            tensor<fp16, [1280]> obj_57_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_57_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199071552)))];
+            tensor<fp16, []> obj_57_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_57_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
+            tensor<int32, [2]> var_1865 = const()[name = tensor<string, []>("op_1865"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1867 = const()[name = tensor<string, []>("op_1867"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199074176))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199893440))), name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199893568)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_29_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = var_1867, groups = var_1832, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = var_1865, weight = layers_14_self_attn_q_proj_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
+            tensor<int32, [2]> var_1871 = const()[name = tensor<string, []>("op_1871"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1873 = const()[name = tensor<string, []>("op_1873"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_29_pad_type_0 = const()[name = tensor<string, []>("key_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_29_pad_0 = const()[name = tensor<string, []>("key_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199896192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200715456))), name = tensor<string, []>("layers_14_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_29_cast_fp16 = conv(dilations = var_1873, groups = var_1832, pad = key_29_pad_0, pad_type = key_29_pad_type_0, strides = var_1871, weight = layers_14_self_attn_k_proj_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<int32, [2]> var_1878 = const()[name = tensor<string, []>("op_1878"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1880 = const()[name = tensor<string, []>("op_1880"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_29_pad_type_0 = const()[name = tensor<string, []>("value_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_29_pad_0 = const()[name = tensor<string, []>("value_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200715584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201534848))), name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201534976)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = var_1880, groups = var_1832, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = var_1878, weight = layers_14_self_attn_v_proj_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
+            tensor<int32, [4]> var_1884 = const()[name = tensor<string, []>("op_1884"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1885_cast_fp16 = reshape(shape = var_1884, x = query_29_cast_fp16)[name = tensor<string, []>("op_1885_cast_fp16")];
+            tensor<fp16, []> var_1886_to_fp16 = const()[name = tensor<string, []>("op_1886_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_1887_cast_fp16 = mul(x = var_1885_cast_fp16, y = var_1886_to_fp16)[name = tensor<string, []>("op_1887_cast_fp16")];
+            tensor<int32, [4]> var_1888 = const()[name = tensor<string, []>("op_1888"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1889_cast_fp16 = reshape(shape = var_1888, x = key_29_cast_fp16)[name = tensor<string, []>("op_1889_cast_fp16")];
+            tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1887_cast_fp16, y = var_1889_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_1892_cast_fp16 = softmax(axis = var_1830, x = mh_w_29_cast_fp16)[name = tensor<string, []>("op_1892_cast_fp16")];
+            tensor<int32, [4]> var_1893 = const()[name = tensor<string, []>("op_1893"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_1894_cast_fp16 = reshape(shape = var_1893, x = value_29_cast_fp16)[name = tensor<string, []>("op_1894_cast_fp16")];
+            tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1894_cast_fp16, y = var_1892_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
+            tensor<int32, [4]> var_1897 = const()[name = tensor<string, []>("op_1897"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_113_cast_fp16 = reshape(shape = var_1897, x = attn_29_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
+            tensor<int32, [2]> var_1901 = const()[name = tensor<string, []>("op_1901"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1903 = const()[name = tensor<string, []>("op_1903"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_59_pad_type_0 = const()[name = tensor<string, []>("obj_59_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_59_pad_0 = const()[name = tensor<string, []>("obj_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_14_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201537600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202356864))), name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202356992)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_59_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = var_1903, groups = var_1832, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = var_1901, weight = layers_14_self_attn_o_proj_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor<string, []>("obj_59_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
+            tensor<int32, [1]> var_1909 = const()[name = tensor<string, []>("op_1909"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_59_cast_fp16 = reduce_mean(axes = var_1909, keep_dims = var_1833, x = inputs_59_cast_fp16)[name = tensor<string, []>("channels_mean_59_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor<string, []>("zero_mean_59_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor<string, []>("zero_mean_sq_59_cast_fp16")];
+            tensor<int32, [1]> var_1913 = const()[name = tensor<string, []>("op_1913"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1914_cast_fp16 = reduce_mean(axes = var_1913, keep_dims = var_1833, x = zero_mean_sq_59_cast_fp16)[name = tensor<string, []>("op_1914_cast_fp16")];
+            tensor<fp16, []> var_1915_to_fp16 = const()[name = tensor<string, []>("op_1915_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1916_cast_fp16 = add(x = var_1914_cast_fp16, y = var_1915_to_fp16)[name = tensor<string, []>("op_1916_cast_fp16")];
+            tensor<fp32, []> denom_59_epsilon_0 = const()[name = tensor<string, []>("denom_59_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0, x = var_1916_cast_fp16)[name = tensor<string, []>("denom_59_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
+            tensor<fp16, [1280]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202359616)))];
+            tensor<fp16, [1280]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202362240)))];
+            tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [2]> var_1927 = const()[name = tensor<string, []>("op_1927"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1929 = const()[name = tensor<string, []>("op_1929"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_14_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202364864))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205641728))), name = tensor<string, []>("layers_14_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_14_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205641856)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_117_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = var_1929, groups = var_1832, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = var_1927, weight = layers_14_fc1_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<string, []> input_119_mode_0 = const()[name = tensor<string, []>("input_119_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
+            tensor<int32, [2]> var_1935 = const()[name = tensor<string, []>("op_1935"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1937 = const()[name = tensor<string, []>("op_1937"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_33_pad_type_0 = const()[name = tensor<string, []>("hidden_states_33_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = tensor<string, []>("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_14_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205652160))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208929024))), name = tensor<string, []>("layers_14_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_14_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208929152)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_33_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = var_1937, groups = var_1832, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = var_1935, weight = layers_14_fc2_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = tensor<string, []>("hidden_states_33_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
+            tensor<int32, []> var_1948 = const()[name = tensor<string, []>("op_1948"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_1950 = const()[name = tensor<string, []>("op_1950"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1951 = const()[name = tensor<string, []>("op_1951"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_1961 = const()[name = tensor<string, []>("op_1961"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_61_cast_fp16 = reduce_mean(axes = var_1961, keep_dims = var_1951, x = inputs_61_cast_fp16)[name = tensor<string, []>("channels_mean_61_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor<string, []>("zero_mean_61_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor<string, []>("zero_mean_sq_61_cast_fp16")];
+            tensor<int32, [1]> var_1965 = const()[name = tensor<string, []>("op_1965"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1966_cast_fp16 = reduce_mean(axes = var_1965, keep_dims = var_1951, x = zero_mean_sq_61_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
+            tensor<fp16, []> var_1967_to_fp16 = const()[name = tensor<string, []>("op_1967_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_1968_cast_fp16 = add(x = var_1966_cast_fp16, y = var_1967_to_fp16)[name = tensor<string, []>("op_1968_cast_fp16")];
+            tensor<fp32, []> denom_61_epsilon_0 = const()[name = tensor<string, []>("denom_61_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0, x = var_1968_cast_fp16)[name = tensor<string, []>("denom_61_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
+            tensor<fp16, [1280]> obj_61_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_61_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208931776)))];
+            tensor<fp16, [1280]> obj_61_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_61_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208934400)))];
+            tensor<fp16, []> obj_61_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_61_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("obj_61_cast_fp16")];
+            tensor<int32, [2]> var_1983 = const()[name = tensor<string, []>("op_1983"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_31_pad_type_0 = const()[name = tensor<string, []>("query_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_31_pad_0 = const()[name = tensor<string, []>("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208937024))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209756288))), name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209756416)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_31_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = var_1985, groups = var_1950, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = var_1983, weight = layers_15_self_attn_q_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
+            tensor<int32, [2]> var_1989 = const()[name = tensor<string, []>("op_1989"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1991 = const()[name = tensor<string, []>("op_1991"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209759040))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210578304))), name = tensor<string, []>("layers_15_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_31_cast_fp16 = conv(dilations = var_1991, groups = var_1950, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = var_1989, weight = layers_15_self_attn_k_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<int32, [2]> var_1996 = const()[name = tensor<string, []>("op_1996"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_1998 = const()[name = tensor<string, []>("op_1998"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_31_pad_type_0 = const()[name = tensor<string, []>("value_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_31_pad_0 = const()[name = tensor<string, []>("value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210578432))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211397696))), name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211397824)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = var_1998, groups = var_1950, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = var_1996, weight = layers_15_self_attn_v_proj_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor<string, []>("value_31_cast_fp16")];
+            tensor<int32, [4]> var_2002 = const()[name = tensor<string, []>("op_2002"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2003_cast_fp16 = reshape(shape = var_2002, x = query_31_cast_fp16)[name = tensor<string, []>("op_2003_cast_fp16")];
+            tensor<fp16, []> var_2004_to_fp16 = const()[name = tensor<string, []>("op_2004_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2005_cast_fp16 = mul(x = var_2003_cast_fp16, y = var_2004_to_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
+            tensor<int32, [4]> var_2006 = const()[name = tensor<string, []>("op_2006"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2007_cast_fp16 = reshape(shape = var_2006, x = key_31_cast_fp16)[name = tensor<string, []>("op_2007_cast_fp16")];
+            tensor<bool, []> mh_w_31_transpose_x_0 = const()[name = tensor<string, []>("mh_w_31_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_31_transpose_y_0 = const()[name = tensor<string, []>("mh_w_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2005_cast_fp16, y = var_2007_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2010_cast_fp16 = softmax(axis = var_1948, x = mh_w_31_cast_fp16)[name = tensor<string, []>("op_2010_cast_fp16")];
+            tensor<int32, [4]> var_2011 = const()[name = tensor<string, []>("op_2011"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2012_cast_fp16 = reshape(shape = var_2011, x = value_31_cast_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
+            tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2012_cast_fp16, y = var_2010_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
+            tensor<int32, [4]> var_2015 = const()[name = tensor<string, []>("op_2015"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_121_cast_fp16 = reshape(shape = var_2015, x = attn_31_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<int32, [2]> var_2019 = const()[name = tensor<string, []>("op_2019"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2021 = const()[name = tensor<string, []>("op_2021"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_63_pad_type_0 = const()[name = tensor<string, []>("obj_63_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_63_pad_0 = const()[name = tensor<string, []>("obj_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_15_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211400448))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212219712))), name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212219840)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_63_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = var_2021, groups = var_1950, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = var_2019, weight = layers_15_self_attn_o_proj_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
+            tensor<int32, [1]> var_2027 = const()[name = tensor<string, []>("op_2027"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_63_cast_fp16 = reduce_mean(axes = var_2027, keep_dims = var_1951, x = inputs_63_cast_fp16)[name = tensor<string, []>("channels_mean_63_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor<string, []>("zero_mean_63_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor<string, []>("zero_mean_sq_63_cast_fp16")];
+            tensor<int32, [1]> var_2031 = const()[name = tensor<string, []>("op_2031"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2032_cast_fp16 = reduce_mean(axes = var_2031, keep_dims = var_1951, x = zero_mean_sq_63_cast_fp16)[name = tensor<string, []>("op_2032_cast_fp16")];
+            tensor<fp16, []> var_2033_to_fp16 = const()[name = tensor<string, []>("op_2033_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2034_cast_fp16 = add(x = var_2032_cast_fp16, y = var_2033_to_fp16)[name = tensor<string, []>("op_2034_cast_fp16")];
+            tensor<fp32, []> denom_63_epsilon_0 = const()[name = tensor<string, []>("denom_63_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0, x = var_2034_cast_fp16)[name = tensor<string, []>("denom_63_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
+            tensor<fp16, [1280]> input_123_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_123_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212222464)))];
+            tensor<fp16, [1280]> input_123_beta_0_to_fp16 = const()[name = tensor<string, []>("input_123_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212225088)))];
+            tensor<fp16, []> input_123_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_123_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_123_cast_fp16 = batch_norm(beta = input_123_beta_0_to_fp16, epsilon = input_123_epsilon_0_to_fp16, gamma = input_123_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
+            tensor<int32, [2]> var_2045 = const()[name = tensor<string, []>("op_2045"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2047 = const()[name = tensor<string, []>("op_2047"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_125_pad_type_0 = const()[name = tensor<string, []>("input_125_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_125_pad_0 = const()[name = tensor<string, []>("input_125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_15_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212227712))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215504576))), name = tensor<string, []>("layers_15_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_15_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215504704)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_125_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = var_2047, groups = var_1950, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = var_2045, weight = layers_15_fc1_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
+            tensor<string, []> input_127_mode_0 = const()[name = tensor<string, []>("input_127_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = input_125_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [2]> var_2053 = const()[name = tensor<string, []>("op_2053"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2055 = const()[name = tensor<string, []>("op_2055"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_35_pad_type_0 = const()[name = tensor<string, []>("hidden_states_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = tensor<string, []>("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_15_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215515008))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218791872))), name = tensor<string, []>("layers_15_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_15_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218792000)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_35_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = var_2055, groups = var_1950, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = var_2053, weight = layers_15_fc2_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor<string, []>("hidden_states_35_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")];
+            tensor<int32, []> var_2066 = const()[name = tensor<string, []>("op_2066"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2068 = const()[name = tensor<string, []>("op_2068"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2069 = const()[name = tensor<string, []>("op_2069"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2079 = const()[name = tensor<string, []>("op_2079"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_65_cast_fp16 = reduce_mean(axes = var_2079, keep_dims = var_2069, x = inputs_65_cast_fp16)[name = tensor<string, []>("channels_mean_65_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor<string, []>("zero_mean_65_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor<string, []>("zero_mean_sq_65_cast_fp16")];
+            tensor<int32, [1]> var_2083 = const()[name = tensor<string, []>("op_2083"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2084_cast_fp16 = reduce_mean(axes = var_2083, keep_dims = var_2069, x = zero_mean_sq_65_cast_fp16)[name = tensor<string, []>("op_2084_cast_fp16")];
+            tensor<fp16, []> var_2085_to_fp16 = const()[name = tensor<string, []>("op_2085_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2086_cast_fp16 = add(x = var_2084_cast_fp16, y = var_2085_to_fp16)[name = tensor<string, []>("op_2086_cast_fp16")];
+            tensor<fp32, []> denom_65_epsilon_0 = const()[name = tensor<string, []>("denom_65_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0, x = var_2086_cast_fp16)[name = tensor<string, []>("denom_65_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
+            tensor<fp16, [1280]> obj_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_65_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218794624)))];
+            tensor<fp16, [1280]> obj_65_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_65_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218797248)))];
+            tensor<fp16, []> obj_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")];
+            tensor<int32, [2]> var_2101 = const()[name = tensor<string, []>("op_2101"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2103 = const()[name = tensor<string, []>("op_2103"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218799872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219619136))), name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219619264)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_33_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = var_2103, groups = var_2068, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = var_2101, weight = layers_16_self_attn_q_proj_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
+            tensor<int32, [2]> var_2107 = const()[name = tensor<string, []>("op_2107"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2109 = const()[name = tensor<string, []>("op_2109"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_33_pad_type_0 = const()[name = tensor<string, []>("key_33_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_33_pad_0 = const()[name = tensor<string, []>("key_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219621888))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220441152))), name = tensor<string, []>("layers_16_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_33_cast_fp16 = conv(dilations = var_2109, groups = var_2068, pad = key_33_pad_0, pad_type = key_33_pad_type_0, strides = var_2107, weight = layers_16_self_attn_k_proj_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
+            tensor<int32, [2]> var_2114 = const()[name = tensor<string, []>("op_2114"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2116 = const()[name = tensor<string, []>("op_2116"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_33_pad_type_0 = const()[name = tensor<string, []>("value_33_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_33_pad_0 = const()[name = tensor<string, []>("value_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220441280))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221260544))), name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221260672)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = var_2116, groups = var_2068, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = var_2114, weight = layers_16_self_attn_v_proj_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
+            tensor<int32, [4]> var_2120 = const()[name = tensor<string, []>("op_2120"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2121_cast_fp16 = reshape(shape = var_2120, x = query_33_cast_fp16)[name = tensor<string, []>("op_2121_cast_fp16")];
+            tensor<fp16, []> var_2122_to_fp16 = const()[name = tensor<string, []>("op_2122_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2123_cast_fp16 = mul(x = var_2121_cast_fp16, y = var_2122_to_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
+            tensor<int32, [4]> var_2124 = const()[name = tensor<string, []>("op_2124"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2125_cast_fp16 = reshape(shape = var_2124, x = key_33_cast_fp16)[name = tensor<string, []>("op_2125_cast_fp16")];
+            tensor<bool, []> mh_w_33_transpose_x_0 = const()[name = tensor<string, []>("mh_w_33_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_33_transpose_y_0 = const()[name = tensor<string, []>("mh_w_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2123_cast_fp16, y = var_2125_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2128_cast_fp16 = softmax(axis = var_2066, x = mh_w_33_cast_fp16)[name = tensor<string, []>("op_2128_cast_fp16")];
+            tensor<int32, [4]> var_2129 = const()[name = tensor<string, []>("op_2129"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2130_cast_fp16 = reshape(shape = var_2129, x = value_33_cast_fp16)[name = tensor<string, []>("op_2130_cast_fp16")];
+            tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2130_cast_fp16, y = var_2128_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
+            tensor<int32, [4]> var_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_129_cast_fp16 = reshape(shape = var_2133, x = attn_33_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [2]> var_2137 = const()[name = tensor<string, []>("op_2137"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2139 = const()[name = tensor<string, []>("op_2139"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_67_pad_type_0 = const()[name = tensor<string, []>("obj_67_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_67_pad_0 = const()[name = tensor<string, []>("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_16_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221263296))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222082560))), name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222082688)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_67_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = var_2139, groups = var_2068, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_2137, weight = layers_16_self_attn_o_proj_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")];
+            tensor<int32, [1]> var_2145 = const()[name = tensor<string, []>("op_2145"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_67_cast_fp16 = reduce_mean(axes = var_2145, keep_dims = var_2069, x = inputs_67_cast_fp16)[name = tensor<string, []>("channels_mean_67_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor<string, []>("zero_mean_67_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor<string, []>("zero_mean_sq_67_cast_fp16")];
+            tensor<int32, [1]> var_2149 = const()[name = tensor<string, []>("op_2149"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2150_cast_fp16 = reduce_mean(axes = var_2149, keep_dims = var_2069, x = zero_mean_sq_67_cast_fp16)[name = tensor<string, []>("op_2150_cast_fp16")];
+            tensor<fp16, []> var_2151_to_fp16 = const()[name = tensor<string, []>("op_2151_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2152_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2151_to_fp16)[name = tensor<string, []>("op_2152_cast_fp16")];
+            tensor<fp32, []> denom_67_epsilon_0 = const()[name = tensor<string, []>("denom_67_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0, x = var_2152_cast_fp16)[name = tensor<string, []>("denom_67_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
+            tensor<fp16, [1280]> input_131_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_131_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222085312)))];
+            tensor<fp16, [1280]> input_131_beta_0_to_fp16 = const()[name = tensor<string, []>("input_131_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222087936)))];
+            tensor<fp16, []> input_131_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_131_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<int32, [2]> var_2163 = const()[name = tensor<string, []>("op_2163"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2165 = const()[name = tensor<string, []>("op_2165"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_133_pad_type_0 = const()[name = tensor<string, []>("input_133_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_133_pad_0 = const()[name = tensor<string, []>("input_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_16_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222090560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225367424))), name = tensor<string, []>("layers_16_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_16_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225367552)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_133_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = var_2165, groups = var_2068, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = var_2163, weight = layers_16_fc1_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
+            tensor<string, []> input_135_mode_0 = const()[name = tensor<string, []>("input_135_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_135_cast_fp16 = gelu(mode = input_135_mode_0, x = input_133_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<int32, [2]> var_2171 = const()[name = tensor<string, []>("op_2171"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2173 = const()[name = tensor<string, []>("op_2173"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_37_pad_type_0 = const()[name = tensor<string, []>("hidden_states_37_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = tensor<string, []>("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_16_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225377856))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228654720))), name = tensor<string, []>("layers_16_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_16_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228654848)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_37_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = var_2173, groups = var_2068, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = var_2171, weight = layers_16_fc2_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = tensor<string, []>("hidden_states_37_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
+            tensor<int32, []> var_2184 = const()[name = tensor<string, []>("op_2184"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2186 = const()[name = tensor<string, []>("op_2186"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2187 = const()[name = tensor<string, []>("op_2187"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2197 = const()[name = tensor<string, []>("op_2197"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_69_cast_fp16 = reduce_mean(axes = var_2197, keep_dims = var_2187, x = inputs_69_cast_fp16)[name = tensor<string, []>("channels_mean_69_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor<string, []>("zero_mean_69_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor<string, []>("zero_mean_sq_69_cast_fp16")];
+            tensor<int32, [1]> var_2201 = const()[name = tensor<string, []>("op_2201"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2202_cast_fp16 = reduce_mean(axes = var_2201, keep_dims = var_2187, x = zero_mean_sq_69_cast_fp16)[name = tensor<string, []>("op_2202_cast_fp16")];
+            tensor<fp16, []> var_2203_to_fp16 = const()[name = tensor<string, []>("op_2203_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2204_cast_fp16 = add(x = var_2202_cast_fp16, y = var_2203_to_fp16)[name = tensor<string, []>("op_2204_cast_fp16")];
+            tensor<fp32, []> denom_69_epsilon_0 = const()[name = tensor<string, []>("denom_69_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0, x = var_2204_cast_fp16)[name = tensor<string, []>("denom_69_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
+            tensor<fp16, [1280]> obj_69_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_69_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228657472)))];
+            tensor<fp16, [1280]> obj_69_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_69_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228660096)))];
+            tensor<fp16, []> obj_69_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_69_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
+            tensor<int32, [2]> var_2219 = const()[name = tensor<string, []>("op_2219"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2221 = const()[name = tensor<string, []>("op_2221"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_35_pad_type_0 = const()[name = tensor<string, []>("query_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_35_pad_0 = const()[name = tensor<string, []>("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228662720))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229481984))), name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229482112)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_35_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = var_2221, groups = var_2186, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = var_2219, weight = layers_17_self_attn_q_proj_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
+            tensor<int32, [2]> var_2225 = const()[name = tensor<string, []>("op_2225"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2227 = const()[name = tensor<string, []>("op_2227"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229484736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230304000))), name = tensor<string, []>("layers_17_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_35_cast_fp16 = conv(dilations = var_2227, groups = var_2186, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = var_2225, weight = layers_17_self_attn_k_proj_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor<string, []>("key_35_cast_fp16")];
+            tensor<int32, [2]> var_2232 = const()[name = tensor<string, []>("op_2232"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2234 = const()[name = tensor<string, []>("op_2234"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_35_pad_type_0 = const()[name = tensor<string, []>("value_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_35_pad_0 = const()[name = tensor<string, []>("value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230304128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231123392))), name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231123520)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = var_2234, groups = var_2186, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = var_2232, weight = layers_17_self_attn_v_proj_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor<string, []>("value_35_cast_fp16")];
+            tensor<int32, [4]> var_2238 = const()[name = tensor<string, []>("op_2238"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2239_cast_fp16 = reshape(shape = var_2238, x = query_35_cast_fp16)[name = tensor<string, []>("op_2239_cast_fp16")];
+            tensor<fp16, []> var_2240_to_fp16 = const()[name = tensor<string, []>("op_2240_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2241_cast_fp16 = mul(x = var_2239_cast_fp16, y = var_2240_to_fp16)[name = tensor<string, []>("op_2241_cast_fp16")];
+            tensor<int32, [4]> var_2242 = const()[name = tensor<string, []>("op_2242"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2243_cast_fp16 = reshape(shape = var_2242, x = key_35_cast_fp16)[name = tensor<string, []>("op_2243_cast_fp16")];
+            tensor<bool, []> mh_w_35_transpose_x_0 = const()[name = tensor<string, []>("mh_w_35_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_35_transpose_y_0 = const()[name = tensor<string, []>("mh_w_35_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2241_cast_fp16, y = var_2243_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2246_cast_fp16 = softmax(axis = var_2184, x = mh_w_35_cast_fp16)[name = tensor<string, []>("op_2246_cast_fp16")];
+            tensor<int32, [4]> var_2247 = const()[name = tensor<string, []>("op_2247"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2248_cast_fp16 = reshape(shape = var_2247, x = value_35_cast_fp16)[name = tensor<string, []>("op_2248_cast_fp16")];
+            tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2248_cast_fp16, y = var_2246_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
+            tensor<int32, [4]> var_2251 = const()[name = tensor<string, []>("op_2251"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_137_cast_fp16 = reshape(shape = var_2251, x = attn_35_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
+            tensor<int32, [2]> var_2255 = const()[name = tensor<string, []>("op_2255"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2257 = const()[name = tensor<string, []>("op_2257"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_71_pad_type_0 = const()[name = tensor<string, []>("obj_71_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_71_pad_0 = const()[name = tensor<string, []>("obj_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_17_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231126144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231945408))), name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231945536)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_71_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = var_2257, groups = var_2186, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = var_2255, weight = layers_17_self_attn_o_proj_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
+            tensor<int32, [1]> var_2263 = const()[name = tensor<string, []>("op_2263"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_71_cast_fp16 = reduce_mean(axes = var_2263, keep_dims = var_2187, x = inputs_71_cast_fp16)[name = tensor<string, []>("channels_mean_71_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor<string, []>("zero_mean_71_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor<string, []>("zero_mean_sq_71_cast_fp16")];
+            tensor<int32, [1]> var_2267 = const()[name = tensor<string, []>("op_2267"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2268_cast_fp16 = reduce_mean(axes = var_2267, keep_dims = var_2187, x = zero_mean_sq_71_cast_fp16)[name = tensor<string, []>("op_2268_cast_fp16")];
+            tensor<fp16, []> var_2269_to_fp16 = const()[name = tensor<string, []>("op_2269_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2270_cast_fp16 = add(x = var_2268_cast_fp16, y = var_2269_to_fp16)[name = tensor<string, []>("op_2270_cast_fp16")];
+            tensor<fp32, []> denom_71_epsilon_0 = const()[name = tensor<string, []>("denom_71_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0, x = var_2270_cast_fp16)[name = tensor<string, []>("denom_71_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
+            tensor<fp16, [1280]> input_139_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_139_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231948160)))];
+            tensor<fp16, [1280]> input_139_beta_0_to_fp16 = const()[name = tensor<string, []>("input_139_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231950784)))];
+            tensor<fp16, []> input_139_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_139_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_139_cast_fp16 = batch_norm(beta = input_139_beta_0_to_fp16, epsilon = input_139_epsilon_0_to_fp16, gamma = input_139_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
+            tensor<int32, [2]> var_2281 = const()[name = tensor<string, []>("op_2281"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2283 = const()[name = tensor<string, []>("op_2283"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_141_pad_type_0 = const()[name = tensor<string, []>("input_141_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_141_pad_0 = const()[name = tensor<string, []>("input_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_17_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231953408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235230272))), name = tensor<string, []>("layers_17_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_17_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235230400)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_141_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = var_2283, groups = var_2186, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = var_2281, weight = layers_17_fc1_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<string, []> input_143_mode_0 = const()[name = tensor<string, []>("input_143_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = input_141_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<int32, [2]> var_2289 = const()[name = tensor<string, []>("op_2289"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2291 = const()[name = tensor<string, []>("op_2291"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_39_pad_type_0 = const()[name = tensor<string, []>("hidden_states_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = tensor<string, []>("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_17_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235240704))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238517568))), name = tensor<string, []>("layers_17_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_17_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238517696)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_39_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = var_2291, groups = var_2186, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = var_2289, weight = layers_17_fc2_weight_to_fp16_palettized, x = input_143_cast_fp16)[name = tensor<string, []>("hidden_states_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
+            tensor<int32, []> var_2302 = const()[name = tensor<string, []>("op_2302"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2304 = const()[name = tensor<string, []>("op_2304"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2305 = const()[name = tensor<string, []>("op_2305"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2315 = const()[name = tensor<string, []>("op_2315"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_73_cast_fp16 = reduce_mean(axes = var_2315, keep_dims = var_2305, x = inputs_73_cast_fp16)[name = tensor<string, []>("channels_mean_73_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor<string, []>("zero_mean_73_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor<string, []>("zero_mean_sq_73_cast_fp16")];
+            tensor<int32, [1]> var_2319 = const()[name = tensor<string, []>("op_2319"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2320_cast_fp16 = reduce_mean(axes = var_2319, keep_dims = var_2305, x = zero_mean_sq_73_cast_fp16)[name = tensor<string, []>("op_2320_cast_fp16")];
+            tensor<fp16, []> var_2321_to_fp16 = const()[name = tensor<string, []>("op_2321_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2322_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2321_to_fp16)[name = tensor<string, []>("op_2322_cast_fp16")];
+            tensor<fp32, []> denom_73_epsilon_0 = const()[name = tensor<string, []>("denom_73_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0, x = var_2322_cast_fp16)[name = tensor<string, []>("denom_73_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
+            tensor<fp16, [1280]> obj_73_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_73_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238520320)))];
+            tensor<fp16, [1280]> obj_73_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_73_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238522944)))];
+            tensor<fp16, []> obj_73_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_73_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_73_cast_fp16")];
+            tensor<int32, [2]> var_2337 = const()[name = tensor<string, []>("op_2337"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2339 = const()[name = tensor<string, []>("op_2339"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238525568))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239344832))), name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239344960)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_37_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = var_2339, groups = var_2304, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = var_2337, weight = layers_18_self_attn_q_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
+            tensor<int32, [2]> var_2343 = const()[name = tensor<string, []>("op_2343"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2345 = const()[name = tensor<string, []>("op_2345"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_37_pad_type_0 = const()[name = tensor<string, []>("key_37_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_37_pad_0 = const()[name = tensor<string, []>("key_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239347584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240166848))), name = tensor<string, []>("layers_18_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_37_cast_fp16 = conv(dilations = var_2345, groups = var_2304, pad = key_37_pad_0, pad_type = key_37_pad_type_0, strides = var_2343, weight = layers_18_self_attn_k_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
+            tensor<int32, [2]> var_2350 = const()[name = tensor<string, []>("op_2350"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2352 = const()[name = tensor<string, []>("op_2352"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_37_pad_type_0 = const()[name = tensor<string, []>("value_37_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_37_pad_0 = const()[name = tensor<string, []>("value_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240166976))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240986240))), name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240986368)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = var_2352, groups = var_2304, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = var_2350, weight = layers_18_self_attn_v_proj_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
+            tensor<int32, [4]> var_2356 = const()[name = tensor<string, []>("op_2356"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2357_cast_fp16 = reshape(shape = var_2356, x = query_37_cast_fp16)[name = tensor<string, []>("op_2357_cast_fp16")];
+            tensor<fp16, []> var_2358_to_fp16 = const()[name = tensor<string, []>("op_2358_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2359_cast_fp16 = mul(x = var_2357_cast_fp16, y = var_2358_to_fp16)[name = tensor<string, []>("op_2359_cast_fp16")];
+            tensor<int32, [4]> var_2360 = const()[name = tensor<string, []>("op_2360"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2361_cast_fp16 = reshape(shape = var_2360, x = key_37_cast_fp16)[name = tensor<string, []>("op_2361_cast_fp16")];
+            tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2359_cast_fp16, y = var_2361_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2364_cast_fp16 = softmax(axis = var_2302, x = mh_w_37_cast_fp16)[name = tensor<string, []>("op_2364_cast_fp16")];
+            tensor<int32, [4]> var_2365 = const()[name = tensor<string, []>("op_2365"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2366_cast_fp16 = reshape(shape = var_2365, x = value_37_cast_fp16)[name = tensor<string, []>("op_2366_cast_fp16")];
+            tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2366_cast_fp16, y = var_2364_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
+            tensor<int32, [4]> var_2369 = const()[name = tensor<string, []>("op_2369"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_145_cast_fp16 = reshape(shape = var_2369, x = attn_37_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<int32, [2]> var_2373 = const()[name = tensor<string, []>("op_2373"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2375 = const()[name = tensor<string, []>("op_2375"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_75_pad_type_0 = const()[name = tensor<string, []>("obj_75_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_75_pad_0 = const()[name = tensor<string, []>("obj_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_18_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240988992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241808256))), name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241808384)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_75_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = var_2375, groups = var_2304, pad = obj_75_pad_0, pad_type = obj_75_pad_type_0, strides = var_2373, weight = layers_18_self_attn_o_proj_weight_to_fp16_palettized, x = input_145_cast_fp16)[name = tensor<string, []>("obj_75_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = tensor<string, []>("inputs_75_cast_fp16")];
+            tensor<int32, [1]> var_2381 = const()[name = tensor<string, []>("op_2381"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_75_cast_fp16 = reduce_mean(axes = var_2381, keep_dims = var_2305, x = inputs_75_cast_fp16)[name = tensor<string, []>("channels_mean_75_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor<string, []>("zero_mean_75_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor<string, []>("zero_mean_sq_75_cast_fp16")];
+            tensor<int32, [1]> var_2385 = const()[name = tensor<string, []>("op_2385"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2386_cast_fp16 = reduce_mean(axes = var_2385, keep_dims = var_2305, x = zero_mean_sq_75_cast_fp16)[name = tensor<string, []>("op_2386_cast_fp16")];
+            tensor<fp16, []> var_2387_to_fp16 = const()[name = tensor<string, []>("op_2387_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2388_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = tensor<string, []>("op_2388_cast_fp16")];
+            tensor<fp32, []> denom_75_epsilon_0 = const()[name = tensor<string, []>("denom_75_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0, x = var_2388_cast_fp16)[name = tensor<string, []>("denom_75_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
+            tensor<fp16, [1280]> input_147_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_147_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241811008)))];
+            tensor<fp16, [1280]> input_147_beta_0_to_fp16 = const()[name = tensor<string, []>("input_147_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241813632)))];
+            tensor<fp16, []> input_147_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_147_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_147_cast_fp16 = batch_norm(beta = input_147_beta_0_to_fp16, epsilon = input_147_epsilon_0_to_fp16, gamma = input_147_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
+            tensor<int32, [2]> var_2399 = const()[name = tensor<string, []>("op_2399"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2401 = const()[name = tensor<string, []>("op_2401"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_149_pad_type_0 = const()[name = tensor<string, []>("input_149_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_149_pad_0 = const()[name = tensor<string, []>("input_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_18_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241816256))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245093120))), name = tensor<string, []>("layers_18_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_18_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245093248)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_149_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = var_2401, groups = var_2304, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = var_2399, weight = layers_18_fc1_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
+            tensor<string, []> input_151_mode_0 = const()[name = tensor<string, []>("input_151_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [2]> var_2407 = const()[name = tensor<string, []>("op_2407"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2409 = const()[name = tensor<string, []>("op_2409"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_41_pad_type_0 = const()[name = tensor<string, []>("hidden_states_41_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_41_pad_0 = const()[name = tensor<string, []>("hidden_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_18_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245103552))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248380416))), name = tensor<string, []>("layers_18_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_18_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248380544)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_41_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = var_2409, groups = var_2304, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = var_2407, weight = layers_18_fc2_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = tensor<string, []>("hidden_states_41_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor<string, []>("inputs_77_cast_fp16")];
+            tensor<int32, []> var_2420 = const()[name = tensor<string, []>("op_2420"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2422 = const()[name = tensor<string, []>("op_2422"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2423 = const()[name = tensor<string, []>("op_2423"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2433 = const()[name = tensor<string, []>("op_2433"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_77_cast_fp16 = reduce_mean(axes = var_2433, keep_dims = var_2423, x = inputs_77_cast_fp16)[name = tensor<string, []>("channels_mean_77_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor<string, []>("zero_mean_77_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor<string, []>("zero_mean_sq_77_cast_fp16")];
+            tensor<int32, [1]> var_2437 = const()[name = tensor<string, []>("op_2437"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2438_cast_fp16 = reduce_mean(axes = var_2437, keep_dims = var_2423, x = zero_mean_sq_77_cast_fp16)[name = tensor<string, []>("op_2438_cast_fp16")];
+            tensor<fp16, []> var_2439_to_fp16 = const()[name = tensor<string, []>("op_2439_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2440_cast_fp16 = add(x = var_2438_cast_fp16, y = var_2439_to_fp16)[name = tensor<string, []>("op_2440_cast_fp16")];
+            tensor<fp32, []> denom_77_epsilon_0 = const()[name = tensor<string, []>("denom_77_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0, x = var_2440_cast_fp16)[name = tensor<string, []>("denom_77_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
+            tensor<fp16, [1280]> obj_77_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_77_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248383168)))];
+            tensor<fp16, [1280]> obj_77_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_77_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248385792)))];
+            tensor<fp16, []> obj_77_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_77_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")];
+            tensor<int32, [2]> var_2455 = const()[name = tensor<string, []>("op_2455"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2457 = const()[name = tensor<string, []>("op_2457"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_39_pad_type_0 = const()[name = tensor<string, []>("query_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_39_pad_0 = const()[name = tensor<string, []>("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248388416))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249207680))), name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249207808)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_39_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = var_2457, groups = var_2422, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = var_2455, weight = layers_19_self_attn_q_proj_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
+            tensor<int32, [2]> var_2461 = const()[name = tensor<string, []>("op_2461"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2463 = const()[name = tensor<string, []>("op_2463"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249210432))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250029696))), name = tensor<string, []>("layers_19_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_39_cast_fp16 = conv(dilations = var_2463, groups = var_2422, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = var_2461, weight = layers_19_self_attn_k_proj_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor<string, []>("key_39_cast_fp16")];
+            tensor<int32, [2]> var_2468 = const()[name = tensor<string, []>("op_2468"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2470 = const()[name = tensor<string, []>("op_2470"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_39_pad_type_0 = const()[name = tensor<string, []>("value_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_39_pad_0 = const()[name = tensor<string, []>("value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250029824))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250849088))), name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250849216)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = var_2470, groups = var_2422, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = var_2468, weight = layers_19_self_attn_v_proj_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor<string, []>("value_39_cast_fp16")];
+            tensor<int32, [4]> var_2474 = const()[name = tensor<string, []>("op_2474"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2475_cast_fp16 = reshape(shape = var_2474, x = query_39_cast_fp16)[name = tensor<string, []>("op_2475_cast_fp16")];
+            tensor<fp16, []> var_2476_to_fp16 = const()[name = tensor<string, []>("op_2476_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2477_cast_fp16 = mul(x = var_2475_cast_fp16, y = var_2476_to_fp16)[name = tensor<string, []>("op_2477_cast_fp16")];
+            tensor<int32, [4]> var_2478 = const()[name = tensor<string, []>("op_2478"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2479_cast_fp16 = reshape(shape = var_2478, x = key_39_cast_fp16)[name = tensor<string, []>("op_2479_cast_fp16")];
+            tensor<bool, []> mh_w_39_transpose_x_0 = const()[name = tensor<string, []>("mh_w_39_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_39_transpose_y_0 = const()[name = tensor<string, []>("mh_w_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_2477_cast_fp16, y = var_2479_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2482_cast_fp16 = softmax(axis = var_2420, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_2482_cast_fp16")];
+            tensor<int32, [4]> var_2483 = const()[name = tensor<string, []>("op_2483"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2484_cast_fp16 = reshape(shape = var_2483, x = value_39_cast_fp16)[name = tensor<string, []>("op_2484_cast_fp16")];
+            tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2484_cast_fp16, y = var_2482_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
+            tensor<int32, [4]> var_2487 = const()[name = tensor<string, []>("op_2487"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_153_cast_fp16 = reshape(shape = var_2487, x = attn_39_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
+            tensor<int32, [2]> var_2491 = const()[name = tensor<string, []>("op_2491"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2493 = const()[name = tensor<string, []>("op_2493"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_79_pad_type_0 = const()[name = tensor<string, []>("obj_79_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_79_pad_0 = const()[name = tensor<string, []>("obj_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250851840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251671104))), name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251671232)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_79_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = var_2493, groups = var_2422, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = var_2491, weight = layers_19_self_attn_o_proj_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
+            tensor<int32, [1]> var_2499 = const()[name = tensor<string, []>("op_2499"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_79_cast_fp16 = reduce_mean(axes = var_2499, keep_dims = var_2423, x = inputs_79_cast_fp16)[name = tensor<string, []>("channels_mean_79_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor<string, []>("zero_mean_79_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor<string, []>("zero_mean_sq_79_cast_fp16")];
+            tensor<int32, [1]> var_2503 = const()[name = tensor<string, []>("op_2503"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2504_cast_fp16 = reduce_mean(axes = var_2503, keep_dims = var_2423, x = zero_mean_sq_79_cast_fp16)[name = tensor<string, []>("op_2504_cast_fp16")];
+            tensor<fp16, []> var_2505_to_fp16 = const()[name = tensor<string, []>("op_2505_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2506_cast_fp16 = add(x = var_2504_cast_fp16, y = var_2505_to_fp16)[name = tensor<string, []>("op_2506_cast_fp16")];
+            tensor<fp32, []> denom_79_epsilon_0 = const()[name = tensor<string, []>("denom_79_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0, x = var_2506_cast_fp16)[name = tensor<string, []>("denom_79_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
+            tensor<fp16, [1280]> input_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251673856)))];
+            tensor<fp16, [1280]> input_155_beta_0_to_fp16 = const()[name = tensor<string, []>("input_155_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251676480)))];
+            tensor<fp16, []> input_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, [2]> var_2517 = const()[name = tensor<string, []>("op_2517"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2519 = const()[name = tensor<string, []>("op_2519"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_19_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251679104))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254955968))), name = tensor<string, []>("layers_19_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_19_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254956096)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_157_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = var_2519, groups = var_2422, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = var_2517, weight = layers_19_fc1_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<string, []> input_159_mode_0 = const()[name = tensor<string, []>("input_159_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
+            tensor<int32, [2]> var_2525 = const()[name = tensor<string, []>("op_2525"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2527 = const()[name = tensor<string, []>("op_2527"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_43_pad_type_0 = const()[name = tensor<string, []>("hidden_states_43_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = tensor<string, []>("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_19_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254966400)))];
+            tensor<fp16, [1280]> layers_19_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268073664)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_43_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = var_2527, groups = var_2422, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = var_2525, weight = layers_19_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
+            tensor<int32, []> var_2538 = const()[name = tensor<string, []>("op_2538"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2540 = const()[name = tensor<string, []>("op_2540"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2541 = const()[name = tensor<string, []>("op_2541"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2551 = const()[name = tensor<string, []>("op_2551"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_81_cast_fp16 = reduce_mean(axes = var_2551, keep_dims = var_2541, x = inputs_81_cast_fp16)[name = tensor<string, []>("channels_mean_81_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor<string, []>("zero_mean_81_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor<string, []>("zero_mean_sq_81_cast_fp16")];
+            tensor<int32, [1]> var_2555 = const()[name = tensor<string, []>("op_2555"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2556_cast_fp16 = reduce_mean(axes = var_2555, keep_dims = var_2541, x = zero_mean_sq_81_cast_fp16)[name = tensor<string, []>("op_2556_cast_fp16")];
+            tensor<fp16, []> var_2557_to_fp16 = const()[name = tensor<string, []>("op_2557_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2558_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2557_to_fp16)[name = tensor<string, []>("op_2558_cast_fp16")];
+            tensor<fp32, []> denom_81_epsilon_0 = const()[name = tensor<string, []>("denom_81_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0, x = var_2558_cast_fp16)[name = tensor<string, []>("denom_81_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
+            tensor<fp16, [1280]> obj_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268076288)))];
+            tensor<fp16, [1280]> obj_81_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268078912)))];
+            tensor<fp16, []> obj_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
+            tensor<int32, [2]> var_2573 = const()[name = tensor<string, []>("op_2573"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2575 = const()[name = tensor<string, []>("op_2575"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268081536))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268900800))), name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268900928)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_41_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = var_2575, groups = var_2540, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = var_2573, weight = layers_20_self_attn_q_proj_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
+            tensor<int32, [2]> var_2579 = const()[name = tensor<string, []>("op_2579"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2581 = const()[name = tensor<string, []>("op_2581"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_41_pad_type_0 = const()[name = tensor<string, []>("key_41_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_41_pad_0 = const()[name = tensor<string, []>("key_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268903552))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269722816))), name = tensor<string, []>("layers_20_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_41_cast_fp16 = conv(dilations = var_2581, groups = var_2540, pad = key_41_pad_0, pad_type = key_41_pad_type_0, strides = var_2579, weight = layers_20_self_attn_k_proj_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
+            tensor<int32, [2]> var_2586 = const()[name = tensor<string, []>("op_2586"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2588 = const()[name = tensor<string, []>("op_2588"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_41_pad_type_0 = const()[name = tensor<string, []>("value_41_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_41_pad_0 = const()[name = tensor<string, []>("value_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269722944))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270542208))), name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270542336)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = var_2588, groups = var_2540, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = var_2586, weight = layers_20_self_attn_v_proj_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
+            tensor<int32, [4]> var_2592 = const()[name = tensor<string, []>("op_2592"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2593_cast_fp16 = reshape(shape = var_2592, x = query_41_cast_fp16)[name = tensor<string, []>("op_2593_cast_fp16")];
+            tensor<fp16, []> var_2594_to_fp16 = const()[name = tensor<string, []>("op_2594_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2595_cast_fp16 = mul(x = var_2593_cast_fp16, y = var_2594_to_fp16)[name = tensor<string, []>("op_2595_cast_fp16")];
+            tensor<int32, [4]> var_2596 = const()[name = tensor<string, []>("op_2596"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2597_cast_fp16 = reshape(shape = var_2596, x = key_41_cast_fp16)[name = tensor<string, []>("op_2597_cast_fp16")];
+            tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_2595_cast_fp16, y = var_2597_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2600_cast_fp16 = softmax(axis = var_2538, x = mh_w_41_cast_fp16)[name = tensor<string, []>("op_2600_cast_fp16")];
+            tensor<int32, [4]> var_2601 = const()[name = tensor<string, []>("op_2601"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2602_cast_fp16 = reshape(shape = var_2601, x = value_41_cast_fp16)[name = tensor<string, []>("op_2602_cast_fp16")];
+            tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2602_cast_fp16, y = var_2600_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
+            tensor<int32, [4]> var_2605 = const()[name = tensor<string, []>("op_2605"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_161_cast_fp16 = reshape(shape = var_2605, x = attn_41_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
+            tensor<int32, [2]> var_2609 = const()[name = tensor<string, []>("op_2609"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2611 = const()[name = tensor<string, []>("op_2611"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_83_pad_type_0 = const()[name = tensor<string, []>("obj_83_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_83_pad_0 = const()[name = tensor<string, []>("obj_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_20_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270544960))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271364224))), name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271364352)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_83_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = var_2611, groups = var_2540, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = var_2609, weight = layers_20_self_attn_o_proj_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
+            tensor<int32, [1]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_83_cast_fp16 = reduce_mean(axes = var_2617, keep_dims = var_2541, x = inputs_83_cast_fp16)[name = tensor<string, []>("channels_mean_83_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor<string, []>("zero_mean_83_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor<string, []>("zero_mean_sq_83_cast_fp16")];
+            tensor<int32, [1]> var_2621 = const()[name = tensor<string, []>("op_2621"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2622_cast_fp16 = reduce_mean(axes = var_2621, keep_dims = var_2541, x = zero_mean_sq_83_cast_fp16)[name = tensor<string, []>("op_2622_cast_fp16")];
+            tensor<fp16, []> var_2623_to_fp16 = const()[name = tensor<string, []>("op_2623_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2624_cast_fp16 = add(x = var_2622_cast_fp16, y = var_2623_to_fp16)[name = tensor<string, []>("op_2624_cast_fp16")];
+            tensor<fp32, []> denom_83_epsilon_0 = const()[name = tensor<string, []>("denom_83_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0, x = var_2624_cast_fp16)[name = tensor<string, []>("denom_83_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
+            tensor<fp16, [1280]> input_163_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_163_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271366976)))];
+            tensor<fp16, [1280]> input_163_beta_0_to_fp16 = const()[name = tensor<string, []>("input_163_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271369600)))];
+            tensor<fp16, []> input_163_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_163_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_163_cast_fp16 = batch_norm(beta = input_163_beta_0_to_fp16, epsilon = input_163_epsilon_0_to_fp16, gamma = input_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<int32, [2]> var_2635 = const()[name = tensor<string, []>("op_2635"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2637 = const()[name = tensor<string, []>("op_2637"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_165_pad_type_0 = const()[name = tensor<string, []>("input_165_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_165_pad_0 = const()[name = tensor<string, []>("input_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_20_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271372224))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274649088))), name = tensor<string, []>("layers_20_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_20_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274649216)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_165_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = var_2637, groups = var_2540, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = var_2635, weight = layers_20_fc1_weight_to_fp16_palettized, x = input_163_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
+            tensor<string, []> input_167_mode_0 = const()[name = tensor<string, []>("input_167_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [2]> var_2643 = const()[name = tensor<string, []>("op_2643"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2645 = const()[name = tensor<string, []>("op_2645"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_45_pad_type_0 = const()[name = tensor<string, []>("hidden_states_45_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = tensor<string, []>("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_20_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274659520))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277936384))), name = tensor<string, []>("layers_20_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_20_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277936512)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_45_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = var_2645, groups = var_2540, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = var_2643, weight = layers_20_fc2_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor<string, []>("hidden_states_45_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor<string, []>("inputs_85_cast_fp16")];
+            tensor<int32, []> var_2656 = const()[name = tensor<string, []>("op_2656"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2658 = const()[name = tensor<string, []>("op_2658"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2659 = const()[name = tensor<string, []>("op_2659"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2669 = const()[name = tensor<string, []>("op_2669"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_85_cast_fp16 = reduce_mean(axes = var_2669, keep_dims = var_2659, x = inputs_85_cast_fp16)[name = tensor<string, []>("channels_mean_85_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor<string, []>("zero_mean_85_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor<string, []>("zero_mean_sq_85_cast_fp16")];
+            tensor<int32, [1]> var_2673 = const()[name = tensor<string, []>("op_2673"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2674_cast_fp16 = reduce_mean(axes = var_2673, keep_dims = var_2659, x = zero_mean_sq_85_cast_fp16)[name = tensor<string, []>("op_2674_cast_fp16")];
+            tensor<fp16, []> var_2675_to_fp16 = const()[name = tensor<string, []>("op_2675_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2676_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2675_to_fp16)[name = tensor<string, []>("op_2676_cast_fp16")];
+            tensor<fp32, []> denom_85_epsilon_0 = const()[name = tensor<string, []>("denom_85_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0, x = var_2676_cast_fp16)[name = tensor<string, []>("denom_85_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
+            tensor<fp16, [1280]> obj_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277939136)))];
+            tensor<fp16, [1280]> obj_85_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_85_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277941760)))];
+            tensor<fp16, []> obj_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
+            tensor<int32, [2]> var_2691 = const()[name = tensor<string, []>("op_2691"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2693 = const()[name = tensor<string, []>("op_2693"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_43_pad_type_0 = const()[name = tensor<string, []>("query_43_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_43_pad_0 = const()[name = tensor<string, []>("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277944384))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278763648))), name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278763776)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_43_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = var_2693, groups = var_2658, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = var_2691, weight = layers_21_self_attn_q_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
+            tensor<int32, [2]> var_2697 = const()[name = tensor<string, []>("op_2697"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2699 = const()[name = tensor<string, []>("op_2699"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278766400))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279585664))), name = tensor<string, []>("layers_21_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_43_cast_fp16 = conv(dilations = var_2699, groups = var_2658, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = var_2697, weight = layers_21_self_attn_k_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("key_43_cast_fp16")];
+            tensor<int32, [2]> var_2704 = const()[name = tensor<string, []>("op_2704"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2706 = const()[name = tensor<string, []>("op_2706"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_43_pad_type_0 = const()[name = tensor<string, []>("value_43_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_43_pad_0 = const()[name = tensor<string, []>("value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279585792))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280405056))), name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280405184)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = var_2706, groups = var_2658, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = var_2704, weight = layers_21_self_attn_v_proj_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor<string, []>("value_43_cast_fp16")];
+            tensor<int32, [4]> var_2710 = const()[name = tensor<string, []>("op_2710"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2711_cast_fp16 = reshape(shape = var_2710, x = query_43_cast_fp16)[name = tensor<string, []>("op_2711_cast_fp16")];
+            tensor<fp16, []> var_2712_to_fp16 = const()[name = tensor<string, []>("op_2712_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2713_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_2712_to_fp16)[name = tensor<string, []>("op_2713_cast_fp16")];
+            tensor<int32, [4]> var_2714 = const()[name = tensor<string, []>("op_2714"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2715_cast_fp16 = reshape(shape = var_2714, x = key_43_cast_fp16)[name = tensor<string, []>("op_2715_cast_fp16")];
+            tensor<bool, []> mh_w_43_transpose_x_0 = const()[name = tensor<string, []>("mh_w_43_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_43_transpose_y_0 = const()[name = tensor<string, []>("mh_w_43_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_2713_cast_fp16, y = var_2715_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2718_cast_fp16 = softmax(axis = var_2656, x = mh_w_43_cast_fp16)[name = tensor<string, []>("op_2718_cast_fp16")];
+            tensor<int32, [4]> var_2719 = const()[name = tensor<string, []>("op_2719"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2720_cast_fp16 = reshape(shape = var_2719, x = value_43_cast_fp16)[name = tensor<string, []>("op_2720_cast_fp16")];
+            tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2720_cast_fp16, y = var_2718_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
+            tensor<int32, [4]> var_2723 = const()[name = tensor<string, []>("op_2723"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_169_cast_fp16 = reshape(shape = var_2723, x = attn_43_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<int32, [2]> var_2727 = const()[name = tensor<string, []>("op_2727"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2729 = const()[name = tensor<string, []>("op_2729"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_87_pad_type_0 = const()[name = tensor<string, []>("obj_87_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_87_pad_0 = const()[name = tensor<string, []>("obj_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_21_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280407808))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281227072))), name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281227200)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_87_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = var_2729, groups = var_2658, pad = obj_87_pad_0, pad_type = obj_87_pad_type_0, strides = var_2727, weight = layers_21_self_attn_o_proj_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor<string, []>("obj_87_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = tensor<string, []>("inputs_87_cast_fp16")];
+            tensor<int32, [1]> var_2735 = const()[name = tensor<string, []>("op_2735"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_87_cast_fp16 = reduce_mean(axes = var_2735, keep_dims = var_2659, x = inputs_87_cast_fp16)[name = tensor<string, []>("channels_mean_87_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor<string, []>("zero_mean_87_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor<string, []>("zero_mean_sq_87_cast_fp16")];
+            tensor<int32, [1]> var_2739 = const()[name = tensor<string, []>("op_2739"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2740_cast_fp16 = reduce_mean(axes = var_2739, keep_dims = var_2659, x = zero_mean_sq_87_cast_fp16)[name = tensor<string, []>("op_2740_cast_fp16")];
+            tensor<fp16, []> var_2741_to_fp16 = const()[name = tensor<string, []>("op_2741_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2742_cast_fp16 = add(x = var_2740_cast_fp16, y = var_2741_to_fp16)[name = tensor<string, []>("op_2742_cast_fp16")];
+            tensor<fp32, []> denom_87_epsilon_0 = const()[name = tensor<string, []>("denom_87_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0, x = var_2742_cast_fp16)[name = tensor<string, []>("denom_87_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
+            tensor<fp16, [1280]> input_171_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_171_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281229824)))];
+            tensor<fp16, [1280]> input_171_beta_0_to_fp16 = const()[name = tensor<string, []>("input_171_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281232448)))];
+            tensor<fp16, []> input_171_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_171_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
+            tensor<int32, [2]> var_2753 = const()[name = tensor<string, []>("op_2753"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2755 = const()[name = tensor<string, []>("op_2755"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_173_pad_type_0 = const()[name = tensor<string, []>("input_173_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_173_pad_0 = const()[name = tensor<string, []>("input_173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_21_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281235072))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284511936))), name = tensor<string, []>("layers_21_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_21_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284512064)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_173_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = var_2755, groups = var_2658, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = var_2753, weight = layers_21_fc1_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
+            tensor<string, []> input_175_mode_0 = const()[name = tensor<string, []>("input_175_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
+            tensor<int32, [2]> var_2761 = const()[name = tensor<string, []>("op_2761"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2763 = const()[name = tensor<string, []>("op_2763"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_47_pad_type_0 = const()[name = tensor<string, []>("hidden_states_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = tensor<string, []>("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_21_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284522368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287799232))), name = tensor<string, []>("layers_21_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_21_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287799360)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_47_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = var_2763, groups = var_2658, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = var_2761, weight = layers_21_fc2_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = tensor<string, []>("hidden_states_47_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
+            tensor<int32, []> var_2774 = const()[name = tensor<string, []>("op_2774"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2776 = const()[name = tensor<string, []>("op_2776"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2777 = const()[name = tensor<string, []>("op_2777"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2787 = const()[name = tensor<string, []>("op_2787"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_89_cast_fp16 = reduce_mean(axes = var_2787, keep_dims = var_2777, x = inputs_89_cast_fp16)[name = tensor<string, []>("channels_mean_89_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor<string, []>("zero_mean_89_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor<string, []>("zero_mean_sq_89_cast_fp16")];
+            tensor<int32, [1]> var_2791 = const()[name = tensor<string, []>("op_2791"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2792_cast_fp16 = reduce_mean(axes = var_2791, keep_dims = var_2777, x = zero_mean_sq_89_cast_fp16)[name = tensor<string, []>("op_2792_cast_fp16")];
+            tensor<fp16, []> var_2793_to_fp16 = const()[name = tensor<string, []>("op_2793_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2794_cast_fp16 = add(x = var_2792_cast_fp16, y = var_2793_to_fp16)[name = tensor<string, []>("op_2794_cast_fp16")];
+            tensor<fp32, []> denom_89_epsilon_0 = const()[name = tensor<string, []>("denom_89_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0, x = var_2794_cast_fp16)[name = tensor<string, []>("denom_89_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
+            tensor<fp16, [1280]> obj_89_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_89_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287801984)))];
+            tensor<fp16, [1280]> obj_89_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_89_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287804608)))];
+            tensor<fp16, []> obj_89_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_89_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("obj_89_cast_fp16")];
+            tensor<int32, [2]> var_2809 = const()[name = tensor<string, []>("op_2809"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2811 = const()[name = tensor<string, []>("op_2811"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287807232))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288626496))), name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288626624)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_45_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = var_2811, groups = var_2776, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = var_2809, weight = layers_22_self_attn_q_proj_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
+            tensor<int32, [2]> var_2815 = const()[name = tensor<string, []>("op_2815"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2817 = const()[name = tensor<string, []>("op_2817"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_45_pad_type_0 = const()[name = tensor<string, []>("key_45_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_45_pad_0 = const()[name = tensor<string, []>("key_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288629248))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289448512))), name = tensor<string, []>("layers_22_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_45_cast_fp16 = conv(dilations = var_2817, groups = var_2776, pad = key_45_pad_0, pad_type = key_45_pad_type_0, strides = var_2815, weight = layers_22_self_attn_k_proj_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
+            tensor<int32, [2]> var_2822 = const()[name = tensor<string, []>("op_2822"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2824 = const()[name = tensor<string, []>("op_2824"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_45_pad_type_0 = const()[name = tensor<string, []>("value_45_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_45_pad_0 = const()[name = tensor<string, []>("value_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289448640))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290267904))), name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290268032)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = var_2824, groups = var_2776, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = var_2822, weight = layers_22_self_attn_v_proj_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
+            tensor<int32, [4]> var_2828 = const()[name = tensor<string, []>("op_2828"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2829_cast_fp16 = reshape(shape = var_2828, x = query_45_cast_fp16)[name = tensor<string, []>("op_2829_cast_fp16")];
+            tensor<fp16, []> var_2830_to_fp16 = const()[name = tensor<string, []>("op_2830_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2831_cast_fp16 = mul(x = var_2829_cast_fp16, y = var_2830_to_fp16)[name = tensor<string, []>("op_2831_cast_fp16")];
+            tensor<int32, [4]> var_2832 = const()[name = tensor<string, []>("op_2832"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2833_cast_fp16 = reshape(shape = var_2832, x = key_45_cast_fp16)[name = tensor<string, []>("op_2833_cast_fp16")];
+            tensor<bool, []> mh_w_45_transpose_x_0 = const()[name = tensor<string, []>("mh_w_45_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_45_transpose_y_0 = const()[name = tensor<string, []>("mh_w_45_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_2831_cast_fp16, y = var_2833_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2836_cast_fp16 = softmax(axis = var_2774, x = mh_w_45_cast_fp16)[name = tensor<string, []>("op_2836_cast_fp16")];
+            tensor<int32, [4]> var_2837 = const()[name = tensor<string, []>("op_2837"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2838_cast_fp16 = reshape(shape = var_2837, x = value_45_cast_fp16)[name = tensor<string, []>("op_2838_cast_fp16")];
+            tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2838_cast_fp16, y = var_2836_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
+            tensor<int32, [4]> var_2841 = const()[name = tensor<string, []>("op_2841"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_177_cast_fp16 = reshape(shape = var_2841, x = attn_45_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
+            tensor<int32, [2]> var_2845 = const()[name = tensor<string, []>("op_2845"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2847 = const()[name = tensor<string, []>("op_2847"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_91_pad_type_0 = const()[name = tensor<string, []>("obj_91_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_91_pad_0 = const()[name = tensor<string, []>("obj_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290270656))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291089920))), name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291090048)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_91_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = var_2847, groups = var_2776, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = var_2845, weight = layers_22_self_attn_o_proj_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
+            tensor<int32, [1]> var_2853 = const()[name = tensor<string, []>("op_2853"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_91_cast_fp16 = reduce_mean(axes = var_2853, keep_dims = var_2777, x = inputs_91_cast_fp16)[name = tensor<string, []>("channels_mean_91_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor<string, []>("zero_mean_91_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor<string, []>("zero_mean_sq_91_cast_fp16")];
+            tensor<int32, [1]> var_2857 = const()[name = tensor<string, []>("op_2857"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2858_cast_fp16 = reduce_mean(axes = var_2857, keep_dims = var_2777, x = zero_mean_sq_91_cast_fp16)[name = tensor<string, []>("op_2858_cast_fp16")];
+            tensor<fp16, []> var_2859_to_fp16 = const()[name = tensor<string, []>("op_2859_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2860_cast_fp16 = add(x = var_2858_cast_fp16, y = var_2859_to_fp16)[name = tensor<string, []>("op_2860_cast_fp16")];
+            tensor<fp32, []> denom_91_epsilon_0 = const()[name = tensor<string, []>("denom_91_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0, x = var_2860_cast_fp16)[name = tensor<string, []>("denom_91_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
+            tensor<fp16, [1280]> input_179_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_179_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291092672)))];
+            tensor<fp16, [1280]> input_179_beta_0_to_fp16 = const()[name = tensor<string, []>("input_179_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291095296)))];
+            tensor<fp16, []> input_179_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_179_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_179_cast_fp16 = batch_norm(beta = input_179_beta_0_to_fp16, epsilon = input_179_epsilon_0_to_fp16, gamma = input_179_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [2]> var_2871 = const()[name = tensor<string, []>("op_2871"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2873 = const()[name = tensor<string, []>("op_2873"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_181_pad_type_0 = const()[name = tensor<string, []>("input_181_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_181_pad_0 = const()[name = tensor<string, []>("input_181_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_22_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291097920)))];
+            tensor<fp16, [5120]> layers_22_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304205184)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_181_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = var_2873, groups = var_2776, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = var_2871, weight = layers_22_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<string, []> input_183_mode_0 = const()[name = tensor<string, []>("input_183_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
+            tensor<int32, [2]> var_2879 = const()[name = tensor<string, []>("op_2879"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2881 = const()[name = tensor<string, []>("op_2881"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_49_pad_type_0 = const()[name = tensor<string, []>("hidden_states_49_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_49_pad_0 = const()[name = tensor<string, []>("hidden_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_22_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304215488))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307492352))), name = tensor<string, []>("layers_22_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_22_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307492480)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_49_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = var_2881, groups = var_2776, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = var_2879, weight = layers_22_fc2_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
+            tensor<int32, []> var_2892 = const()[name = tensor<string, []>("op_2892"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_2894 = const()[name = tensor<string, []>("op_2894"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_2895 = const()[name = tensor<string, []>("op_2895"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_2905 = const()[name = tensor<string, []>("op_2905"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_93_cast_fp16 = reduce_mean(axes = var_2905, keep_dims = var_2895, x = inputs_93_cast_fp16)[name = tensor<string, []>("channels_mean_93_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor<string, []>("zero_mean_93_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor<string, []>("zero_mean_sq_93_cast_fp16")];
+            tensor<int32, [1]> var_2909 = const()[name = tensor<string, []>("op_2909"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2910_cast_fp16 = reduce_mean(axes = var_2909, keep_dims = var_2895, x = zero_mean_sq_93_cast_fp16)[name = tensor<string, []>("op_2910_cast_fp16")];
+            tensor<fp16, []> var_2911_to_fp16 = const()[name = tensor<string, []>("op_2911_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2912_cast_fp16 = add(x = var_2910_cast_fp16, y = var_2911_to_fp16)[name = tensor<string, []>("op_2912_cast_fp16")];
+            tensor<fp32, []> denom_93_epsilon_0 = const()[name = tensor<string, []>("denom_93_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0, x = var_2912_cast_fp16)[name = tensor<string, []>("denom_93_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
+            tensor<fp16, [1280]> obj_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307495104)))];
+            tensor<fp16, [1280]> obj_93_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_93_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307497728)))];
+            tensor<fp16, []> obj_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
+            tensor<int32, [2]> var_2927 = const()[name = tensor<string, []>("op_2927"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2929 = const()[name = tensor<string, []>("op_2929"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_47_pad_type_0 = const()[name = tensor<string, []>("query_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_47_pad_0 = const()[name = tensor<string, []>("query_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307500352))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308319616))), name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308319744)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_47_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = var_2929, groups = var_2894, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = var_2927, weight = layers_23_self_attn_q_proj_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
+            tensor<int32, [2]> var_2933 = const()[name = tensor<string, []>("op_2933"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2935 = const()[name = tensor<string, []>("op_2935"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_47_pad_type_0 = const()[name = tensor<string, []>("key_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_47_pad_0 = const()[name = tensor<string, []>("key_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308322368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309141632))), name = tensor<string, []>("layers_23_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_47_cast_fp16 = conv(dilations = var_2935, groups = var_2894, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = var_2933, weight = layers_23_self_attn_k_proj_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor<string, []>("key_47_cast_fp16")];
+            tensor<int32, [2]> var_2940 = const()[name = tensor<string, []>("op_2940"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2942 = const()[name = tensor<string, []>("op_2942"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_47_pad_type_0 = const()[name = tensor<string, []>("value_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_47_pad_0 = const()[name = tensor<string, []>("value_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309141760))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309961024))), name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309961152)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = var_2942, groups = var_2894, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = var_2940, weight = layers_23_self_attn_v_proj_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor<string, []>("value_47_cast_fp16")];
+            tensor<int32, [4]> var_2946 = const()[name = tensor<string, []>("op_2946"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2947_cast_fp16 = reshape(shape = var_2946, x = query_47_cast_fp16)[name = tensor<string, []>("op_2947_cast_fp16")];
+            tensor<fp16, []> var_2948_to_fp16 = const()[name = tensor<string, []>("op_2948_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_2949_cast_fp16 = mul(x = var_2947_cast_fp16, y = var_2948_to_fp16)[name = tensor<string, []>("op_2949_cast_fp16")];
+            tensor<int32, [4]> var_2950 = const()[name = tensor<string, []>("op_2950"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2951_cast_fp16 = reshape(shape = var_2950, x = key_47_cast_fp16)[name = tensor<string, []>("op_2951_cast_fp16")];
+            tensor<bool, []> mh_w_47_transpose_x_0 = const()[name = tensor<string, []>("mh_w_47_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_47_transpose_y_0 = const()[name = tensor<string, []>("mh_w_47_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_2949_cast_fp16, y = var_2951_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_2954_cast_fp16 = softmax(axis = var_2892, x = mh_w_47_cast_fp16)[name = tensor<string, []>("op_2954_cast_fp16")];
+            tensor<int32, [4]> var_2955 = const()[name = tensor<string, []>("op_2955"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_2956_cast_fp16 = reshape(shape = var_2955, x = value_47_cast_fp16)[name = tensor<string, []>("op_2956_cast_fp16")];
+            tensor<bool, []> attn_47_transpose_x_0 = const()[name = tensor<string, []>("attn_47_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_47_transpose_y_0 = const()[name = tensor<string, []>("attn_47_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2956_cast_fp16, y = var_2954_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
+            tensor<int32, [4]> var_2959 = const()[name = tensor<string, []>("op_2959"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_185_cast_fp16 = reshape(shape = var_2959, x = attn_47_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
+            tensor<int32, [2]> var_2963 = const()[name = tensor<string, []>("op_2963"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2965 = const()[name = tensor<string, []>("op_2965"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_95_pad_type_0 = const()[name = tensor<string, []>("obj_95_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_95_pad_0 = const()[name = tensor<string, []>("obj_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_23_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309963776))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310783040))), name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310783168)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_95_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = var_2965, groups = var_2894, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = var_2963, weight = layers_23_self_attn_o_proj_weight_to_fp16_palettized, x = input_185_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = tensor<string, []>("inputs_95_cast_fp16")];
+            tensor<int32, [1]> var_2971 = const()[name = tensor<string, []>("op_2971"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_95_cast_fp16 = reduce_mean(axes = var_2971, keep_dims = var_2895, x = inputs_95_cast_fp16)[name = tensor<string, []>("channels_mean_95_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor<string, []>("zero_mean_95_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor<string, []>("zero_mean_sq_95_cast_fp16")];
+            tensor<int32, [1]> var_2975 = const()[name = tensor<string, []>("op_2975"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_2976_cast_fp16 = reduce_mean(axes = var_2975, keep_dims = var_2895, x = zero_mean_sq_95_cast_fp16)[name = tensor<string, []>("op_2976_cast_fp16")];
+            tensor<fp16, []> var_2977_to_fp16 = const()[name = tensor<string, []>("op_2977_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_2978_cast_fp16 = add(x = var_2976_cast_fp16, y = var_2977_to_fp16)[name = tensor<string, []>("op_2978_cast_fp16")];
+            tensor<fp32, []> denom_95_epsilon_0 = const()[name = tensor<string, []>("denom_95_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0, x = var_2978_cast_fp16)[name = tensor<string, []>("denom_95_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
+            tensor<fp16, [1280]> input_187_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_187_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310785792)))];
+            tensor<fp16, [1280]> input_187_beta_0_to_fp16 = const()[name = tensor<string, []>("input_187_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310788416)))];
+            tensor<fp16, []> input_187_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_187_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_187_cast_fp16 = batch_norm(beta = input_187_beta_0_to_fp16, epsilon = input_187_epsilon_0_to_fp16, gamma = input_187_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
+            tensor<int32, [2]> var_2989 = const()[name = tensor<string, []>("op_2989"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2991 = const()[name = tensor<string, []>("op_2991"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_189_pad_type_0 = const()[name = tensor<string, []>("input_189_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_189_pad_0 = const()[name = tensor<string, []>("input_189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_23_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310791040))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314067904))), name = tensor<string, []>("layers_23_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_23_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314068032)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_189_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = var_2991, groups = var_2894, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = var_2989, weight = layers_23_fc1_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
+            tensor<string, []> input_191_mode_0 = const()[name = tensor<string, []>("input_191_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
+            tensor<int32, [2]> var_2997 = const()[name = tensor<string, []>("op_2997"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_2999 = const()[name = tensor<string, []>("op_2999"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_51_pad_type_0 = const()[name = tensor<string, []>("hidden_states_51_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_51_pad_0 = const()[name = tensor<string, []>("hidden_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_23_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314078336))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317355200))), name = tensor<string, []>("layers_23_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_23_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317355328)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_51_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = var_2999, groups = var_2894, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = var_2997, weight = layers_23_fc2_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor<string, []>("hidden_states_51_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor<string, []>("inputs_97_cast_fp16")];
+            tensor<int32, []> var_3010 = const()[name = tensor<string, []>("op_3010"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3012 = const()[name = tensor<string, []>("op_3012"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3013 = const()[name = tensor<string, []>("op_3013"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3023 = const()[name = tensor<string, []>("op_3023"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_97_cast_fp16 = reduce_mean(axes = var_3023, keep_dims = var_3013, x = inputs_97_cast_fp16)[name = tensor<string, []>("channels_mean_97_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor<string, []>("zero_mean_97_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor<string, []>("zero_mean_sq_97_cast_fp16")];
+            tensor<int32, [1]> var_3027 = const()[name = tensor<string, []>("op_3027"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3028_cast_fp16 = reduce_mean(axes = var_3027, keep_dims = var_3013, x = zero_mean_sq_97_cast_fp16)[name = tensor<string, []>("op_3028_cast_fp16")];
+            tensor<fp16, []> var_3029_to_fp16 = const()[name = tensor<string, []>("op_3029_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3030_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3029_to_fp16)[name = tensor<string, []>("op_3030_cast_fp16")];
+            tensor<fp32, []> denom_97_epsilon_0 = const()[name = tensor<string, []>("denom_97_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0, x = var_3030_cast_fp16)[name = tensor<string, []>("denom_97_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
+            tensor<fp16, [1280]> obj_97_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_97_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317357952)))];
+            tensor<fp16, [1280]> obj_97_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_97_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317360576)))];
+            tensor<fp16, []> obj_97_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_97_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("obj_97_cast_fp16")];
+            tensor<int32, [2]> var_3045 = const()[name = tensor<string, []>("op_3045"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3047 = const()[name = tensor<string, []>("op_3047"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_49_pad_type_0 = const()[name = tensor<string, []>("query_49_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_49_pad_0 = const()[name = tensor<string, []>("query_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317363200))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318182464))), name = tensor<string, []>("layers_24_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318182592)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_49_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = var_3047, groups = var_3012, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = var_3045, weight = layers_24_self_attn_q_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
+            tensor<int32, [2]> var_3051 = const()[name = tensor<string, []>("op_3051"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3053 = const()[name = tensor<string, []>("op_3053"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_49_pad_type_0 = const()[name = tensor<string, []>("key_49_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_49_pad_0 = const()[name = tensor<string, []>("key_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318185216))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319004480))), name = tensor<string, []>("layers_24_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_49_cast_fp16 = conv(dilations = var_3053, groups = var_3012, pad = key_49_pad_0, pad_type = key_49_pad_type_0, strides = var_3051, weight = layers_24_self_attn_k_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("key_49_cast_fp16")];
+            tensor<int32, [2]> var_3058 = const()[name = tensor<string, []>("op_3058"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3060 = const()[name = tensor<string, []>("op_3060"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_49_pad_type_0 = const()[name = tensor<string, []>("value_49_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_49_pad_0 = const()[name = tensor<string, []>("value_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319004608))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319823872))), name = tensor<string, []>("layers_24_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319824000)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = var_3060, groups = var_3012, pad = value_49_pad_0, pad_type = value_49_pad_type_0, strides = var_3058, weight = layers_24_self_attn_v_proj_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor<string, []>("value_49_cast_fp16")];
+            tensor<int32, [4]> var_3064 = const()[name = tensor<string, []>("op_3064"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3065_cast_fp16 = reshape(shape = var_3064, x = query_49_cast_fp16)[name = tensor<string, []>("op_3065_cast_fp16")];
+            tensor<fp16, []> var_3066_to_fp16 = const()[name = tensor<string, []>("op_3066_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3067_cast_fp16 = mul(x = var_3065_cast_fp16, y = var_3066_to_fp16)[name = tensor<string, []>("op_3067_cast_fp16")];
+            tensor<int32, [4]> var_3068 = const()[name = tensor<string, []>("op_3068"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3069_cast_fp16 = reshape(shape = var_3068, x = key_49_cast_fp16)[name = tensor<string, []>("op_3069_cast_fp16")];
+            tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3067_cast_fp16, y = var_3069_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3072_cast_fp16 = softmax(axis = var_3010, x = mh_w_49_cast_fp16)[name = tensor<string, []>("op_3072_cast_fp16")];
+            tensor<int32, [4]> var_3073 = const()[name = tensor<string, []>("op_3073"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3074_cast_fp16 = reshape(shape = var_3073, x = value_49_cast_fp16)[name = tensor<string, []>("op_3074_cast_fp16")];
+            tensor<bool, []> attn_49_transpose_x_0 = const()[name = tensor<string, []>("attn_49_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_49_transpose_y_0 = const()[name = tensor<string, []>("attn_49_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3074_cast_fp16, y = var_3072_cast_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
+            tensor<int32, [4]> var_3077 = const()[name = tensor<string, []>("op_3077"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_193_cast_fp16 = reshape(shape = var_3077, x = attn_49_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [2]> var_3081 = const()[name = tensor<string, []>("op_3081"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3083 = const()[name = tensor<string, []>("op_3083"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_99_pad_type_0 = const()[name = tensor<string, []>("obj_99_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_99_pad_0 = const()[name = tensor<string, []>("obj_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_24_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319826624))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320645888))), name = tensor<string, []>("layers_24_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320646016)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_99_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = var_3083, groups = var_3012, pad = obj_99_pad_0, pad_type = obj_99_pad_type_0, strides = var_3081, weight = layers_24_self_attn_o_proj_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = tensor<string, []>("obj_99_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
+            tensor<int32, [1]> var_3089 = const()[name = tensor<string, []>("op_3089"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_99_cast_fp16 = reduce_mean(axes = var_3089, keep_dims = var_3013, x = inputs_99_cast_fp16)[name = tensor<string, []>("channels_mean_99_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor<string, []>("zero_mean_99_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor<string, []>("zero_mean_sq_99_cast_fp16")];
+            tensor<int32, [1]> var_3093 = const()[name = tensor<string, []>("op_3093"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3094_cast_fp16 = reduce_mean(axes = var_3093, keep_dims = var_3013, x = zero_mean_sq_99_cast_fp16)[name = tensor<string, []>("op_3094_cast_fp16")];
+            tensor<fp16, []> var_3095_to_fp16 = const()[name = tensor<string, []>("op_3095_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3096_cast_fp16 = add(x = var_3094_cast_fp16, y = var_3095_to_fp16)[name = tensor<string, []>("op_3096_cast_fp16")];
+            tensor<fp32, []> denom_99_epsilon_0 = const()[name = tensor<string, []>("denom_99_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0, x = var_3096_cast_fp16)[name = tensor<string, []>("denom_99_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
+            tensor<fp16, [1280]> input_195_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320648640)))];
+            tensor<fp16, [1280]> input_195_beta_0_to_fp16 = const()[name = tensor<string, []>("input_195_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320651264)))];
+            tensor<fp16, []> input_195_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_195_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<int32, [2]> var_3107 = const()[name = tensor<string, []>("op_3107"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3109 = const()[name = tensor<string, []>("op_3109"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_197_pad_type_0 = const()[name = tensor<string, []>("input_197_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_197_pad_0 = const()[name = tensor<string, []>("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_24_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320653888))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323930752))), name = tensor<string, []>("layers_24_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_24_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323930880)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_197_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = var_3109, groups = var_3012, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = var_3107, weight = layers_24_fc1_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<string, []> input_199_mode_0 = const()[name = tensor<string, []>("input_199_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
+            tensor<int32, [2]> var_3115 = const()[name = tensor<string, []>("op_3115"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3117 = const()[name = tensor<string, []>("op_3117"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_53_pad_type_0 = const()[name = tensor<string, []>("hidden_states_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = tensor<string, []>("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_24_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323941184))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327218048))), name = tensor<string, []>("layers_24_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_24_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_24_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327218176)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_53_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = var_3117, groups = var_3012, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = var_3115, weight = layers_24_fc2_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_53_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
+            tensor<int32, []> var_3128 = const()[name = tensor<string, []>("op_3128"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3130 = const()[name = tensor<string, []>("op_3130"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3131 = const()[name = tensor<string, []>("op_3131"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3141 = const()[name = tensor<string, []>("op_3141"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_101_cast_fp16 = reduce_mean(axes = var_3141, keep_dims = var_3131, x = inputs_101_cast_fp16)[name = tensor<string, []>("channels_mean_101_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor<string, []>("zero_mean_101_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor<string, []>("zero_mean_sq_101_cast_fp16")];
+            tensor<int32, [1]> var_3145 = const()[name = tensor<string, []>("op_3145"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3146_cast_fp16 = reduce_mean(axes = var_3145, keep_dims = var_3131, x = zero_mean_sq_101_cast_fp16)[name = tensor<string, []>("op_3146_cast_fp16")];
+            tensor<fp16, []> var_3147_to_fp16 = const()[name = tensor<string, []>("op_3147_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3148_cast_fp16 = add(x = var_3146_cast_fp16, y = var_3147_to_fp16)[name = tensor<string, []>("op_3148_cast_fp16")];
+            tensor<fp32, []> denom_101_epsilon_0 = const()[name = tensor<string, []>("denom_101_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0, x = var_3148_cast_fp16)[name = tensor<string, []>("denom_101_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
+            tensor<fp16, [1280]> obj_101_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_101_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327220800)))];
+            tensor<fp16, [1280]> obj_101_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_101_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327223424)))];
+            tensor<fp16, []> obj_101_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_101_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("obj_101_cast_fp16")];
+            tensor<int32, [2]> var_3163 = const()[name = tensor<string, []>("op_3163"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3165 = const()[name = tensor<string, []>("op_3165"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_51_pad_type_0 = const()[name = tensor<string, []>("query_51_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_51_pad_0 = const()[name = tensor<string, []>("query_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327226048))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328045312))), name = tensor<string, []>("layers_25_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328045440)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_51_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = var_3165, groups = var_3130, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = var_3163, weight = layers_25_self_attn_q_proj_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
+            tensor<int32, [2]> var_3169 = const()[name = tensor<string, []>("op_3169"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3171 = const()[name = tensor<string, []>("op_3171"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_51_pad_type_0 = const()[name = tensor<string, []>("key_51_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_51_pad_0 = const()[name = tensor<string, []>("key_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328048064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328867328))), name = tensor<string, []>("layers_25_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_51_cast_fp16 = conv(dilations = var_3171, groups = var_3130, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = var_3169, weight = layers_25_self_attn_k_proj_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor<string, []>("key_51_cast_fp16")];
+            tensor<int32, [2]> var_3176 = const()[name = tensor<string, []>("op_3176"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3178 = const()[name = tensor<string, []>("op_3178"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_51_pad_type_0 = const()[name = tensor<string, []>("value_51_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_51_pad_0 = const()[name = tensor<string, []>("value_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328867456))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329686720))), name = tensor<string, []>("layers_25_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329686848)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = var_3178, groups = var_3130, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = var_3176, weight = layers_25_self_attn_v_proj_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor<string, []>("value_51_cast_fp16")];
+            tensor<int32, [4]> var_3182 = const()[name = tensor<string, []>("op_3182"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3183_cast_fp16 = reshape(shape = var_3182, x = query_51_cast_fp16)[name = tensor<string, []>("op_3183_cast_fp16")];
+            tensor<fp16, []> var_3184_to_fp16 = const()[name = tensor<string, []>("op_3184_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3185_cast_fp16 = mul(x = var_3183_cast_fp16, y = var_3184_to_fp16)[name = tensor<string, []>("op_3185_cast_fp16")];
+            tensor<int32, [4]> var_3186 = const()[name = tensor<string, []>("op_3186"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3187_cast_fp16 = reshape(shape = var_3186, x = key_51_cast_fp16)[name = tensor<string, []>("op_3187_cast_fp16")];
+            tensor<bool, []> mh_w_51_transpose_x_0 = const()[name = tensor<string, []>("mh_w_51_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_51_transpose_y_0 = const()[name = tensor<string, []>("mh_w_51_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_3185_cast_fp16, y = var_3187_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3190_cast_fp16 = softmax(axis = var_3128, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_3190_cast_fp16")];
+            tensor<int32, [4]> var_3191 = const()[name = tensor<string, []>("op_3191"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3192_cast_fp16 = reshape(shape = var_3191, x = value_51_cast_fp16)[name = tensor<string, []>("op_3192_cast_fp16")];
+            tensor<bool, []> attn_51_transpose_x_0 = const()[name = tensor<string, []>("attn_51_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_51_transpose_y_0 = const()[name = tensor<string, []>("attn_51_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3192_cast_fp16, y = var_3190_cast_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
+            tensor<int32, [4]> var_3195 = const()[name = tensor<string, []>("op_3195"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_201_cast_fp16 = reshape(shape = var_3195, x = attn_51_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
+            tensor<int32, [2]> var_3199 = const()[name = tensor<string, []>("op_3199"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3201 = const()[name = tensor<string, []>("op_3201"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_103_pad_type_0 = const()[name = tensor<string, []>("obj_103_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_103_pad_0 = const()[name = tensor<string, []>("obj_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_25_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329689472))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330508736))), name = tensor<string, []>("layers_25_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330508864)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_103_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = var_3201, groups = var_3130, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = var_3199, weight = layers_25_self_attn_o_proj_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = tensor<string, []>("obj_103_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
+            tensor<int32, [1]> var_3207 = const()[name = tensor<string, []>("op_3207"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_103_cast_fp16 = reduce_mean(axes = var_3207, keep_dims = var_3131, x = inputs_103_cast_fp16)[name = tensor<string, []>("channels_mean_103_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor<string, []>("zero_mean_103_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor<string, []>("zero_mean_sq_103_cast_fp16")];
+            tensor<int32, [1]> var_3211 = const()[name = tensor<string, []>("op_3211"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3212_cast_fp16 = reduce_mean(axes = var_3211, keep_dims = var_3131, x = zero_mean_sq_103_cast_fp16)[name = tensor<string, []>("op_3212_cast_fp16")];
+            tensor<fp16, []> var_3213_to_fp16 = const()[name = tensor<string, []>("op_3213_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3214_cast_fp16 = add(x = var_3212_cast_fp16, y = var_3213_to_fp16)[name = tensor<string, []>("op_3214_cast_fp16")];
+            tensor<fp32, []> denom_103_epsilon_0 = const()[name = tensor<string, []>("denom_103_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0, x = var_3214_cast_fp16)[name = tensor<string, []>("denom_103_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
+            tensor<fp16, [1280]> input_203_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_203_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330511488)))];
+            tensor<fp16, [1280]> input_203_beta_0_to_fp16 = const()[name = tensor<string, []>("input_203_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330514112)))];
+            tensor<fp16, []> input_203_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_203_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_203_cast_fp16 = batch_norm(beta = input_203_beta_0_to_fp16, epsilon = input_203_epsilon_0_to_fp16, gamma = input_203_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [2]> var_3225 = const()[name = tensor<string, []>("op_3225"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3227 = const()[name = tensor<string, []>("op_3227"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_205_pad_type_0 = const()[name = tensor<string, []>("input_205_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_205_pad_0 = const()[name = tensor<string, []>("input_205_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_25_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330516736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333793600))), name = tensor<string, []>("layers_25_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_25_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333793728)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_205_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = var_3227, groups = var_3130, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = var_3225, weight = layers_25_fc1_weight_to_fp16_palettized, x = input_203_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
+            tensor<string, []> input_207_mode_0 = const()[name = tensor<string, []>("input_207_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, [2]> var_3233 = const()[name = tensor<string, []>("op_3233"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3235 = const()[name = tensor<string, []>("op_3235"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_55_pad_type_0 = const()[name = tensor<string, []>("hidden_states_55_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_55_pad_0 = const()[name = tensor<string, []>("hidden_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_25_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333804032))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337080896))), name = tensor<string, []>("layers_25_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_25_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_25_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337081024)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_55_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = var_3235, groups = var_3130, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = var_3233, weight = layers_25_fc2_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor<string, []>("hidden_states_55_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor<string, []>("inputs_105_cast_fp16")];
+            tensor<int32, []> var_3246 = const()[name = tensor<string, []>("op_3246"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3248 = const()[name = tensor<string, []>("op_3248"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3249 = const()[name = tensor<string, []>("op_3249"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3259 = const()[name = tensor<string, []>("op_3259"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_105_cast_fp16 = reduce_mean(axes = var_3259, keep_dims = var_3249, x = inputs_105_cast_fp16)[name = tensor<string, []>("channels_mean_105_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor<string, []>("zero_mean_105_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor<string, []>("zero_mean_sq_105_cast_fp16")];
+            tensor<int32, [1]> var_3263 = const()[name = tensor<string, []>("op_3263"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3264_cast_fp16 = reduce_mean(axes = var_3263, keep_dims = var_3249, x = zero_mean_sq_105_cast_fp16)[name = tensor<string, []>("op_3264_cast_fp16")];
+            tensor<fp16, []> var_3265_to_fp16 = const()[name = tensor<string, []>("op_3265_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3266_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3265_to_fp16)[name = tensor<string, []>("op_3266_cast_fp16")];
+            tensor<fp32, []> denom_105_epsilon_0 = const()[name = tensor<string, []>("denom_105_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0, x = var_3266_cast_fp16)[name = tensor<string, []>("denom_105_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
+            tensor<fp16, [1280]> obj_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_105_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337083648)))];
+            tensor<fp16, [1280]> obj_105_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_105_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337086272)))];
+            tensor<fp16, []> obj_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("obj_105_cast_fp16")];
+            tensor<int32, [2]> var_3281 = const()[name = tensor<string, []>("op_3281"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3283 = const()[name = tensor<string, []>("op_3283"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_53_pad_type_0 = const()[name = tensor<string, []>("query_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_53_pad_0 = const()[name = tensor<string, []>("query_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337088896))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337908160))), name = tensor<string, []>("layers_26_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337908288)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_53_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = var_3283, groups = var_3248, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = var_3281, weight = layers_26_self_attn_q_proj_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
+            tensor<int32, [2]> var_3287 = const()[name = tensor<string, []>("op_3287"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3289 = const()[name = tensor<string, []>("op_3289"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_53_pad_type_0 = const()[name = tensor<string, []>("key_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_53_pad_0 = const()[name = tensor<string, []>("key_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337910912))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338730176))), name = tensor<string, []>("layers_26_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_53_cast_fp16 = conv(dilations = var_3289, groups = var_3248, pad = key_53_pad_0, pad_type = key_53_pad_type_0, strides = var_3287, weight = layers_26_self_attn_k_proj_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor<string, []>("key_53_cast_fp16")];
+            tensor<int32, [2]> var_3294 = const()[name = tensor<string, []>("op_3294"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3296 = const()[name = tensor<string, []>("op_3296"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_53_pad_type_0 = const()[name = tensor<string, []>("value_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_53_pad_0 = const()[name = tensor<string, []>("value_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338730304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339549568))), name = tensor<string, []>("layers_26_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339549696)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = var_3296, groups = var_3248, pad = value_53_pad_0, pad_type = value_53_pad_type_0, strides = var_3294, weight = layers_26_self_attn_v_proj_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor<string, []>("value_53_cast_fp16")];
+            tensor<int32, [4]> var_3300 = const()[name = tensor<string, []>("op_3300"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3301_cast_fp16 = reshape(shape = var_3300, x = query_53_cast_fp16)[name = tensor<string, []>("op_3301_cast_fp16")];
+            tensor<fp16, []> var_3302_to_fp16 = const()[name = tensor<string, []>("op_3302_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3303_cast_fp16 = mul(x = var_3301_cast_fp16, y = var_3302_to_fp16)[name = tensor<string, []>("op_3303_cast_fp16")];
+            tensor<int32, [4]> var_3304 = const()[name = tensor<string, []>("op_3304"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3305_cast_fp16 = reshape(shape = var_3304, x = key_53_cast_fp16)[name = tensor<string, []>("op_3305_cast_fp16")];
+            tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3303_cast_fp16, y = var_3305_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3308_cast_fp16 = softmax(axis = var_3246, x = mh_w_53_cast_fp16)[name = tensor<string, []>("op_3308_cast_fp16")];
+            tensor<int32, [4]> var_3309 = const()[name = tensor<string, []>("op_3309"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3310_cast_fp16 = reshape(shape = var_3309, x = value_53_cast_fp16)[name = tensor<string, []>("op_3310_cast_fp16")];
+            tensor<bool, []> attn_53_transpose_x_0 = const()[name = tensor<string, []>("attn_53_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_53_transpose_y_0 = const()[name = tensor<string, []>("attn_53_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3310_cast_fp16, y = var_3308_cast_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
+            tensor<int32, [4]> var_3313 = const()[name = tensor<string, []>("op_3313"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_209_cast_fp16 = reshape(shape = var_3313, x = attn_53_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<int32, [2]> var_3317 = const()[name = tensor<string, []>("op_3317"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3319 = const()[name = tensor<string, []>("op_3319"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_107_pad_type_0 = const()[name = tensor<string, []>("obj_107_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_107_pad_0 = const()[name = tensor<string, []>("obj_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_26_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339552320))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340371584))), name = tensor<string, []>("layers_26_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340371712)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_107_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = var_3319, groups = var_3248, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = var_3317, weight = layers_26_self_attn_o_proj_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = tensor<string, []>("obj_107_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = tensor<string, []>("inputs_107_cast_fp16")];
+            tensor<int32, [1]> var_3325 = const()[name = tensor<string, []>("op_3325"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_107_cast_fp16 = reduce_mean(axes = var_3325, keep_dims = var_3249, x = inputs_107_cast_fp16)[name = tensor<string, []>("channels_mean_107_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor<string, []>("zero_mean_107_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor<string, []>("zero_mean_sq_107_cast_fp16")];
+            tensor<int32, [1]> var_3329 = const()[name = tensor<string, []>("op_3329"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3330_cast_fp16 = reduce_mean(axes = var_3329, keep_dims = var_3249, x = zero_mean_sq_107_cast_fp16)[name = tensor<string, []>("op_3330_cast_fp16")];
+            tensor<fp16, []> var_3331_to_fp16 = const()[name = tensor<string, []>("op_3331_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3332_cast_fp16 = add(x = var_3330_cast_fp16, y = var_3331_to_fp16)[name = tensor<string, []>("op_3332_cast_fp16")];
+            tensor<fp32, []> denom_107_epsilon_0 = const()[name = tensor<string, []>("denom_107_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0, x = var_3332_cast_fp16)[name = tensor<string, []>("denom_107_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
+            tensor<fp16, [1280]> input_211_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_211_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340374336)))];
+            tensor<fp16, [1280]> input_211_beta_0_to_fp16 = const()[name = tensor<string, []>("input_211_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340376960)))];
+            tensor<fp16, []> input_211_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_211_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_211_cast_fp16 = batch_norm(beta = input_211_beta_0_to_fp16, epsilon = input_211_epsilon_0_to_fp16, gamma = input_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
+            tensor<int32, [2]> var_3343 = const()[name = tensor<string, []>("op_3343"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3345 = const()[name = tensor<string, []>("op_3345"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_213_pad_type_0 = const()[name = tensor<string, []>("input_213_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_213_pad_0 = const()[name = tensor<string, []>("input_213_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_26_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340379584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343656448))), name = tensor<string, []>("layers_26_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_26_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343656576)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_213_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = var_3345, groups = var_3248, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = var_3343, weight = layers_26_fc1_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
+            tensor<string, []> input_215_mode_0 = const()[name = tensor<string, []>("input_215_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<int32, [2]> var_3351 = const()[name = tensor<string, []>("op_3351"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3353 = const()[name = tensor<string, []>("op_3353"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_57_pad_type_0 = const()[name = tensor<string, []>("hidden_states_57_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = tensor<string, []>("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_26_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343666880))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346943744))), name = tensor<string, []>("layers_26_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_26_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_26_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346943872)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_57_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = var_3353, groups = var_3248, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = var_3351, weight = layers_26_fc2_weight_to_fp16_palettized, x = input_215_cast_fp16)[name = tensor<string, []>("hidden_states_57_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
+            tensor<int32, []> var_3364 = const()[name = tensor<string, []>("op_3364"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3366 = const()[name = tensor<string, []>("op_3366"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3367 = const()[name = tensor<string, []>("op_3367"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3377 = const()[name = tensor<string, []>("op_3377"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_109_cast_fp16 = reduce_mean(axes = var_3377, keep_dims = var_3367, x = inputs_109_cast_fp16)[name = tensor<string, []>("channels_mean_109_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor<string, []>("zero_mean_109_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor<string, []>("zero_mean_sq_109_cast_fp16")];
+            tensor<int32, [1]> var_3381 = const()[name = tensor<string, []>("op_3381"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3382_cast_fp16 = reduce_mean(axes = var_3381, keep_dims = var_3367, x = zero_mean_sq_109_cast_fp16)[name = tensor<string, []>("op_3382_cast_fp16")];
+            tensor<fp16, []> var_3383_to_fp16 = const()[name = tensor<string, []>("op_3383_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3384_cast_fp16 = add(x = var_3382_cast_fp16, y = var_3383_to_fp16)[name = tensor<string, []>("op_3384_cast_fp16")];
+            tensor<fp32, []> denom_109_epsilon_0 = const()[name = tensor<string, []>("denom_109_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0, x = var_3384_cast_fp16)[name = tensor<string, []>("denom_109_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
+            tensor<fp16, [1280]> obj_109_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_109_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346946496)))];
+            tensor<fp16, [1280]> obj_109_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_109_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346949120)))];
+            tensor<fp16, []> obj_109_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_109_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("obj_109_cast_fp16")];
+            tensor<int32, [2]> var_3399 = const()[name = tensor<string, []>("op_3399"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3401 = const()[name = tensor<string, []>("op_3401"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_55_pad_type_0 = const()[name = tensor<string, []>("query_55_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_55_pad_0 = const()[name = tensor<string, []>("query_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346951744))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347771008))), name = tensor<string, []>("layers_27_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347771136)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_55_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = var_3401, groups = var_3366, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = var_3399, weight = layers_27_self_attn_q_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
+            tensor<int32, [2]> var_3405 = const()[name = tensor<string, []>("op_3405"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3407 = const()[name = tensor<string, []>("op_3407"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_55_pad_type_0 = const()[name = tensor<string, []>("key_55_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_55_pad_0 = const()[name = tensor<string, []>("key_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347773760))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348593024))), name = tensor<string, []>("layers_27_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_55_cast_fp16 = conv(dilations = var_3407, groups = var_3366, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = var_3405, weight = layers_27_self_attn_k_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("key_55_cast_fp16")];
+            tensor<int32, [2]> var_3412 = const()[name = tensor<string, []>("op_3412"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3414 = const()[name = tensor<string, []>("op_3414"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_55_pad_type_0 = const()[name = tensor<string, []>("value_55_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_55_pad_0 = const()[name = tensor<string, []>("value_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348593152))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349412416))), name = tensor<string, []>("layers_27_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349412544)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = var_3414, groups = var_3366, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = var_3412, weight = layers_27_self_attn_v_proj_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor<string, []>("value_55_cast_fp16")];
+            tensor<int32, [4]> var_3418 = const()[name = tensor<string, []>("op_3418"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3419_cast_fp16 = reshape(shape = var_3418, x = query_55_cast_fp16)[name = tensor<string, []>("op_3419_cast_fp16")];
+            tensor<fp16, []> var_3420_to_fp16 = const()[name = tensor<string, []>("op_3420_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3421_cast_fp16 = mul(x = var_3419_cast_fp16, y = var_3420_to_fp16)[name = tensor<string, []>("op_3421_cast_fp16")];
+            tensor<int32, [4]> var_3422 = const()[name = tensor<string, []>("op_3422"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3423_cast_fp16 = reshape(shape = var_3422, x = key_55_cast_fp16)[name = tensor<string, []>("op_3423_cast_fp16")];
+            tensor<bool, []> mh_w_55_transpose_x_0 = const()[name = tensor<string, []>("mh_w_55_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_55_transpose_y_0 = const()[name = tensor<string, []>("mh_w_55_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_3421_cast_fp16, y = var_3423_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3426_cast_fp16 = softmax(axis = var_3364, x = mh_w_55_cast_fp16)[name = tensor<string, []>("op_3426_cast_fp16")];
+            tensor<int32, [4]> var_3427 = const()[name = tensor<string, []>("op_3427"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3428_cast_fp16 = reshape(shape = var_3427, x = value_55_cast_fp16)[name = tensor<string, []>("op_3428_cast_fp16")];
+            tensor<bool, []> attn_55_transpose_x_0 = const()[name = tensor<string, []>("attn_55_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_55_transpose_y_0 = const()[name = tensor<string, []>("attn_55_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3428_cast_fp16, y = var_3426_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
+            tensor<int32, [4]> var_3431 = const()[name = tensor<string, []>("op_3431"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_217_cast_fp16 = reshape(shape = var_3431, x = attn_55_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
+            tensor<int32, [2]> var_3435 = const()[name = tensor<string, []>("op_3435"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3437 = const()[name = tensor<string, []>("op_3437"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_111_pad_type_0 = const()[name = tensor<string, []>("obj_111_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_111_pad_0 = const()[name = tensor<string, []>("obj_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_27_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349415168))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350234432))), name = tensor<string, []>("layers_27_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350234560)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_111_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = var_3437, groups = var_3366, pad = obj_111_pad_0, pad_type = obj_111_pad_type_0, strides = var_3435, weight = layers_27_self_attn_o_proj_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor<string, []>("obj_111_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
+            tensor<int32, [1]> var_3443 = const()[name = tensor<string, []>("op_3443"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_111_cast_fp16 = reduce_mean(axes = var_3443, keep_dims = var_3367, x = inputs_111_cast_fp16)[name = tensor<string, []>("channels_mean_111_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor<string, []>("zero_mean_111_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor<string, []>("zero_mean_sq_111_cast_fp16")];
+            tensor<int32, [1]> var_3447 = const()[name = tensor<string, []>("op_3447"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3448_cast_fp16 = reduce_mean(axes = var_3447, keep_dims = var_3367, x = zero_mean_sq_111_cast_fp16)[name = tensor<string, []>("op_3448_cast_fp16")];
+            tensor<fp16, []> var_3449_to_fp16 = const()[name = tensor<string, []>("op_3449_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3450_cast_fp16 = add(x = var_3448_cast_fp16, y = var_3449_to_fp16)[name = tensor<string, []>("op_3450_cast_fp16")];
+            tensor<fp32, []> denom_111_epsilon_0 = const()[name = tensor<string, []>("denom_111_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0, x = var_3450_cast_fp16)[name = tensor<string, []>("denom_111_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
+            tensor<fp16, [1280]> input_219_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_219_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350237184)))];
+            tensor<fp16, [1280]> input_219_beta_0_to_fp16 = const()[name = tensor<string, []>("input_219_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350239808)))];
+            tensor<fp16, []> input_219_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_219_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_219_cast_fp16 = batch_norm(beta = input_219_beta_0_to_fp16, epsilon = input_219_epsilon_0_to_fp16, gamma = input_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [2]> var_3461 = const()[name = tensor<string, []>("op_3461"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3463 = const()[name = tensor<string, []>("op_3463"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_221_pad_type_0 = const()[name = tensor<string, []>("input_221_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_221_pad_0 = const()[name = tensor<string, []>("input_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_27_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350242432))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353519296))), name = tensor<string, []>("layers_27_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_27_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353519424)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_221_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = var_3463, groups = var_3366, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = var_3461, weight = layers_27_fc1_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<string, []> input_223_mode_0 = const()[name = tensor<string, []>("input_223_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = input_221_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
+            tensor<int32, [2]> var_3469 = const()[name = tensor<string, []>("op_3469"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3471 = const()[name = tensor<string, []>("op_3471"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_59_pad_type_0 = const()[name = tensor<string, []>("hidden_states_59_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_59_pad_0 = const()[name = tensor<string, []>("hidden_states_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_27_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353529728)))];
+            tensor<fp16, [1280]> layers_27_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_27_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366636992)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_59_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = var_3471, groups = var_3366, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = var_3469, weight = layers_27_fc2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("hidden_states_59_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
+            tensor<int32, []> var_3482 = const()[name = tensor<string, []>("op_3482"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3484 = const()[name = tensor<string, []>("op_3484"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3485 = const()[name = tensor<string, []>("op_3485"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3495 = const()[name = tensor<string, []>("op_3495"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_113_cast_fp16 = reduce_mean(axes = var_3495, keep_dims = var_3485, x = inputs_113_cast_fp16)[name = tensor<string, []>("channels_mean_113_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor<string, []>("zero_mean_113_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor<string, []>("zero_mean_sq_113_cast_fp16")];
+            tensor<int32, [1]> var_3499 = const()[name = tensor<string, []>("op_3499"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3500_cast_fp16 = reduce_mean(axes = var_3499, keep_dims = var_3485, x = zero_mean_sq_113_cast_fp16)[name = tensor<string, []>("op_3500_cast_fp16")];
+            tensor<fp16, []> var_3501_to_fp16 = const()[name = tensor<string, []>("op_3501_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3502_cast_fp16 = add(x = var_3500_cast_fp16, y = var_3501_to_fp16)[name = tensor<string, []>("op_3502_cast_fp16")];
+            tensor<fp32, []> denom_113_epsilon_0 = const()[name = tensor<string, []>("denom_113_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0, x = var_3502_cast_fp16)[name = tensor<string, []>("denom_113_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
+            tensor<fp16, [1280]> obj_113_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_113_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366639616)))];
+            tensor<fp16, [1280]> obj_113_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_113_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366642240)))];
+            tensor<fp16, []> obj_113_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_113_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("obj_113_cast_fp16")];
+            tensor<int32, [2]> var_3517 = const()[name = tensor<string, []>("op_3517"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3519 = const()[name = tensor<string, []>("op_3519"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_57_pad_type_0 = const()[name = tensor<string, []>("query_57_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_57_pad_0 = const()[name = tensor<string, []>("query_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366644864))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367464128))), name = tensor<string, []>("layers_28_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367464256)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_57_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = var_3519, groups = var_3484, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = var_3517, weight = layers_28_self_attn_q_proj_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
+            tensor<int32, [2]> var_3523 = const()[name = tensor<string, []>("op_3523"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3525 = const()[name = tensor<string, []>("op_3525"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_57_pad_type_0 = const()[name = tensor<string, []>("key_57_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_57_pad_0 = const()[name = tensor<string, []>("key_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367466880))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368286144))), name = tensor<string, []>("layers_28_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_57_cast_fp16 = conv(dilations = var_3525, groups = var_3484, pad = key_57_pad_0, pad_type = key_57_pad_type_0, strides = var_3523, weight = layers_28_self_attn_k_proj_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor<string, []>("key_57_cast_fp16")];
+            tensor<int32, [2]> var_3530 = const()[name = tensor<string, []>("op_3530"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3532 = const()[name = tensor<string, []>("op_3532"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_57_pad_type_0 = const()[name = tensor<string, []>("value_57_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_57_pad_0 = const()[name = tensor<string, []>("value_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368286272))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369105536))), name = tensor<string, []>("layers_28_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369105664)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = var_3532, groups = var_3484, pad = value_57_pad_0, pad_type = value_57_pad_type_0, strides = var_3530, weight = layers_28_self_attn_v_proj_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor<string, []>("value_57_cast_fp16")];
+            tensor<int32, [4]> var_3536 = const()[name = tensor<string, []>("op_3536"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3537_cast_fp16 = reshape(shape = var_3536, x = query_57_cast_fp16)[name = tensor<string, []>("op_3537_cast_fp16")];
+            tensor<fp16, []> var_3538_to_fp16 = const()[name = tensor<string, []>("op_3538_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3539_cast_fp16 = mul(x = var_3537_cast_fp16, y = var_3538_to_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
+            tensor<int32, [4]> var_3540 = const()[name = tensor<string, []>("op_3540"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3541_cast_fp16 = reshape(shape = var_3540, x = key_57_cast_fp16)[name = tensor<string, []>("op_3541_cast_fp16")];
+            tensor<bool, []> mh_w_57_transpose_x_0 = const()[name = tensor<string, []>("mh_w_57_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_57_transpose_y_0 = const()[name = tensor<string, []>("mh_w_57_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_3539_cast_fp16, y = var_3541_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3544_cast_fp16 = softmax(axis = var_3482, x = mh_w_57_cast_fp16)[name = tensor<string, []>("op_3544_cast_fp16")];
+            tensor<int32, [4]> var_3545 = const()[name = tensor<string, []>("op_3545"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3546_cast_fp16 = reshape(shape = var_3545, x = value_57_cast_fp16)[name = tensor<string, []>("op_3546_cast_fp16")];
+            tensor<bool, []> attn_57_transpose_x_0 = const()[name = tensor<string, []>("attn_57_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_57_transpose_y_0 = const()[name = tensor<string, []>("attn_57_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3546_cast_fp16, y = var_3544_cast_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
+            tensor<int32, [4]> var_3549 = const()[name = tensor<string, []>("op_3549"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_225_cast_fp16 = reshape(shape = var_3549, x = attn_57_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
+            tensor<int32, [2]> var_3553 = const()[name = tensor<string, []>("op_3553"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3555 = const()[name = tensor<string, []>("op_3555"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_115_pad_type_0 = const()[name = tensor<string, []>("obj_115_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_115_pad_0 = const()[name = tensor<string, []>("obj_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_28_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369108288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369927552))), name = tensor<string, []>("layers_28_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369927680)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_115_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = var_3555, groups = var_3484, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = var_3553, weight = layers_28_self_attn_o_proj_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor<string, []>("obj_115_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = tensor<string, []>("inputs_115_cast_fp16")];
+            tensor<int32, [1]> var_3561 = const()[name = tensor<string, []>("op_3561"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_115_cast_fp16 = reduce_mean(axes = var_3561, keep_dims = var_3485, x = inputs_115_cast_fp16)[name = tensor<string, []>("channels_mean_115_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor<string, []>("zero_mean_115_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor<string, []>("zero_mean_sq_115_cast_fp16")];
+            tensor<int32, [1]> var_3565 = const()[name = tensor<string, []>("op_3565"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3566_cast_fp16 = reduce_mean(axes = var_3565, keep_dims = var_3485, x = zero_mean_sq_115_cast_fp16)[name = tensor<string, []>("op_3566_cast_fp16")];
+            tensor<fp16, []> var_3567_to_fp16 = const()[name = tensor<string, []>("op_3567_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3568_cast_fp16 = add(x = var_3566_cast_fp16, y = var_3567_to_fp16)[name = tensor<string, []>("op_3568_cast_fp16")];
+            tensor<fp32, []> denom_115_epsilon_0 = const()[name = tensor<string, []>("denom_115_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0, x = var_3568_cast_fp16)[name = tensor<string, []>("denom_115_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
+            tensor<fp16, [1280]> input_227_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_227_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369930304)))];
+            tensor<fp16, [1280]> input_227_beta_0_to_fp16 = const()[name = tensor<string, []>("input_227_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369932928)))];
+            tensor<fp16, []> input_227_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_227_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_227_cast_fp16 = batch_norm(beta = input_227_beta_0_to_fp16, epsilon = input_227_epsilon_0_to_fp16, gamma = input_227_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
+            tensor<int32, [2]> var_3579 = const()[name = tensor<string, []>("op_3579"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3581 = const()[name = tensor<string, []>("op_3581"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_229_pad_type_0 = const()[name = tensor<string, []>("input_229_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_229_pad_0 = const()[name = tensor<string, []>("input_229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_28_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_28_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369935552)))];
+            tensor<fp16, [5120]> layers_28_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383042816)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_229_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = var_3581, groups = var_3484, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = var_3579, weight = layers_28_fc1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
+            tensor<string, []> input_231_mode_0 = const()[name = tensor<string, []>("input_231_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [2]> var_3587 = const()[name = tensor<string, []>("op_3587"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3589 = const()[name = tensor<string, []>("op_3589"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_61_pad_type_0 = const()[name = tensor<string, []>("hidden_states_61_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_61_pad_0 = const()[name = tensor<string, []>("hidden_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_28_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383053120))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386329984))), name = tensor<string, []>("layers_28_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_28_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_28_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386330112)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_61_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = var_3589, groups = var_3484, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = var_3587, weight = layers_28_fc2_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor<string, []>("hidden_states_61_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor<string, []>("inputs_117_cast_fp16")];
+            tensor<int32, []> var_3600 = const()[name = tensor<string, []>("op_3600"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3602 = const()[name = tensor<string, []>("op_3602"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3603 = const()[name = tensor<string, []>("op_3603"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3613 = const()[name = tensor<string, []>("op_3613"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_117_cast_fp16 = reduce_mean(axes = var_3613, keep_dims = var_3603, x = inputs_117_cast_fp16)[name = tensor<string, []>("channels_mean_117_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor<string, []>("zero_mean_117_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor<string, []>("zero_mean_sq_117_cast_fp16")];
+            tensor<int32, [1]> var_3617 = const()[name = tensor<string, []>("op_3617"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3618_cast_fp16 = reduce_mean(axes = var_3617, keep_dims = var_3603, x = zero_mean_sq_117_cast_fp16)[name = tensor<string, []>("op_3618_cast_fp16")];
+            tensor<fp16, []> var_3619_to_fp16 = const()[name = tensor<string, []>("op_3619_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3620_cast_fp16 = add(x = var_3618_cast_fp16, y = var_3619_to_fp16)[name = tensor<string, []>("op_3620_cast_fp16")];
+            tensor<fp32, []> denom_117_epsilon_0 = const()[name = tensor<string, []>("denom_117_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0, x = var_3620_cast_fp16)[name = tensor<string, []>("denom_117_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
+            tensor<fp16, [1280]> obj_117_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_117_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386332736)))];
+            tensor<fp16, [1280]> obj_117_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_117_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386335360)))];
+            tensor<fp16, []> obj_117_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_117_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("obj_117_cast_fp16")];
+            tensor<int32, [2]> var_3635 = const()[name = tensor<string, []>("op_3635"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3637 = const()[name = tensor<string, []>("op_3637"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_59_pad_type_0 = const()[name = tensor<string, []>("query_59_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_59_pad_0 = const()[name = tensor<string, []>("query_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386337984))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387157248))), name = tensor<string, []>("layers_29_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387157376)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_59_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = var_3637, groups = var_3602, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = var_3635, weight = layers_29_self_attn_q_proj_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
+            tensor<int32, [2]> var_3641 = const()[name = tensor<string, []>("op_3641"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3643 = const()[name = tensor<string, []>("op_3643"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_59_pad_type_0 = const()[name = tensor<string, []>("key_59_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_59_pad_0 = const()[name = tensor<string, []>("key_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387160000))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387979264))), name = tensor<string, []>("layers_29_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_59_cast_fp16 = conv(dilations = var_3643, groups = var_3602, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = var_3641, weight = layers_29_self_attn_k_proj_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor<string, []>("key_59_cast_fp16")];
+            tensor<int32, [2]> var_3648 = const()[name = tensor<string, []>("op_3648"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3650 = const()[name = tensor<string, []>("op_3650"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_59_pad_type_0 = const()[name = tensor<string, []>("value_59_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_59_pad_0 = const()[name = tensor<string, []>("value_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387979392))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388798656))), name = tensor<string, []>("layers_29_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388798784)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = var_3650, groups = var_3602, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = var_3648, weight = layers_29_self_attn_v_proj_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor<string, []>("value_59_cast_fp16")];
+            tensor<int32, [4]> var_3654 = const()[name = tensor<string, []>("op_3654"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3655_cast_fp16 = reshape(shape = var_3654, x = query_59_cast_fp16)[name = tensor<string, []>("op_3655_cast_fp16")];
+            tensor<fp16, []> var_3656_to_fp16 = const()[name = tensor<string, []>("op_3656_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3657_cast_fp16 = mul(x = var_3655_cast_fp16, y = var_3656_to_fp16)[name = tensor<string, []>("op_3657_cast_fp16")];
+            tensor<int32, [4]> var_3658 = const()[name = tensor<string, []>("op_3658"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3659_cast_fp16 = reshape(shape = var_3658, x = key_59_cast_fp16)[name = tensor<string, []>("op_3659_cast_fp16")];
+            tensor<bool, []> mh_w_59_transpose_x_0 = const()[name = tensor<string, []>("mh_w_59_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_59_transpose_y_0 = const()[name = tensor<string, []>("mh_w_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_3657_cast_fp16, y = var_3659_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3662_cast_fp16 = softmax(axis = var_3600, x = mh_w_59_cast_fp16)[name = tensor<string, []>("op_3662_cast_fp16")];
+            tensor<int32, [4]> var_3663 = const()[name = tensor<string, []>("op_3663"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3664_cast_fp16 = reshape(shape = var_3663, x = value_59_cast_fp16)[name = tensor<string, []>("op_3664_cast_fp16")];
+            tensor<bool, []> attn_59_transpose_x_0 = const()[name = tensor<string, []>("attn_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_59_transpose_y_0 = const()[name = tensor<string, []>("attn_59_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3664_cast_fp16, y = var_3662_cast_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
+            tensor<int32, [4]> var_3667 = const()[name = tensor<string, []>("op_3667"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_233_cast_fp16 = reshape(shape = var_3667, x = attn_59_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [2]> var_3671 = const()[name = tensor<string, []>("op_3671"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3673 = const()[name = tensor<string, []>("op_3673"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_119_pad_type_0 = const()[name = tensor<string, []>("obj_119_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_119_pad_0 = const()[name = tensor<string, []>("obj_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_29_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388801408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389620672))), name = tensor<string, []>("layers_29_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389620800)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_119_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = var_3673, groups = var_3602, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = var_3671, weight = layers_29_self_attn_o_proj_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = tensor<string, []>("obj_119_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
+            tensor<int32, [1]> var_3679 = const()[name = tensor<string, []>("op_3679"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_119_cast_fp16 = reduce_mean(axes = var_3679, keep_dims = var_3603, x = inputs_119_cast_fp16)[name = tensor<string, []>("channels_mean_119_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor<string, []>("zero_mean_119_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor<string, []>("zero_mean_sq_119_cast_fp16")];
+            tensor<int32, [1]> var_3683 = const()[name = tensor<string, []>("op_3683"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3684_cast_fp16 = reduce_mean(axes = var_3683, keep_dims = var_3603, x = zero_mean_sq_119_cast_fp16)[name = tensor<string, []>("op_3684_cast_fp16")];
+            tensor<fp16, []> var_3685_to_fp16 = const()[name = tensor<string, []>("op_3685_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3686_cast_fp16 = add(x = var_3684_cast_fp16, y = var_3685_to_fp16)[name = tensor<string, []>("op_3686_cast_fp16")];
+            tensor<fp32, []> denom_119_epsilon_0 = const()[name = tensor<string, []>("denom_119_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0, x = var_3686_cast_fp16)[name = tensor<string, []>("denom_119_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
+            tensor<fp16, [1280]> input_235_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389623424)))];
+            tensor<fp16, [1280]> input_235_beta_0_to_fp16 = const()[name = tensor<string, []>("input_235_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389626048)))];
+            tensor<fp16, []> input_235_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_235_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
+            tensor<int32, [2]> var_3697 = const()[name = tensor<string, []>("op_3697"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3699 = const()[name = tensor<string, []>("op_3699"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_237_pad_type_0 = const()[name = tensor<string, []>("input_237_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_237_pad_0 = const()[name = tensor<string, []>("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_29_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_29_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389628672)))];
+            tensor<fp16, [5120]> layers_29_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402735936)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_237_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = var_3699, groups = var_3602, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = var_3697, weight = layers_29_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
+            tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
+            tensor<int32, [2]> var_3705 = const()[name = tensor<string, []>("op_3705"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3707 = const()[name = tensor<string, []>("op_3707"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_63_pad_type_0 = const()[name = tensor<string, []>("hidden_states_63_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = tensor<string, []>("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_29_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402746240))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406023104))), name = tensor<string, []>("layers_29_fc2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])];
+            tensor<fp16, [1280]> layers_29_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_29_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406023232)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_63_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = var_3707, groups = var_3602, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = var_3705, weight = layers_29_fc2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = tensor<string, []>("hidden_states_63_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
+            tensor<int32, []> var_3718 = const()[name = tensor<string, []>("op_3718"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3720 = const()[name = tensor<string, []>("op_3720"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3721 = const()[name = tensor<string, []>("op_3721"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3731 = const()[name = tensor<string, []>("op_3731"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_121_cast_fp16 = reduce_mean(axes = var_3731, keep_dims = var_3721, x = inputs_121_cast_fp16)[name = tensor<string, []>("channels_mean_121_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor<string, []>("zero_mean_121_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor<string, []>("zero_mean_sq_121_cast_fp16")];
+            tensor<int32, [1]> var_3735 = const()[name = tensor<string, []>("op_3735"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3736_cast_fp16 = reduce_mean(axes = var_3735, keep_dims = var_3721, x = zero_mean_sq_121_cast_fp16)[name = tensor<string, []>("op_3736_cast_fp16")];
+            tensor<fp16, []> var_3737_to_fp16 = const()[name = tensor<string, []>("op_3737_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3738_cast_fp16 = add(x = var_3736_cast_fp16, y = var_3737_to_fp16)[name = tensor<string, []>("op_3738_cast_fp16")];
+            tensor<fp32, []> denom_121_epsilon_0 = const()[name = tensor<string, []>("denom_121_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0, x = var_3738_cast_fp16)[name = tensor<string, []>("denom_121_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
+            tensor<fp16, [1280]> obj_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406025856)))];
+            tensor<fp16, [1280]> obj_121_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_121_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406028480)))];
+            tensor<fp16, []> obj_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("obj_121_cast_fp16")];
+            tensor<int32, [2]> var_3753 = const()[name = tensor<string, []>("op_3753"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3755 = const()[name = tensor<string, []>("op_3755"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_61_pad_type_0 = const()[name = tensor<string, []>("query_61_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_61_pad_0 = const()[name = tensor<string, []>("query_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406031104))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406850368))), name = tensor<string, []>("layers_30_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406850496)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_61_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = var_3755, groups = var_3720, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = var_3753, weight = layers_30_self_attn_q_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
+            tensor<int32, [2]> var_3759 = const()[name = tensor<string, []>("op_3759"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3761 = const()[name = tensor<string, []>("op_3761"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_61_pad_type_0 = const()[name = tensor<string, []>("key_61_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_61_pad_0 = const()[name = tensor<string, []>("key_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406853120))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407672384))), name = tensor<string, []>("layers_30_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_61_cast_fp16 = conv(dilations = var_3761, groups = var_3720, pad = key_61_pad_0, pad_type = key_61_pad_type_0, strides = var_3759, weight = layers_30_self_attn_k_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("key_61_cast_fp16")];
+            tensor<int32, [2]> var_3766 = const()[name = tensor<string, []>("op_3766"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3768 = const()[name = tensor<string, []>("op_3768"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_61_pad_type_0 = const()[name = tensor<string, []>("value_61_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_61_pad_0 = const()[name = tensor<string, []>("value_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407672512))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408491776))), name = tensor<string, []>("layers_30_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408491904)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = var_3768, groups = var_3720, pad = value_61_pad_0, pad_type = value_61_pad_type_0, strides = var_3766, weight = layers_30_self_attn_v_proj_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor<string, []>("value_61_cast_fp16")];
+            tensor<int32, [4]> var_3772 = const()[name = tensor<string, []>("op_3772"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3773_cast_fp16 = reshape(shape = var_3772, x = query_61_cast_fp16)[name = tensor<string, []>("op_3773_cast_fp16")];
+            tensor<fp16, []> var_3774_to_fp16 = const()[name = tensor<string, []>("op_3774_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3775_cast_fp16 = mul(x = var_3773_cast_fp16, y = var_3774_to_fp16)[name = tensor<string, []>("op_3775_cast_fp16")];
+            tensor<int32, [4]> var_3776 = const()[name = tensor<string, []>("op_3776"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3777_cast_fp16 = reshape(shape = var_3776, x = key_61_cast_fp16)[name = tensor<string, []>("op_3777_cast_fp16")];
+            tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_3775_cast_fp16, y = var_3777_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3780_cast_fp16 = softmax(axis = var_3718, x = mh_w_61_cast_fp16)[name = tensor<string, []>("op_3780_cast_fp16")];
+            tensor<int32, [4]> var_3781 = const()[name = tensor<string, []>("op_3781"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3782_cast_fp16 = reshape(shape = var_3781, x = value_61_cast_fp16)[name = tensor<string, []>("op_3782_cast_fp16")];
+            tensor<bool, []> attn_61_transpose_x_0 = const()[name = tensor<string, []>("attn_61_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_61_transpose_y_0 = const()[name = tensor<string, []>("attn_61_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3782_cast_fp16, y = var_3780_cast_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
+            tensor<int32, [4]> var_3785 = const()[name = tensor<string, []>("op_3785"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_241_cast_fp16 = reshape(shape = var_3785, x = attn_61_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
+            tensor<int32, [2]> var_3789 = const()[name = tensor<string, []>("op_3789"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3791 = const()[name = tensor<string, []>("op_3791"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_123_pad_type_0 = const()[name = tensor<string, []>("obj_123_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_123_pad_0 = const()[name = tensor<string, []>("obj_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_30_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408494528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409313792))), name = tensor<string, []>("layers_30_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409313920)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_123_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = var_3791, groups = var_3720, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = var_3789, weight = layers_30_self_attn_o_proj_weight_to_fp16_palettized, x = input_241_cast_fp16)[name = tensor<string, []>("obj_123_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
+            tensor<int32, [1]> var_3797 = const()[name = tensor<string, []>("op_3797"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_123_cast_fp16 = reduce_mean(axes = var_3797, keep_dims = var_3721, x = inputs_123_cast_fp16)[name = tensor<string, []>("channels_mean_123_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor<string, []>("zero_mean_123_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor<string, []>("zero_mean_sq_123_cast_fp16")];
+            tensor<int32, [1]> var_3801 = const()[name = tensor<string, []>("op_3801"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3802_cast_fp16 = reduce_mean(axes = var_3801, keep_dims = var_3721, x = zero_mean_sq_123_cast_fp16)[name = tensor<string, []>("op_3802_cast_fp16")];
+            tensor<fp16, []> var_3803_to_fp16 = const()[name = tensor<string, []>("op_3803_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3804_cast_fp16 = add(x = var_3802_cast_fp16, y = var_3803_to_fp16)[name = tensor<string, []>("op_3804_cast_fp16")];
+            tensor<fp32, []> denom_123_epsilon_0 = const()[name = tensor<string, []>("denom_123_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0, x = var_3804_cast_fp16)[name = tensor<string, []>("denom_123_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
+            tensor<fp16, [1280]> input_243_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_243_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409316544)))];
+            tensor<fp16, [1280]> input_243_beta_0_to_fp16 = const()[name = tensor<string, []>("input_243_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409319168)))];
+            tensor<fp16, []> input_243_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_243_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_243_cast_fp16 = batch_norm(beta = input_243_beta_0_to_fp16, epsilon = input_243_epsilon_0_to_fp16, gamma = input_243_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
+            tensor<int32, [2]> var_3815 = const()[name = tensor<string, []>("op_3815"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3817 = const()[name = tensor<string, []>("op_3817"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_245_pad_type_0 = const()[name = tensor<string, []>("input_245_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_245_pad_0 = const()[name = tensor<string, []>("input_245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_30_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409321792)))];
+            tensor<fp16, [5120]> layers_30_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422429056)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_245_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = var_3817, groups = var_3720, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = var_3815, weight = layers_30_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<string, []> input_247_mode_0 = const()[name = tensor<string, []>("input_247_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<int32, [2]> var_3823 = const()[name = tensor<string, []>("op_3823"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3825 = const()[name = tensor<string, []>("op_3825"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_65_pad_type_0 = const()[name = tensor<string, []>("hidden_states_65_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_65_pad_0 = const()[name = tensor<string, []>("hidden_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_30_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_30_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422439360)))];
+            tensor<fp16, [1280]> layers_30_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_30_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435546624)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_65_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = var_3825, groups = var_3720, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = var_3823, weight = layers_30_fc2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("hidden_states_65_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor<string, []>("inputs_125_cast_fp16")];
+            tensor<int32, []> var_3836 = const()[name = tensor<string, []>("op_3836"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_3838 = const()[name = tensor<string, []>("op_3838"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_3839 = const()[name = tensor<string, []>("op_3839"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3849 = const()[name = tensor<string, []>("op_3849"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_125_cast_fp16 = reduce_mean(axes = var_3849, keep_dims = var_3839, x = inputs_125_cast_fp16)[name = tensor<string, []>("channels_mean_125_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor<string, []>("zero_mean_125_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor<string, []>("zero_mean_sq_125_cast_fp16")];
+            tensor<int32, [1]> var_3853 = const()[name = tensor<string, []>("op_3853"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3854_cast_fp16 = reduce_mean(axes = var_3853, keep_dims = var_3839, x = zero_mean_sq_125_cast_fp16)[name = tensor<string, []>("op_3854_cast_fp16")];
+            tensor<fp16, []> var_3855_to_fp16 = const()[name = tensor<string, []>("op_3855_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3856_cast_fp16 = add(x = var_3854_cast_fp16, y = var_3855_to_fp16)[name = tensor<string, []>("op_3856_cast_fp16")];
+            tensor<fp32, []> denom_125_epsilon_0 = const()[name = tensor<string, []>("denom_125_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0, x = var_3856_cast_fp16)[name = tensor<string, []>("denom_125_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
+            tensor<fp16, [1280]> obj_125_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_125_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435549248)))];
+            tensor<fp16, [1280]> obj_125_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_125_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435551872)))];
+            tensor<fp16, []> obj_125_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_125_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("obj_125_cast_fp16")];
+            tensor<int32, [2]> var_3871 = const()[name = tensor<string, []>("op_3871"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3873 = const()[name = tensor<string, []>("op_3873"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435554496))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436373760))), name = tensor<string, []>("layers_31_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436373888)))];
+            tensor<fp16, [1, 1280, 1, 1500]> query_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = var_3873, groups = var_3838, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_3871, weight = layers_31_self_attn_q_proj_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
+            tensor<int32, [2]> var_3877 = const()[name = tensor<string, []>("op_3877"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3879 = const()[name = tensor<string, []>("op_3879"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436376512))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437195776))), name = tensor<string, []>("layers_31_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = var_3879, groups = var_3838, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_3877, weight = layers_31_self_attn_k_proj_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<int32, [2]> var_3884 = const()[name = tensor<string, []>("op_3884"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3886 = const()[name = tensor<string, []>("op_3886"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437195904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438015168))), name = tensor<string, []>("layers_31_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438015296)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = var_3886, groups = var_3838, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_3884, weight = layers_31_self_attn_v_proj_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
+            tensor<int32, [4]> var_3890 = const()[name = tensor<string, []>("op_3890"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3891_cast_fp16 = reshape(shape = var_3890, x = query_cast_fp16)[name = tensor<string, []>("op_3891_cast_fp16")];
+            tensor<fp16, []> var_3892_to_fp16 = const()[name = tensor<string, []>("op_3892_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1500]> var_3893_cast_fp16 = mul(x = var_3891_cast_fp16, y = var_3892_to_fp16)[name = tensor<string, []>("op_3893_cast_fp16")];
+            tensor<int32, [4]> var_3894 = const()[name = tensor<string, []>("op_3894"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3895_cast_fp16 = reshape(shape = var_3894, x = key_cast_fp16)[name = tensor<string, []>("op_3895_cast_fp16")];
+            tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1500, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_3893_cast_fp16, y = var_3895_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
+            tensor<fp16, [1, 20, 1500, 1500]> var_3898_cast_fp16 = softmax(axis = var_3836, x = mh_w_cast_fp16)[name = tensor<string, []>("op_3898_cast_fp16")];
+            tensor<int32, [4]> var_3899 = const()[name = tensor<string, []>("op_3899"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_3900_cast_fp16 = reshape(shape = var_3899, x = value_cast_fp16)[name = tensor<string, []>("op_3900_cast_fp16")];
+            tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1500]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_3900_cast_fp16, y = var_3898_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
+            tensor<int32, [4]> var_3903 = const()[name = tensor<string, []>("op_3903"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1500]> input_249_cast_fp16 = reshape(shape = var_3903, x = attn_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<int32, [2]> var_3907 = const()[name = tensor<string, []>("op_3907"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3909 = const()[name = tensor<string, []>("op_3909"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_pad_type_0 = const()[name = tensor<string, []>("obj_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_pad_0 = const()[name = tensor<string, []>("obj_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_31_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438017920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438837184))), name = tensor<string, []>("layers_31_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])];
+            tensor<fp16, [1280]> layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438837312)))];
+            tensor<fp16, [1, 1280, 1, 1500]> obj_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = var_3909, groups = var_3838, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = var_3907, weight = layers_31_self_attn_o_proj_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_127_cast_fp16")];
+            tensor<int32, [1]> var_3915 = const()[name = tensor<string, []>("op_3915"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_127_cast_fp16 = reduce_mean(axes = var_3915, keep_dims = var_3839, x = inputs_127_cast_fp16)[name = tensor<string, []>("channels_mean_127_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor<string, []>("zero_mean_127_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor<string, []>("zero_mean_sq_127_cast_fp16")];
+            tensor<int32, [1]> var_3919 = const()[name = tensor<string, []>("op_3919"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3920_cast_fp16 = reduce_mean(axes = var_3919, keep_dims = var_3839, x = zero_mean_sq_127_cast_fp16)[name = tensor<string, []>("op_3920_cast_fp16")];
+            tensor<fp16, []> var_3921_to_fp16 = const()[name = tensor<string, []>("op_3921_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3922_cast_fp16 = add(x = var_3920_cast_fp16, y = var_3921_to_fp16)[name = tensor<string, []>("op_3922_cast_fp16")];
+            tensor<fp32, []> denom_127_epsilon_0 = const()[name = tensor<string, []>("denom_127_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0, x = var_3922_cast_fp16)[name = tensor<string, []>("denom_127_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
+            tensor<fp16, [1280]> input_251_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_251_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438839936)))];
+            tensor<fp16, [1280]> input_251_beta_0_to_fp16 = const()[name = tensor<string, []>("input_251_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438842560)))];
+            tensor<fp16, []> input_251_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_251_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
+            tensor<int32, [2]> var_3933 = const()[name = tensor<string, []>("op_3933"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3935 = const()[name = tensor<string, []>("op_3935"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_253_pad_type_0 = const()[name = tensor<string, []>("input_253_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_253_pad_0 = const()[name = tensor<string, []>("input_253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_31_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438845184))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442122048))), name = tensor<string, []>("layers_31_fc1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])];
+            tensor<fp16, [5120]> layers_31_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442122176)))];
+            tensor<fp16, [1, 5120, 1, 1500]> input_253_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = var_3935, groups = var_3838, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = var_3933, weight = layers_31_fc1_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
+            tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_253_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [2]> var_3941 = const()[name = tensor<string, []>("op_3941"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_3943 = const()[name = tensor<string, []>("op_3943"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_pad_type_0 = const()[name = tensor<string, []>("hidden_states_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_pad_0 = const()[name = tensor<string, []>("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_31_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_31_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442132480)))];
+            tensor<fp16, [1280]> layers_31_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_31_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455239744)))];
+            tensor<fp16, [1, 1280, 1, 1500]> hidden_states_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = var_3943, groups = var_3838, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = var_3941, weight = layers_31_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
+            tensor<bool, []> var_3949 = const()[name = tensor<string, []>("op_3949"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_3953 = const()[name = tensor<string, []>("op_3953"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> channels_mean_cast_fp16 = reduce_mean(axes = var_3953, keep_dims = var_3949, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
+            tensor<int32, [1]> var_3957 = const()[name = tensor<string, []>("op_3957"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1500]> var_3958_cast_fp16 = reduce_mean(axes = var_3957, keep_dims = var_3949, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_3958_cast_fp16")];
+            tensor<fp16, []> var_3959_to_fp16 = const()[name = tensor<string, []>("op_3959_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1500]> var_3960_cast_fp16 = add(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor<string, []>("op_3960_cast_fp16")];
+            tensor<fp32, []> denom_epsilon_0 = const()[name = tensor<string, []>("denom_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp16, [1, 1, 1, 1500]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_3960_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1500]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
+            tensor<fp16, [1280]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455242368)))];
+            tensor<fp16, [1280]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455244992)))];
+            tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("encoder_output_embeds_type_fp32_cast_fp16")];
+        } -> (encoder_output_embeds);
+}
\ No newline at end of file