program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3400.43.1"}, {"coremlc-version", "3400.58.2"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
{
    func main<ios18>(state<tensor<fp16, [24, 1, 448, 1024]>> k_cache1, state<tensor<fp16, [24, 1, 1500, 1024]>> k_cache2, tensor<fp16, [1, ?]> offset_mask, tensor<int32, [1, ?]> token_data, state<tensor<fp16, [24, 1, 448, 1024]>> v_cache1, state<tensor<fp16, [24, 1, 1500, 1024]>> v_cache2) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"offset_mask", [1, 1]}, {"token_data", [1, 1]}}), ("RangeDims", {{"offset_mask", [[1, 1], [1, 448]]}, {"token_data", [[1, 1], [1, 448]]}})))] {
            tensor<int32, [2]> var_62_shape_cast_fp16 = shape(x = offset_mask)[name = string("op_62_shape_cast_fp16")];
            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
            string var_62_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_62_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)];
            tensor<int16, [2]> var_62_shape_cast_fp16_to_int16 = cast(dtype = var_62_shape_cast_fp16_to_int16_dtype_0, x = var_62_shape_cast_fp16)[name = string("cast_298")];
            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_62_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_66_shape = shape(x = token_data)[name = string("op_66_shape")];
            int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)];
            int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)];
            bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)];
            string var_66_shape_to_uint16_dtype_0 = const()[name = string("op_66_shape_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)];
            tensor<uint16, [2]> var_66_shape_to_uint16 = cast(dtype = var_66_shape_to_uint16_dtype_0, x = var_66_shape)[name = string("cast_296")];
            uint16 gather_1_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_1_to_uint16, validate_indices = gather_1_validate_indices_0, x = var_66_shape_to_uint16)[name = string("gather_1_cast_uint16")];
            string gather_1_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_1_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_1_cast_uint16_to_int32 = cast(dtype = gather_1_cast_uint16_to_int32_dtype_0, x = gather_1_cast_uint16)[name = string("cast_295")];
            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_297")];
            int32 offset = sub(x = gather_0_cast_uint16_to_int32, y = gather_1_cast_uint16_to_int32)[name = string("offset")];
            int32 var_122_axis_0 = const()[name = string("op_122_axis_0"), val = int32(0)];
            int32 var_122_batch_dims_0 = const()[name = string("op_122_batch_dims_0"), val = int32(0)];
            bool var_122_validate_indices_0 = const()[name = string("op_122_validate_indices_0"), val = bool(false)];
            tensor<fp16, [51865, 1024]> token_embedding_weight_to_fp16 = const()[name = string("token_embedding_weight_to_fp16"), val = tensor<fp16, [51865, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, ?, 1024]> var_122_cast_fp16 = gather(axis = var_122_axis_0, batch_dims = var_122_batch_dims_0, indices = token_data, validate_indices = var_122_validate_indices_0, x = token_embedding_weight_to_fp16)[name = string("op_122_cast_fp16")];
            int32 concat_0_values1_0 = const()[name = string("concat_0_values1_0"), val = int32(0)];
            int32 concat_0_axis_0 = const()[name = string("concat_0_axis_0"), val = int32(0)];
            bool concat_0_interleave_0 = const()[name = string("concat_0_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (offset, concat_0_values1_0))[name = string("concat_0")];
            int32 concat_1_values1_0 = const()[name = string("concat_1_values1_0"), val = int32(1024)];
            int32 concat_1_axis_0 = const()[name = string("concat_1_axis_0"), val = int32(0)];
            bool concat_1_interleave_0 = const()[name = string("concat_1_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_1 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = (gather_0_cast_uint16_to_int32, concat_1_values1_0))[name = string("concat_1")];
            tensor<bool, [2]> var_125_end_mask_0 = const()[name = string("op_125_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [448, 1024]> positional_embedding_to_fp16 = const()[name = string("positional_embedding_to_fp16"), val = tensor<fp16, [448, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106219648)))];
            tensor<fp16, [?, ?]> var_125_cast_fp16 = slice_by_index(begin = concat_0, end = concat_1, end_mask = var_125_end_mask_0, x = positional_embedding_to_fp16)[name = string("op_125_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_3_cast_fp16 = add(x = var_122_cast_fp16, y = var_125_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [24, 1, 448, 1024]> read_state_0 = read_state(input = k_cache1)[name = string("read_state_0")];
            tensor<int32, [4]> k_cache_1_begin_0 = const()[name = string("k_cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_cache_1_end_0 = const()[name = string("k_cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_1_end_mask_0 = const()[name = string("k_cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_1_squeeze_mask_0 = const()[name = string("k_cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_1_cast_fp16 = slice_by_index(begin = k_cache_1_begin_0, end = k_cache_1_end_0, end_mask = k_cache_1_end_mask_0, squeeze_mask = k_cache_1_squeeze_mask_0, x = read_state_0)[name = string("k_cache_1_cast_fp16")];
            tensor<fp16, [24, 1, 448, 1024]> read_state_1 = read_state(input = v_cache1)[name = string("read_state_1")];
            tensor<int32, [4]> v_cache_1_begin_0 = const()[name = string("v_cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> v_cache_1_end_0 = const()[name = string("v_cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_1_end_mask_0 = const()[name = string("v_cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_1_squeeze_mask_0 = const()[name = string("v_cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_1_cast_fp16 = slice_by_index(begin = v_cache_1_begin_0, end = v_cache_1_end_0, end_mask = v_cache_1_end_mask_0, squeeze_mask = v_cache_1_squeeze_mask_0, x = read_state_1)[name = string("v_cache_1_cast_fp16")];
            tensor<fp16, [24, 1, 1500, 1024]> read_state_2 = read_state(input = k_cache2)[name = string("read_state_2")];
            tensor<int32, [4]> k_cache_3_begin_0 = const()[name = string("k_cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_cache_3_end_0 = const()[name = string("k_cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_3_end_mask_0 = const()[name = string("k_cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_3_squeeze_mask_0 = const()[name = string("k_cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_3_cast_fp16 = slice_by_index(begin = k_cache_3_begin_0, end = k_cache_3_end_0, end_mask = k_cache_3_end_mask_0, squeeze_mask = k_cache_3_squeeze_mask_0, x = read_state_2)[name = string("k_cache_3_cast_fp16")];
            tensor<fp16, [24, 1, 1500, 1024]> read_state_3 = read_state(input = v_cache2)[name = string("read_state_3")];
            tensor<int32, [4]> v_cache_3_begin_0 = const()[name = string("v_cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> v_cache_3_end_0 = const()[name = string("v_cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_3_end_mask_0 = const()[name = string("v_cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_3_squeeze_mask_0 = const()[name = string("v_cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_3_cast_fp16 = slice_by_index(begin = v_cache_3_begin_0, end = v_cache_3_end_0, end_mask = v_cache_3_end_mask_0, squeeze_mask = v_cache_3_squeeze_mask_0, x = read_state_3)[name = string("v_cache_3_cast_fp16")];
            int32 var_148 = const()[name = string("op_148"), val = int32(-1)];
            tensor<int32, [1]> var_166_axes_0 = const()[name = string("op_166_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_0_attn_ln_weight_to_fp16 = const()[name = string("blocks_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107137216)))];
            tensor<fp16, [1024]> blocks_0_attn_ln_bias_to_fp16 = const()[name = string("blocks_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107139328)))];
            fp16 var_154_to_fp16 = const()[name = string("op_154_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_166_cast_fp16 = layer_norm(axes = var_166_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_154_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = x_3_cast_fp16)[name = string("op_166_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_177_to_fp16 = const()[name = string("op_177_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107141440)))];
            tensor<fp16, [1024]> var_178_to_fp16 = const()[name = string("op_178_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109238656)))];
            tensor<fp16, [1, ?, 1024]> linear_0_cast_fp16 = linear(bias = var_178_to_fp16, weight = var_177_to_fp16, x = var_166_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_181_to_fp16 = const()[name = string("op_181_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109240768)))];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111337984)))];
            tensor<fp16, [1, ?, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_181_to_fp16, x = var_166_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_185_to_fp16 = const()[name = string("op_185_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111340096)))];
            tensor<fp16, [1024]> var_186_to_fp16 = const()[name = string("op_186_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113437312)))];
            tensor<fp16, [1, ?, 1024]> linear_2_cast_fp16 = linear(bias = var_186_to_fp16, weight = var_185_to_fp16, x = var_166_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> var_188_shape_cast_fp16 = shape(x = linear_0_cast_fp16)[name = string("op_188_shape_cast_fp16")];
            int32 gather_2_axis_0 = const()[name = string("gather_2_axis_0"), val = int32(0)];
            int32 gather_2_batch_dims_0 = const()[name = string("gather_2_batch_dims_0"), val = int32(0)];
            bool gather_2_validate_indices_0 = const()[name = string("gather_2_validate_indices_0"), val = bool(false)];
            string var_188_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_188_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_2_to_uint16 = const()[name = string("select_2_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_188_shape_cast_fp16_to_uint16 = cast(dtype = var_188_shape_cast_fp16_to_uint16_dtype_0, x = var_188_shape_cast_fp16)[name = string("cast_294")];
            uint16 gather_2_cast_uint16 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = select_2_to_uint16, validate_indices = gather_2_validate_indices_0, x = var_188_shape_cast_fp16_to_uint16)[name = string("gather_2_cast_uint16")];
            string gather_2_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_2_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_2_cast_uint16_to_int32 = cast(dtype = gather_2_cast_uint16_to_int32_dtype_0, x = gather_2_cast_uint16)[name = string("cast_293")];
            int32 end_step_3 = add(x = offset, y = gather_2_cast_uint16_to_int32)[name = string("end_step_3")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1_axes_0 = const()[name = string("expand_dims_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = expand_dims(axes = expand_dims_1_axes_0, x = offset)[name = string("expand_dims_1")];
            tensor<int32, [1]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3_axes_0 = const()[name = string("expand_dims_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = expand_dims(axes = expand_dims_3_axes_0, x = end_step_3)[name = string("expand_dims_3")];
            tensor<int32, [1]> concat_4_values0_0 = const()[name = string("concat_4_values0_0"), val = tensor<int32, [1]>([0])];
            int32 concat_4_axis_0 = const()[name = string("concat_4_axis_0"), val = int32(0)];
            bool concat_4_interleave_0 = const()[name = string("concat_4_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_4 = concat(axis = concat_4_axis_0, interleave = concat_4_interleave_0, values = (concat_4_values0_0, expand_dims_0, expand_dims_1, expand_dims_2))[name = string("concat_4")];
            tensor<int32, [1]> concat_5_values0_0 = const()[name = string("concat_5_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_5_values1_0 = const()[name = string("concat_5_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_5_values3_0 = const()[name = string("concat_5_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (concat_5_values0_0, concat_5_values1_0, expand_dims_3, concat_5_values3_0))[name = string("concat_5")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_1_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_4, begin_mask = k_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_5, end_mask = k_cache1_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_1_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_1_stride_0, update = linear_1_cast_fp16, x = read_state_0)[name = string("k_cache1_internal_tensor_assign_1_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_1_cast_fp16, input = k_cache1)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_48 = read_state(input = k_cache1)[name = string("coreml_update_state_48")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_1_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_4, begin_mask = v_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_5, end_mask = v_cache1_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_1_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_1_stride_0, update = linear_2_cast_fp16, x = read_state_1)[name = string("v_cache1_internal_tensor_assign_1_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_1_cast_fp16, input = v_cache1)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_49 = read_state(input = v_cache1)[name = string("coreml_update_state_49")];
            int32 concat_10_values0_0 = const()[name = string("concat_10_values0_0"), val = int32(1)];
            int32 concat_10_values2_0 = const()[name = string("concat_10_values2_0"), val = int32(1024)];
            int32 concat_10_axis_0 = const()[name = string("concat_10_axis_0"), val = int32(0)];
            bool concat_10_interleave_0 = const()[name = string("concat_10_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (concat_10_values0_0, end_step_3, concat_10_values2_0))[name = string("concat_10")];
            tensor<int32, [3]> var_204_begin_0 = const()[name = string("op_204_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_204_end_mask_0 = const()[name = string("op_204_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_204_cast_fp16 = slice_by_index(begin = var_204_begin_0, end = concat_10, end_mask = var_204_end_mask_0, x = k_cache_1_cast_fp16)[name = string("op_204_cast_fp16")];
            tensor<int32, [3]> var_207_begin_0 = const()[name = string("op_207_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_207_end_mask_0 = const()[name = string("op_207_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_207_cast_fp16 = slice_by_index(begin = var_207_begin_0, end = concat_10, end_mask = var_207_end_mask_0, x = v_cache_1_cast_fp16)[name = string("op_207_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_217_cast_fp16 = reshape(shape = concat_12x, x = linear_0_cast_fp16)[name = string("op_217_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_120_to_fp16 = const()[name = string("const_120_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_3_cast_fp16 = mul(x = var_217_cast_fp16, y = const_120_to_fp16)[name = string("q_3_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_224_cast_fp16 = reshape(shape = concat_13x, x = var_204_cast_fp16)[name = string("op_224_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_121_to_fp16 = const()[name = string("const_121_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_5_cast_fp16 = mul(x = var_224_cast_fp16, y = const_121_to_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_231_cast_fp16 = reshape(shape = concat_14x, x = var_207_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [4]> var_232 = const()[name = string("op_232"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_1_transpose_x_0 = const()[name = string("qk_1_transpose_x_0"), val = bool(false)];
            bool qk_1_transpose_y_0 = const()[name = string("qk_1_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_193_perm_0 = const()[name = string("transpose_193_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_194_perm_0 = const()[name = string("transpose_194_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_194 = transpose(perm = transpose_194_perm_0, x = k_5_cast_fp16)[name = string("transpose_478")];
            tensor<fp16, [1, 16, ?, 64]> transpose_193 = transpose(perm = transpose_193_perm_0, x = q_3_cast_fp16)[name = string("transpose_479")];
            tensor<fp16, [1, 16, ?, ?]> qk_1_cast_fp16 = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_193, y = transpose_194)[name = string("qk_1_cast_fp16")];
            int32 concat_15_values1_0 = const()[name = string("concat_15_values1_0"), val = int32(448)];
            int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)];
            bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (gather_2_cast_uint16_to_int32, concat_15_values1_0))[name = string("concat_15")];
            tensor<int32, [2]> var_235_begin_0 = const()[name = string("op_235_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_235_end_mask_0 = const()[name = string("op_235_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [448, 448]> mask_to_fp16 = const()[name = string("mask_to_fp16"), val = tensor<fp16, [448, 448]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113439424)))];
            tensor<fp16, [?, 448]> var_235_cast_fp16 = slice_by_index(begin = var_235_begin_0, end = concat_15, end_mask = var_235_end_mask_0, x = mask_to_fp16)[name = string("op_235_cast_fp16")];
            int32 concat_16_values0_0 = const()[name = string("concat_16_values0_0"), val = int32(0)];
            int32 concat_16_axis_0 = const()[name = string("concat_16_axis_0"), val = int32(0)];
            bool concat_16_interleave_0 = const()[name = string("concat_16_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_16 = concat(axis = concat_16_axis_0, interleave = concat_16_interleave_0, values = (concat_16_values0_0, gather_2_cast_uint16_to_int32))[name = string("concat_16")];
            tensor<int32, [2]> var_236_begin_0 = const()[name = string("op_236_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_236_end_mask_0 = const()[name = string("op_236_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = concat_16, end_mask = var_236_end_mask_0, x = var_235_cast_fp16)[name = string("op_236_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_3_cast_fp16 = add(x = qk_1_cast_fp16, y = var_236_cast_fp16)[name = string("qk_3_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_239_cast_fp16 = softmax(axis = var_148, x = qk_3_cast_fp16)[name = string("op_239_cast_fp16")];
            bool var_241_transpose_x_0 = const()[name = string("op_241_transpose_x_0"), val = bool(false)];
            bool var_241_transpose_y_0 = const()[name = string("op_241_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_5_cast_fp16 = transpose(perm = var_232, x = var_231_cast_fp16)[name = string("transpose_480")];
            tensor<fp16, [1, 16, ?, 64]> var_241_cast_fp16 = matmul(transpose_x = var_241_transpose_x_0, transpose_y = var_241_transpose_y_0, x = var_239_cast_fp16, y = v_5_cast_fp16)[name = string("op_241_cast_fp16")];
            tensor<int32, [4]> var_242 = const()[name = string("op_242"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_17x = const()[name = string("concat_17x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_243_cast_fp16 = transpose(perm = var_242, x = var_241_cast_fp16)[name = string("transpose_477")];
            tensor<fp16, [1, ?, 1024]> x_7_cast_fp16 = reshape(shape = concat_17x, x = var_243_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_247_to_fp16 = const()[name = string("op_247_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113840896)))];
            tensor<fp16, [1024]> var_248_to_fp16 = const()[name = string("op_248_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115938112)))];
            tensor<fp16, [1, ?, 1024]> linear_3_cast_fp16 = linear(bias = var_248_to_fp16, weight = var_247_to_fp16, x = x_7_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_9_cast_fp16 = add(x = x_3_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [1]> var_255_axes_0 = const()[name = string("op_255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_0_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_0_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115940224)))];
            tensor<fp16, [1024]> blocks_0_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_0_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115942336)))];
            tensor<fp16, [1, ?, 1024]> var_255_cast_fp16 = layer_norm(axes = var_255_axes_0, beta = blocks_0_cross_attn_ln_bias_to_fp16, epsilon = var_154_to_fp16, gamma = blocks_0_cross_attn_ln_weight_to_fp16, x = x_9_cast_fp16)[name = string("op_255_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_264_to_fp16 = const()[name = string("op_264_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115944448)))];
            tensor<fp16, [1024]> var_265_to_fp16 = const()[name = string("op_265_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118041664)))];
            tensor<fp16, [1, ?, 1024]> linear_4_cast_fp16 = linear(bias = var_265_to_fp16, weight = var_264_to_fp16, x = var_255_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_7_internal_tensor_assign_1_stride_0 = const()[name = string("k_7_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_7_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_7_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_7_to_fp16 = const()[name = string("k_7_to_fp16"), val = tensor<fp16, [1, 1500, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118043776)))];
            tensor<fp16, [1, 1500, 1024]> k_7_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_18, begin_mask = k_7_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = k_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_7_internal_tensor_assign_1_squeeze_mask_0, stride = k_7_internal_tensor_assign_1_stride_0, update = k_cache_3_cast_fp16, x = k_7_to_fp16)[name = string("k_7_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_7_internal_tensor_assign_1_stride_0 = const()[name = string("v_7_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_7_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_7_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_7_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_20, begin_mask = v_7_internal_tensor_assign_1_begin_mask_0, end = concat_21, end_mask = v_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_7_internal_tensor_assign_1_squeeze_mask_0, stride = v_7_internal_tensor_assign_1_stride_0, update = v_cache_3_cast_fp16, x = k_7_to_fp16)[name = string("v_7_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_22x = const()[name = string("concat_22x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_285_cast_fp16 = reshape(shape = concat_22x, x = linear_4_cast_fp16)[name = string("op_285_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_122_to_fp16 = const()[name = string("const_122_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_7_cast_fp16 = mul(x = var_285_cast_fp16, y = const_122_to_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_291 = const()[name = string("op_291"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_292_cast_fp16 = reshape(shape = var_291, x = k_7_internal_tensor_assign_1_cast_fp16)[name = string("op_292_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_123_to_fp16 = const()[name = string("const_123_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_9_cast_fp16 = mul(x = var_292_cast_fp16, y = const_123_to_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_298 = const()[name = string("op_298"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_299_cast_fp16 = reshape(shape = var_298, x = v_7_internal_tensor_assign_1_cast_fp16)[name = string("op_299_cast_fp16")];
            tensor<int32, [4]> var_300 = const()[name = string("op_300"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_5_transpose_x_0 = const()[name = string("qk_5_transpose_x_0"), val = bool(false)];
            bool qk_5_transpose_y_0 = const()[name = string("qk_5_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_195_perm_0 = const()[name = string("transpose_195_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_196_perm_0 = const()[name = string("transpose_196_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_196 = transpose(perm = transpose_196_perm_0, x = k_9_cast_fp16)[name = string("transpose_474")];
            tensor<fp16, [1, 16, ?, 64]> transpose_195 = transpose(perm = transpose_195_perm_0, x = q_7_cast_fp16)[name = string("transpose_475")];
            tensor<fp16, [1, 16, ?, 1500]> qk_5_cast_fp16 = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_195, y = transpose_196)[name = string("qk_5_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_304_cast_fp16 = softmax(axis = var_148, x = qk_5_cast_fp16)[name = string("op_304_cast_fp16")];
            bool var_306_transpose_x_0 = const()[name = string("op_306_transpose_x_0"), val = bool(false)];
            bool var_306_transpose_y_0 = const()[name = string("op_306_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_9_cast_fp16 = transpose(perm = var_300, x = var_299_cast_fp16)[name = string("transpose_476")];
            tensor<fp16, [1, 16, ?, 64]> var_306_cast_fp16 = matmul(transpose_x = var_306_transpose_x_0, transpose_y = var_306_transpose_y_0, x = var_304_cast_fp16, y = v_9_cast_fp16)[name = string("op_306_cast_fp16")];
            tensor<int32, [4]> var_307 = const()[name = string("op_307"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_308_cast_fp16 = transpose(perm = var_307, x = var_306_cast_fp16)[name = string("transpose_473")];
            tensor<fp16, [1, ?, 1024]> x_13_cast_fp16 = reshape(shape = concat_23x, x = var_308_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_312_to_fp16 = const()[name = string("op_312_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121115840)))];
            tensor<fp16, [1024]> var_313_to_fp16 = const()[name = string("op_313_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123213056)))];
            tensor<fp16, [1, ?, 1024]> linear_5_cast_fp16 = linear(bias = var_313_to_fp16, weight = var_312_to_fp16, x = x_13_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_15_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [1]> var_320_axes_0 = const()[name = string("op_320_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_0_mlp_ln_weight_to_fp16 = const()[name = string("blocks_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123215168)))];
            tensor<fp16, [1024]> blocks_0_mlp_ln_bias_to_fp16 = const()[name = string("blocks_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123217280)))];
            tensor<fp16, [1, ?, 1024]> var_320_cast_fp16 = layer_norm(axes = var_320_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_154_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_15_cast_fp16)[name = string("op_320_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_329_to_fp16 = const()[name = string("op_329_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123219392)))];
            tensor<fp16, [4096]> var_330_to_fp16 = const()[name = string("op_330_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131608064)))];
            tensor<fp16, [1, ?, 4096]> linear_6_cast_fp16 = linear(bias = var_330_to_fp16, weight = var_329_to_fp16, x = var_320_cast_fp16)[name = string("linear_6_cast_fp16")];
            string x_19_mode_0 = const()[name = string("x_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_19_cast_fp16 = gelu(mode = x_19_mode_0, x = linear_6_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_335_to_fp16 = const()[name = string("op_335_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131616320)))];
            tensor<fp16, [1024]> var_336_to_fp16 = const()[name = string("op_336_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140004992)))];
            tensor<fp16, [1, ?, 1024]> linear_7_cast_fp16 = linear(bias = var_336_to_fp16, weight = var_335_to_fp16, x = x_19_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_21_cast_fp16 = add(x = x_15_cast_fp16, y = linear_7_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> k_cache_5_begin_0 = const()[name = string("k_cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> k_cache_5_end_0 = const()[name = string("k_cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_5_end_mask_0 = const()[name = string("k_cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_5_squeeze_mask_0 = const()[name = string("k_cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_5_cast_fp16 = slice_by_index(begin = k_cache_5_begin_0, end = k_cache_5_end_0, end_mask = k_cache_5_end_mask_0, squeeze_mask = k_cache_5_squeeze_mask_0, x = coreml_update_state_48)[name = string("k_cache_5_cast_fp16")];
            tensor<int32, [4]> v_cache_5_begin_0 = const()[name = string("v_cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> v_cache_5_end_0 = const()[name = string("v_cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_5_end_mask_0 = const()[name = string("v_cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_5_squeeze_mask_0 = const()[name = string("v_cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_5_cast_fp16 = slice_by_index(begin = v_cache_5_begin_0, end = v_cache_5_end_0, end_mask = v_cache_5_end_mask_0, squeeze_mask = v_cache_5_squeeze_mask_0, x = coreml_update_state_49)[name = string("v_cache_5_cast_fp16")];
            tensor<int32, [4]> k_cache_7_begin_0 = const()[name = string("k_cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> k_cache_7_end_0 = const()[name = string("k_cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_7_end_mask_0 = const()[name = string("k_cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_7_squeeze_mask_0 = const()[name = string("k_cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_7_cast_fp16 = slice_by_index(begin = k_cache_7_begin_0, end = k_cache_7_end_0, end_mask = k_cache_7_end_mask_0, squeeze_mask = k_cache_7_squeeze_mask_0, x = read_state_2)[name = string("k_cache_7_cast_fp16")];
            tensor<int32, [4]> v_cache_7_begin_0 = const()[name = string("v_cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> v_cache_7_end_0 = const()[name = string("v_cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_7_end_mask_0 = const()[name = string("v_cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_7_squeeze_mask_0 = const()[name = string("v_cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_7_cast_fp16 = slice_by_index(begin = v_cache_7_begin_0, end = v_cache_7_end_0, end_mask = v_cache_7_end_mask_0, squeeze_mask = v_cache_7_squeeze_mask_0, x = read_state_3)[name = string("v_cache_7_cast_fp16")];
            int32 var_359 = const()[name = string("op_359"), val = int32(-1)];
            tensor<int32, [1]> var_377_axes_0 = const()[name = string("op_377_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_1_attn_ln_weight_to_fp16 = const()[name = string("blocks_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140007104)))];
            tensor<fp16, [1024]> blocks_1_attn_ln_bias_to_fp16 = const()[name = string("blocks_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140009216)))];
            fp16 var_365_to_fp16 = const()[name = string("op_365_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_377_cast_fp16 = layer_norm(axes = var_377_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_365_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_21_cast_fp16)[name = string("op_377_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_388_to_fp16 = const()[name = string("op_388_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140011328)))];
            tensor<fp16, [1024]> var_389_to_fp16 = const()[name = string("op_389_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142108544)))];
            tensor<fp16, [1, ?, 1024]> linear_8_cast_fp16 = linear(bias = var_389_to_fp16, weight = var_388_to_fp16, x = var_377_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_392_to_fp16 = const()[name = string("op_392_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142110656)))];
            tensor<fp16, [1, ?, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_392_to_fp16, x = var_377_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144207872)))];
            tensor<fp16, [1024]> var_397_to_fp16 = const()[name = string("op_397_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146305088)))];
            tensor<fp16, [1, ?, 1024]> linear_10_cast_fp16 = linear(bias = var_397_to_fp16, weight = var_396_to_fp16, x = var_377_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> var_399_shape_cast_fp16 = shape(x = linear_8_cast_fp16)[name = string("op_399_shape_cast_fp16")];
            int32 gather_14_axis_0 = const()[name = string("gather_14_axis_0"), val = int32(0)];
            int32 gather_14_batch_dims_0 = const()[name = string("gather_14_batch_dims_0"), val = int32(0)];
            bool gather_14_validate_indices_0 = const()[name = string("gather_14_validate_indices_0"), val = bool(false)];
            string var_399_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_399_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_14_to_uint16 = const()[name = string("select_14_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_399_shape_cast_fp16_to_uint16 = cast(dtype = var_399_shape_cast_fp16_to_uint16_dtype_0, x = var_399_shape_cast_fp16)[name = string("cast_292")];
            uint16 gather_14_cast_uint16 = gather(axis = gather_14_axis_0, batch_dims = gather_14_batch_dims_0, indices = select_14_to_uint16, validate_indices = gather_14_validate_indices_0, x = var_399_shape_cast_fp16_to_uint16)[name = string("gather_14_cast_uint16")];
            string gather_14_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_14_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_14_cast_uint16_to_int32 = cast(dtype = gather_14_cast_uint16_to_int32_dtype_0, x = gather_14_cast_uint16)[name = string("cast_291")];
            int32 end_step_5 = add(x = offset, y = gather_14_cast_uint16_to_int32)[name = string("end_step_5")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_19_axes_0 = const()[name = string("expand_dims_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_19 = expand_dims(axes = expand_dims_19_axes_0, x = end_step_5)[name = string("expand_dims_19")];
            tensor<int32, [1]> concat_26_values0_0 = const()[name = string("concat_26_values0_0"), val = tensor<int32, [1]>([1])];
            int32 concat_26_axis_0 = const()[name = string("concat_26_axis_0"), val = int32(0)];
            bool concat_26_interleave_0 = const()[name = string("concat_26_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_26 = concat(axis = concat_26_axis_0, interleave = concat_26_interleave_0, values = (concat_26_values0_0, expand_dims_16, expand_dims_1, expand_dims_18))[name = string("concat_26")];
            tensor<int32, [1]> concat_27_values0_0 = const()[name = string("concat_27_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_27_values1_0 = const()[name = string("concat_27_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_27_values3_0 = const()[name = string("concat_27_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_27_axis_0 = const()[name = string("concat_27_axis_0"), val = int32(0)];
            bool concat_27_interleave_0 = const()[name = string("concat_27_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_27 = concat(axis = concat_27_axis_0, interleave = concat_27_interleave_0, values = (concat_27_values0_0, concat_27_values1_0, expand_dims_19, concat_27_values3_0))[name = string("concat_27")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_2_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_2_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_2_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_26, begin_mask = k_cache1_internal_tensor_assign_2_begin_mask_0, end = concat_27, end_mask = k_cache1_internal_tensor_assign_2_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_2_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_2_stride_0, update = linear_9_cast_fp16, x = coreml_update_state_48)[name = string("k_cache1_internal_tensor_assign_2_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_2_cast_fp16, input = k_cache1)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_50 = read_state(input = k_cache1)[name = string("coreml_update_state_50")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_2_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_2_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_2_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_26, begin_mask = v_cache1_internal_tensor_assign_2_begin_mask_0, end = concat_27, end_mask = v_cache1_internal_tensor_assign_2_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_2_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_2_stride_0, update = linear_10_cast_fp16, x = coreml_update_state_49)[name = string("v_cache1_internal_tensor_assign_2_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_2_cast_fp16, input = v_cache1)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_51 = read_state(input = v_cache1)[name = string("coreml_update_state_51")];
            int32 concat_32_values0_0 = const()[name = string("concat_32_values0_0"), val = int32(1)];
            int32 concat_32_values2_0 = const()[name = string("concat_32_values2_0"), val = int32(1024)];
            int32 concat_32_axis_0 = const()[name = string("concat_32_axis_0"), val = int32(0)];
            bool concat_32_interleave_0 = const()[name = string("concat_32_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (concat_32_values0_0, end_step_5, concat_32_values2_0))[name = string("concat_32")];
            tensor<int32, [3]> var_415_begin_0 = const()[name = string("op_415_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_415_end_mask_0 = const()[name = string("op_415_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_415_cast_fp16 = slice_by_index(begin = var_415_begin_0, end = concat_32, end_mask = var_415_end_mask_0, x = k_cache_5_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<int32, [3]> var_418_begin_0 = const()[name = string("op_418_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_418_end_mask_0 = const()[name = string("op_418_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_418_cast_fp16 = slice_by_index(begin = var_418_begin_0, end = concat_32, end_mask = var_418_end_mask_0, x = v_cache_5_cast_fp16)[name = string("op_418_cast_fp16")];
            tensor<int32, [4]> concat_34x = const()[name = string("concat_34x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_428_cast_fp16 = reshape(shape = concat_34x, x = linear_8_cast_fp16)[name = string("op_428_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_124_to_fp16 = const()[name = string("const_124_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_11_cast_fp16 = mul(x = var_428_cast_fp16, y = const_124_to_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_435_cast_fp16 = reshape(shape = concat_35x, x = var_415_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_125_to_fp16 = const()[name = string("const_125_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_15_cast_fp16 = mul(x = var_435_cast_fp16, y = const_125_to_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_442_cast_fp16 = reshape(shape = concat_36x, x = var_418_cast_fp16)[name = string("op_442_cast_fp16")];
            tensor<int32, [4]> var_443 = const()[name = string("op_443"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_7_transpose_x_0 = const()[name = string("qk_7_transpose_x_0"), val = bool(false)];
            bool qk_7_transpose_y_0 = const()[name = string("qk_7_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_197_perm_0 = const()[name = string("transpose_197_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_198_perm_0 = const()[name = string("transpose_198_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_198 = transpose(perm = transpose_198_perm_0, x = k_15_cast_fp16)[name = string("transpose_470")];
            tensor<fp16, [1, 16, ?, 64]> transpose_197 = transpose(perm = transpose_197_perm_0, x = q_11_cast_fp16)[name = string("transpose_471")];
            tensor<fp16, [1, 16, ?, ?]> qk_7_cast_fp16 = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_197, y = transpose_198)[name = string("qk_7_cast_fp16")];
            int32 concat_37_values1_0 = const()[name = string("concat_37_values1_0"), val = int32(448)];
            int32 concat_37_axis_0 = const()[name = string("concat_37_axis_0"), val = int32(0)];
            bool concat_37_interleave_0 = const()[name = string("concat_37_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_37 = concat(axis = concat_37_axis_0, interleave = concat_37_interleave_0, values = (gather_14_cast_uint16_to_int32, concat_37_values1_0))[name = string("concat_37")];
            tensor<int32, [2]> var_446_begin_0 = const()[name = string("op_446_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_446_end_mask_0 = const()[name = string("op_446_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_446_cast_fp16 = slice_by_index(begin = var_446_begin_0, end = concat_37, end_mask = var_446_end_mask_0, x = mask_to_fp16)[name = string("op_446_cast_fp16")];
            int32 concat_38_values0_0 = const()[name = string("concat_38_values0_0"), val = int32(0)];
            int32 concat_38_axis_0 = const()[name = string("concat_38_axis_0"), val = int32(0)];
            bool concat_38_interleave_0 = const()[name = string("concat_38_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_38 = concat(axis = concat_38_axis_0, interleave = concat_38_interleave_0, values = (concat_38_values0_0, gather_14_cast_uint16_to_int32))[name = string("concat_38")];
            tensor<int32, [2]> var_447_begin_0 = const()[name = string("op_447_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_447_end_mask_0 = const()[name = string("op_447_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_447_cast_fp16 = slice_by_index(begin = var_447_begin_0, end = concat_38, end_mask = var_447_end_mask_0, x = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_9_cast_fp16 = add(x = qk_7_cast_fp16, y = var_447_cast_fp16)[name = string("qk_9_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_450_cast_fp16 = softmax(axis = var_359, x = qk_9_cast_fp16)[name = string("op_450_cast_fp16")];
            bool var_452_transpose_x_0 = const()[name = string("op_452_transpose_x_0"), val = bool(false)];
            bool var_452_transpose_y_0 = const()[name = string("op_452_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_15_cast_fp16 = transpose(perm = var_443, x = var_442_cast_fp16)[name = string("transpose_472")];
            tensor<fp16, [1, 16, ?, 64]> var_452_cast_fp16 = matmul(transpose_x = var_452_transpose_x_0, transpose_y = var_452_transpose_y_0, x = var_450_cast_fp16, y = v_15_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [4]> var_453 = const()[name = string("op_453"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_454_cast_fp16 = transpose(perm = var_453, x = var_452_cast_fp16)[name = string("transpose_469")];
            tensor<fp16, [1, ?, 1024]> x_25_cast_fp16 = reshape(shape = concat_39x, x = var_454_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_458_to_fp16 = const()[name = string("op_458_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146307200)))];
            tensor<fp16, [1024]> var_459_to_fp16 = const()[name = string("op_459_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148404416)))];
            tensor<fp16, [1, ?, 1024]> linear_11_cast_fp16 = linear(bias = var_459_to_fp16, weight = var_458_to_fp16, x = x_25_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_27_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<int32, [1]> var_466_axes_0 = const()[name = string("op_466_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_1_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_1_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148406528)))];
            tensor<fp16, [1024]> blocks_1_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_1_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148408640)))];
            tensor<fp16, [1, ?, 1024]> var_466_cast_fp16 = layer_norm(axes = var_466_axes_0, beta = blocks_1_cross_attn_ln_bias_to_fp16, epsilon = var_365_to_fp16, gamma = blocks_1_cross_attn_ln_weight_to_fp16, x = x_27_cast_fp16)[name = string("op_466_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_475_to_fp16 = const()[name = string("op_475_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148410752)))];
            tensor<fp16, [1024]> var_476_to_fp16 = const()[name = string("op_476_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150507968)))];
            tensor<fp16, [1, ?, 1024]> linear_12_cast_fp16 = linear(bias = var_476_to_fp16, weight = var_475_to_fp16, x = var_466_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_17_internal_tensor_assign_1_stride_0 = const()[name = string("k_17_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_17_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_17_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_17_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_40, begin_mask = k_17_internal_tensor_assign_1_begin_mask_0, end = concat_41, end_mask = k_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_17_internal_tensor_assign_1_squeeze_mask_0, stride = k_17_internal_tensor_assign_1_stride_0, update = k_cache_7_cast_fp16, x = k_7_to_fp16)[name = string("k_17_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_17_internal_tensor_assign_1_stride_0 = const()[name = string("v_17_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_17_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_17_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_17_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_42, begin_mask = v_17_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = v_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_17_internal_tensor_assign_1_squeeze_mask_0, stride = v_17_internal_tensor_assign_1_stride_0, update = v_cache_7_cast_fp16, x = k_7_to_fp16)[name = string("v_17_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_44x = const()[name = string("concat_44x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_496_cast_fp16 = reshape(shape = concat_44x, x = linear_12_cast_fp16)[name = string("op_496_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_126_to_fp16 = const()[name = string("const_126_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_15_cast_fp16 = mul(x = var_496_cast_fp16, y = const_126_to_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> var_502 = const()[name = string("op_502"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_503_cast_fp16 = reshape(shape = var_502, x = k_17_internal_tensor_assign_1_cast_fp16)[name = string("op_503_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_127_to_fp16 = const()[name = string("const_127_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_19_cast_fp16 = mul(x = var_503_cast_fp16, y = const_127_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_509 = const()[name = string("op_509"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_510_cast_fp16 = reshape(shape = var_509, x = v_17_internal_tensor_assign_1_cast_fp16)[name = string("op_510_cast_fp16")];
            tensor<int32, [4]> var_511 = const()[name = string("op_511"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_11_transpose_x_0 = const()[name = string("qk_11_transpose_x_0"), val = bool(false)];
            bool qk_11_transpose_y_0 = const()[name = string("qk_11_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_199_perm_0 = const()[name = string("transpose_199_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_200_perm_0 = const()[name = string("transpose_200_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_200 = transpose(perm = transpose_200_perm_0, x = k_19_cast_fp16)[name = string("transpose_466")];
            tensor<fp16, [1, 16, ?, 64]> transpose_199 = transpose(perm = transpose_199_perm_0, x = q_15_cast_fp16)[name = string("transpose_467")];
            tensor<fp16, [1, 16, ?, 1500]> qk_11_cast_fp16 = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_199, y = transpose_200)[name = string("qk_11_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_515_cast_fp16 = softmax(axis = var_359, x = qk_11_cast_fp16)[name = string("op_515_cast_fp16")];
            bool var_517_transpose_x_0 = const()[name = string("op_517_transpose_x_0"), val = bool(false)];
            bool var_517_transpose_y_0 = const()[name = string("op_517_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_19_cast_fp16 = transpose(perm = var_511, x = var_510_cast_fp16)[name = string("transpose_468")];
            tensor<fp16, [1, 16, ?, 64]> var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = var_515_cast_fp16, y = v_19_cast_fp16)[name = string("op_517_cast_fp16")];
            tensor<int32, [4]> var_518 = const()[name = string("op_518"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_45x = const()[name = string("concat_45x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_519_cast_fp16 = transpose(perm = var_518, x = var_517_cast_fp16)[name = string("transpose_465")];
            tensor<fp16, [1, ?, 1024]> x_31_cast_fp16 = reshape(shape = concat_45x, x = var_519_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_523_to_fp16 = const()[name = string("op_523_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150510080)))];
            tensor<fp16, [1024]> var_524_to_fp16 = const()[name = string("op_524_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152607296)))];
            tensor<fp16, [1, ?, 1024]> linear_13_cast_fp16 = linear(bias = var_524_to_fp16, weight = var_523_to_fp16, x = x_31_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_33_cast_fp16 = add(x = x_27_cast_fp16, y = linear_13_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [1]> var_531_axes_0 = const()[name = string("op_531_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_1_mlp_ln_weight_to_fp16 = const()[name = string("blocks_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152609408)))];
            tensor<fp16, [1024]> blocks_1_mlp_ln_bias_to_fp16 = const()[name = string("blocks_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152611520)))];
            tensor<fp16, [1, ?, 1024]> var_531_cast_fp16 = layer_norm(axes = var_531_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_365_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_33_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_540_to_fp16 = const()[name = string("op_540_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152613632)))];
            tensor<fp16, [4096]> var_541_to_fp16 = const()[name = string("op_541_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161002304)))];
            tensor<fp16, [1, ?, 4096]> linear_14_cast_fp16 = linear(bias = var_541_to_fp16, weight = var_540_to_fp16, x = var_531_cast_fp16)[name = string("linear_14_cast_fp16")];
            string x_37_mode_0 = const()[name = string("x_37_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_37_cast_fp16 = gelu(mode = x_37_mode_0, x = linear_14_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_546_to_fp16 = const()[name = string("op_546_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161010560)))];
            tensor<fp16, [1024]> var_547_to_fp16 = const()[name = string("op_547_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169399232)))];
            tensor<fp16, [1, ?, 1024]> linear_15_cast_fp16 = linear(bias = var_547_to_fp16, weight = var_546_to_fp16, x = x_37_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_39_cast_fp16 = add(x = x_33_cast_fp16, y = linear_15_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [4]> k_cache_9_begin_0 = const()[name = string("k_cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> k_cache_9_end_0 = const()[name = string("k_cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_9_end_mask_0 = const()[name = string("k_cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_9_squeeze_mask_0 = const()[name = string("k_cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_9_cast_fp16 = slice_by_index(begin = k_cache_9_begin_0, end = k_cache_9_end_0, end_mask = k_cache_9_end_mask_0, squeeze_mask = k_cache_9_squeeze_mask_0, x = coreml_update_state_50)[name = string("k_cache_9_cast_fp16")];
            tensor<int32, [4]> v_cache_9_begin_0 = const()[name = string("v_cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> v_cache_9_end_0 = const()[name = string("v_cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_9_end_mask_0 = const()[name = string("v_cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_9_squeeze_mask_0 = const()[name = string("v_cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_9_cast_fp16 = slice_by_index(begin = v_cache_9_begin_0, end = v_cache_9_end_0, end_mask = v_cache_9_end_mask_0, squeeze_mask = v_cache_9_squeeze_mask_0, x = coreml_update_state_51)[name = string("v_cache_9_cast_fp16")];
            tensor<int32, [4]> k_cache_11_begin_0 = const()[name = string("k_cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> k_cache_11_end_0 = const()[name = string("k_cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_11_end_mask_0 = const()[name = string("k_cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_11_squeeze_mask_0 = const()[name = string("k_cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_11_cast_fp16 = slice_by_index(begin = k_cache_11_begin_0, end = k_cache_11_end_0, end_mask = k_cache_11_end_mask_0, squeeze_mask = k_cache_11_squeeze_mask_0, x = read_state_2)[name = string("k_cache_11_cast_fp16")];
            tensor<int32, [4]> v_cache_11_begin_0 = const()[name = string("v_cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> v_cache_11_end_0 = const()[name = string("v_cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_11_end_mask_0 = const()[name = string("v_cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_11_squeeze_mask_0 = const()[name = string("v_cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_11_cast_fp16 = slice_by_index(begin = v_cache_11_begin_0, end = v_cache_11_end_0, end_mask = v_cache_11_end_mask_0, squeeze_mask = v_cache_11_squeeze_mask_0, x = read_state_3)[name = string("v_cache_11_cast_fp16")];
            int32 var_570 = const()[name = string("op_570"), val = int32(-1)];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_2_attn_ln_weight_to_fp16 = const()[name = string("blocks_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169401344)))];
            tensor<fp16, [1024]> blocks_2_attn_ln_bias_to_fp16 = const()[name = string("blocks_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169403456)))];
            fp16 var_576_to_fp16 = const()[name = string("op_576_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_588_cast_fp16 = layer_norm(axes = var_588_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_576_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_39_cast_fp16)[name = string("op_588_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_599_to_fp16 = const()[name = string("op_599_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169405568)))];
            tensor<fp16, [1024]> var_600_to_fp16 = const()[name = string("op_600_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171502784)))];
            tensor<fp16, [1, ?, 1024]> linear_16_cast_fp16 = linear(bias = var_600_to_fp16, weight = var_599_to_fp16, x = var_588_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_603_to_fp16 = const()[name = string("op_603_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171504896)))];
            tensor<fp16, [1, ?, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_603_to_fp16, x = var_588_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_607_to_fp16 = const()[name = string("op_607_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173602112)))];
            tensor<fp16, [1024]> var_608_to_fp16 = const()[name = string("op_608_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175699328)))];
            tensor<fp16, [1, ?, 1024]> linear_18_cast_fp16 = linear(bias = var_608_to_fp16, weight = var_607_to_fp16, x = var_588_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> var_610_shape_cast_fp16 = shape(x = linear_16_cast_fp16)[name = string("op_610_shape_cast_fp16")];
            int32 gather_26_axis_0 = const()[name = string("gather_26_axis_0"), val = int32(0)];
            int32 gather_26_batch_dims_0 = const()[name = string("gather_26_batch_dims_0"), val = int32(0)];
            bool gather_26_validate_indices_0 = const()[name = string("gather_26_validate_indices_0"), val = bool(false)];
            string var_610_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_610_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_26_to_uint16 = const()[name = string("select_26_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_610_shape_cast_fp16_to_uint16 = cast(dtype = var_610_shape_cast_fp16_to_uint16_dtype_0, x = var_610_shape_cast_fp16)[name = string("cast_290")];
            uint16 gather_26_cast_uint16 = gather(axis = gather_26_axis_0, batch_dims = gather_26_batch_dims_0, indices = select_26_to_uint16, validate_indices = gather_26_validate_indices_0, x = var_610_shape_cast_fp16_to_uint16)[name = string("gather_26_cast_uint16")];
            string gather_26_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_26_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_26_cast_uint16_to_int32 = cast(dtype = gather_26_cast_uint16_to_int32_dtype_0, x = gather_26_cast_uint16)[name = string("cast_289")];
            int32 end_step_7 = add(x = offset, y = gather_26_cast_uint16_to_int32)[name = string("end_step_7")];
            tensor<int32, [1]> expand_dims_32 = const()[name = string("expand_dims_32"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_35_axes_0 = const()[name = string("expand_dims_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_35 = expand_dims(axes = expand_dims_35_axes_0, x = end_step_7)[name = string("expand_dims_35")];
            tensor<int32, [1]> concat_48_values0_0 = const()[name = string("concat_48_values0_0"), val = tensor<int32, [1]>([2])];
            int32 concat_48_axis_0 = const()[name = string("concat_48_axis_0"), val = int32(0)];
            bool concat_48_interleave_0 = const()[name = string("concat_48_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_48 = concat(axis = concat_48_axis_0, interleave = concat_48_interleave_0, values = (concat_48_values0_0, expand_dims_32, expand_dims_1, expand_dims_34))[name = string("concat_48")];
            tensor<int32, [1]> concat_49_values0_0 = const()[name = string("concat_49_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_49_values1_0 = const()[name = string("concat_49_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_49_values3_0 = const()[name = string("concat_49_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_49_axis_0 = const()[name = string("concat_49_axis_0"), val = int32(0)];
            bool concat_49_interleave_0 = const()[name = string("concat_49_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_49 = concat(axis = concat_49_axis_0, interleave = concat_49_interleave_0, values = (concat_49_values0_0, concat_49_values1_0, expand_dims_35, concat_49_values3_0))[name = string("concat_49")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_3_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_3_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_3_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_48, begin_mask = k_cache1_internal_tensor_assign_3_begin_mask_0, end = concat_49, end_mask = k_cache1_internal_tensor_assign_3_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_3_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_3_stride_0, update = linear_17_cast_fp16, x = coreml_update_state_50)[name = string("k_cache1_internal_tensor_assign_3_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_3_cast_fp16, input = k_cache1)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_52 = read_state(input = k_cache1)[name = string("coreml_update_state_52")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_3_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_3_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_3_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_48, begin_mask = v_cache1_internal_tensor_assign_3_begin_mask_0, end = concat_49, end_mask = v_cache1_internal_tensor_assign_3_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_3_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_3_stride_0, update = linear_18_cast_fp16, x = coreml_update_state_51)[name = string("v_cache1_internal_tensor_assign_3_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_3_cast_fp16, input = v_cache1)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_53 = read_state(input = v_cache1)[name = string("coreml_update_state_53")];
            int32 concat_54_values0_0 = const()[name = string("concat_54_values0_0"), val = int32(1)];
            int32 concat_54_values2_0 = const()[name = string("concat_54_values2_0"), val = int32(1024)];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (concat_54_values0_0, end_step_7, concat_54_values2_0))[name = string("concat_54")];
            tensor<int32, [3]> var_626_begin_0 = const()[name = string("op_626_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_626_end_mask_0 = const()[name = string("op_626_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_626_cast_fp16 = slice_by_index(begin = var_626_begin_0, end = concat_54, end_mask = var_626_end_mask_0, x = k_cache_9_cast_fp16)[name = string("op_626_cast_fp16")];
            tensor<int32, [3]> var_629_begin_0 = const()[name = string("op_629_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_629_end_mask_0 = const()[name = string("op_629_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_629_cast_fp16 = slice_by_index(begin = var_629_begin_0, end = concat_54, end_mask = var_629_end_mask_0, x = v_cache_9_cast_fp16)[name = string("op_629_cast_fp16")];
            tensor<int32, [4]> concat_56x = const()[name = string("concat_56x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_639_cast_fp16 = reshape(shape = concat_56x, x = linear_16_cast_fp16)[name = string("op_639_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_128_to_fp16 = const()[name = string("const_128_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_19_cast_fp16 = mul(x = var_639_cast_fp16, y = const_128_to_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> concat_57x = const()[name = string("concat_57x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_646_cast_fp16 = reshape(shape = concat_57x, x = var_626_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_129_to_fp16 = const()[name = string("const_129_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_25_cast_fp16 = mul(x = var_646_cast_fp16, y = const_129_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> concat_58x = const()[name = string("concat_58x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_653_cast_fp16 = reshape(shape = concat_58x, x = var_629_cast_fp16)[name = string("op_653_cast_fp16")];
            tensor<int32, [4]> var_654 = const()[name = string("op_654"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_13_transpose_x_0 = const()[name = string("qk_13_transpose_x_0"), val = bool(false)];
            bool qk_13_transpose_y_0 = const()[name = string("qk_13_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_201_perm_0 = const()[name = string("transpose_201_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_202_perm_0 = const()[name = string("transpose_202_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_202 = transpose(perm = transpose_202_perm_0, x = k_25_cast_fp16)[name = string("transpose_462")];
            tensor<fp16, [1, 16, ?, 64]> transpose_201 = transpose(perm = transpose_201_perm_0, x = q_19_cast_fp16)[name = string("transpose_463")];
            tensor<fp16, [1, 16, ?, ?]> qk_13_cast_fp16 = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_201, y = transpose_202)[name = string("qk_13_cast_fp16")];
            int32 concat_59_values1_0 = const()[name = string("concat_59_values1_0"), val = int32(448)];
            int32 concat_59_axis_0 = const()[name = string("concat_59_axis_0"), val = int32(0)];
            bool concat_59_interleave_0 = const()[name = string("concat_59_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_59 = concat(axis = concat_59_axis_0, interleave = concat_59_interleave_0, values = (gather_26_cast_uint16_to_int32, concat_59_values1_0))[name = string("concat_59")];
            tensor<int32, [2]> var_657_begin_0 = const()[name = string("op_657_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_657_end_mask_0 = const()[name = string("op_657_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_657_cast_fp16 = slice_by_index(begin = var_657_begin_0, end = concat_59, end_mask = var_657_end_mask_0, x = mask_to_fp16)[name = string("op_657_cast_fp16")];
            int32 concat_60_values0_0 = const()[name = string("concat_60_values0_0"), val = int32(0)];
            int32 concat_60_axis_0 = const()[name = string("concat_60_axis_0"), val = int32(0)];
            bool concat_60_interleave_0 = const()[name = string("concat_60_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (concat_60_values0_0, gather_26_cast_uint16_to_int32))[name = string("concat_60")];
            tensor<int32, [2]> var_658_begin_0 = const()[name = string("op_658_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_658_end_mask_0 = const()[name = string("op_658_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_658_cast_fp16 = slice_by_index(begin = var_658_begin_0, end = concat_60, end_mask = var_658_end_mask_0, x = var_657_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_15_cast_fp16 = add(x = qk_13_cast_fp16, y = var_658_cast_fp16)[name = string("qk_15_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_661_cast_fp16 = softmax(axis = var_570, x = qk_15_cast_fp16)[name = string("op_661_cast_fp16")];
            bool var_663_transpose_x_0 = const()[name = string("op_663_transpose_x_0"), val = bool(false)];
            bool var_663_transpose_y_0 = const()[name = string("op_663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_25_cast_fp16 = transpose(perm = var_654, x = var_653_cast_fp16)[name = string("transpose_464")];
            tensor<fp16, [1, 16, ?, 64]> var_663_cast_fp16 = matmul(transpose_x = var_663_transpose_x_0, transpose_y = var_663_transpose_y_0, x = var_661_cast_fp16, y = v_25_cast_fp16)[name = string("op_663_cast_fp16")];
            tensor<int32, [4]> var_664 = const()[name = string("op_664"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_665_cast_fp16 = transpose(perm = var_664, x = var_663_cast_fp16)[name = string("transpose_461")];
            tensor<fp16, [1, ?, 1024]> x_43_cast_fp16 = reshape(shape = concat_61x, x = var_665_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175701440)))];
            tensor<fp16, [1024]> var_670_to_fp16 = const()[name = string("op_670_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177798656)))];
            tensor<fp16, [1, ?, 1024]> linear_19_cast_fp16 = linear(bias = var_670_to_fp16, weight = var_669_to_fp16, x = x_43_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_45_cast_fp16 = add(x = x_39_cast_fp16, y = linear_19_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [1]> var_677_axes_0 = const()[name = string("op_677_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_2_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_2_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177800768)))];
            tensor<fp16, [1024]> blocks_2_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_2_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177802880)))];
            tensor<fp16, [1, ?, 1024]> var_677_cast_fp16 = layer_norm(axes = var_677_axes_0, beta = blocks_2_cross_attn_ln_bias_to_fp16, epsilon = var_576_to_fp16, gamma = blocks_2_cross_attn_ln_weight_to_fp16, x = x_45_cast_fp16)[name = string("op_677_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_686_to_fp16 = const()[name = string("op_686_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177804992)))];
            tensor<fp16, [1024]> var_687_to_fp16 = const()[name = string("op_687_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179902208)))];
            tensor<fp16, [1, ?, 1024]> linear_20_cast_fp16 = linear(bias = var_687_to_fp16, weight = var_686_to_fp16, x = var_677_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_27_internal_tensor_assign_1_stride_0 = const()[name = string("k_27_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_27_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_27_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_27_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_62, begin_mask = k_27_internal_tensor_assign_1_begin_mask_0, end = concat_63, end_mask = k_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_27_internal_tensor_assign_1_squeeze_mask_0, stride = k_27_internal_tensor_assign_1_stride_0, update = k_cache_11_cast_fp16, x = k_7_to_fp16)[name = string("k_27_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_27_internal_tensor_assign_1_stride_0 = const()[name = string("v_27_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_27_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_27_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_27_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_64, begin_mask = v_27_internal_tensor_assign_1_begin_mask_0, end = concat_65, end_mask = v_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_27_internal_tensor_assign_1_squeeze_mask_0, stride = v_27_internal_tensor_assign_1_stride_0, update = v_cache_11_cast_fp16, x = k_7_to_fp16)[name = string("v_27_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_66x = const()[name = string("concat_66x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_707_cast_fp16 = reshape(shape = concat_66x, x = linear_20_cast_fp16)[name = string("op_707_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_130_to_fp16 = const()[name = string("const_130_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_23_cast_fp16 = mul(x = var_707_cast_fp16, y = const_130_to_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> var_713 = const()[name = string("op_713"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_714_cast_fp16 = reshape(shape = var_713, x = k_27_internal_tensor_assign_1_cast_fp16)[name = string("op_714_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_29_cast_fp16 = mul(x = var_714_cast_fp16, y = const_131_to_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_720 = const()[name = string("op_720"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_721_cast_fp16 = reshape(shape = var_720, x = v_27_internal_tensor_assign_1_cast_fp16)[name = string("op_721_cast_fp16")];
            tensor<int32, [4]> var_722 = const()[name = string("op_722"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_17_transpose_x_0 = const()[name = string("qk_17_transpose_x_0"), val = bool(false)];
            bool qk_17_transpose_y_0 = const()[name = string("qk_17_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_203_perm_0 = const()[name = string("transpose_203_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_204_perm_0 = const()[name = string("transpose_204_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_204 = transpose(perm = transpose_204_perm_0, x = k_29_cast_fp16)[name = string("transpose_458")];
            tensor<fp16, [1, 16, ?, 64]> transpose_203 = transpose(perm = transpose_203_perm_0, x = q_23_cast_fp16)[name = string("transpose_459")];
            tensor<fp16, [1, 16, ?, 1500]> qk_17_cast_fp16 = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_203, y = transpose_204)[name = string("qk_17_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_726_cast_fp16 = softmax(axis = var_570, x = qk_17_cast_fp16)[name = string("op_726_cast_fp16")];
            bool var_728_transpose_x_0 = const()[name = string("op_728_transpose_x_0"), val = bool(false)];
            bool var_728_transpose_y_0 = const()[name = string("op_728_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_29_cast_fp16 = transpose(perm = var_722, x = var_721_cast_fp16)[name = string("transpose_460")];
            tensor<fp16, [1, 16, ?, 64]> var_728_cast_fp16 = matmul(transpose_x = var_728_transpose_x_0, transpose_y = var_728_transpose_y_0, x = var_726_cast_fp16, y = v_29_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<int32, [4]> var_729 = const()[name = string("op_729"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_67x = const()[name = string("concat_67x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_730_cast_fp16 = transpose(perm = var_729, x = var_728_cast_fp16)[name = string("transpose_457")];
            tensor<fp16, [1, ?, 1024]> x_49_cast_fp16 = reshape(shape = concat_67x, x = var_730_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_734_to_fp16 = const()[name = string("op_734_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179904320)))];
            tensor<fp16, [1024]> var_735_to_fp16 = const()[name = string("op_735_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182001536)))];
            tensor<fp16, [1, ?, 1024]> linear_21_cast_fp16 = linear(bias = var_735_to_fp16, weight = var_734_to_fp16, x = x_49_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_21_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [1]> var_742_axes_0 = const()[name = string("op_742_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_2_mlp_ln_weight_to_fp16 = const()[name = string("blocks_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182003648)))];
            tensor<fp16, [1024]> blocks_2_mlp_ln_bias_to_fp16 = const()[name = string("blocks_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182005760)))];
            tensor<fp16, [1, ?, 1024]> var_742_cast_fp16 = layer_norm(axes = var_742_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_576_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_51_cast_fp16)[name = string("op_742_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_751_to_fp16 = const()[name = string("op_751_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182007872)))];
            tensor<fp16, [4096]> var_752_to_fp16 = const()[name = string("op_752_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190396544)))];
            tensor<fp16, [1, ?, 4096]> linear_22_cast_fp16 = linear(bias = var_752_to_fp16, weight = var_751_to_fp16, x = var_742_cast_fp16)[name = string("linear_22_cast_fp16")];
            string x_55_mode_0 = const()[name = string("x_55_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_55_cast_fp16 = gelu(mode = x_55_mode_0, x = linear_22_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_757_to_fp16 = const()[name = string("op_757_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190404800)))];
            tensor<fp16, [1024]> var_758_to_fp16 = const()[name = string("op_758_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198793472)))];
            tensor<fp16, [1, ?, 1024]> linear_23_cast_fp16 = linear(bias = var_758_to_fp16, weight = var_757_to_fp16, x = x_55_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_57_cast_fp16 = add(x = x_51_cast_fp16, y = linear_23_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<int32, [4]> k_cache_13_begin_0 = const()[name = string("k_cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> k_cache_13_end_0 = const()[name = string("k_cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_13_end_mask_0 = const()[name = string("k_cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_13_squeeze_mask_0 = const()[name = string("k_cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_13_cast_fp16 = slice_by_index(begin = k_cache_13_begin_0, end = k_cache_13_end_0, end_mask = k_cache_13_end_mask_0, squeeze_mask = k_cache_13_squeeze_mask_0, x = coreml_update_state_52)[name = string("k_cache_13_cast_fp16")];
            tensor<int32, [4]> v_cache_13_begin_0 = const()[name = string("v_cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> v_cache_13_end_0 = const()[name = string("v_cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_13_end_mask_0 = const()[name = string("v_cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_13_squeeze_mask_0 = const()[name = string("v_cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_13_cast_fp16 = slice_by_index(begin = v_cache_13_begin_0, end = v_cache_13_end_0, end_mask = v_cache_13_end_mask_0, squeeze_mask = v_cache_13_squeeze_mask_0, x = coreml_update_state_53)[name = string("v_cache_13_cast_fp16")];
            tensor<int32, [4]> k_cache_15_begin_0 = const()[name = string("k_cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> k_cache_15_end_0 = const()[name = string("k_cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_15_end_mask_0 = const()[name = string("k_cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_15_squeeze_mask_0 = const()[name = string("k_cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_15_cast_fp16 = slice_by_index(begin = k_cache_15_begin_0, end = k_cache_15_end_0, end_mask = k_cache_15_end_mask_0, squeeze_mask = k_cache_15_squeeze_mask_0, x = read_state_2)[name = string("k_cache_15_cast_fp16")];
            tensor<int32, [4]> v_cache_15_begin_0 = const()[name = string("v_cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> v_cache_15_end_0 = const()[name = string("v_cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_15_end_mask_0 = const()[name = string("v_cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_15_squeeze_mask_0 = const()[name = string("v_cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_15_cast_fp16 = slice_by_index(begin = v_cache_15_begin_0, end = v_cache_15_end_0, end_mask = v_cache_15_end_mask_0, squeeze_mask = v_cache_15_squeeze_mask_0, x = read_state_3)[name = string("v_cache_15_cast_fp16")];
            int32 var_781 = const()[name = string("op_781"), val = int32(-1)];
            tensor<int32, [1]> var_799_axes_0 = const()[name = string("op_799_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_3_attn_ln_weight_to_fp16 = const()[name = string("blocks_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198795584)))];
            tensor<fp16, [1024]> blocks_3_attn_ln_bias_to_fp16 = const()[name = string("blocks_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198797696)))];
            fp16 var_787_to_fp16 = const()[name = string("op_787_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_799_cast_fp16 = layer_norm(axes = var_799_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_787_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_57_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_810_to_fp16 = const()[name = string("op_810_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198799808)))];
            tensor<fp16, [1024]> var_811_to_fp16 = const()[name = string("op_811_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200897024)))];
            tensor<fp16, [1, ?, 1024]> linear_24_cast_fp16 = linear(bias = var_811_to_fp16, weight = var_810_to_fp16, x = var_799_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200899136)))];
            tensor<fp16, [1, ?, 1024]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_814_to_fp16, x = var_799_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_818_to_fp16 = const()[name = string("op_818_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202996352)))];
            tensor<fp16, [1024]> var_819_to_fp16 = const()[name = string("op_819_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205093568)))];
            tensor<fp16, [1, ?, 1024]> linear_26_cast_fp16 = linear(bias = var_819_to_fp16, weight = var_818_to_fp16, x = var_799_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> var_821_shape_cast_fp16 = shape(x = linear_24_cast_fp16)[name = string("op_821_shape_cast_fp16")];
            int32 gather_38_axis_0 = const()[name = string("gather_38_axis_0"), val = int32(0)];
            int32 gather_38_batch_dims_0 = const()[name = string("gather_38_batch_dims_0"), val = int32(0)];
            bool gather_38_validate_indices_0 = const()[name = string("gather_38_validate_indices_0"), val = bool(false)];
            string var_821_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_821_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_38_to_uint16 = const()[name = string("select_38_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_821_shape_cast_fp16_to_uint16 = cast(dtype = var_821_shape_cast_fp16_to_uint16_dtype_0, x = var_821_shape_cast_fp16)[name = string("cast_288")];
            uint16 gather_38_cast_uint16 = gather(axis = gather_38_axis_0, batch_dims = gather_38_batch_dims_0, indices = select_38_to_uint16, validate_indices = gather_38_validate_indices_0, x = var_821_shape_cast_fp16_to_uint16)[name = string("gather_38_cast_uint16")];
            string gather_38_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_38_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_38_cast_uint16_to_int32 = cast(dtype = gather_38_cast_uint16_to_int32_dtype_0, x = gather_38_cast_uint16)[name = string("cast_287")];
            int32 end_step_9 = add(x = offset, y = gather_38_cast_uint16_to_int32)[name = string("end_step_9")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_50 = const()[name = string("expand_dims_50"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51_axes_0 = const()[name = string("expand_dims_51_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = expand_dims(axes = expand_dims_51_axes_0, x = end_step_9)[name = string("expand_dims_51")];
            tensor<int32, [1]> concat_70_values0_0 = const()[name = string("concat_70_values0_0"), val = tensor<int32, [1]>([3])];
            int32 concat_70_axis_0 = const()[name = string("concat_70_axis_0"), val = int32(0)];
            bool concat_70_interleave_0 = const()[name = string("concat_70_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_70 = concat(axis = concat_70_axis_0, interleave = concat_70_interleave_0, values = (concat_70_values0_0, expand_dims_48, expand_dims_1, expand_dims_50))[name = string("concat_70")];
            tensor<int32, [1]> concat_71_values0_0 = const()[name = string("concat_71_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_71_values1_0 = const()[name = string("concat_71_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_71_values3_0 = const()[name = string("concat_71_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_71_axis_0 = const()[name = string("concat_71_axis_0"), val = int32(0)];
            bool concat_71_interleave_0 = const()[name = string("concat_71_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_71 = concat(axis = concat_71_axis_0, interleave = concat_71_interleave_0, values = (concat_71_values0_0, concat_71_values1_0, expand_dims_51, concat_71_values3_0))[name = string("concat_71")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_4_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_4_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_4_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_70, begin_mask = k_cache1_internal_tensor_assign_4_begin_mask_0, end = concat_71, end_mask = k_cache1_internal_tensor_assign_4_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_4_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_4_stride_0, update = linear_25_cast_fp16, x = coreml_update_state_52)[name = string("k_cache1_internal_tensor_assign_4_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_4_cast_fp16, input = k_cache1)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_54 = read_state(input = k_cache1)[name = string("coreml_update_state_54")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_4_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_4_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_4_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_70, begin_mask = v_cache1_internal_tensor_assign_4_begin_mask_0, end = concat_71, end_mask = v_cache1_internal_tensor_assign_4_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_4_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_4_stride_0, update = linear_26_cast_fp16, x = coreml_update_state_53)[name = string("v_cache1_internal_tensor_assign_4_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_4_cast_fp16, input = v_cache1)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_55 = read_state(input = v_cache1)[name = string("coreml_update_state_55")];
            int32 concat_76_values0_0 = const()[name = string("concat_76_values0_0"), val = int32(1)];
            int32 concat_76_values2_0 = const()[name = string("concat_76_values2_0"), val = int32(1024)];
            int32 concat_76_axis_0 = const()[name = string("concat_76_axis_0"), val = int32(0)];
            bool concat_76_interleave_0 = const()[name = string("concat_76_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_76 = concat(axis = concat_76_axis_0, interleave = concat_76_interleave_0, values = (concat_76_values0_0, end_step_9, concat_76_values2_0))[name = string("concat_76")];
            tensor<int32, [3]> var_837_begin_0 = const()[name = string("op_837_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_837_end_mask_0 = const()[name = string("op_837_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_837_cast_fp16 = slice_by_index(begin = var_837_begin_0, end = concat_76, end_mask = var_837_end_mask_0, x = k_cache_13_cast_fp16)[name = string("op_837_cast_fp16")];
            tensor<int32, [3]> var_840_begin_0 = const()[name = string("op_840_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_840_end_mask_0 = const()[name = string("op_840_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_840_cast_fp16 = slice_by_index(begin = var_840_begin_0, end = concat_76, end_mask = var_840_end_mask_0, x = v_cache_13_cast_fp16)[name = string("op_840_cast_fp16")];
            tensor<int32, [4]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_850_cast_fp16 = reshape(shape = concat_78x, x = linear_24_cast_fp16)[name = string("op_850_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_132_to_fp16 = const()[name = string("const_132_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_27_cast_fp16 = mul(x = var_850_cast_fp16, y = const_132_to_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> concat_79x = const()[name = string("concat_79x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_857_cast_fp16 = reshape(shape = concat_79x, x = var_837_cast_fp16)[name = string("op_857_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_133_to_fp16 = const()[name = string("const_133_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_35_cast_fp16 = mul(x = var_857_cast_fp16, y = const_133_to_fp16)[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> concat_80x = const()[name = string("concat_80x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_864_cast_fp16 = reshape(shape = concat_80x, x = var_840_cast_fp16)[name = string("op_864_cast_fp16")];
            tensor<int32, [4]> var_865 = const()[name = string("op_865"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_19_transpose_x_0 = const()[name = string("qk_19_transpose_x_0"), val = bool(false)];
            bool qk_19_transpose_y_0 = const()[name = string("qk_19_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_205_perm_0 = const()[name = string("transpose_205_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_206_perm_0 = const()[name = string("transpose_206_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_206 = transpose(perm = transpose_206_perm_0, x = k_35_cast_fp16)[name = string("transpose_454")];
            tensor<fp16, [1, 16, ?, 64]> transpose_205 = transpose(perm = transpose_205_perm_0, x = q_27_cast_fp16)[name = string("transpose_455")];
            tensor<fp16, [1, 16, ?, ?]> qk_19_cast_fp16 = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_205, y = transpose_206)[name = string("qk_19_cast_fp16")];
            int32 concat_81_values1_0 = const()[name = string("concat_81_values1_0"), val = int32(448)];
            int32 concat_81_axis_0 = const()[name = string("concat_81_axis_0"), val = int32(0)];
            bool concat_81_interleave_0 = const()[name = string("concat_81_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_81 = concat(axis = concat_81_axis_0, interleave = concat_81_interleave_0, values = (gather_38_cast_uint16_to_int32, concat_81_values1_0))[name = string("concat_81")];
            tensor<int32, [2]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = concat_81, end_mask = var_868_end_mask_0, x = mask_to_fp16)[name = string("op_868_cast_fp16")];
            int32 concat_82_values0_0 = const()[name = string("concat_82_values0_0"), val = int32(0)];
            int32 concat_82_axis_0 = const()[name = string("concat_82_axis_0"), val = int32(0)];
            bool concat_82_interleave_0 = const()[name = string("concat_82_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_82 = concat(axis = concat_82_axis_0, interleave = concat_82_interleave_0, values = (concat_82_values0_0, gather_38_cast_uint16_to_int32))[name = string("concat_82")];
            tensor<int32, [2]> var_869_begin_0 = const()[name = string("op_869_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_869_end_mask_0 = const()[name = string("op_869_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_869_cast_fp16 = slice_by_index(begin = var_869_begin_0, end = concat_82, end_mask = var_869_end_mask_0, x = var_868_cast_fp16)[name = string("op_869_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_21_cast_fp16 = add(x = qk_19_cast_fp16, y = var_869_cast_fp16)[name = string("qk_21_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_872_cast_fp16 = softmax(axis = var_781, x = qk_21_cast_fp16)[name = string("op_872_cast_fp16")];
            bool var_874_transpose_x_0 = const()[name = string("op_874_transpose_x_0"), val = bool(false)];
            bool var_874_transpose_y_0 = const()[name = string("op_874_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_35_cast_fp16 = transpose(perm = var_865, x = var_864_cast_fp16)[name = string("transpose_456")];
            tensor<fp16, [1, 16, ?, 64]> var_874_cast_fp16 = matmul(transpose_x = var_874_transpose_x_0, transpose_y = var_874_transpose_y_0, x = var_872_cast_fp16, y = v_35_cast_fp16)[name = string("op_874_cast_fp16")];
            tensor<int32, [4]> var_875 = const()[name = string("op_875"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_876_cast_fp16 = transpose(perm = var_875, x = var_874_cast_fp16)[name = string("transpose_453")];
            tensor<fp16, [1, ?, 1024]> x_61_cast_fp16 = reshape(shape = concat_83x, x = var_876_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_880_to_fp16 = const()[name = string("op_880_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205095680)))];
            tensor<fp16, [1024]> var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207192896)))];
            tensor<fp16, [1, ?, 1024]> linear_27_cast_fp16 = linear(bias = var_881_to_fp16, weight = var_880_to_fp16, x = x_61_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_63_cast_fp16 = add(x = x_57_cast_fp16, y = linear_27_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<int32, [1]> var_888_axes_0 = const()[name = string("op_888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_3_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_3_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207195008)))];
            tensor<fp16, [1024]> blocks_3_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_3_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207197120)))];
            tensor<fp16, [1, ?, 1024]> var_888_cast_fp16 = layer_norm(axes = var_888_axes_0, beta = blocks_3_cross_attn_ln_bias_to_fp16, epsilon = var_787_to_fp16, gamma = blocks_3_cross_attn_ln_weight_to_fp16, x = x_63_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_897_to_fp16 = const()[name = string("op_897_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207199232)))];
            tensor<fp16, [1024]> var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209296448)))];
            tensor<fp16, [1, ?, 1024]> linear_28_cast_fp16 = linear(bias = var_898_to_fp16, weight = var_897_to_fp16, x = var_888_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_37_internal_tensor_assign_1_stride_0 = const()[name = string("k_37_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_37_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_37_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_37_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_84, begin_mask = k_37_internal_tensor_assign_1_begin_mask_0, end = concat_85, end_mask = k_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_37_internal_tensor_assign_1_squeeze_mask_0, stride = k_37_internal_tensor_assign_1_stride_0, update = k_cache_15_cast_fp16, x = k_7_to_fp16)[name = string("k_37_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_37_internal_tensor_assign_1_stride_0 = const()[name = string("v_37_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_37_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_37_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_37_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_86, begin_mask = v_37_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = v_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_37_internal_tensor_assign_1_squeeze_mask_0, stride = v_37_internal_tensor_assign_1_stride_0, update = v_cache_15_cast_fp16, x = k_7_to_fp16)[name = string("v_37_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_88x = const()[name = string("concat_88x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_918_cast_fp16 = reshape(shape = concat_88x, x = linear_28_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_134_to_fp16 = const()[name = string("const_134_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_31_cast_fp16 = mul(x = var_918_cast_fp16, y = const_134_to_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> var_924 = const()[name = string("op_924"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_925_cast_fp16 = reshape(shape = var_924, x = k_37_internal_tensor_assign_1_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_39_cast_fp16 = mul(x = var_925_cast_fp16, y = const_135_to_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_931 = const()[name = string("op_931"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_932_cast_fp16 = reshape(shape = var_931, x = v_37_internal_tensor_assign_1_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<int32, [4]> var_933 = const()[name = string("op_933"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_23_transpose_x_0 = const()[name = string("qk_23_transpose_x_0"), val = bool(false)];
            bool qk_23_transpose_y_0 = const()[name = string("qk_23_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_207_perm_0 = const()[name = string("transpose_207_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_208_perm_0 = const()[name = string("transpose_208_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_208 = transpose(perm = transpose_208_perm_0, x = k_39_cast_fp16)[name = string("transpose_450")];
            tensor<fp16, [1, 16, ?, 64]> transpose_207 = transpose(perm = transpose_207_perm_0, x = q_31_cast_fp16)[name = string("transpose_451")];
            tensor<fp16, [1, 16, ?, 1500]> qk_23_cast_fp16 = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_207, y = transpose_208)[name = string("qk_23_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_937_cast_fp16 = softmax(axis = var_781, x = qk_23_cast_fp16)[name = string("op_937_cast_fp16")];
            bool var_939_transpose_x_0 = const()[name = string("op_939_transpose_x_0"), val = bool(false)];
            bool var_939_transpose_y_0 = const()[name = string("op_939_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_39_cast_fp16 = transpose(perm = var_933, x = var_932_cast_fp16)[name = string("transpose_452")];
            tensor<fp16, [1, 16, ?, 64]> var_939_cast_fp16 = matmul(transpose_x = var_939_transpose_x_0, transpose_y = var_939_transpose_y_0, x = var_937_cast_fp16, y = v_39_cast_fp16)[name = string("op_939_cast_fp16")];
            tensor<int32, [4]> var_940 = const()[name = string("op_940"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_89x = const()[name = string("concat_89x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_941_cast_fp16 = transpose(perm = var_940, x = var_939_cast_fp16)[name = string("transpose_449")];
            tensor<fp16, [1, ?, 1024]> x_67_cast_fp16 = reshape(shape = concat_89x, x = var_941_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_945_to_fp16 = const()[name = string("op_945_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209298560)))];
            tensor<fp16, [1024]> var_946_to_fp16 = const()[name = string("op_946_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211395776)))];
            tensor<fp16, [1, ?, 1024]> linear_29_cast_fp16 = linear(bias = var_946_to_fp16, weight = var_945_to_fp16, x = x_67_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_69_cast_fp16 = add(x = x_63_cast_fp16, y = linear_29_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [1]> var_953_axes_0 = const()[name = string("op_953_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_3_mlp_ln_weight_to_fp16 = const()[name = string("blocks_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211397888)))];
            tensor<fp16, [1024]> blocks_3_mlp_ln_bias_to_fp16 = const()[name = string("blocks_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211400000)))];
            tensor<fp16, [1, ?, 1024]> var_953_cast_fp16 = layer_norm(axes = var_953_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_787_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_69_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_962_to_fp16 = const()[name = string("op_962_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211402112)))];
            tensor<fp16, [4096]> var_963_to_fp16 = const()[name = string("op_963_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219790784)))];
            tensor<fp16, [1, ?, 4096]> linear_30_cast_fp16 = linear(bias = var_963_to_fp16, weight = var_962_to_fp16, x = var_953_cast_fp16)[name = string("linear_30_cast_fp16")];
            string x_73_mode_0 = const()[name = string("x_73_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_73_cast_fp16 = gelu(mode = x_73_mode_0, x = linear_30_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_968_to_fp16 = const()[name = string("op_968_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219799040)))];
            tensor<fp16, [1024]> var_969_to_fp16 = const()[name = string("op_969_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228187712)))];
            tensor<fp16, [1, ?, 1024]> linear_31_cast_fp16 = linear(bias = var_969_to_fp16, weight = var_968_to_fp16, x = x_73_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_75_cast_fp16 = add(x = x_69_cast_fp16, y = linear_31_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<int32, [4]> k_cache_17_begin_0 = const()[name = string("k_cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> k_cache_17_end_0 = const()[name = string("k_cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_17_end_mask_0 = const()[name = string("k_cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_17_squeeze_mask_0 = const()[name = string("k_cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_17_cast_fp16 = slice_by_index(begin = k_cache_17_begin_0, end = k_cache_17_end_0, end_mask = k_cache_17_end_mask_0, squeeze_mask = k_cache_17_squeeze_mask_0, x = coreml_update_state_54)[name = string("k_cache_17_cast_fp16")];
            tensor<int32, [4]> v_cache_17_begin_0 = const()[name = string("v_cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> v_cache_17_end_0 = const()[name = string("v_cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_17_end_mask_0 = const()[name = string("v_cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_17_squeeze_mask_0 = const()[name = string("v_cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_17_cast_fp16 = slice_by_index(begin = v_cache_17_begin_0, end = v_cache_17_end_0, end_mask = v_cache_17_end_mask_0, squeeze_mask = v_cache_17_squeeze_mask_0, x = coreml_update_state_55)[name = string("v_cache_17_cast_fp16")];
            tensor<int32, [4]> k_cache_19_begin_0 = const()[name = string("k_cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> k_cache_19_end_0 = const()[name = string("k_cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_19_end_mask_0 = const()[name = string("k_cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_19_squeeze_mask_0 = const()[name = string("k_cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_19_cast_fp16 = slice_by_index(begin = k_cache_19_begin_0, end = k_cache_19_end_0, end_mask = k_cache_19_end_mask_0, squeeze_mask = k_cache_19_squeeze_mask_0, x = read_state_2)[name = string("k_cache_19_cast_fp16")];
            tensor<int32, [4]> v_cache_19_begin_0 = const()[name = string("v_cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> v_cache_19_end_0 = const()[name = string("v_cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_19_end_mask_0 = const()[name = string("v_cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_19_squeeze_mask_0 = const()[name = string("v_cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_19_cast_fp16 = slice_by_index(begin = v_cache_19_begin_0, end = v_cache_19_end_0, end_mask = v_cache_19_end_mask_0, squeeze_mask = v_cache_19_squeeze_mask_0, x = read_state_3)[name = string("v_cache_19_cast_fp16")];
            int32 var_992 = const()[name = string("op_992"), val = int32(-1)];
            tensor<int32, [1]> var_1010_axes_0 = const()[name = string("op_1010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_4_attn_ln_weight_to_fp16 = const()[name = string("blocks_4_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228189824)))];
            tensor<fp16, [1024]> blocks_4_attn_ln_bias_to_fp16 = const()[name = string("blocks_4_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228191936)))];
            fp16 var_998_to_fp16 = const()[name = string("op_998_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_1010_cast_fp16 = layer_norm(axes = var_1010_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_998_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_75_cast_fp16)[name = string("op_1010_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1021_to_fp16 = const()[name = string("op_1021_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228194048)))];
            tensor<fp16, [1024]> var_1022_to_fp16 = const()[name = string("op_1022_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230291264)))];
            tensor<fp16, [1, ?, 1024]> linear_32_cast_fp16 = linear(bias = var_1022_to_fp16, weight = var_1021_to_fp16, x = var_1010_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1025_to_fp16 = const()[name = string("op_1025_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230293376)))];
            tensor<fp16, [1, ?, 1024]> linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1025_to_fp16, x = var_1010_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1029_to_fp16 = const()[name = string("op_1029_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232390592)))];
            tensor<fp16, [1024]> var_1030_to_fp16 = const()[name = string("op_1030_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234487808)))];
            tensor<fp16, [1, ?, 1024]> linear_34_cast_fp16 = linear(bias = var_1030_to_fp16, weight = var_1029_to_fp16, x = var_1010_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<int32, [3]> var_1032_shape_cast_fp16 = shape(x = linear_32_cast_fp16)[name = string("op_1032_shape_cast_fp16")];
            int32 gather_50_axis_0 = const()[name = string("gather_50_axis_0"), val = int32(0)];
            int32 gather_50_batch_dims_0 = const()[name = string("gather_50_batch_dims_0"), val = int32(0)];
            bool gather_50_validate_indices_0 = const()[name = string("gather_50_validate_indices_0"), val = bool(false)];
            string var_1032_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1032_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_50_to_uint16 = const()[name = string("select_50_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1032_shape_cast_fp16_to_uint16 = cast(dtype = var_1032_shape_cast_fp16_to_uint16_dtype_0, x = var_1032_shape_cast_fp16)[name = string("cast_286")];
            uint16 gather_50_cast_uint16 = gather(axis = gather_50_axis_0, batch_dims = gather_50_batch_dims_0, indices = select_50_to_uint16, validate_indices = gather_50_validate_indices_0, x = var_1032_shape_cast_fp16_to_uint16)[name = string("gather_50_cast_uint16")];
            string gather_50_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_50_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_50_cast_uint16_to_int32 = cast(dtype = gather_50_cast_uint16_to_int32_dtype_0, x = gather_50_cast_uint16)[name = string("cast_285")];
            int32 end_step_11 = add(x = offset, y = gather_50_cast_uint16_to_int32)[name = string("end_step_11")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_66 = const()[name = string("expand_dims_66"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_67_axes_0 = const()[name = string("expand_dims_67_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_67 = expand_dims(axes = expand_dims_67_axes_0, x = end_step_11)[name = string("expand_dims_67")];
            tensor<int32, [1]> concat_92_values0_0 = const()[name = string("concat_92_values0_0"), val = tensor<int32, [1]>([4])];
            int32 concat_92_axis_0 = const()[name = string("concat_92_axis_0"), val = int32(0)];
            bool concat_92_interleave_0 = const()[name = string("concat_92_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_92 = concat(axis = concat_92_axis_0, interleave = concat_92_interleave_0, values = (concat_92_values0_0, expand_dims_64, expand_dims_1, expand_dims_66))[name = string("concat_92")];
            tensor<int32, [1]> concat_93_values0_0 = const()[name = string("concat_93_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_93_values1_0 = const()[name = string("concat_93_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_93_values3_0 = const()[name = string("concat_93_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_93_axis_0 = const()[name = string("concat_93_axis_0"), val = int32(0)];
            bool concat_93_interleave_0 = const()[name = string("concat_93_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_93 = concat(axis = concat_93_axis_0, interleave = concat_93_interleave_0, values = (concat_93_values0_0, concat_93_values1_0, expand_dims_67, concat_93_values3_0))[name = string("concat_93")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_5_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_5_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_5_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_92, begin_mask = k_cache1_internal_tensor_assign_5_begin_mask_0, end = concat_93, end_mask = k_cache1_internal_tensor_assign_5_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_5_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_5_stride_0, update = linear_33_cast_fp16, x = coreml_update_state_54)[name = string("k_cache1_internal_tensor_assign_5_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_5_cast_fp16, input = k_cache1)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_56 = read_state(input = k_cache1)[name = string("coreml_update_state_56")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_5_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_5_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_5_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_92, begin_mask = v_cache1_internal_tensor_assign_5_begin_mask_0, end = concat_93, end_mask = v_cache1_internal_tensor_assign_5_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_5_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_5_stride_0, update = linear_34_cast_fp16, x = coreml_update_state_55)[name = string("v_cache1_internal_tensor_assign_5_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_5_cast_fp16, input = v_cache1)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_57 = read_state(input = v_cache1)[name = string("coreml_update_state_57")];
            int32 concat_98_values0_0 = const()[name = string("concat_98_values0_0"), val = int32(1)];
            int32 concat_98_values2_0 = const()[name = string("concat_98_values2_0"), val = int32(1024)];
            int32 concat_98_axis_0 = const()[name = string("concat_98_axis_0"), val = int32(0)];
            bool concat_98_interleave_0 = const()[name = string("concat_98_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_98 = concat(axis = concat_98_axis_0, interleave = concat_98_interleave_0, values = (concat_98_values0_0, end_step_11, concat_98_values2_0))[name = string("concat_98")];
            tensor<int32, [3]> var_1048_begin_0 = const()[name = string("op_1048_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1048_end_mask_0 = const()[name = string("op_1048_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1048_cast_fp16 = slice_by_index(begin = var_1048_begin_0, end = concat_98, end_mask = var_1048_end_mask_0, x = k_cache_17_cast_fp16)[name = string("op_1048_cast_fp16")];
            tensor<int32, [3]> var_1051_begin_0 = const()[name = string("op_1051_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1051_end_mask_0 = const()[name = string("op_1051_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1051_cast_fp16 = slice_by_index(begin = var_1051_begin_0, end = concat_98, end_mask = var_1051_end_mask_0, x = v_cache_17_cast_fp16)[name = string("op_1051_cast_fp16")];
            tensor<int32, [4]> concat_100x = const()[name = string("concat_100x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1061_cast_fp16 = reshape(shape = concat_100x, x = linear_32_cast_fp16)[name = string("op_1061_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_35_cast_fp16 = mul(x = var_1061_cast_fp16, y = const_136_to_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> concat_101x = const()[name = string("concat_101x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1068_cast_fp16 = reshape(shape = concat_101x, x = var_1048_cast_fp16)[name = string("op_1068_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_137_to_fp16 = const()[name = string("const_137_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_45_cast_fp16 = mul(x = var_1068_cast_fp16, y = const_137_to_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> concat_102x = const()[name = string("concat_102x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1075_cast_fp16 = reshape(shape = concat_102x, x = var_1051_cast_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_25_transpose_x_0 = const()[name = string("qk_25_transpose_x_0"), val = bool(false)];
            bool qk_25_transpose_y_0 = const()[name = string("qk_25_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_209_perm_0 = const()[name = string("transpose_209_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_210_perm_0 = const()[name = string("transpose_210_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_210 = transpose(perm = transpose_210_perm_0, x = k_45_cast_fp16)[name = string("transpose_446")];
            tensor<fp16, [1, 16, ?, 64]> transpose_209 = transpose(perm = transpose_209_perm_0, x = q_35_cast_fp16)[name = string("transpose_447")];
            tensor<fp16, [1, 16, ?, ?]> qk_25_cast_fp16 = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_209, y = transpose_210)[name = string("qk_25_cast_fp16")];
            int32 concat_103_values1_0 = const()[name = string("concat_103_values1_0"), val = int32(448)];
            int32 concat_103_axis_0 = const()[name = string("concat_103_axis_0"), val = int32(0)];
            bool concat_103_interleave_0 = const()[name = string("concat_103_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_103 = concat(axis = concat_103_axis_0, interleave = concat_103_interleave_0, values = (gather_50_cast_uint16_to_int32, concat_103_values1_0))[name = string("concat_103")];
            tensor<int32, [2]> var_1079_begin_0 = const()[name = string("op_1079_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1079_end_mask_0 = const()[name = string("op_1079_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_1079_cast_fp16 = slice_by_index(begin = var_1079_begin_0, end = concat_103, end_mask = var_1079_end_mask_0, x = mask_to_fp16)[name = string("op_1079_cast_fp16")];
            int32 concat_104_values0_0 = const()[name = string("concat_104_values0_0"), val = int32(0)];
            int32 concat_104_axis_0 = const()[name = string("concat_104_axis_0"), val = int32(0)];
            bool concat_104_interleave_0 = const()[name = string("concat_104_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_104 = concat(axis = concat_104_axis_0, interleave = concat_104_interleave_0, values = (concat_104_values0_0, gather_50_cast_uint16_to_int32))[name = string("concat_104")];
            tensor<int32, [2]> var_1080_begin_0 = const()[name = string("op_1080_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1080_end_mask_0 = const()[name = string("op_1080_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_1080_cast_fp16 = slice_by_index(begin = var_1080_begin_0, end = concat_104, end_mask = var_1080_end_mask_0, x = var_1079_cast_fp16)[name = string("op_1080_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_27_cast_fp16 = add(x = qk_25_cast_fp16, y = var_1080_cast_fp16)[name = string("qk_27_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_1083_cast_fp16 = softmax(axis = var_992, x = qk_27_cast_fp16)[name = string("op_1083_cast_fp16")];
            bool var_1085_transpose_x_0 = const()[name = string("op_1085_transpose_x_0"), val = bool(false)];
            bool var_1085_transpose_y_0 = const()[name = string("op_1085_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_45_cast_fp16 = transpose(perm = var_1076, x = var_1075_cast_fp16)[name = string("transpose_448")];
            tensor<fp16, [1, 16, ?, 64]> var_1085_cast_fp16 = matmul(transpose_x = var_1085_transpose_x_0, transpose_y = var_1085_transpose_y_0, x = var_1083_cast_fp16, y = v_45_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<int32, [4]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_105x = const()[name = string("concat_105x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1087_cast_fp16 = transpose(perm = var_1086, x = var_1085_cast_fp16)[name = string("transpose_445")];
            tensor<fp16, [1, ?, 1024]> x_79_cast_fp16 = reshape(shape = concat_105x, x = var_1087_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1091_to_fp16 = const()[name = string("op_1091_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234489920)))];
            tensor<fp16, [1024]> var_1092_to_fp16 = const()[name = string("op_1092_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236587136)))];
            tensor<fp16, [1, ?, 1024]> linear_35_cast_fp16 = linear(bias = var_1092_to_fp16, weight = var_1091_to_fp16, x = x_79_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_81_cast_fp16 = add(x = x_75_cast_fp16, y = linear_35_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [1]> var_1099_axes_0 = const()[name = string("op_1099_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_4_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_4_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236589248)))];
            tensor<fp16, [1024]> blocks_4_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_4_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236591360)))];
            tensor<fp16, [1, ?, 1024]> var_1099_cast_fp16 = layer_norm(axes = var_1099_axes_0, beta = blocks_4_cross_attn_ln_bias_to_fp16, epsilon = var_998_to_fp16, gamma = blocks_4_cross_attn_ln_weight_to_fp16, x = x_81_cast_fp16)[name = string("op_1099_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1108_to_fp16 = const()[name = string("op_1108_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236593472)))];
            tensor<fp16, [1024]> var_1109_to_fp16 = const()[name = string("op_1109_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238690688)))];
            tensor<fp16, [1, ?, 1024]> linear_36_cast_fp16 = linear(bias = var_1109_to_fp16, weight = var_1108_to_fp16, x = var_1099_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [3]> concat_106 = const()[name = string("concat_106"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_107 = const()[name = string("concat_107"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_47_internal_tensor_assign_1_stride_0 = const()[name = string("k_47_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_47_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_47_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_47_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_47_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_47_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_47_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_47_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_106, begin_mask = k_47_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = k_47_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_47_internal_tensor_assign_1_squeeze_mask_0, stride = k_47_internal_tensor_assign_1_stride_0, update = k_cache_19_cast_fp16, x = k_7_to_fp16)[name = string("k_47_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_108 = const()[name = string("concat_108"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_109 = const()[name = string("concat_109"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_47_internal_tensor_assign_1_stride_0 = const()[name = string("v_47_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_47_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_47_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_47_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_47_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_47_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_47_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_47_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_108, begin_mask = v_47_internal_tensor_assign_1_begin_mask_0, end = concat_109, end_mask = v_47_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_47_internal_tensor_assign_1_squeeze_mask_0, stride = v_47_internal_tensor_assign_1_stride_0, update = v_cache_19_cast_fp16, x = k_7_to_fp16)[name = string("v_47_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1129_cast_fp16 = reshape(shape = concat_110x, x = linear_36_cast_fp16)[name = string("op_1129_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_138_to_fp16 = const()[name = string("const_138_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_39_cast_fp16 = mul(x = var_1129_cast_fp16, y = const_138_to_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> var_1135 = const()[name = string("op_1135"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1136_cast_fp16 = reshape(shape = var_1135, x = k_47_internal_tensor_assign_1_cast_fp16)[name = string("op_1136_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_139_to_fp16 = const()[name = string("const_139_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_49_cast_fp16 = mul(x = var_1136_cast_fp16, y = const_139_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_1142 = const()[name = string("op_1142"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1143_cast_fp16 = reshape(shape = var_1142, x = v_47_internal_tensor_assign_1_cast_fp16)[name = string("op_1143_cast_fp16")];
            tensor<int32, [4]> var_1144 = const()[name = string("op_1144"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_29_transpose_x_0 = const()[name = string("qk_29_transpose_x_0"), val = bool(false)];
            bool qk_29_transpose_y_0 = const()[name = string("qk_29_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_211_perm_0 = const()[name = string("transpose_211_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_212_perm_0 = const()[name = string("transpose_212_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_212 = transpose(perm = transpose_212_perm_0, x = k_49_cast_fp16)[name = string("transpose_442")];
            tensor<fp16, [1, 16, ?, 64]> transpose_211 = transpose(perm = transpose_211_perm_0, x = q_39_cast_fp16)[name = string("transpose_443")];
            tensor<fp16, [1, 16, ?, 1500]> qk_29_cast_fp16 = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_211, y = transpose_212)[name = string("qk_29_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_1148_cast_fp16 = softmax(axis = var_992, x = qk_29_cast_fp16)[name = string("op_1148_cast_fp16")];
            bool var_1150_transpose_x_0 = const()[name = string("op_1150_transpose_x_0"), val = bool(false)];
            bool var_1150_transpose_y_0 = const()[name = string("op_1150_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_49_cast_fp16 = transpose(perm = var_1144, x = var_1143_cast_fp16)[name = string("transpose_444")];
            tensor<fp16, [1, 16, ?, 64]> var_1150_cast_fp16 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = var_1148_cast_fp16, y = v_49_cast_fp16)[name = string("op_1150_cast_fp16")];
            tensor<int32, [4]> var_1151 = const()[name = string("op_1151"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_111x = const()[name = string("concat_111x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1152_cast_fp16 = transpose(perm = var_1151, x = var_1150_cast_fp16)[name = string("transpose_441")];
            tensor<fp16, [1, ?, 1024]> x_85_cast_fp16 = reshape(shape = concat_111x, x = var_1152_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1156_to_fp16 = const()[name = string("op_1156_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238692800)))];
            tensor<fp16, [1024]> var_1157_to_fp16 = const()[name = string("op_1157_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240790016)))];
            tensor<fp16, [1, ?, 1024]> linear_37_cast_fp16 = linear(bias = var_1157_to_fp16, weight = var_1156_to_fp16, x = x_85_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_87_cast_fp16 = add(x = x_81_cast_fp16, y = linear_37_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [1]> var_1164_axes_0 = const()[name = string("op_1164_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_4_mlp_ln_weight_to_fp16 = const()[name = string("blocks_4_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240792128)))];
            tensor<fp16, [1024]> blocks_4_mlp_ln_bias_to_fp16 = const()[name = string("blocks_4_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240794240)))];
            tensor<fp16, [1, ?, 1024]> var_1164_cast_fp16 = layer_norm(axes = var_1164_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_998_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_87_cast_fp16)[name = string("op_1164_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_1173_to_fp16 = const()[name = string("op_1173_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240796352)))];
            tensor<fp16, [4096]> var_1174_to_fp16 = const()[name = string("op_1174_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249185024)))];
            tensor<fp16, [1, ?, 4096]> linear_38_cast_fp16 = linear(bias = var_1174_to_fp16, weight = var_1173_to_fp16, x = var_1164_cast_fp16)[name = string("linear_38_cast_fp16")];
            string x_91_mode_0 = const()[name = string("x_91_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_91_cast_fp16 = gelu(mode = x_91_mode_0, x = linear_38_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_1179_to_fp16 = const()[name = string("op_1179_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249193280)))];
            tensor<fp16, [1024]> var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257581952)))];
            tensor<fp16, [1, ?, 1024]> linear_39_cast_fp16 = linear(bias = var_1180_to_fp16, weight = var_1179_to_fp16, x = x_91_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_93_cast_fp16 = add(x = x_87_cast_fp16, y = linear_39_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [4]> k_cache_21_begin_0 = const()[name = string("k_cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> k_cache_21_end_0 = const()[name = string("k_cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_21_end_mask_0 = const()[name = string("k_cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_21_squeeze_mask_0 = const()[name = string("k_cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_21_cast_fp16 = slice_by_index(begin = k_cache_21_begin_0, end = k_cache_21_end_0, end_mask = k_cache_21_end_mask_0, squeeze_mask = k_cache_21_squeeze_mask_0, x = coreml_update_state_56)[name = string("k_cache_21_cast_fp16")];
            tensor<int32, [4]> v_cache_21_begin_0 = const()[name = string("v_cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> v_cache_21_end_0 = const()[name = string("v_cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_21_end_mask_0 = const()[name = string("v_cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_21_squeeze_mask_0 = const()[name = string("v_cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_21_cast_fp16 = slice_by_index(begin = v_cache_21_begin_0, end = v_cache_21_end_0, end_mask = v_cache_21_end_mask_0, squeeze_mask = v_cache_21_squeeze_mask_0, x = coreml_update_state_57)[name = string("v_cache_21_cast_fp16")];
            tensor<int32, [4]> k_cache_23_begin_0 = const()[name = string("k_cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> k_cache_23_end_0 = const()[name = string("k_cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_23_end_mask_0 = const()[name = string("k_cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_23_squeeze_mask_0 = const()[name = string("k_cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_23_cast_fp16 = slice_by_index(begin = k_cache_23_begin_0, end = k_cache_23_end_0, end_mask = k_cache_23_end_mask_0, squeeze_mask = k_cache_23_squeeze_mask_0, x = read_state_2)[name = string("k_cache_23_cast_fp16")];
            tensor<int32, [4]> v_cache_23_begin_0 = const()[name = string("v_cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> v_cache_23_end_0 = const()[name = string("v_cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_23_end_mask_0 = const()[name = string("v_cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_23_squeeze_mask_0 = const()[name = string("v_cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_23_cast_fp16 = slice_by_index(begin = v_cache_23_begin_0, end = v_cache_23_end_0, end_mask = v_cache_23_end_mask_0, squeeze_mask = v_cache_23_squeeze_mask_0, x = read_state_3)[name = string("v_cache_23_cast_fp16")];
            int32 var_1203 = const()[name = string("op_1203"), val = int32(-1)];
            tensor<int32, [1]> var_1221_axes_0 = const()[name = string("op_1221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_5_attn_ln_weight_to_fp16 = const()[name = string("blocks_5_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257584064)))];
            tensor<fp16, [1024]> blocks_5_attn_ln_bias_to_fp16 = const()[name = string("blocks_5_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257586176)))];
            fp16 var_1209_to_fp16 = const()[name = string("op_1209_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_1221_cast_fp16 = layer_norm(axes = var_1221_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_1209_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_93_cast_fp16)[name = string("op_1221_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1232_to_fp16 = const()[name = string("op_1232_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257588288)))];
            tensor<fp16, [1024]> var_1233_to_fp16 = const()[name = string("op_1233_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259685504)))];
            tensor<fp16, [1, ?, 1024]> linear_40_cast_fp16 = linear(bias = var_1233_to_fp16, weight = var_1232_to_fp16, x = var_1221_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1236_to_fp16 = const()[name = string("op_1236_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259687616)))];
            tensor<fp16, [1, ?, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1236_to_fp16, x = var_1221_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1240_to_fp16 = const()[name = string("op_1240_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261784832)))];
            tensor<fp16, [1024]> var_1241_to_fp16 = const()[name = string("op_1241_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263882048)))];
            tensor<fp16, [1, ?, 1024]> linear_42_cast_fp16 = linear(bias = var_1241_to_fp16, weight = var_1240_to_fp16, x = var_1221_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [3]> var_1243_shape_cast_fp16 = shape(x = linear_40_cast_fp16)[name = string("op_1243_shape_cast_fp16")];
            int32 gather_62_axis_0 = const()[name = string("gather_62_axis_0"), val = int32(0)];
            int32 gather_62_batch_dims_0 = const()[name = string("gather_62_batch_dims_0"), val = int32(0)];
            bool gather_62_validate_indices_0 = const()[name = string("gather_62_validate_indices_0"), val = bool(false)];
            string var_1243_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1243_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_62_to_uint16 = const()[name = string("select_62_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1243_shape_cast_fp16_to_uint16 = cast(dtype = var_1243_shape_cast_fp16_to_uint16_dtype_0, x = var_1243_shape_cast_fp16)[name = string("cast_284")];
            uint16 gather_62_cast_uint16 = gather(axis = gather_62_axis_0, batch_dims = gather_62_batch_dims_0, indices = select_62_to_uint16, validate_indices = gather_62_validate_indices_0, x = var_1243_shape_cast_fp16_to_uint16)[name = string("gather_62_cast_uint16")];
            string gather_62_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_62_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_62_cast_uint16_to_int32 = cast(dtype = gather_62_cast_uint16_to_int32_dtype_0, x = gather_62_cast_uint16)[name = string("cast_283")];
            int32 end_step_13 = add(x = offset, y = gather_62_cast_uint16_to_int32)[name = string("end_step_13")];
            tensor<int32, [1]> expand_dims_80 = const()[name = string("expand_dims_80"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_82 = const()[name = string("expand_dims_82"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_83_axes_0 = const()[name = string("expand_dims_83_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_83 = expand_dims(axes = expand_dims_83_axes_0, x = end_step_13)[name = string("expand_dims_83")];
            tensor<int32, [1]> concat_114_values0_0 = const()[name = string("concat_114_values0_0"), val = tensor<int32, [1]>([5])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (concat_114_values0_0, expand_dims_80, expand_dims_1, expand_dims_82))[name = string("concat_114")];
            tensor<int32, [1]> concat_115_values0_0 = const()[name = string("concat_115_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_115_values1_0 = const()[name = string("concat_115_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_115_values3_0 = const()[name = string("concat_115_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_115_axis_0 = const()[name = string("concat_115_axis_0"), val = int32(0)];
            bool concat_115_interleave_0 = const()[name = string("concat_115_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_115 = concat(axis = concat_115_axis_0, interleave = concat_115_interleave_0, values = (concat_115_values0_0, concat_115_values1_0, expand_dims_83, concat_115_values3_0))[name = string("concat_115")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_6_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_6_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_6_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_114, begin_mask = k_cache1_internal_tensor_assign_6_begin_mask_0, end = concat_115, end_mask = k_cache1_internal_tensor_assign_6_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_6_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_6_stride_0, update = linear_41_cast_fp16, x = coreml_update_state_56)[name = string("k_cache1_internal_tensor_assign_6_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_6_cast_fp16, input = k_cache1)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_58 = read_state(input = k_cache1)[name = string("coreml_update_state_58")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_6_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_6_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_6_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_114, begin_mask = v_cache1_internal_tensor_assign_6_begin_mask_0, end = concat_115, end_mask = v_cache1_internal_tensor_assign_6_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_6_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_6_stride_0, update = linear_42_cast_fp16, x = coreml_update_state_57)[name = string("v_cache1_internal_tensor_assign_6_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_6_cast_fp16, input = v_cache1)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_59 = read_state(input = v_cache1)[name = string("coreml_update_state_59")];
            int32 concat_120_values0_0 = const()[name = string("concat_120_values0_0"), val = int32(1)];
            int32 concat_120_values2_0 = const()[name = string("concat_120_values2_0"), val = int32(1024)];
            int32 concat_120_axis_0 = const()[name = string("concat_120_axis_0"), val = int32(0)];
            bool concat_120_interleave_0 = const()[name = string("concat_120_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_120 = concat(axis = concat_120_axis_0, interleave = concat_120_interleave_0, values = (concat_120_values0_0, end_step_13, concat_120_values2_0))[name = string("concat_120")];
            tensor<int32, [3]> var_1259_begin_0 = const()[name = string("op_1259_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1259_end_mask_0 = const()[name = string("op_1259_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1259_cast_fp16 = slice_by_index(begin = var_1259_begin_0, end = concat_120, end_mask = var_1259_end_mask_0, x = k_cache_21_cast_fp16)[name = string("op_1259_cast_fp16")];
            tensor<int32, [3]> var_1262_begin_0 = const()[name = string("op_1262_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1262_end_mask_0 = const()[name = string("op_1262_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = concat_120, end_mask = var_1262_end_mask_0, x = v_cache_21_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1272_cast_fp16 = reshape(shape = concat_122x, x = linear_40_cast_fp16)[name = string("op_1272_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_140_to_fp16 = const()[name = string("const_140_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_43_cast_fp16 = mul(x = var_1272_cast_fp16, y = const_140_to_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> concat_123x = const()[name = string("concat_123x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1279_cast_fp16 = reshape(shape = concat_123x, x = var_1259_cast_fp16)[name = string("op_1279_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_141_to_fp16 = const()[name = string("const_141_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_55_cast_fp16 = mul(x = var_1279_cast_fp16, y = const_141_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<int32, [4]> concat_124x = const()[name = string("concat_124x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1286_cast_fp16 = reshape(shape = concat_124x, x = var_1262_cast_fp16)[name = string("op_1286_cast_fp16")];
            tensor<int32, [4]> var_1287 = const()[name = string("op_1287"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_31_transpose_x_0 = const()[name = string("qk_31_transpose_x_0"), val = bool(false)];
            bool qk_31_transpose_y_0 = const()[name = string("qk_31_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_213_perm_0 = const()[name = string("transpose_213_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_214_perm_0 = const()[name = string("transpose_214_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_214 = transpose(perm = transpose_214_perm_0, x = k_55_cast_fp16)[name = string("transpose_438")];
            tensor<fp16, [1, 16, ?, 64]> transpose_213 = transpose(perm = transpose_213_perm_0, x = q_43_cast_fp16)[name = string("transpose_439")];
            tensor<fp16, [1, 16, ?, ?]> qk_31_cast_fp16 = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_213, y = transpose_214)[name = string("qk_31_cast_fp16")];
            int32 concat_125_values1_0 = const()[name = string("concat_125_values1_0"), val = int32(448)];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (gather_62_cast_uint16_to_int32, concat_125_values1_0))[name = string("concat_125")];
            tensor<int32, [2]> var_1290_begin_0 = const()[name = string("op_1290_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1290_end_mask_0 = const()[name = string("op_1290_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_1290_cast_fp16 = slice_by_index(begin = var_1290_begin_0, end = concat_125, end_mask = var_1290_end_mask_0, x = mask_to_fp16)[name = string("op_1290_cast_fp16")];
            int32 concat_126_values0_0 = const()[name = string("concat_126_values0_0"), val = int32(0)];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (concat_126_values0_0, gather_62_cast_uint16_to_int32))[name = string("concat_126")];
            tensor<int32, [2]> var_1291_begin_0 = const()[name = string("op_1291_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1291_end_mask_0 = const()[name = string("op_1291_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_1291_cast_fp16 = slice_by_index(begin = var_1291_begin_0, end = concat_126, end_mask = var_1291_end_mask_0, x = var_1290_cast_fp16)[name = string("op_1291_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_33_cast_fp16 = add(x = qk_31_cast_fp16, y = var_1291_cast_fp16)[name = string("qk_33_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_1294_cast_fp16 = softmax(axis = var_1203, x = qk_33_cast_fp16)[name = string("op_1294_cast_fp16")];
            bool var_1296_transpose_x_0 = const()[name = string("op_1296_transpose_x_0"), val = bool(false)];
            bool var_1296_transpose_y_0 = const()[name = string("op_1296_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_55_cast_fp16 = transpose(perm = var_1287, x = var_1286_cast_fp16)[name = string("transpose_440")];
            tensor<fp16, [1, 16, ?, 64]> var_1296_cast_fp16 = matmul(transpose_x = var_1296_transpose_x_0, transpose_y = var_1296_transpose_y_0, x = var_1294_cast_fp16, y = v_55_cast_fp16)[name = string("op_1296_cast_fp16")];
            tensor<int32, [4]> var_1297 = const()[name = string("op_1297"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_127x = const()[name = string("concat_127x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1298_cast_fp16 = transpose(perm = var_1297, x = var_1296_cast_fp16)[name = string("transpose_437")];
            tensor<fp16, [1, ?, 1024]> x_97_cast_fp16 = reshape(shape = concat_127x, x = var_1298_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1302_to_fp16 = const()[name = string("op_1302_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263884160)))];
            tensor<fp16, [1024]> var_1303_to_fp16 = const()[name = string("op_1303_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265981376)))];
            tensor<fp16, [1, ?, 1024]> linear_43_cast_fp16 = linear(bias = var_1303_to_fp16, weight = var_1302_to_fp16, x = x_97_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_99_cast_fp16 = add(x = x_93_cast_fp16, y = linear_43_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_5_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_5_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265983488)))];
            tensor<fp16, [1024]> blocks_5_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_5_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265985600)))];
            tensor<fp16, [1, ?, 1024]> var_1310_cast_fp16 = layer_norm(axes = var_1310_axes_0, beta = blocks_5_cross_attn_ln_bias_to_fp16, epsilon = var_1209_to_fp16, gamma = blocks_5_cross_attn_ln_weight_to_fp16, x = x_99_cast_fp16)[name = string("op_1310_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1319_to_fp16 = const()[name = string("op_1319_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265987712)))];
            tensor<fp16, [1024]> var_1320_to_fp16 = const()[name = string("op_1320_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268084928)))];
            tensor<fp16, [1, ?, 1024]> linear_44_cast_fp16 = linear(bias = var_1320_to_fp16, weight = var_1319_to_fp16, x = var_1310_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> concat_128 = const()[name = string("concat_128"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_129 = const()[name = string("concat_129"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_57_internal_tensor_assign_1_stride_0 = const()[name = string("k_57_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_57_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_57_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_57_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_57_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_57_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_57_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_57_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_128, begin_mask = k_57_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = k_57_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_57_internal_tensor_assign_1_squeeze_mask_0, stride = k_57_internal_tensor_assign_1_stride_0, update = k_cache_23_cast_fp16, x = k_7_to_fp16)[name = string("k_57_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_130 = const()[name = string("concat_130"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_131 = const()[name = string("concat_131"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_57_internal_tensor_assign_1_stride_0 = const()[name = string("v_57_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_57_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_57_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_57_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_57_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_57_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_57_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_57_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_130, begin_mask = v_57_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = v_57_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_57_internal_tensor_assign_1_squeeze_mask_0, stride = v_57_internal_tensor_assign_1_stride_0, update = v_cache_23_cast_fp16, x = k_7_to_fp16)[name = string("v_57_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1340_cast_fp16 = reshape(shape = concat_132x, x = linear_44_cast_fp16)[name = string("op_1340_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_142_to_fp16 = const()[name = string("const_142_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_47_cast_fp16 = mul(x = var_1340_cast_fp16, y = const_142_to_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> var_1346 = const()[name = string("op_1346"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1347_cast_fp16 = reshape(shape = var_1346, x = k_57_internal_tensor_assign_1_cast_fp16)[name = string("op_1347_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_143_to_fp16 = const()[name = string("const_143_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_59_cast_fp16 = mul(x = var_1347_cast_fp16, y = const_143_to_fp16)[name = string("k_59_cast_fp16")];
            tensor<int32, [4]> var_1353 = const()[name = string("op_1353"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1354_cast_fp16 = reshape(shape = var_1353, x = v_57_internal_tensor_assign_1_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<int32, [4]> var_1355 = const()[name = string("op_1355"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_35_transpose_x_0 = const()[name = string("qk_35_transpose_x_0"), val = bool(false)];
            bool qk_35_transpose_y_0 = const()[name = string("qk_35_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_215_perm_0 = const()[name = string("transpose_215_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_216_perm_0 = const()[name = string("transpose_216_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_216 = transpose(perm = transpose_216_perm_0, x = k_59_cast_fp16)[name = string("transpose_434")];
            tensor<fp16, [1, 16, ?, 64]> transpose_215 = transpose(perm = transpose_215_perm_0, x = q_47_cast_fp16)[name = string("transpose_435")];
            tensor<fp16, [1, 16, ?, 1500]> qk_35_cast_fp16 = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_215, y = transpose_216)[name = string("qk_35_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_1359_cast_fp16 = softmax(axis = var_1203, x = qk_35_cast_fp16)[name = string("op_1359_cast_fp16")];
            bool var_1361_transpose_x_0 = const()[name = string("op_1361_transpose_x_0"), val = bool(false)];
            bool var_1361_transpose_y_0 = const()[name = string("op_1361_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_59_cast_fp16 = transpose(perm = var_1355, x = var_1354_cast_fp16)[name = string("transpose_436")];
            tensor<fp16, [1, 16, ?, 64]> var_1361_cast_fp16 = matmul(transpose_x = var_1361_transpose_x_0, transpose_y = var_1361_transpose_y_0, x = var_1359_cast_fp16, y = v_59_cast_fp16)[name = string("op_1361_cast_fp16")];
            tensor<int32, [4]> var_1362 = const()[name = string("op_1362"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1363_cast_fp16 = transpose(perm = var_1362, x = var_1361_cast_fp16)[name = string("transpose_433")];
            tensor<fp16, [1, ?, 1024]> x_103_cast_fp16 = reshape(shape = concat_133x, x = var_1363_cast_fp16)[name = string("x_103_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1367_to_fp16 = const()[name = string("op_1367_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268087040)))];
            tensor<fp16, [1024]> var_1368_to_fp16 = const()[name = string("op_1368_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270184256)))];
            tensor<fp16, [1, ?, 1024]> linear_45_cast_fp16 = linear(bias = var_1368_to_fp16, weight = var_1367_to_fp16, x = x_103_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_105_cast_fp16 = add(x = x_99_cast_fp16, y = linear_45_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [1]> var_1375_axes_0 = const()[name = string("op_1375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_5_mlp_ln_weight_to_fp16 = const()[name = string("blocks_5_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270186368)))];
            tensor<fp16, [1024]> blocks_5_mlp_ln_bias_to_fp16 = const()[name = string("blocks_5_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270188480)))];
            tensor<fp16, [1, ?, 1024]> var_1375_cast_fp16 = layer_norm(axes = var_1375_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_1209_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_105_cast_fp16)[name = string("op_1375_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_1384_to_fp16 = const()[name = string("op_1384_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270190592)))];
            tensor<fp16, [4096]> var_1385_to_fp16 = const()[name = string("op_1385_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278579264)))];
            tensor<fp16, [1, ?, 4096]> linear_46_cast_fp16 = linear(bias = var_1385_to_fp16, weight = var_1384_to_fp16, x = var_1375_cast_fp16)[name = string("linear_46_cast_fp16")];
            string x_109_mode_0 = const()[name = string("x_109_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_109_cast_fp16 = gelu(mode = x_109_mode_0, x = linear_46_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_1390_to_fp16 = const()[name = string("op_1390_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278587520)))];
            tensor<fp16, [1024]> var_1391_to_fp16 = const()[name = string("op_1391_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286976192)))];
            tensor<fp16, [1, ?, 1024]> linear_47_cast_fp16 = linear(bias = var_1391_to_fp16, weight = var_1390_to_fp16, x = x_109_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_111_cast_fp16 = add(x = x_105_cast_fp16, y = linear_47_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> k_cache_25_begin_0 = const()[name = string("k_cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> k_cache_25_end_0 = const()[name = string("k_cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_25_end_mask_0 = const()[name = string("k_cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_25_squeeze_mask_0 = const()[name = string("k_cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_25_cast_fp16 = slice_by_index(begin = k_cache_25_begin_0, end = k_cache_25_end_0, end_mask = k_cache_25_end_mask_0, squeeze_mask = k_cache_25_squeeze_mask_0, x = coreml_update_state_58)[name = string("k_cache_25_cast_fp16")];
            tensor<int32, [4]> v_cache_25_begin_0 = const()[name = string("v_cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> v_cache_25_end_0 = const()[name = string("v_cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_25_end_mask_0 = const()[name = string("v_cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_25_squeeze_mask_0 = const()[name = string("v_cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_25_cast_fp16 = slice_by_index(begin = v_cache_25_begin_0, end = v_cache_25_end_0, end_mask = v_cache_25_end_mask_0, squeeze_mask = v_cache_25_squeeze_mask_0, x = coreml_update_state_59)[name = string("v_cache_25_cast_fp16")];
            tensor<int32, [4]> k_cache_27_begin_0 = const()[name = string("k_cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> k_cache_27_end_0 = const()[name = string("k_cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_27_end_mask_0 = const()[name = string("k_cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_27_squeeze_mask_0 = const()[name = string("k_cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_27_cast_fp16 = slice_by_index(begin = k_cache_27_begin_0, end = k_cache_27_end_0, end_mask = k_cache_27_end_mask_0, squeeze_mask = k_cache_27_squeeze_mask_0, x = read_state_2)[name = string("k_cache_27_cast_fp16")];
            tensor<int32, [4]> v_cache_27_begin_0 = const()[name = string("v_cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> v_cache_27_end_0 = const()[name = string("v_cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_27_end_mask_0 = const()[name = string("v_cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_27_squeeze_mask_0 = const()[name = string("v_cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_27_cast_fp16 = slice_by_index(begin = v_cache_27_begin_0, end = v_cache_27_end_0, end_mask = v_cache_27_end_mask_0, squeeze_mask = v_cache_27_squeeze_mask_0, x = read_state_3)[name = string("v_cache_27_cast_fp16")];
            int32 var_1414 = const()[name = string("op_1414"), val = int32(-1)];
            tensor<int32, [1]> var_1432_axes_0 = const()[name = string("op_1432_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_6_attn_ln_weight_to_fp16 = const()[name = string("blocks_6_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286978304)))];
            tensor<fp16, [1024]> blocks_6_attn_ln_bias_to_fp16 = const()[name = string("blocks_6_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286980416)))];
            fp16 var_1420_to_fp16 = const()[name = string("op_1420_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_1432_cast_fp16 = layer_norm(axes = var_1432_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_1420_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_111_cast_fp16)[name = string("op_1432_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1443_to_fp16 = const()[name = string("op_1443_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286982528)))];
            tensor<fp16, [1024]> var_1444_to_fp16 = const()[name = string("op_1444_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289079744)))];
            tensor<fp16, [1, ?, 1024]> linear_48_cast_fp16 = linear(bias = var_1444_to_fp16, weight = var_1443_to_fp16, x = var_1432_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289081856)))];
            tensor<fp16, [1, ?, 1024]> linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1447_to_fp16, x = var_1432_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1451_to_fp16 = const()[name = string("op_1451_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291179072)))];
            tensor<fp16, [1024]> var_1452_to_fp16 = const()[name = string("op_1452_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293276288)))];
            tensor<fp16, [1, ?, 1024]> linear_50_cast_fp16 = linear(bias = var_1452_to_fp16, weight = var_1451_to_fp16, x = var_1432_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [3]> var_1454_shape_cast_fp16 = shape(x = linear_48_cast_fp16)[name = string("op_1454_shape_cast_fp16")];
            int32 gather_74_axis_0 = const()[name = string("gather_74_axis_0"), val = int32(0)];
            int32 gather_74_batch_dims_0 = const()[name = string("gather_74_batch_dims_0"), val = int32(0)];
            bool gather_74_validate_indices_0 = const()[name = string("gather_74_validate_indices_0"), val = bool(false)];
            string var_1454_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1454_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_74_to_uint16 = const()[name = string("select_74_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1454_shape_cast_fp16_to_uint16 = cast(dtype = var_1454_shape_cast_fp16_to_uint16_dtype_0, x = var_1454_shape_cast_fp16)[name = string("cast_282")];
            uint16 gather_74_cast_uint16 = gather(axis = gather_74_axis_0, batch_dims = gather_74_batch_dims_0, indices = select_74_to_uint16, validate_indices = gather_74_validate_indices_0, x = var_1454_shape_cast_fp16_to_uint16)[name = string("gather_74_cast_uint16")];
            string gather_74_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_74_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_74_cast_uint16_to_int32 = cast(dtype = gather_74_cast_uint16_to_int32_dtype_0, x = gather_74_cast_uint16)[name = string("cast_281")];
            int32 end_step_15 = add(x = offset, y = gather_74_cast_uint16_to_int32)[name = string("end_step_15")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_98 = const()[name = string("expand_dims_98"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99_axes_0 = const()[name = string("expand_dims_99_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = expand_dims(axes = expand_dims_99_axes_0, x = end_step_15)[name = string("expand_dims_99")];
            tensor<int32, [1]> concat_136_values0_0 = const()[name = string("concat_136_values0_0"), val = tensor<int32, [1]>([6])];
            int32 concat_136_axis_0 = const()[name = string("concat_136_axis_0"), val = int32(0)];
            bool concat_136_interleave_0 = const()[name = string("concat_136_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_136 = concat(axis = concat_136_axis_0, interleave = concat_136_interleave_0, values = (concat_136_values0_0, expand_dims_96, expand_dims_1, expand_dims_98))[name = string("concat_136")];
            tensor<int32, [1]> concat_137_values0_0 = const()[name = string("concat_137_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_137_values1_0 = const()[name = string("concat_137_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_137_values3_0 = const()[name = string("concat_137_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (concat_137_values0_0, concat_137_values1_0, expand_dims_99, concat_137_values3_0))[name = string("concat_137")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_7_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_7_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_7_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_136, begin_mask = k_cache1_internal_tensor_assign_7_begin_mask_0, end = concat_137, end_mask = k_cache1_internal_tensor_assign_7_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_7_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_7_stride_0, update = linear_49_cast_fp16, x = coreml_update_state_58)[name = string("k_cache1_internal_tensor_assign_7_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_7_cast_fp16, input = k_cache1)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_60 = read_state(input = k_cache1)[name = string("coreml_update_state_60")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_7_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_7_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_7_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_136, begin_mask = v_cache1_internal_tensor_assign_7_begin_mask_0, end = concat_137, end_mask = v_cache1_internal_tensor_assign_7_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_7_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_7_stride_0, update = linear_50_cast_fp16, x = coreml_update_state_59)[name = string("v_cache1_internal_tensor_assign_7_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_7_cast_fp16, input = v_cache1)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_61 = read_state(input = v_cache1)[name = string("coreml_update_state_61")];
            int32 concat_142_values0_0 = const()[name = string("concat_142_values0_0"), val = int32(1)];
            int32 concat_142_values2_0 = const()[name = string("concat_142_values2_0"), val = int32(1024)];
            int32 concat_142_axis_0 = const()[name = string("concat_142_axis_0"), val = int32(0)];
            bool concat_142_interleave_0 = const()[name = string("concat_142_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_142 = concat(axis = concat_142_axis_0, interleave = concat_142_interleave_0, values = (concat_142_values0_0, end_step_15, concat_142_values2_0))[name = string("concat_142")];
            tensor<int32, [3]> var_1470_begin_0 = const()[name = string("op_1470_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1470_end_mask_0 = const()[name = string("op_1470_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1470_cast_fp16 = slice_by_index(begin = var_1470_begin_0, end = concat_142, end_mask = var_1470_end_mask_0, x = k_cache_25_cast_fp16)[name = string("op_1470_cast_fp16")];
            tensor<int32, [3]> var_1473_begin_0 = const()[name = string("op_1473_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1473_end_mask_0 = const()[name = string("op_1473_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1473_cast_fp16 = slice_by_index(begin = var_1473_begin_0, end = concat_142, end_mask = var_1473_end_mask_0, x = v_cache_25_cast_fp16)[name = string("op_1473_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1483_cast_fp16 = reshape(shape = concat_144x, x = linear_48_cast_fp16)[name = string("op_1483_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_51_cast_fp16 = mul(x = var_1483_cast_fp16, y = const_144_to_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1490_cast_fp16 = reshape(shape = concat_145x, x = var_1470_cast_fp16)[name = string("op_1490_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_145_to_fp16 = const()[name = string("const_145_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_65_cast_fp16 = mul(x = var_1490_cast_fp16, y = const_145_to_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1497_cast_fp16 = reshape(shape = concat_146x, x = var_1473_cast_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [4]> var_1498 = const()[name = string("op_1498"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_37_transpose_x_0 = const()[name = string("qk_37_transpose_x_0"), val = bool(false)];
            bool qk_37_transpose_y_0 = const()[name = string("qk_37_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_217_perm_0 = const()[name = string("transpose_217_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_218_perm_0 = const()[name = string("transpose_218_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_218 = transpose(perm = transpose_218_perm_0, x = k_65_cast_fp16)[name = string("transpose_430")];
            tensor<fp16, [1, 16, ?, 64]> transpose_217 = transpose(perm = transpose_217_perm_0, x = q_51_cast_fp16)[name = string("transpose_431")];
            tensor<fp16, [1, 16, ?, ?]> qk_37_cast_fp16 = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_217, y = transpose_218)[name = string("qk_37_cast_fp16")];
            int32 concat_147_values1_0 = const()[name = string("concat_147_values1_0"), val = int32(448)];
            int32 concat_147_axis_0 = const()[name = string("concat_147_axis_0"), val = int32(0)];
            bool concat_147_interleave_0 = const()[name = string("concat_147_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_147 = concat(axis = concat_147_axis_0, interleave = concat_147_interleave_0, values = (gather_74_cast_uint16_to_int32, concat_147_values1_0))[name = string("concat_147")];
            tensor<int32, [2]> var_1501_begin_0 = const()[name = string("op_1501_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1501_end_mask_0 = const()[name = string("op_1501_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_1501_cast_fp16 = slice_by_index(begin = var_1501_begin_0, end = concat_147, end_mask = var_1501_end_mask_0, x = mask_to_fp16)[name = string("op_1501_cast_fp16")];
            int32 concat_148_values0_0 = const()[name = string("concat_148_values0_0"), val = int32(0)];
            int32 concat_148_axis_0 = const()[name = string("concat_148_axis_0"), val = int32(0)];
            bool concat_148_interleave_0 = const()[name = string("concat_148_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_148 = concat(axis = concat_148_axis_0, interleave = concat_148_interleave_0, values = (concat_148_values0_0, gather_74_cast_uint16_to_int32))[name = string("concat_148")];
            tensor<int32, [2]> var_1502_begin_0 = const()[name = string("op_1502_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1502_end_mask_0 = const()[name = string("op_1502_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_1502_cast_fp16 = slice_by_index(begin = var_1502_begin_0, end = concat_148, end_mask = var_1502_end_mask_0, x = var_1501_cast_fp16)[name = string("op_1502_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_39_cast_fp16 = add(x = qk_37_cast_fp16, y = var_1502_cast_fp16)[name = string("qk_39_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_1505_cast_fp16 = softmax(axis = var_1414, x = qk_39_cast_fp16)[name = string("op_1505_cast_fp16")];
            bool var_1507_transpose_x_0 = const()[name = string("op_1507_transpose_x_0"), val = bool(false)];
            bool var_1507_transpose_y_0 = const()[name = string("op_1507_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_65_cast_fp16 = transpose(perm = var_1498, x = var_1497_cast_fp16)[name = string("transpose_432")];
            tensor<fp16, [1, 16, ?, 64]> var_1507_cast_fp16 = matmul(transpose_x = var_1507_transpose_x_0, transpose_y = var_1507_transpose_y_0, x = var_1505_cast_fp16, y = v_65_cast_fp16)[name = string("op_1507_cast_fp16")];
            tensor<int32, [4]> var_1508 = const()[name = string("op_1508"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_149x = const()[name = string("concat_149x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1509_cast_fp16 = transpose(perm = var_1508, x = var_1507_cast_fp16)[name = string("transpose_429")];
            tensor<fp16, [1, ?, 1024]> x_115_cast_fp16 = reshape(shape = concat_149x, x = var_1509_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1513_to_fp16 = const()[name = string("op_1513_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293278400)))];
            tensor<fp16, [1024]> var_1514_to_fp16 = const()[name = string("op_1514_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295375616)))];
            tensor<fp16, [1, ?, 1024]> linear_51_cast_fp16 = linear(bias = var_1514_to_fp16, weight = var_1513_to_fp16, x = x_115_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_117_cast_fp16 = add(x = x_111_cast_fp16, y = linear_51_cast_fp16)[name = string("x_117_cast_fp16")];
            tensor<int32, [1]> var_1521_axes_0 = const()[name = string("op_1521_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_6_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_6_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295377728)))];
            tensor<fp16, [1024]> blocks_6_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_6_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295379840)))];
            tensor<fp16, [1, ?, 1024]> var_1521_cast_fp16 = layer_norm(axes = var_1521_axes_0, beta = blocks_6_cross_attn_ln_bias_to_fp16, epsilon = var_1420_to_fp16, gamma = blocks_6_cross_attn_ln_weight_to_fp16, x = x_117_cast_fp16)[name = string("op_1521_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1530_to_fp16 = const()[name = string("op_1530_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295381952)))];
            tensor<fp16, [1024]> var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297479168)))];
            tensor<fp16, [1, ?, 1024]> linear_52_cast_fp16 = linear(bias = var_1531_to_fp16, weight = var_1530_to_fp16, x = var_1521_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [3]> concat_150 = const()[name = string("concat_150"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_151 = const()[name = string("concat_151"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_67_internal_tensor_assign_1_stride_0 = const()[name = string("k_67_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_67_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_67_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_67_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_67_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_67_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_67_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_67_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_150, begin_mask = k_67_internal_tensor_assign_1_begin_mask_0, end = concat_151, end_mask = k_67_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_67_internal_tensor_assign_1_squeeze_mask_0, stride = k_67_internal_tensor_assign_1_stride_0, update = k_cache_27_cast_fp16, x = k_7_to_fp16)[name = string("k_67_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_152 = const()[name = string("concat_152"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_153 = const()[name = string("concat_153"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_67_internal_tensor_assign_1_stride_0 = const()[name = string("v_67_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_67_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_67_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_67_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_67_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_67_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_67_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_67_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_152, begin_mask = v_67_internal_tensor_assign_1_begin_mask_0, end = concat_153, end_mask = v_67_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_67_internal_tensor_assign_1_squeeze_mask_0, stride = v_67_internal_tensor_assign_1_stride_0, update = v_cache_27_cast_fp16, x = k_7_to_fp16)[name = string("v_67_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_154x = const()[name = string("concat_154x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1551_cast_fp16 = reshape(shape = concat_154x, x = linear_52_cast_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_146_to_fp16 = const()[name = string("const_146_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_55_cast_fp16 = mul(x = var_1551_cast_fp16, y = const_146_to_fp16)[name = string("q_55_cast_fp16")];
            tensor<int32, [4]> var_1557 = const()[name = string("op_1557"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1558_cast_fp16 = reshape(shape = var_1557, x = k_67_internal_tensor_assign_1_cast_fp16)[name = string("op_1558_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_147_to_fp16 = const()[name = string("const_147_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_69_cast_fp16 = mul(x = var_1558_cast_fp16, y = const_147_to_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> var_1564 = const()[name = string("op_1564"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1565_cast_fp16 = reshape(shape = var_1564, x = v_67_internal_tensor_assign_1_cast_fp16)[name = string("op_1565_cast_fp16")];
            tensor<int32, [4]> var_1566 = const()[name = string("op_1566"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_41_transpose_x_0 = const()[name = string("qk_41_transpose_x_0"), val = bool(false)];
            bool qk_41_transpose_y_0 = const()[name = string("qk_41_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_219_perm_0 = const()[name = string("transpose_219_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_220_perm_0 = const()[name = string("transpose_220_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_220 = transpose(perm = transpose_220_perm_0, x = k_69_cast_fp16)[name = string("transpose_426")];
            tensor<fp16, [1, 16, ?, 64]> transpose_219 = transpose(perm = transpose_219_perm_0, x = q_55_cast_fp16)[name = string("transpose_427")];
            tensor<fp16, [1, 16, ?, 1500]> qk_41_cast_fp16 = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_219, y = transpose_220)[name = string("qk_41_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_1570_cast_fp16 = softmax(axis = var_1414, x = qk_41_cast_fp16)[name = string("op_1570_cast_fp16")];
            bool var_1572_transpose_x_0 = const()[name = string("op_1572_transpose_x_0"), val = bool(false)];
            bool var_1572_transpose_y_0 = const()[name = string("op_1572_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_69_cast_fp16 = transpose(perm = var_1566, x = var_1565_cast_fp16)[name = string("transpose_428")];
            tensor<fp16, [1, 16, ?, 64]> var_1572_cast_fp16 = matmul(transpose_x = var_1572_transpose_x_0, transpose_y = var_1572_transpose_y_0, x = var_1570_cast_fp16, y = v_69_cast_fp16)[name = string("op_1572_cast_fp16")];
            tensor<int32, [4]> var_1573 = const()[name = string("op_1573"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1574_cast_fp16 = transpose(perm = var_1573, x = var_1572_cast_fp16)[name = string("transpose_425")];
            tensor<fp16, [1, ?, 1024]> x_121_cast_fp16 = reshape(shape = concat_155x, x = var_1574_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1578_to_fp16 = const()[name = string("op_1578_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297481280)))];
            tensor<fp16, [1024]> var_1579_to_fp16 = const()[name = string("op_1579_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299578496)))];
            tensor<fp16, [1, ?, 1024]> linear_53_cast_fp16 = linear(bias = var_1579_to_fp16, weight = var_1578_to_fp16, x = x_121_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_123_cast_fp16 = add(x = x_117_cast_fp16, y = linear_53_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [1]> var_1586_axes_0 = const()[name = string("op_1586_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_6_mlp_ln_weight_to_fp16 = const()[name = string("blocks_6_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299580608)))];
            tensor<fp16, [1024]> blocks_6_mlp_ln_bias_to_fp16 = const()[name = string("blocks_6_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299582720)))];
            tensor<fp16, [1, ?, 1024]> var_1586_cast_fp16 = layer_norm(axes = var_1586_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_1420_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_123_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_1595_to_fp16 = const()[name = string("op_1595_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299584832)))];
            tensor<fp16, [4096]> var_1596_to_fp16 = const()[name = string("op_1596_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307973504)))];
            tensor<fp16, [1, ?, 4096]> linear_54_cast_fp16 = linear(bias = var_1596_to_fp16, weight = var_1595_to_fp16, x = var_1586_cast_fp16)[name = string("linear_54_cast_fp16")];
            string x_127_mode_0 = const()[name = string("x_127_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_127_cast_fp16 = gelu(mode = x_127_mode_0, x = linear_54_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_1601_to_fp16 = const()[name = string("op_1601_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307981760)))];
            tensor<fp16, [1024]> var_1602_to_fp16 = const()[name = string("op_1602_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316370432)))];
            tensor<fp16, [1, ?, 1024]> linear_55_cast_fp16 = linear(bias = var_1602_to_fp16, weight = var_1601_to_fp16, x = x_127_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_129_cast_fp16 = add(x = x_123_cast_fp16, y = linear_55_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [4]> k_cache_29_begin_0 = const()[name = string("k_cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> k_cache_29_end_0 = const()[name = string("k_cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_29_end_mask_0 = const()[name = string("k_cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_29_squeeze_mask_0 = const()[name = string("k_cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_29_cast_fp16 = slice_by_index(begin = k_cache_29_begin_0, end = k_cache_29_end_0, end_mask = k_cache_29_end_mask_0, squeeze_mask = k_cache_29_squeeze_mask_0, x = coreml_update_state_60)[name = string("k_cache_29_cast_fp16")];
            tensor<int32, [4]> v_cache_29_begin_0 = const()[name = string("v_cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> v_cache_29_end_0 = const()[name = string("v_cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_29_end_mask_0 = const()[name = string("v_cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_29_squeeze_mask_0 = const()[name = string("v_cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_29_cast_fp16 = slice_by_index(begin = v_cache_29_begin_0, end = v_cache_29_end_0, end_mask = v_cache_29_end_mask_0, squeeze_mask = v_cache_29_squeeze_mask_0, x = coreml_update_state_61)[name = string("v_cache_29_cast_fp16")];
            tensor<int32, [4]> k_cache_31_begin_0 = const()[name = string("k_cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> k_cache_31_end_0 = const()[name = string("k_cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_31_end_mask_0 = const()[name = string("k_cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_31_squeeze_mask_0 = const()[name = string("k_cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_31_cast_fp16 = slice_by_index(begin = k_cache_31_begin_0, end = k_cache_31_end_0, end_mask = k_cache_31_end_mask_0, squeeze_mask = k_cache_31_squeeze_mask_0, x = read_state_2)[name = string("k_cache_31_cast_fp16")];
            tensor<int32, [4]> v_cache_31_begin_0 = const()[name = string("v_cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> v_cache_31_end_0 = const()[name = string("v_cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_31_end_mask_0 = const()[name = string("v_cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_31_squeeze_mask_0 = const()[name = string("v_cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_31_cast_fp16 = slice_by_index(begin = v_cache_31_begin_0, end = v_cache_31_end_0, end_mask = v_cache_31_end_mask_0, squeeze_mask = v_cache_31_squeeze_mask_0, x = read_state_3)[name = string("v_cache_31_cast_fp16")];
            int32 var_1625 = const()[name = string("op_1625"), val = int32(-1)];
            tensor<int32, [1]> var_1643_axes_0 = const()[name = string("op_1643_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_7_attn_ln_weight_to_fp16 = const()[name = string("blocks_7_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316372544)))];
            tensor<fp16, [1024]> blocks_7_attn_ln_bias_to_fp16 = const()[name = string("blocks_7_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316374656)))];
            fp16 var_1631_to_fp16 = const()[name = string("op_1631_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_1643_cast_fp16 = layer_norm(axes = var_1643_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_1631_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_129_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1654_to_fp16 = const()[name = string("op_1654_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316376768)))];
            tensor<fp16, [1024]> var_1655_to_fp16 = const()[name = string("op_1655_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318473984)))];
            tensor<fp16, [1, ?, 1024]> linear_56_cast_fp16 = linear(bias = var_1655_to_fp16, weight = var_1654_to_fp16, x = var_1643_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1658_to_fp16 = const()[name = string("op_1658_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318476096)))];
            tensor<fp16, [1, ?, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1658_to_fp16, x = var_1643_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1662_to_fp16 = const()[name = string("op_1662_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320573312)))];
            tensor<fp16, [1024]> var_1663_to_fp16 = const()[name = string("op_1663_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322670528)))];
            tensor<fp16, [1, ?, 1024]> linear_58_cast_fp16 = linear(bias = var_1663_to_fp16, weight = var_1662_to_fp16, x = var_1643_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [3]> var_1665_shape_cast_fp16 = shape(x = linear_56_cast_fp16)[name = string("op_1665_shape_cast_fp16")];
            int32 gather_86_axis_0 = const()[name = string("gather_86_axis_0"), val = int32(0)];
            int32 gather_86_batch_dims_0 = const()[name = string("gather_86_batch_dims_0"), val = int32(0)];
            bool gather_86_validate_indices_0 = const()[name = string("gather_86_validate_indices_0"), val = bool(false)];
            string var_1665_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1665_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_86_to_uint16 = const()[name = string("select_86_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1665_shape_cast_fp16_to_uint16 = cast(dtype = var_1665_shape_cast_fp16_to_uint16_dtype_0, x = var_1665_shape_cast_fp16)[name = string("cast_280")];
            uint16 gather_86_cast_uint16 = gather(axis = gather_86_axis_0, batch_dims = gather_86_batch_dims_0, indices = select_86_to_uint16, validate_indices = gather_86_validate_indices_0, x = var_1665_shape_cast_fp16_to_uint16)[name = string("gather_86_cast_uint16")];
            string gather_86_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_86_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_86_cast_uint16_to_int32 = cast(dtype = gather_86_cast_uint16_to_int32_dtype_0, x = gather_86_cast_uint16)[name = string("cast_279")];
            int32 end_step_17 = add(x = offset, y = gather_86_cast_uint16_to_int32)[name = string("end_step_17")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_114 = const()[name = string("expand_dims_114"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_115_axes_0 = const()[name = string("expand_dims_115_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_115 = expand_dims(axes = expand_dims_115_axes_0, x = end_step_17)[name = string("expand_dims_115")];
            tensor<int32, [1]> concat_158_values0_0 = const()[name = string("concat_158_values0_0"), val = tensor<int32, [1]>([7])];
            int32 concat_158_axis_0 = const()[name = string("concat_158_axis_0"), val = int32(0)];
            bool concat_158_interleave_0 = const()[name = string("concat_158_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_158 = concat(axis = concat_158_axis_0, interleave = concat_158_interleave_0, values = (concat_158_values0_0, expand_dims_112, expand_dims_1, expand_dims_114))[name = string("concat_158")];
            tensor<int32, [1]> concat_159_values0_0 = const()[name = string("concat_159_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_159_values1_0 = const()[name = string("concat_159_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_159_values3_0 = const()[name = string("concat_159_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_159_axis_0 = const()[name = string("concat_159_axis_0"), val = int32(0)];
            bool concat_159_interleave_0 = const()[name = string("concat_159_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_159 = concat(axis = concat_159_axis_0, interleave = concat_159_interleave_0, values = (concat_159_values0_0, concat_159_values1_0, expand_dims_115, concat_159_values3_0))[name = string("concat_159")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_8_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_8_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_8_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_158, begin_mask = k_cache1_internal_tensor_assign_8_begin_mask_0, end = concat_159, end_mask = k_cache1_internal_tensor_assign_8_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_8_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_8_stride_0, update = linear_57_cast_fp16, x = coreml_update_state_60)[name = string("k_cache1_internal_tensor_assign_8_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_8_cast_fp16, input = k_cache1)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_62 = read_state(input = k_cache1)[name = string("coreml_update_state_62")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_8_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_8_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_8_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_158, begin_mask = v_cache1_internal_tensor_assign_8_begin_mask_0, end = concat_159, end_mask = v_cache1_internal_tensor_assign_8_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_8_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_8_stride_0, update = linear_58_cast_fp16, x = coreml_update_state_61)[name = string("v_cache1_internal_tensor_assign_8_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_8_cast_fp16, input = v_cache1)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_63 = read_state(input = v_cache1)[name = string("coreml_update_state_63")];
            int32 concat_164_values0_0 = const()[name = string("concat_164_values0_0"), val = int32(1)];
            int32 concat_164_values2_0 = const()[name = string("concat_164_values2_0"), val = int32(1024)];
            int32 concat_164_axis_0 = const()[name = string("concat_164_axis_0"), val = int32(0)];
            bool concat_164_interleave_0 = const()[name = string("concat_164_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_164 = concat(axis = concat_164_axis_0, interleave = concat_164_interleave_0, values = (concat_164_values0_0, end_step_17, concat_164_values2_0))[name = string("concat_164")];
            tensor<int32, [3]> var_1681_begin_0 = const()[name = string("op_1681_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1681_end_mask_0 = const()[name = string("op_1681_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1681_cast_fp16 = slice_by_index(begin = var_1681_begin_0, end = concat_164, end_mask = var_1681_end_mask_0, x = k_cache_29_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [3]> var_1684_begin_0 = const()[name = string("op_1684_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1684_end_mask_0 = const()[name = string("op_1684_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1684_cast_fp16 = slice_by_index(begin = var_1684_begin_0, end = concat_164, end_mask = var_1684_end_mask_0, x = v_cache_29_cast_fp16)[name = string("op_1684_cast_fp16")];
            tensor<int32, [4]> concat_166x = const()[name = string("concat_166x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1694_cast_fp16 = reshape(shape = concat_166x, x = linear_56_cast_fp16)[name = string("op_1694_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_148_to_fp16 = const()[name = string("const_148_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_59_cast_fp16 = mul(x = var_1694_cast_fp16, y = const_148_to_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1701_cast_fp16 = reshape(shape = concat_167x, x = var_1681_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_149_to_fp16 = const()[name = string("const_149_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_75_cast_fp16 = mul(x = var_1701_cast_fp16, y = const_149_to_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> concat_168x = const()[name = string("concat_168x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1708_cast_fp16 = reshape(shape = concat_168x, x = var_1684_cast_fp16)[name = string("op_1708_cast_fp16")];
            tensor<int32, [4]> var_1709 = const()[name = string("op_1709"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_43_transpose_x_0 = const()[name = string("qk_43_transpose_x_0"), val = bool(false)];
            bool qk_43_transpose_y_0 = const()[name = string("qk_43_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_221_perm_0 = const()[name = string("transpose_221_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_222_perm_0 = const()[name = string("transpose_222_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_222 = transpose(perm = transpose_222_perm_0, x = k_75_cast_fp16)[name = string("transpose_422")];
            tensor<fp16, [1, 16, ?, 64]> transpose_221 = transpose(perm = transpose_221_perm_0, x = q_59_cast_fp16)[name = string("transpose_423")];
            tensor<fp16, [1, 16, ?, ?]> qk_43_cast_fp16 = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_221, y = transpose_222)[name = string("qk_43_cast_fp16")];
            int32 concat_169_values1_0 = const()[name = string("concat_169_values1_0"), val = int32(448)];
            int32 concat_169_axis_0 = const()[name = string("concat_169_axis_0"), val = int32(0)];
            bool concat_169_interleave_0 = const()[name = string("concat_169_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_169 = concat(axis = concat_169_axis_0, interleave = concat_169_interleave_0, values = (gather_86_cast_uint16_to_int32, concat_169_values1_0))[name = string("concat_169")];
            tensor<int32, [2]> var_1712_begin_0 = const()[name = string("op_1712_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1712_end_mask_0 = const()[name = string("op_1712_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_1712_cast_fp16 = slice_by_index(begin = var_1712_begin_0, end = concat_169, end_mask = var_1712_end_mask_0, x = mask_to_fp16)[name = string("op_1712_cast_fp16")];
            int32 concat_170_values0_0 = const()[name = string("concat_170_values0_0"), val = int32(0)];
            int32 concat_170_axis_0 = const()[name = string("concat_170_axis_0"), val = int32(0)];
            bool concat_170_interleave_0 = const()[name = string("concat_170_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_170 = concat(axis = concat_170_axis_0, interleave = concat_170_interleave_0, values = (concat_170_values0_0, gather_86_cast_uint16_to_int32))[name = string("concat_170")];
            tensor<int32, [2]> var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_1713_cast_fp16 = slice_by_index(begin = var_1713_begin_0, end = concat_170, end_mask = var_1713_end_mask_0, x = var_1712_cast_fp16)[name = string("op_1713_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_45_cast_fp16 = add(x = qk_43_cast_fp16, y = var_1713_cast_fp16)[name = string("qk_45_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_1716_cast_fp16 = softmax(axis = var_1625, x = qk_45_cast_fp16)[name = string("op_1716_cast_fp16")];
            bool var_1718_transpose_x_0 = const()[name = string("op_1718_transpose_x_0"), val = bool(false)];
            bool var_1718_transpose_y_0 = const()[name = string("op_1718_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_75_cast_fp16 = transpose(perm = var_1709, x = var_1708_cast_fp16)[name = string("transpose_424")];
            tensor<fp16, [1, 16, ?, 64]> var_1718_cast_fp16 = matmul(transpose_x = var_1718_transpose_x_0, transpose_y = var_1718_transpose_y_0, x = var_1716_cast_fp16, y = v_75_cast_fp16)[name = string("op_1718_cast_fp16")];
            tensor<int32, [4]> var_1719 = const()[name = string("op_1719"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_171x = const()[name = string("concat_171x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1720_cast_fp16 = transpose(perm = var_1719, x = var_1718_cast_fp16)[name = string("transpose_421")];
            tensor<fp16, [1, ?, 1024]> x_133_cast_fp16 = reshape(shape = concat_171x, x = var_1720_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1724_to_fp16 = const()[name = string("op_1724_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322672640)))];
            tensor<fp16, [1024]> var_1725_to_fp16 = const()[name = string("op_1725_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324769856)))];
            tensor<fp16, [1, ?, 1024]> linear_59_cast_fp16 = linear(bias = var_1725_to_fp16, weight = var_1724_to_fp16, x = x_133_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_135_cast_fp16 = add(x = x_129_cast_fp16, y = linear_59_cast_fp16)[name = string("x_135_cast_fp16")];
            tensor<int32, [1]> var_1732_axes_0 = const()[name = string("op_1732_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_7_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_7_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324771968)))];
            tensor<fp16, [1024]> blocks_7_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_7_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324774080)))];
            tensor<fp16, [1, ?, 1024]> var_1732_cast_fp16 = layer_norm(axes = var_1732_axes_0, beta = blocks_7_cross_attn_ln_bias_to_fp16, epsilon = var_1631_to_fp16, gamma = blocks_7_cross_attn_ln_weight_to_fp16, x = x_135_cast_fp16)[name = string("op_1732_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1741_to_fp16 = const()[name = string("op_1741_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324776192)))];
            tensor<fp16, [1024]> var_1742_to_fp16 = const()[name = string("op_1742_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326873408)))];
            tensor<fp16, [1, ?, 1024]> linear_60_cast_fp16 = linear(bias = var_1742_to_fp16, weight = var_1741_to_fp16, x = var_1732_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [3]> concat_172 = const()[name = string("concat_172"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_173 = const()[name = string("concat_173"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_77_internal_tensor_assign_1_stride_0 = const()[name = string("k_77_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_77_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_77_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_77_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_77_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_77_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_77_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_77_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_172, begin_mask = k_77_internal_tensor_assign_1_begin_mask_0, end = concat_173, end_mask = k_77_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_77_internal_tensor_assign_1_squeeze_mask_0, stride = k_77_internal_tensor_assign_1_stride_0, update = k_cache_31_cast_fp16, x = k_7_to_fp16)[name = string("k_77_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_174 = const()[name = string("concat_174"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_175 = const()[name = string("concat_175"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_77_internal_tensor_assign_1_stride_0 = const()[name = string("v_77_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_77_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_77_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_77_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_77_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_77_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_77_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_77_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_174, begin_mask = v_77_internal_tensor_assign_1_begin_mask_0, end = concat_175, end_mask = v_77_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_77_internal_tensor_assign_1_squeeze_mask_0, stride = v_77_internal_tensor_assign_1_stride_0, update = v_cache_31_cast_fp16, x = k_7_to_fp16)[name = string("v_77_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_176x = const()[name = string("concat_176x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1762_cast_fp16 = reshape(shape = concat_176x, x = linear_60_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_150_to_fp16 = const()[name = string("const_150_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_63_cast_fp16 = mul(x = var_1762_cast_fp16, y = const_150_to_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [4]> var_1768 = const()[name = string("op_1768"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1769_cast_fp16 = reshape(shape = var_1768, x = k_77_internal_tensor_assign_1_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_151_to_fp16 = const()[name = string("const_151_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_79_cast_fp16 = mul(x = var_1769_cast_fp16, y = const_151_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<int32, [4]> var_1775 = const()[name = string("op_1775"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1776_cast_fp16 = reshape(shape = var_1775, x = v_77_internal_tensor_assign_1_cast_fp16)[name = string("op_1776_cast_fp16")];
            tensor<int32, [4]> var_1777 = const()[name = string("op_1777"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_47_transpose_x_0 = const()[name = string("qk_47_transpose_x_0"), val = bool(false)];
            bool qk_47_transpose_y_0 = const()[name = string("qk_47_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_223_perm_0 = const()[name = string("transpose_223_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_224_perm_0 = const()[name = string("transpose_224_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_224 = transpose(perm = transpose_224_perm_0, x = k_79_cast_fp16)[name = string("transpose_418")];
            tensor<fp16, [1, 16, ?, 64]> transpose_223 = transpose(perm = transpose_223_perm_0, x = q_63_cast_fp16)[name = string("transpose_419")];
            tensor<fp16, [1, 16, ?, 1500]> qk_47_cast_fp16 = matmul(transpose_x = qk_47_transpose_x_0, transpose_y = qk_47_transpose_y_0, x = transpose_223, y = transpose_224)[name = string("qk_47_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_1781_cast_fp16 = softmax(axis = var_1625, x = qk_47_cast_fp16)[name = string("op_1781_cast_fp16")];
            bool var_1783_transpose_x_0 = const()[name = string("op_1783_transpose_x_0"), val = bool(false)];
            bool var_1783_transpose_y_0 = const()[name = string("op_1783_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_79_cast_fp16 = transpose(perm = var_1777, x = var_1776_cast_fp16)[name = string("transpose_420")];
            tensor<fp16, [1, 16, ?, 64]> var_1783_cast_fp16 = matmul(transpose_x = var_1783_transpose_x_0, transpose_y = var_1783_transpose_y_0, x = var_1781_cast_fp16, y = v_79_cast_fp16)[name = string("op_1783_cast_fp16")];
            tensor<int32, [4]> var_1784 = const()[name = string("op_1784"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_177x = const()[name = string("concat_177x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1785_cast_fp16 = transpose(perm = var_1784, x = var_1783_cast_fp16)[name = string("transpose_417")];
            tensor<fp16, [1, ?, 1024]> x_139_cast_fp16 = reshape(shape = concat_177x, x = var_1785_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1789_to_fp16 = const()[name = string("op_1789_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326875520)))];
            tensor<fp16, [1024]> var_1790_to_fp16 = const()[name = string("op_1790_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328972736)))];
            tensor<fp16, [1, ?, 1024]> linear_61_cast_fp16 = linear(bias = var_1790_to_fp16, weight = var_1789_to_fp16, x = x_139_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_141_cast_fp16 = add(x = x_135_cast_fp16, y = linear_61_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [1]> var_1797_axes_0 = const()[name = string("op_1797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_7_mlp_ln_weight_to_fp16 = const()[name = string("blocks_7_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328974848)))];
            tensor<fp16, [1024]> blocks_7_mlp_ln_bias_to_fp16 = const()[name = string("blocks_7_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328976960)))];
            tensor<fp16, [1, ?, 1024]> var_1797_cast_fp16 = layer_norm(axes = var_1797_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_1631_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_141_cast_fp16)[name = string("op_1797_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328979072)))];
            tensor<fp16, [4096]> var_1807_to_fp16 = const()[name = string("op_1807_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337367744)))];
            tensor<fp16, [1, ?, 4096]> linear_62_cast_fp16 = linear(bias = var_1807_to_fp16, weight = var_1806_to_fp16, x = var_1797_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_145_mode_0 = const()[name = string("x_145_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_145_cast_fp16 = gelu(mode = x_145_mode_0, x = linear_62_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_1812_to_fp16 = const()[name = string("op_1812_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337376000)))];
            tensor<fp16, [1024]> var_1813_to_fp16 = const()[name = string("op_1813_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345764672)))];
            tensor<fp16, [1, ?, 1024]> linear_63_cast_fp16 = linear(bias = var_1813_to_fp16, weight = var_1812_to_fp16, x = x_145_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_147_cast_fp16 = add(x = x_141_cast_fp16, y = linear_63_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<int32, [4]> k_cache_33_begin_0 = const()[name = string("k_cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> k_cache_33_end_0 = const()[name = string("k_cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_33_end_mask_0 = const()[name = string("k_cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_33_squeeze_mask_0 = const()[name = string("k_cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_33_cast_fp16 = slice_by_index(begin = k_cache_33_begin_0, end = k_cache_33_end_0, end_mask = k_cache_33_end_mask_0, squeeze_mask = k_cache_33_squeeze_mask_0, x = coreml_update_state_62)[name = string("k_cache_33_cast_fp16")];
            tensor<int32, [4]> v_cache_33_begin_0 = const()[name = string("v_cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> v_cache_33_end_0 = const()[name = string("v_cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_33_end_mask_0 = const()[name = string("v_cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_33_squeeze_mask_0 = const()[name = string("v_cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_33_cast_fp16 = slice_by_index(begin = v_cache_33_begin_0, end = v_cache_33_end_0, end_mask = v_cache_33_end_mask_0, squeeze_mask = v_cache_33_squeeze_mask_0, x = coreml_update_state_63)[name = string("v_cache_33_cast_fp16")];
            tensor<int32, [4]> k_cache_35_begin_0 = const()[name = string("k_cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> k_cache_35_end_0 = const()[name = string("k_cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_35_end_mask_0 = const()[name = string("k_cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_35_squeeze_mask_0 = const()[name = string("k_cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_35_cast_fp16 = slice_by_index(begin = k_cache_35_begin_0, end = k_cache_35_end_0, end_mask = k_cache_35_end_mask_0, squeeze_mask = k_cache_35_squeeze_mask_0, x = read_state_2)[name = string("k_cache_35_cast_fp16")];
            tensor<int32, [4]> v_cache_35_begin_0 = const()[name = string("v_cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> v_cache_35_end_0 = const()[name = string("v_cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_35_end_mask_0 = const()[name = string("v_cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_35_squeeze_mask_0 = const()[name = string("v_cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_35_cast_fp16 = slice_by_index(begin = v_cache_35_begin_0, end = v_cache_35_end_0, end_mask = v_cache_35_end_mask_0, squeeze_mask = v_cache_35_squeeze_mask_0, x = read_state_3)[name = string("v_cache_35_cast_fp16")];
            int32 var_1836 = const()[name = string("op_1836"), val = int32(-1)];
            tensor<int32, [1]> var_1854_axes_0 = const()[name = string("op_1854_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_8_attn_ln_weight_to_fp16 = const()[name = string("blocks_8_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345766784)))];
            tensor<fp16, [1024]> blocks_8_attn_ln_bias_to_fp16 = const()[name = string("blocks_8_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345768896)))];
            fp16 var_1842_to_fp16 = const()[name = string("op_1842_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_1854_cast_fp16 = layer_norm(axes = var_1854_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_1842_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_147_cast_fp16)[name = string("op_1854_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1865_to_fp16 = const()[name = string("op_1865_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345771008)))];
            tensor<fp16, [1024]> var_1866_to_fp16 = const()[name = string("op_1866_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347868224)))];
            tensor<fp16, [1, ?, 1024]> linear_64_cast_fp16 = linear(bias = var_1866_to_fp16, weight = var_1865_to_fp16, x = var_1854_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1869_to_fp16 = const()[name = string("op_1869_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347870336)))];
            tensor<fp16, [1, ?, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1869_to_fp16, x = var_1854_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1873_to_fp16 = const()[name = string("op_1873_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349967552)))];
            tensor<fp16, [1024]> var_1874_to_fp16 = const()[name = string("op_1874_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352064768)))];
            tensor<fp16, [1, ?, 1024]> linear_66_cast_fp16 = linear(bias = var_1874_to_fp16, weight = var_1873_to_fp16, x = var_1854_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [3]> var_1876_shape_cast_fp16 = shape(x = linear_64_cast_fp16)[name = string("op_1876_shape_cast_fp16")];
            int32 gather_98_axis_0 = const()[name = string("gather_98_axis_0"), val = int32(0)];
            int32 gather_98_batch_dims_0 = const()[name = string("gather_98_batch_dims_0"), val = int32(0)];
            bool gather_98_validate_indices_0 = const()[name = string("gather_98_validate_indices_0"), val = bool(false)];
            string var_1876_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1876_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_98_to_uint16 = const()[name = string("select_98_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1876_shape_cast_fp16_to_uint16 = cast(dtype = var_1876_shape_cast_fp16_to_uint16_dtype_0, x = var_1876_shape_cast_fp16)[name = string("cast_278")];
            uint16 gather_98_cast_uint16 = gather(axis = gather_98_axis_0, batch_dims = gather_98_batch_dims_0, indices = select_98_to_uint16, validate_indices = gather_98_validate_indices_0, x = var_1876_shape_cast_fp16_to_uint16)[name = string("gather_98_cast_uint16")];
            string gather_98_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_98_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_98_cast_uint16_to_int32 = cast(dtype = gather_98_cast_uint16_to_int32_dtype_0, x = gather_98_cast_uint16)[name = string("cast_277")];
            int32 end_step_19 = add(x = offset, y = gather_98_cast_uint16_to_int32)[name = string("end_step_19")];
            tensor<int32, [1]> expand_dims_128 = const()[name = string("expand_dims_128"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_130 = const()[name = string("expand_dims_130"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_131_axes_0 = const()[name = string("expand_dims_131_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_131 = expand_dims(axes = expand_dims_131_axes_0, x = end_step_19)[name = string("expand_dims_131")];
            tensor<int32, [1]> concat_180_values0_0 = const()[name = string("concat_180_values0_0"), val = tensor<int32, [1]>([8])];
            int32 concat_180_axis_0 = const()[name = string("concat_180_axis_0"), val = int32(0)];
            bool concat_180_interleave_0 = const()[name = string("concat_180_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_180 = concat(axis = concat_180_axis_0, interleave = concat_180_interleave_0, values = (concat_180_values0_0, expand_dims_128, expand_dims_1, expand_dims_130))[name = string("concat_180")];
            tensor<int32, [1]> concat_181_values0_0 = const()[name = string("concat_181_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_181_values1_0 = const()[name = string("concat_181_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_181_values3_0 = const()[name = string("concat_181_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_181_axis_0 = const()[name = string("concat_181_axis_0"), val = int32(0)];
            bool concat_181_interleave_0 = const()[name = string("concat_181_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_181 = concat(axis = concat_181_axis_0, interleave = concat_181_interleave_0, values = (concat_181_values0_0, concat_181_values1_0, expand_dims_131, concat_181_values3_0))[name = string("concat_181")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_9_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_9_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_9_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_180, begin_mask = k_cache1_internal_tensor_assign_9_begin_mask_0, end = concat_181, end_mask = k_cache1_internal_tensor_assign_9_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_9_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_9_stride_0, update = linear_65_cast_fp16, x = coreml_update_state_62)[name = string("k_cache1_internal_tensor_assign_9_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_9_cast_fp16, input = k_cache1)[name = string("coreml_update_state_64_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_64 = read_state(input = k_cache1)[name = string("coreml_update_state_64")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_9_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_9_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_9_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_180, begin_mask = v_cache1_internal_tensor_assign_9_begin_mask_0, end = concat_181, end_mask = v_cache1_internal_tensor_assign_9_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_9_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_9_stride_0, update = linear_66_cast_fp16, x = coreml_update_state_63)[name = string("v_cache1_internal_tensor_assign_9_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_9_cast_fp16, input = v_cache1)[name = string("coreml_update_state_65_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_65 = read_state(input = v_cache1)[name = string("coreml_update_state_65")];
            int32 concat_186_values0_0 = const()[name = string("concat_186_values0_0"), val = int32(1)];
            int32 concat_186_values2_0 = const()[name = string("concat_186_values2_0"), val = int32(1024)];
            int32 concat_186_axis_0 = const()[name = string("concat_186_axis_0"), val = int32(0)];
            bool concat_186_interleave_0 = const()[name = string("concat_186_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_186 = concat(axis = concat_186_axis_0, interleave = concat_186_interleave_0, values = (concat_186_values0_0, end_step_19, concat_186_values2_0))[name = string("concat_186")];
            tensor<int32, [3]> var_1892_begin_0 = const()[name = string("op_1892_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1892_end_mask_0 = const()[name = string("op_1892_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1892_cast_fp16 = slice_by_index(begin = var_1892_begin_0, end = concat_186, end_mask = var_1892_end_mask_0, x = k_cache_33_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [3]> var_1895_begin_0 = const()[name = string("op_1895_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_1895_end_mask_0 = const()[name = string("op_1895_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_1895_cast_fp16 = slice_by_index(begin = var_1895_begin_0, end = concat_186, end_mask = var_1895_end_mask_0, x = v_cache_33_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<int32, [4]> concat_188x = const()[name = string("concat_188x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1905_cast_fp16 = reshape(shape = concat_188x, x = linear_64_cast_fp16)[name = string("op_1905_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_152_to_fp16 = const()[name = string("const_152_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_67_cast_fp16 = mul(x = var_1905_cast_fp16, y = const_152_to_fp16)[name = string("q_67_cast_fp16")];
            tensor<int32, [4]> concat_189x = const()[name = string("concat_189x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1912_cast_fp16 = reshape(shape = concat_189x, x = var_1892_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_85_cast_fp16 = mul(x = var_1912_cast_fp16, y = const_153_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<int32, [4]> concat_190x = const()[name = string("concat_190x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1919_cast_fp16 = reshape(shape = concat_190x, x = var_1895_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<int32, [4]> var_1920 = const()[name = string("op_1920"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_49_transpose_x_0 = const()[name = string("qk_49_transpose_x_0"), val = bool(false)];
            bool qk_49_transpose_y_0 = const()[name = string("qk_49_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_225_perm_0 = const()[name = string("transpose_225_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_226_perm_0 = const()[name = string("transpose_226_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_226 = transpose(perm = transpose_226_perm_0, x = k_85_cast_fp16)[name = string("transpose_414")];
            tensor<fp16, [1, 16, ?, 64]> transpose_225 = transpose(perm = transpose_225_perm_0, x = q_67_cast_fp16)[name = string("transpose_415")];
            tensor<fp16, [1, 16, ?, ?]> qk_49_cast_fp16 = matmul(transpose_x = qk_49_transpose_x_0, transpose_y = qk_49_transpose_y_0, x = transpose_225, y = transpose_226)[name = string("qk_49_cast_fp16")];
            int32 concat_191_values1_0 = const()[name = string("concat_191_values1_0"), val = int32(448)];
            int32 concat_191_axis_0 = const()[name = string("concat_191_axis_0"), val = int32(0)];
            bool concat_191_interleave_0 = const()[name = string("concat_191_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_191 = concat(axis = concat_191_axis_0, interleave = concat_191_interleave_0, values = (gather_98_cast_uint16_to_int32, concat_191_values1_0))[name = string("concat_191")];
            tensor<int32, [2]> var_1923_begin_0 = const()[name = string("op_1923_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1923_end_mask_0 = const()[name = string("op_1923_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_1923_cast_fp16 = slice_by_index(begin = var_1923_begin_0, end = concat_191, end_mask = var_1923_end_mask_0, x = mask_to_fp16)[name = string("op_1923_cast_fp16")];
            int32 concat_192_values0_0 = const()[name = string("concat_192_values0_0"), val = int32(0)];
            int32 concat_192_axis_0 = const()[name = string("concat_192_axis_0"), val = int32(0)];
            bool concat_192_interleave_0 = const()[name = string("concat_192_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_192 = concat(axis = concat_192_axis_0, interleave = concat_192_interleave_0, values = (concat_192_values0_0, gather_98_cast_uint16_to_int32))[name = string("concat_192")];
            tensor<int32, [2]> var_1924_begin_0 = const()[name = string("op_1924_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_1924_end_mask_0 = const()[name = string("op_1924_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_1924_cast_fp16 = slice_by_index(begin = var_1924_begin_0, end = concat_192, end_mask = var_1924_end_mask_0, x = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_51_cast_fp16 = add(x = qk_49_cast_fp16, y = var_1924_cast_fp16)[name = string("qk_51_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_1927_cast_fp16 = softmax(axis = var_1836, x = qk_51_cast_fp16)[name = string("op_1927_cast_fp16")];
            bool var_1929_transpose_x_0 = const()[name = string("op_1929_transpose_x_0"), val = bool(false)];
            bool var_1929_transpose_y_0 = const()[name = string("op_1929_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_85_cast_fp16 = transpose(perm = var_1920, x = var_1919_cast_fp16)[name = string("transpose_416")];
            tensor<fp16, [1, 16, ?, 64]> var_1929_cast_fp16 = matmul(transpose_x = var_1929_transpose_x_0, transpose_y = var_1929_transpose_y_0, x = var_1927_cast_fp16, y = v_85_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<int32, [4]> var_1930 = const()[name = string("op_1930"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_193x = const()[name = string("concat_193x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1931_cast_fp16 = transpose(perm = var_1930, x = var_1929_cast_fp16)[name = string("transpose_413")];
            tensor<fp16, [1, ?, 1024]> x_151_cast_fp16 = reshape(shape = concat_193x, x = var_1931_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1935_to_fp16 = const()[name = string("op_1935_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352066880)))];
            tensor<fp16, [1024]> var_1936_to_fp16 = const()[name = string("op_1936_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354164096)))];
            tensor<fp16, [1, ?, 1024]> linear_67_cast_fp16 = linear(bias = var_1936_to_fp16, weight = var_1935_to_fp16, x = x_151_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_153_cast_fp16 = add(x = x_147_cast_fp16, y = linear_67_cast_fp16)[name = string("x_153_cast_fp16")];
            tensor<int32, [1]> var_1943_axes_0 = const()[name = string("op_1943_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_8_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_8_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354166208)))];
            tensor<fp16, [1024]> blocks_8_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_8_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354168320)))];
            tensor<fp16, [1, ?, 1024]> var_1943_cast_fp16 = layer_norm(axes = var_1943_axes_0, beta = blocks_8_cross_attn_ln_bias_to_fp16, epsilon = var_1842_to_fp16, gamma = blocks_8_cross_attn_ln_weight_to_fp16, x = x_153_cast_fp16)[name = string("op_1943_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_1952_to_fp16 = const()[name = string("op_1952_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354170432)))];
            tensor<fp16, [1024]> var_1953_to_fp16 = const()[name = string("op_1953_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356267648)))];
            tensor<fp16, [1, ?, 1024]> linear_68_cast_fp16 = linear(bias = var_1953_to_fp16, weight = var_1952_to_fp16, x = var_1943_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> concat_194 = const()[name = string("concat_194"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_195 = const()[name = string("concat_195"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_87_internal_tensor_assign_1_stride_0 = const()[name = string("k_87_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_87_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_87_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_87_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_87_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_87_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_87_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_87_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_194, begin_mask = k_87_internal_tensor_assign_1_begin_mask_0, end = concat_195, end_mask = k_87_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_87_internal_tensor_assign_1_squeeze_mask_0, stride = k_87_internal_tensor_assign_1_stride_0, update = k_cache_35_cast_fp16, x = k_7_to_fp16)[name = string("k_87_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_196 = const()[name = string("concat_196"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_197 = const()[name = string("concat_197"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_87_internal_tensor_assign_1_stride_0 = const()[name = string("v_87_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_87_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_87_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_87_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_87_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_87_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_87_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_87_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_196, begin_mask = v_87_internal_tensor_assign_1_begin_mask_0, end = concat_197, end_mask = v_87_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_87_internal_tensor_assign_1_squeeze_mask_0, stride = v_87_internal_tensor_assign_1_stride_0, update = v_cache_35_cast_fp16, x = k_7_to_fp16)[name = string("v_87_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_198x = const()[name = string("concat_198x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1973_cast_fp16 = reshape(shape = concat_198x, x = linear_68_cast_fp16)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_71_cast_fp16 = mul(x = var_1973_cast_fp16, y = const_154_to_fp16)[name = string("q_71_cast_fp16")];
            tensor<int32, [4]> var_1979 = const()[name = string("op_1979"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1980_cast_fp16 = reshape(shape = var_1979, x = k_87_internal_tensor_assign_1_cast_fp16)[name = string("op_1980_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_155_to_fp16 = const()[name = string("const_155_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_89_cast_fp16 = mul(x = var_1980_cast_fp16, y = const_155_to_fp16)[name = string("k_89_cast_fp16")];
            tensor<int32, [4]> var_1986 = const()[name = string("op_1986"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_1987_cast_fp16 = reshape(shape = var_1986, x = v_87_internal_tensor_assign_1_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> var_1988 = const()[name = string("op_1988"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_53_transpose_x_0 = const()[name = string("qk_53_transpose_x_0"), val = bool(false)];
            bool qk_53_transpose_y_0 = const()[name = string("qk_53_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_227_perm_0 = const()[name = string("transpose_227_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_228_perm_0 = const()[name = string("transpose_228_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_228 = transpose(perm = transpose_228_perm_0, x = k_89_cast_fp16)[name = string("transpose_410")];
            tensor<fp16, [1, 16, ?, 64]> transpose_227 = transpose(perm = transpose_227_perm_0, x = q_71_cast_fp16)[name = string("transpose_411")];
            tensor<fp16, [1, 16, ?, 1500]> qk_53_cast_fp16 = matmul(transpose_x = qk_53_transpose_x_0, transpose_y = qk_53_transpose_y_0, x = transpose_227, y = transpose_228)[name = string("qk_53_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_1992_cast_fp16 = softmax(axis = var_1836, x = qk_53_cast_fp16)[name = string("op_1992_cast_fp16")];
            bool var_1994_transpose_x_0 = const()[name = string("op_1994_transpose_x_0"), val = bool(false)];
            bool var_1994_transpose_y_0 = const()[name = string("op_1994_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_89_cast_fp16 = transpose(perm = var_1988, x = var_1987_cast_fp16)[name = string("transpose_412")];
            tensor<fp16, [1, 16, ?, 64]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_0, transpose_y = var_1994_transpose_y_0, x = var_1992_cast_fp16, y = v_89_cast_fp16)[name = string("op_1994_cast_fp16")];
            tensor<int32, [4]> var_1995 = const()[name = string("op_1995"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_199x = const()[name = string("concat_199x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1996_cast_fp16 = transpose(perm = var_1995, x = var_1994_cast_fp16)[name = string("transpose_409")];
            tensor<fp16, [1, ?, 1024]> x_157_cast_fp16 = reshape(shape = concat_199x, x = var_1996_cast_fp16)[name = string("x_157_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356269760)))];
            tensor<fp16, [1024]> var_2001_to_fp16 = const()[name = string("op_2001_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358366976)))];
            tensor<fp16, [1, ?, 1024]> linear_69_cast_fp16 = linear(bias = var_2001_to_fp16, weight = var_2000_to_fp16, x = x_157_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_159_cast_fp16 = add(x = x_153_cast_fp16, y = linear_69_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [1]> var_2008_axes_0 = const()[name = string("op_2008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_8_mlp_ln_weight_to_fp16 = const()[name = string("blocks_8_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358369088)))];
            tensor<fp16, [1024]> blocks_8_mlp_ln_bias_to_fp16 = const()[name = string("blocks_8_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358371200)))];
            tensor<fp16, [1, ?, 1024]> var_2008_cast_fp16 = layer_norm(axes = var_2008_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_1842_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_159_cast_fp16)[name = string("op_2008_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_2017_to_fp16 = const()[name = string("op_2017_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358373312)))];
            tensor<fp16, [4096]> var_2018_to_fp16 = const()[name = string("op_2018_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366761984)))];
            tensor<fp16, [1, ?, 4096]> linear_70_cast_fp16 = linear(bias = var_2018_to_fp16, weight = var_2017_to_fp16, x = var_2008_cast_fp16)[name = string("linear_70_cast_fp16")];
            string x_163_mode_0 = const()[name = string("x_163_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_163_cast_fp16 = gelu(mode = x_163_mode_0, x = linear_70_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_2023_to_fp16 = const()[name = string("op_2023_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366770240)))];
            tensor<fp16, [1024]> var_2024_to_fp16 = const()[name = string("op_2024_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375158912)))];
            tensor<fp16, [1, ?, 1024]> linear_71_cast_fp16 = linear(bias = var_2024_to_fp16, weight = var_2023_to_fp16, x = x_163_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_165_cast_fp16 = add(x = x_159_cast_fp16, y = linear_71_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<int32, [4]> k_cache_37_begin_0 = const()[name = string("k_cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> k_cache_37_end_0 = const()[name = string("k_cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_37_end_mask_0 = const()[name = string("k_cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_37_squeeze_mask_0 = const()[name = string("k_cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_37_cast_fp16 = slice_by_index(begin = k_cache_37_begin_0, end = k_cache_37_end_0, end_mask = k_cache_37_end_mask_0, squeeze_mask = k_cache_37_squeeze_mask_0, x = coreml_update_state_64)[name = string("k_cache_37_cast_fp16")];
            tensor<int32, [4]> v_cache_37_begin_0 = const()[name = string("v_cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> v_cache_37_end_0 = const()[name = string("v_cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_37_end_mask_0 = const()[name = string("v_cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_37_squeeze_mask_0 = const()[name = string("v_cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_37_cast_fp16 = slice_by_index(begin = v_cache_37_begin_0, end = v_cache_37_end_0, end_mask = v_cache_37_end_mask_0, squeeze_mask = v_cache_37_squeeze_mask_0, x = coreml_update_state_65)[name = string("v_cache_37_cast_fp16")];
            tensor<int32, [4]> k_cache_39_begin_0 = const()[name = string("k_cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> k_cache_39_end_0 = const()[name = string("k_cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_39_end_mask_0 = const()[name = string("k_cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_39_squeeze_mask_0 = const()[name = string("k_cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_39_cast_fp16 = slice_by_index(begin = k_cache_39_begin_0, end = k_cache_39_end_0, end_mask = k_cache_39_end_mask_0, squeeze_mask = k_cache_39_squeeze_mask_0, x = read_state_2)[name = string("k_cache_39_cast_fp16")];
            tensor<int32, [4]> v_cache_39_begin_0 = const()[name = string("v_cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> v_cache_39_end_0 = const()[name = string("v_cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_39_end_mask_0 = const()[name = string("v_cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_39_squeeze_mask_0 = const()[name = string("v_cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_39_cast_fp16 = slice_by_index(begin = v_cache_39_begin_0, end = v_cache_39_end_0, end_mask = v_cache_39_end_mask_0, squeeze_mask = v_cache_39_squeeze_mask_0, x = read_state_3)[name = string("v_cache_39_cast_fp16")];
            int32 var_2047 = const()[name = string("op_2047"), val = int32(-1)];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_9_attn_ln_weight_to_fp16 = const()[name = string("blocks_9_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375161024)))];
            tensor<fp16, [1024]> blocks_9_attn_ln_bias_to_fp16 = const()[name = string("blocks_9_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375163136)))];
            fp16 var_2053_to_fp16 = const()[name = string("op_2053_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_2065_cast_fp16 = layer_norm(axes = var_2065_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_2053_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_165_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2076_to_fp16 = const()[name = string("op_2076_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375165248)))];
            tensor<fp16, [1024]> var_2077_to_fp16 = const()[name = string("op_2077_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377262464)))];
            tensor<fp16, [1, ?, 1024]> linear_72_cast_fp16 = linear(bias = var_2077_to_fp16, weight = var_2076_to_fp16, x = var_2065_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2080_to_fp16 = const()[name = string("op_2080_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377264576)))];
            tensor<fp16, [1, ?, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2080_to_fp16, x = var_2065_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2084_to_fp16 = const()[name = string("op_2084_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379361792)))];
            tensor<fp16, [1024]> var_2085_to_fp16 = const()[name = string("op_2085_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381459008)))];
            tensor<fp16, [1, ?, 1024]> linear_74_cast_fp16 = linear(bias = var_2085_to_fp16, weight = var_2084_to_fp16, x = var_2065_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [3]> var_2087_shape_cast_fp16 = shape(x = linear_72_cast_fp16)[name = string("op_2087_shape_cast_fp16")];
            int32 gather_110_axis_0 = const()[name = string("gather_110_axis_0"), val = int32(0)];
            int32 gather_110_batch_dims_0 = const()[name = string("gather_110_batch_dims_0"), val = int32(0)];
            bool gather_110_validate_indices_0 = const()[name = string("gather_110_validate_indices_0"), val = bool(false)];
            string var_2087_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2087_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_110_to_uint16 = const()[name = string("select_110_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2087_shape_cast_fp16_to_uint16 = cast(dtype = var_2087_shape_cast_fp16_to_uint16_dtype_0, x = var_2087_shape_cast_fp16)[name = string("cast_276")];
            uint16 gather_110_cast_uint16 = gather(axis = gather_110_axis_0, batch_dims = gather_110_batch_dims_0, indices = select_110_to_uint16, validate_indices = gather_110_validate_indices_0, x = var_2087_shape_cast_fp16_to_uint16)[name = string("gather_110_cast_uint16")];
            string gather_110_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_110_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_110_cast_uint16_to_int32 = cast(dtype = gather_110_cast_uint16_to_int32_dtype_0, x = gather_110_cast_uint16)[name = string("cast_275")];
            int32 end_step_21 = add(x = offset, y = gather_110_cast_uint16_to_int32)[name = string("end_step_21")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_146 = const()[name = string("expand_dims_146"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147_axes_0 = const()[name = string("expand_dims_147_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = expand_dims(axes = expand_dims_147_axes_0, x = end_step_21)[name = string("expand_dims_147")];
            tensor<int32, [1]> concat_202_values0_0 = const()[name = string("concat_202_values0_0"), val = tensor<int32, [1]>([9])];
            int32 concat_202_axis_0 = const()[name = string("concat_202_axis_0"), val = int32(0)];
            bool concat_202_interleave_0 = const()[name = string("concat_202_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_202 = concat(axis = concat_202_axis_0, interleave = concat_202_interleave_0, values = (concat_202_values0_0, expand_dims_144, expand_dims_1, expand_dims_146))[name = string("concat_202")];
            tensor<int32, [1]> concat_203_values0_0 = const()[name = string("concat_203_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_203_values1_0 = const()[name = string("concat_203_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_203_values3_0 = const()[name = string("concat_203_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_203_axis_0 = const()[name = string("concat_203_axis_0"), val = int32(0)];
            bool concat_203_interleave_0 = const()[name = string("concat_203_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_203 = concat(axis = concat_203_axis_0, interleave = concat_203_interleave_0, values = (concat_203_values0_0, concat_203_values1_0, expand_dims_147, concat_203_values3_0))[name = string("concat_203")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_10_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_10_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_10_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_202, begin_mask = k_cache1_internal_tensor_assign_10_begin_mask_0, end = concat_203, end_mask = k_cache1_internal_tensor_assign_10_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_10_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_10_stride_0, update = linear_73_cast_fp16, x = coreml_update_state_64)[name = string("k_cache1_internal_tensor_assign_10_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_10_cast_fp16, input = k_cache1)[name = string("coreml_update_state_66_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_66 = read_state(input = k_cache1)[name = string("coreml_update_state_66")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_10_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_10_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_10_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_202, begin_mask = v_cache1_internal_tensor_assign_10_begin_mask_0, end = concat_203, end_mask = v_cache1_internal_tensor_assign_10_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_10_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_10_stride_0, update = linear_74_cast_fp16, x = coreml_update_state_65)[name = string("v_cache1_internal_tensor_assign_10_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_10_cast_fp16, input = v_cache1)[name = string("coreml_update_state_67_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_67 = read_state(input = v_cache1)[name = string("coreml_update_state_67")];
            int32 concat_208_values0_0 = const()[name = string("concat_208_values0_0"), val = int32(1)];
            int32 concat_208_values2_0 = const()[name = string("concat_208_values2_0"), val = int32(1024)];
            int32 concat_208_axis_0 = const()[name = string("concat_208_axis_0"), val = int32(0)];
            bool concat_208_interleave_0 = const()[name = string("concat_208_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_208 = concat(axis = concat_208_axis_0, interleave = concat_208_interleave_0, values = (concat_208_values0_0, end_step_21, concat_208_values2_0))[name = string("concat_208")];
            tensor<int32, [3]> var_2103_begin_0 = const()[name = string("op_2103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2103_end_mask_0 = const()[name = string("op_2103_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2103_cast_fp16 = slice_by_index(begin = var_2103_begin_0, end = concat_208, end_mask = var_2103_end_mask_0, x = k_cache_37_cast_fp16)[name = string("op_2103_cast_fp16")];
            tensor<int32, [3]> var_2106_begin_0 = const()[name = string("op_2106_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2106_end_mask_0 = const()[name = string("op_2106_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2106_cast_fp16 = slice_by_index(begin = var_2106_begin_0, end = concat_208, end_mask = var_2106_end_mask_0, x = v_cache_37_cast_fp16)[name = string("op_2106_cast_fp16")];
            tensor<int32, [4]> concat_210x = const()[name = string("concat_210x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2116_cast_fp16 = reshape(shape = concat_210x, x = linear_72_cast_fp16)[name = string("op_2116_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_156_to_fp16 = const()[name = string("const_156_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_75_cast_fp16 = mul(x = var_2116_cast_fp16, y = const_156_to_fp16)[name = string("q_75_cast_fp16")];
            tensor<int32, [4]> concat_211x = const()[name = string("concat_211x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2123_cast_fp16 = reshape(shape = concat_211x, x = var_2103_cast_fp16)[name = string("op_2123_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_95_cast_fp16 = mul(x = var_2123_cast_fp16, y = const_157_to_fp16)[name = string("k_95_cast_fp16")];
            tensor<int32, [4]> concat_212x = const()[name = string("concat_212x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2130_cast_fp16 = reshape(shape = concat_212x, x = var_2106_cast_fp16)[name = string("op_2130_cast_fp16")];
            tensor<int32, [4]> var_2131 = const()[name = string("op_2131"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_55_transpose_x_0 = const()[name = string("qk_55_transpose_x_0"), val = bool(false)];
            bool qk_55_transpose_y_0 = const()[name = string("qk_55_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_229_perm_0 = const()[name = string("transpose_229_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_230_perm_0 = const()[name = string("transpose_230_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_230 = transpose(perm = transpose_230_perm_0, x = k_95_cast_fp16)[name = string("transpose_406")];
            tensor<fp16, [1, 16, ?, 64]> transpose_229 = transpose(perm = transpose_229_perm_0, x = q_75_cast_fp16)[name = string("transpose_407")];
            tensor<fp16, [1, 16, ?, ?]> qk_55_cast_fp16 = matmul(transpose_x = qk_55_transpose_x_0, transpose_y = qk_55_transpose_y_0, x = transpose_229, y = transpose_230)[name = string("qk_55_cast_fp16")];
            int32 concat_213_values1_0 = const()[name = string("concat_213_values1_0"), val = int32(448)];
            int32 concat_213_axis_0 = const()[name = string("concat_213_axis_0"), val = int32(0)];
            bool concat_213_interleave_0 = const()[name = string("concat_213_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_213 = concat(axis = concat_213_axis_0, interleave = concat_213_interleave_0, values = (gather_110_cast_uint16_to_int32, concat_213_values1_0))[name = string("concat_213")];
            tensor<int32, [2]> var_2134_begin_0 = const()[name = string("op_2134_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2134_end_mask_0 = const()[name = string("op_2134_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_2134_cast_fp16 = slice_by_index(begin = var_2134_begin_0, end = concat_213, end_mask = var_2134_end_mask_0, x = mask_to_fp16)[name = string("op_2134_cast_fp16")];
            int32 concat_214_values0_0 = const()[name = string("concat_214_values0_0"), val = int32(0)];
            int32 concat_214_axis_0 = const()[name = string("concat_214_axis_0"), val = int32(0)];
            bool concat_214_interleave_0 = const()[name = string("concat_214_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_214 = concat(axis = concat_214_axis_0, interleave = concat_214_interleave_0, values = (concat_214_values0_0, gather_110_cast_uint16_to_int32))[name = string("concat_214")];
            tensor<int32, [2]> var_2135_begin_0 = const()[name = string("op_2135_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2135_end_mask_0 = const()[name = string("op_2135_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_2135_cast_fp16 = slice_by_index(begin = var_2135_begin_0, end = concat_214, end_mask = var_2135_end_mask_0, x = var_2134_cast_fp16)[name = string("op_2135_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_57_cast_fp16 = add(x = qk_55_cast_fp16, y = var_2135_cast_fp16)[name = string("qk_57_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_2138_cast_fp16 = softmax(axis = var_2047, x = qk_57_cast_fp16)[name = string("op_2138_cast_fp16")];
            bool var_2140_transpose_x_0 = const()[name = string("op_2140_transpose_x_0"), val = bool(false)];
            bool var_2140_transpose_y_0 = const()[name = string("op_2140_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_95_cast_fp16 = transpose(perm = var_2131, x = var_2130_cast_fp16)[name = string("transpose_408")];
            tensor<fp16, [1, 16, ?, 64]> var_2140_cast_fp16 = matmul(transpose_x = var_2140_transpose_x_0, transpose_y = var_2140_transpose_y_0, x = var_2138_cast_fp16, y = v_95_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<int32, [4]> var_2141 = const()[name = string("op_2141"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_215x = const()[name = string("concat_215x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2142_cast_fp16 = transpose(perm = var_2141, x = var_2140_cast_fp16)[name = string("transpose_405")];
            tensor<fp16, [1, ?, 1024]> x_169_cast_fp16 = reshape(shape = concat_215x, x = var_2142_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2146_to_fp16 = const()[name = string("op_2146_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381461120)))];
            tensor<fp16, [1024]> var_2147_to_fp16 = const()[name = string("op_2147_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383558336)))];
            tensor<fp16, [1, ?, 1024]> linear_75_cast_fp16 = linear(bias = var_2147_to_fp16, weight = var_2146_to_fp16, x = x_169_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_171_cast_fp16 = add(x = x_165_cast_fp16, y = linear_75_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<int32, [1]> var_2154_axes_0 = const()[name = string("op_2154_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_9_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_9_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383560448)))];
            tensor<fp16, [1024]> blocks_9_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_9_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383562560)))];
            tensor<fp16, [1, ?, 1024]> var_2154_cast_fp16 = layer_norm(axes = var_2154_axes_0, beta = blocks_9_cross_attn_ln_bias_to_fp16, epsilon = var_2053_to_fp16, gamma = blocks_9_cross_attn_ln_weight_to_fp16, x = x_171_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383564672)))];
            tensor<fp16, [1024]> var_2164_to_fp16 = const()[name = string("op_2164_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385661888)))];
            tensor<fp16, [1, ?, 1024]> linear_76_cast_fp16 = linear(bias = var_2164_to_fp16, weight = var_2163_to_fp16, x = var_2154_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [3]> concat_216 = const()[name = string("concat_216"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_217 = const()[name = string("concat_217"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_97_internal_tensor_assign_1_stride_0 = const()[name = string("k_97_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_97_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_97_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_97_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_97_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_97_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_97_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_97_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_216, begin_mask = k_97_internal_tensor_assign_1_begin_mask_0, end = concat_217, end_mask = k_97_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_97_internal_tensor_assign_1_squeeze_mask_0, stride = k_97_internal_tensor_assign_1_stride_0, update = k_cache_39_cast_fp16, x = k_7_to_fp16)[name = string("k_97_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_218 = const()[name = string("concat_218"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_219 = const()[name = string("concat_219"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_97_internal_tensor_assign_1_stride_0 = const()[name = string("v_97_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_97_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_97_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_97_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_97_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_97_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_97_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_97_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_218, begin_mask = v_97_internal_tensor_assign_1_begin_mask_0, end = concat_219, end_mask = v_97_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_97_internal_tensor_assign_1_squeeze_mask_0, stride = v_97_internal_tensor_assign_1_stride_0, update = v_cache_39_cast_fp16, x = k_7_to_fp16)[name = string("v_97_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_220x = const()[name = string("concat_220x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2184_cast_fp16 = reshape(shape = concat_220x, x = linear_76_cast_fp16)[name = string("op_2184_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_158_to_fp16 = const()[name = string("const_158_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_79_cast_fp16 = mul(x = var_2184_cast_fp16, y = const_158_to_fp16)[name = string("q_79_cast_fp16")];
            tensor<int32, [4]> var_2190 = const()[name = string("op_2190"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2191_cast_fp16 = reshape(shape = var_2190, x = k_97_internal_tensor_assign_1_cast_fp16)[name = string("op_2191_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_159_to_fp16 = const()[name = string("const_159_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_99_cast_fp16 = mul(x = var_2191_cast_fp16, y = const_159_to_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> var_2197 = const()[name = string("op_2197"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2198_cast_fp16 = reshape(shape = var_2197, x = v_97_internal_tensor_assign_1_cast_fp16)[name = string("op_2198_cast_fp16")];
            tensor<int32, [4]> var_2199 = const()[name = string("op_2199"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_59_transpose_x_0 = const()[name = string("qk_59_transpose_x_0"), val = bool(false)];
            bool qk_59_transpose_y_0 = const()[name = string("qk_59_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_231_perm_0 = const()[name = string("transpose_231_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_232_perm_0 = const()[name = string("transpose_232_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_232 = transpose(perm = transpose_232_perm_0, x = k_99_cast_fp16)[name = string("transpose_402")];
            tensor<fp16, [1, 16, ?, 64]> transpose_231 = transpose(perm = transpose_231_perm_0, x = q_79_cast_fp16)[name = string("transpose_403")];
            tensor<fp16, [1, 16, ?, 1500]> qk_59_cast_fp16 = matmul(transpose_x = qk_59_transpose_x_0, transpose_y = qk_59_transpose_y_0, x = transpose_231, y = transpose_232)[name = string("qk_59_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_2203_cast_fp16 = softmax(axis = var_2047, x = qk_59_cast_fp16)[name = string("op_2203_cast_fp16")];
            bool var_2205_transpose_x_0 = const()[name = string("op_2205_transpose_x_0"), val = bool(false)];
            bool var_2205_transpose_y_0 = const()[name = string("op_2205_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_99_cast_fp16 = transpose(perm = var_2199, x = var_2198_cast_fp16)[name = string("transpose_404")];
            tensor<fp16, [1, 16, ?, 64]> var_2205_cast_fp16 = matmul(transpose_x = var_2205_transpose_x_0, transpose_y = var_2205_transpose_y_0, x = var_2203_cast_fp16, y = v_99_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<int32, [4]> var_2206 = const()[name = string("op_2206"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_221x = const()[name = string("concat_221x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2207_cast_fp16 = transpose(perm = var_2206, x = var_2205_cast_fp16)[name = string("transpose_401")];
            tensor<fp16, [1, ?, 1024]> x_175_cast_fp16 = reshape(shape = concat_221x, x = var_2207_cast_fp16)[name = string("x_175_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2211_to_fp16 = const()[name = string("op_2211_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385664000)))];
            tensor<fp16, [1024]> var_2212_to_fp16 = const()[name = string("op_2212_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387761216)))];
            tensor<fp16, [1, ?, 1024]> linear_77_cast_fp16 = linear(bias = var_2212_to_fp16, weight = var_2211_to_fp16, x = x_175_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_177_cast_fp16 = add(x = x_171_cast_fp16, y = linear_77_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [1]> var_2219_axes_0 = const()[name = string("op_2219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_9_mlp_ln_weight_to_fp16 = const()[name = string("blocks_9_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387763328)))];
            tensor<fp16, [1024]> blocks_9_mlp_ln_bias_to_fp16 = const()[name = string("blocks_9_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387765440)))];
            tensor<fp16, [1, ?, 1024]> var_2219_cast_fp16 = layer_norm(axes = var_2219_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_2053_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_177_cast_fp16)[name = string("op_2219_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_2228_to_fp16 = const()[name = string("op_2228_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387767552)))];
            tensor<fp16, [4096]> var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396156224)))];
            tensor<fp16, [1, ?, 4096]> linear_78_cast_fp16 = linear(bias = var_2229_to_fp16, weight = var_2228_to_fp16, x = var_2219_cast_fp16)[name = string("linear_78_cast_fp16")];
            string x_181_mode_0 = const()[name = string("x_181_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_181_cast_fp16 = gelu(mode = x_181_mode_0, x = linear_78_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_2234_to_fp16 = const()[name = string("op_2234_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396164480)))];
            tensor<fp16, [1024]> var_2235_to_fp16 = const()[name = string("op_2235_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404553152)))];
            tensor<fp16, [1, ?, 1024]> linear_79_cast_fp16 = linear(bias = var_2235_to_fp16, weight = var_2234_to_fp16, x = x_181_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_183_cast_fp16 = add(x = x_177_cast_fp16, y = linear_79_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<int32, [4]> k_cache_41_begin_0 = const()[name = string("k_cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> k_cache_41_end_0 = const()[name = string("k_cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_41_end_mask_0 = const()[name = string("k_cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_41_squeeze_mask_0 = const()[name = string("k_cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_41_cast_fp16 = slice_by_index(begin = k_cache_41_begin_0, end = k_cache_41_end_0, end_mask = k_cache_41_end_mask_0, squeeze_mask = k_cache_41_squeeze_mask_0, x = coreml_update_state_66)[name = string("k_cache_41_cast_fp16")];
            tensor<int32, [4]> v_cache_41_begin_0 = const()[name = string("v_cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> v_cache_41_end_0 = const()[name = string("v_cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_41_end_mask_0 = const()[name = string("v_cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_41_squeeze_mask_0 = const()[name = string("v_cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_41_cast_fp16 = slice_by_index(begin = v_cache_41_begin_0, end = v_cache_41_end_0, end_mask = v_cache_41_end_mask_0, squeeze_mask = v_cache_41_squeeze_mask_0, x = coreml_update_state_67)[name = string("v_cache_41_cast_fp16")];
            tensor<int32, [4]> k_cache_43_begin_0 = const()[name = string("k_cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> k_cache_43_end_0 = const()[name = string("k_cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_43_end_mask_0 = const()[name = string("k_cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_43_squeeze_mask_0 = const()[name = string("k_cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_43_cast_fp16 = slice_by_index(begin = k_cache_43_begin_0, end = k_cache_43_end_0, end_mask = k_cache_43_end_mask_0, squeeze_mask = k_cache_43_squeeze_mask_0, x = read_state_2)[name = string("k_cache_43_cast_fp16")];
            tensor<int32, [4]> v_cache_43_begin_0 = const()[name = string("v_cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> v_cache_43_end_0 = const()[name = string("v_cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_43_end_mask_0 = const()[name = string("v_cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_43_squeeze_mask_0 = const()[name = string("v_cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_43_cast_fp16 = slice_by_index(begin = v_cache_43_begin_0, end = v_cache_43_end_0, end_mask = v_cache_43_end_mask_0, squeeze_mask = v_cache_43_squeeze_mask_0, x = read_state_3)[name = string("v_cache_43_cast_fp16")];
            int32 var_2258 = const()[name = string("op_2258"), val = int32(-1)];
            tensor<int32, [1]> var_2276_axes_0 = const()[name = string("op_2276_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_10_attn_ln_weight_to_fp16 = const()[name = string("blocks_10_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404555264)))];
            tensor<fp16, [1024]> blocks_10_attn_ln_bias_to_fp16 = const()[name = string("blocks_10_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404557376)))];
            fp16 var_2264_to_fp16 = const()[name = string("op_2264_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_2276_cast_fp16 = layer_norm(axes = var_2276_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_2264_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_183_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2287_to_fp16 = const()[name = string("op_2287_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404559488)))];
            tensor<fp16, [1024]> var_2288_to_fp16 = const()[name = string("op_2288_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406656704)))];
            tensor<fp16, [1, ?, 1024]> linear_80_cast_fp16 = linear(bias = var_2288_to_fp16, weight = var_2287_to_fp16, x = var_2276_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2291_to_fp16 = const()[name = string("op_2291_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406658816)))];
            tensor<fp16, [1, ?, 1024]> linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2291_to_fp16, x = var_2276_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2295_to_fp16 = const()[name = string("op_2295_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408756032)))];
            tensor<fp16, [1024]> var_2296_to_fp16 = const()[name = string("op_2296_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410853248)))];
            tensor<fp16, [1, ?, 1024]> linear_82_cast_fp16 = linear(bias = var_2296_to_fp16, weight = var_2295_to_fp16, x = var_2276_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [3]> var_2298_shape_cast_fp16 = shape(x = linear_80_cast_fp16)[name = string("op_2298_shape_cast_fp16")];
            int32 gather_122_axis_0 = const()[name = string("gather_122_axis_0"), val = int32(0)];
            int32 gather_122_batch_dims_0 = const()[name = string("gather_122_batch_dims_0"), val = int32(0)];
            bool gather_122_validate_indices_0 = const()[name = string("gather_122_validate_indices_0"), val = bool(false)];
            string var_2298_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2298_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_122_to_uint16 = const()[name = string("select_122_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2298_shape_cast_fp16_to_uint16 = cast(dtype = var_2298_shape_cast_fp16_to_uint16_dtype_0, x = var_2298_shape_cast_fp16)[name = string("cast_274")];
            uint16 gather_122_cast_uint16 = gather(axis = gather_122_axis_0, batch_dims = gather_122_batch_dims_0, indices = select_122_to_uint16, validate_indices = gather_122_validate_indices_0, x = var_2298_shape_cast_fp16_to_uint16)[name = string("gather_122_cast_uint16")];
            string gather_122_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_122_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_122_cast_uint16_to_int32 = cast(dtype = gather_122_cast_uint16_to_int32_dtype_0, x = gather_122_cast_uint16)[name = string("cast_273")];
            int32 end_step_23 = add(x = offset, y = gather_122_cast_uint16_to_int32)[name = string("end_step_23")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_162 = const()[name = string("expand_dims_162"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_163_axes_0 = const()[name = string("expand_dims_163_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_163 = expand_dims(axes = expand_dims_163_axes_0, x = end_step_23)[name = string("expand_dims_163")];
            tensor<int32, [1]> concat_224_values0_0 = const()[name = string("concat_224_values0_0"), val = tensor<int32, [1]>([10])];
            int32 concat_224_axis_0 = const()[name = string("concat_224_axis_0"), val = int32(0)];
            bool concat_224_interleave_0 = const()[name = string("concat_224_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_224 = concat(axis = concat_224_axis_0, interleave = concat_224_interleave_0, values = (concat_224_values0_0, expand_dims_160, expand_dims_1, expand_dims_162))[name = string("concat_224")];
            tensor<int32, [1]> concat_225_values0_0 = const()[name = string("concat_225_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_225_values1_0 = const()[name = string("concat_225_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_225_values3_0 = const()[name = string("concat_225_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_225_axis_0 = const()[name = string("concat_225_axis_0"), val = int32(0)];
            bool concat_225_interleave_0 = const()[name = string("concat_225_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_225 = concat(axis = concat_225_axis_0, interleave = concat_225_interleave_0, values = (concat_225_values0_0, concat_225_values1_0, expand_dims_163, concat_225_values3_0))[name = string("concat_225")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_11_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_11_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_11_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_224, begin_mask = k_cache1_internal_tensor_assign_11_begin_mask_0, end = concat_225, end_mask = k_cache1_internal_tensor_assign_11_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_11_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_11_stride_0, update = linear_81_cast_fp16, x = coreml_update_state_66)[name = string("k_cache1_internal_tensor_assign_11_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_11_cast_fp16, input = k_cache1)[name = string("coreml_update_state_68_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_68 = read_state(input = k_cache1)[name = string("coreml_update_state_68")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_11_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_11_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_11_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_224, begin_mask = v_cache1_internal_tensor_assign_11_begin_mask_0, end = concat_225, end_mask = v_cache1_internal_tensor_assign_11_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_11_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_11_stride_0, update = linear_82_cast_fp16, x = coreml_update_state_67)[name = string("v_cache1_internal_tensor_assign_11_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_11_cast_fp16, input = v_cache1)[name = string("coreml_update_state_69_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_69 = read_state(input = v_cache1)[name = string("coreml_update_state_69")];
            int32 concat_230_values0_0 = const()[name = string("concat_230_values0_0"), val = int32(1)];
            int32 concat_230_values2_0 = const()[name = string("concat_230_values2_0"), val = int32(1024)];
            int32 concat_230_axis_0 = const()[name = string("concat_230_axis_0"), val = int32(0)];
            bool concat_230_interleave_0 = const()[name = string("concat_230_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_230 = concat(axis = concat_230_axis_0, interleave = concat_230_interleave_0, values = (concat_230_values0_0, end_step_23, concat_230_values2_0))[name = string("concat_230")];
            tensor<int32, [3]> var_2314_begin_0 = const()[name = string("op_2314_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2314_end_mask_0 = const()[name = string("op_2314_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2314_cast_fp16 = slice_by_index(begin = var_2314_begin_0, end = concat_230, end_mask = var_2314_end_mask_0, x = k_cache_41_cast_fp16)[name = string("op_2314_cast_fp16")];
            tensor<int32, [3]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = concat_230, end_mask = var_2317_end_mask_0, x = v_cache_41_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<int32, [4]> concat_232x = const()[name = string("concat_232x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2327_cast_fp16 = reshape(shape = concat_232x, x = linear_80_cast_fp16)[name = string("op_2327_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_160_to_fp16 = const()[name = string("const_160_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_83_cast_fp16 = mul(x = var_2327_cast_fp16, y = const_160_to_fp16)[name = string("q_83_cast_fp16")];
            tensor<int32, [4]> concat_233x = const()[name = string("concat_233x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2334_cast_fp16 = reshape(shape = concat_233x, x = var_2314_cast_fp16)[name = string("op_2334_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_161_to_fp16 = const()[name = string("const_161_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_105_cast_fp16 = mul(x = var_2334_cast_fp16, y = const_161_to_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> concat_234x = const()[name = string("concat_234x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2341_cast_fp16 = reshape(shape = concat_234x, x = var_2317_cast_fp16)[name = string("op_2341_cast_fp16")];
            tensor<int32, [4]> var_2342 = const()[name = string("op_2342"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_61_transpose_x_0 = const()[name = string("qk_61_transpose_x_0"), val = bool(false)];
            bool qk_61_transpose_y_0 = const()[name = string("qk_61_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_233_perm_0 = const()[name = string("transpose_233_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_234_perm_0 = const()[name = string("transpose_234_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_234 = transpose(perm = transpose_234_perm_0, x = k_105_cast_fp16)[name = string("transpose_398")];
            tensor<fp16, [1, 16, ?, 64]> transpose_233 = transpose(perm = transpose_233_perm_0, x = q_83_cast_fp16)[name = string("transpose_399")];
            tensor<fp16, [1, 16, ?, ?]> qk_61_cast_fp16 = matmul(transpose_x = qk_61_transpose_x_0, transpose_y = qk_61_transpose_y_0, x = transpose_233, y = transpose_234)[name = string("qk_61_cast_fp16")];
            int32 concat_235_values1_0 = const()[name = string("concat_235_values1_0"), val = int32(448)];
            int32 concat_235_axis_0 = const()[name = string("concat_235_axis_0"), val = int32(0)];
            bool concat_235_interleave_0 = const()[name = string("concat_235_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_235 = concat(axis = concat_235_axis_0, interleave = concat_235_interleave_0, values = (gather_122_cast_uint16_to_int32, concat_235_values1_0))[name = string("concat_235")];
            tensor<int32, [2]> var_2345_begin_0 = const()[name = string("op_2345_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2345_end_mask_0 = const()[name = string("op_2345_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = concat_235, end_mask = var_2345_end_mask_0, x = mask_to_fp16)[name = string("op_2345_cast_fp16")];
            int32 concat_236_values0_0 = const()[name = string("concat_236_values0_0"), val = int32(0)];
            int32 concat_236_axis_0 = const()[name = string("concat_236_axis_0"), val = int32(0)];
            bool concat_236_interleave_0 = const()[name = string("concat_236_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_236 = concat(axis = concat_236_axis_0, interleave = concat_236_interleave_0, values = (concat_236_values0_0, gather_122_cast_uint16_to_int32))[name = string("concat_236")];
            tensor<int32, [2]> var_2346_begin_0 = const()[name = string("op_2346_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2346_end_mask_0 = const()[name = string("op_2346_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_2346_cast_fp16 = slice_by_index(begin = var_2346_begin_0, end = concat_236, end_mask = var_2346_end_mask_0, x = var_2345_cast_fp16)[name = string("op_2346_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_63_cast_fp16 = add(x = qk_61_cast_fp16, y = var_2346_cast_fp16)[name = string("qk_63_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_2349_cast_fp16 = softmax(axis = var_2258, x = qk_63_cast_fp16)[name = string("op_2349_cast_fp16")];
            bool var_2351_transpose_x_0 = const()[name = string("op_2351_transpose_x_0"), val = bool(false)];
            bool var_2351_transpose_y_0 = const()[name = string("op_2351_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_105_cast_fp16 = transpose(perm = var_2342, x = var_2341_cast_fp16)[name = string("transpose_400")];
            tensor<fp16, [1, 16, ?, 64]> var_2351_cast_fp16 = matmul(transpose_x = var_2351_transpose_x_0, transpose_y = var_2351_transpose_y_0, x = var_2349_cast_fp16, y = v_105_cast_fp16)[name = string("op_2351_cast_fp16")];
            tensor<int32, [4]> var_2352 = const()[name = string("op_2352"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_237x = const()[name = string("concat_237x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2353_cast_fp16 = transpose(perm = var_2352, x = var_2351_cast_fp16)[name = string("transpose_397")];
            tensor<fp16, [1, ?, 1024]> x_187_cast_fp16 = reshape(shape = concat_237x, x = var_2353_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2357_to_fp16 = const()[name = string("op_2357_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410855360)))];
            tensor<fp16, [1024]> var_2358_to_fp16 = const()[name = string("op_2358_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412952576)))];
            tensor<fp16, [1, ?, 1024]> linear_83_cast_fp16 = linear(bias = var_2358_to_fp16, weight = var_2357_to_fp16, x = x_187_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_189_cast_fp16 = add(x = x_183_cast_fp16, y = linear_83_cast_fp16)[name = string("x_189_cast_fp16")];
            tensor<int32, [1]> var_2365_axes_0 = const()[name = string("op_2365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_10_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_10_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412954688)))];
            tensor<fp16, [1024]> blocks_10_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_10_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412956800)))];
            tensor<fp16, [1, ?, 1024]> var_2365_cast_fp16 = layer_norm(axes = var_2365_axes_0, beta = blocks_10_cross_attn_ln_bias_to_fp16, epsilon = var_2264_to_fp16, gamma = blocks_10_cross_attn_ln_weight_to_fp16, x = x_189_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2374_to_fp16 = const()[name = string("op_2374_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412958912)))];
            tensor<fp16, [1024]> var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415056128)))];
            tensor<fp16, [1, ?, 1024]> linear_84_cast_fp16 = linear(bias = var_2375_to_fp16, weight = var_2374_to_fp16, x = var_2365_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [3]> concat_238 = const()[name = string("concat_238"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_239 = const()[name = string("concat_239"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_107_internal_tensor_assign_1_stride_0 = const()[name = string("k_107_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_107_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_107_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_107_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_107_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_107_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_107_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_107_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_238, begin_mask = k_107_internal_tensor_assign_1_begin_mask_0, end = concat_239, end_mask = k_107_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_107_internal_tensor_assign_1_squeeze_mask_0, stride = k_107_internal_tensor_assign_1_stride_0, update = k_cache_43_cast_fp16, x = k_7_to_fp16)[name = string("k_107_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_240 = const()[name = string("concat_240"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_241 = const()[name = string("concat_241"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_107_internal_tensor_assign_1_stride_0 = const()[name = string("v_107_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_107_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_107_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_107_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_107_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_107_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_107_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_107_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_240, begin_mask = v_107_internal_tensor_assign_1_begin_mask_0, end = concat_241, end_mask = v_107_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_107_internal_tensor_assign_1_squeeze_mask_0, stride = v_107_internal_tensor_assign_1_stride_0, update = v_cache_43_cast_fp16, x = k_7_to_fp16)[name = string("v_107_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_242x = const()[name = string("concat_242x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2395_cast_fp16 = reshape(shape = concat_242x, x = linear_84_cast_fp16)[name = string("op_2395_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_162_to_fp16 = const()[name = string("const_162_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_87_cast_fp16 = mul(x = var_2395_cast_fp16, y = const_162_to_fp16)[name = string("q_87_cast_fp16")];
            tensor<int32, [4]> var_2401 = const()[name = string("op_2401"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2402_cast_fp16 = reshape(shape = var_2401, x = k_107_internal_tensor_assign_1_cast_fp16)[name = string("op_2402_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_163_to_fp16 = const()[name = string("const_163_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_109_cast_fp16 = mul(x = var_2402_cast_fp16, y = const_163_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<int32, [4]> var_2408 = const()[name = string("op_2408"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2409_cast_fp16 = reshape(shape = var_2408, x = v_107_internal_tensor_assign_1_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<int32, [4]> var_2410 = const()[name = string("op_2410"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_65_transpose_x_0 = const()[name = string("qk_65_transpose_x_0"), val = bool(false)];
            bool qk_65_transpose_y_0 = const()[name = string("qk_65_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_235_perm_0 = const()[name = string("transpose_235_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_236_perm_0 = const()[name = string("transpose_236_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_236 = transpose(perm = transpose_236_perm_0, x = k_109_cast_fp16)[name = string("transpose_394")];
            tensor<fp16, [1, 16, ?, 64]> transpose_235 = transpose(perm = transpose_235_perm_0, x = q_87_cast_fp16)[name = string("transpose_395")];
            tensor<fp16, [1, 16, ?, 1500]> qk_65_cast_fp16 = matmul(transpose_x = qk_65_transpose_x_0, transpose_y = qk_65_transpose_y_0, x = transpose_235, y = transpose_236)[name = string("qk_65_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_2414_cast_fp16 = softmax(axis = var_2258, x = qk_65_cast_fp16)[name = string("op_2414_cast_fp16")];
            bool var_2416_transpose_x_0 = const()[name = string("op_2416_transpose_x_0"), val = bool(false)];
            bool var_2416_transpose_y_0 = const()[name = string("op_2416_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_109_cast_fp16 = transpose(perm = var_2410, x = var_2409_cast_fp16)[name = string("transpose_396")];
            tensor<fp16, [1, 16, ?, 64]> var_2416_cast_fp16 = matmul(transpose_x = var_2416_transpose_x_0, transpose_y = var_2416_transpose_y_0, x = var_2414_cast_fp16, y = v_109_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<int32, [4]> var_2417 = const()[name = string("op_2417"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_243x = const()[name = string("concat_243x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2418_cast_fp16 = transpose(perm = var_2417, x = var_2416_cast_fp16)[name = string("transpose_393")];
            tensor<fp16, [1, ?, 1024]> x_193_cast_fp16 = reshape(shape = concat_243x, x = var_2418_cast_fp16)[name = string("x_193_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2422_to_fp16 = const()[name = string("op_2422_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415058240)))];
            tensor<fp16, [1024]> var_2423_to_fp16 = const()[name = string("op_2423_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417155456)))];
            tensor<fp16, [1, ?, 1024]> linear_85_cast_fp16 = linear(bias = var_2423_to_fp16, weight = var_2422_to_fp16, x = x_193_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_195_cast_fp16 = add(x = x_189_cast_fp16, y = linear_85_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [1]> var_2430_axes_0 = const()[name = string("op_2430_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_10_mlp_ln_weight_to_fp16 = const()[name = string("blocks_10_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417157568)))];
            tensor<fp16, [1024]> blocks_10_mlp_ln_bias_to_fp16 = const()[name = string("blocks_10_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417159680)))];
            tensor<fp16, [1, ?, 1024]> var_2430_cast_fp16 = layer_norm(axes = var_2430_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_2264_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_195_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_2439_to_fp16 = const()[name = string("op_2439_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417161792)))];
            tensor<fp16, [4096]> var_2440_to_fp16 = const()[name = string("op_2440_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425550464)))];
            tensor<fp16, [1, ?, 4096]> linear_86_cast_fp16 = linear(bias = var_2440_to_fp16, weight = var_2439_to_fp16, x = var_2430_cast_fp16)[name = string("linear_86_cast_fp16")];
            string x_199_mode_0 = const()[name = string("x_199_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_199_cast_fp16 = gelu(mode = x_199_mode_0, x = linear_86_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_2445_to_fp16 = const()[name = string("op_2445_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425558720)))];
            tensor<fp16, [1024]> var_2446_to_fp16 = const()[name = string("op_2446_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433947392)))];
            tensor<fp16, [1, ?, 1024]> linear_87_cast_fp16 = linear(bias = var_2446_to_fp16, weight = var_2445_to_fp16, x = x_199_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_201_cast_fp16 = add(x = x_195_cast_fp16, y = linear_87_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [4]> k_cache_45_begin_0 = const()[name = string("k_cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> k_cache_45_end_0 = const()[name = string("k_cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_45_end_mask_0 = const()[name = string("k_cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_45_squeeze_mask_0 = const()[name = string("k_cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_45_cast_fp16 = slice_by_index(begin = k_cache_45_begin_0, end = k_cache_45_end_0, end_mask = k_cache_45_end_mask_0, squeeze_mask = k_cache_45_squeeze_mask_0, x = coreml_update_state_68)[name = string("k_cache_45_cast_fp16")];
            tensor<int32, [4]> v_cache_45_begin_0 = const()[name = string("v_cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> v_cache_45_end_0 = const()[name = string("v_cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_45_end_mask_0 = const()[name = string("v_cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_45_squeeze_mask_0 = const()[name = string("v_cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_45_cast_fp16 = slice_by_index(begin = v_cache_45_begin_0, end = v_cache_45_end_0, end_mask = v_cache_45_end_mask_0, squeeze_mask = v_cache_45_squeeze_mask_0, x = coreml_update_state_69)[name = string("v_cache_45_cast_fp16")];
            tensor<int32, [4]> k_cache_47_begin_0 = const()[name = string("k_cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> k_cache_47_end_0 = const()[name = string("k_cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_47_end_mask_0 = const()[name = string("k_cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_47_squeeze_mask_0 = const()[name = string("k_cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_47_cast_fp16 = slice_by_index(begin = k_cache_47_begin_0, end = k_cache_47_end_0, end_mask = k_cache_47_end_mask_0, squeeze_mask = k_cache_47_squeeze_mask_0, x = read_state_2)[name = string("k_cache_47_cast_fp16")];
            tensor<int32, [4]> v_cache_47_begin_0 = const()[name = string("v_cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> v_cache_47_end_0 = const()[name = string("v_cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_47_end_mask_0 = const()[name = string("v_cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_47_squeeze_mask_0 = const()[name = string("v_cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_47_cast_fp16 = slice_by_index(begin = v_cache_47_begin_0, end = v_cache_47_end_0, end_mask = v_cache_47_end_mask_0, squeeze_mask = v_cache_47_squeeze_mask_0, x = read_state_3)[name = string("v_cache_47_cast_fp16")];
            int32 var_2469 = const()[name = string("op_2469"), val = int32(-1)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_11_attn_ln_weight_to_fp16 = const()[name = string("blocks_11_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433949504)))];
            tensor<fp16, [1024]> blocks_11_attn_ln_bias_to_fp16 = const()[name = string("blocks_11_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433951616)))];
            fp16 var_2475_to_fp16 = const()[name = string("op_2475_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_2487_cast_fp16 = layer_norm(axes = var_2487_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_2475_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_201_cast_fp16)[name = string("op_2487_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2498_to_fp16 = const()[name = string("op_2498_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433953728)))];
            tensor<fp16, [1024]> var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436050944)))];
            tensor<fp16, [1, ?, 1024]> linear_88_cast_fp16 = linear(bias = var_2499_to_fp16, weight = var_2498_to_fp16, x = var_2487_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2502_to_fp16 = const()[name = string("op_2502_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436053056)))];
            tensor<fp16, [1, ?, 1024]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2502_to_fp16, x = var_2487_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2506_to_fp16 = const()[name = string("op_2506_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438150272)))];
            tensor<fp16, [1024]> var_2507_to_fp16 = const()[name = string("op_2507_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440247488)))];
            tensor<fp16, [1, ?, 1024]> linear_90_cast_fp16 = linear(bias = var_2507_to_fp16, weight = var_2506_to_fp16, x = var_2487_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<int32, [3]> var_2509_shape_cast_fp16 = shape(x = linear_88_cast_fp16)[name = string("op_2509_shape_cast_fp16")];
            int32 gather_134_axis_0 = const()[name = string("gather_134_axis_0"), val = int32(0)];
            int32 gather_134_batch_dims_0 = const()[name = string("gather_134_batch_dims_0"), val = int32(0)];
            bool gather_134_validate_indices_0 = const()[name = string("gather_134_validate_indices_0"), val = bool(false)];
            string var_2509_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2509_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_134_to_uint16 = const()[name = string("select_134_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2509_shape_cast_fp16_to_uint16 = cast(dtype = var_2509_shape_cast_fp16_to_uint16_dtype_0, x = var_2509_shape_cast_fp16)[name = string("cast_272")];
            uint16 gather_134_cast_uint16 = gather(axis = gather_134_axis_0, batch_dims = gather_134_batch_dims_0, indices = select_134_to_uint16, validate_indices = gather_134_validate_indices_0, x = var_2509_shape_cast_fp16_to_uint16)[name = string("gather_134_cast_uint16")];
            string gather_134_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_134_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_134_cast_uint16_to_int32 = cast(dtype = gather_134_cast_uint16_to_int32_dtype_0, x = gather_134_cast_uint16)[name = string("cast_271")];
            int32 end_step_25 = add(x = offset, y = gather_134_cast_uint16_to_int32)[name = string("end_step_25")];
            tensor<int32, [1]> expand_dims_176 = const()[name = string("expand_dims_176"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_178 = const()[name = string("expand_dims_178"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_179_axes_0 = const()[name = string("expand_dims_179_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_179 = expand_dims(axes = expand_dims_179_axes_0, x = end_step_25)[name = string("expand_dims_179")];
            tensor<int32, [1]> concat_246_values0_0 = const()[name = string("concat_246_values0_0"), val = tensor<int32, [1]>([11])];
            int32 concat_246_axis_0 = const()[name = string("concat_246_axis_0"), val = int32(0)];
            bool concat_246_interleave_0 = const()[name = string("concat_246_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_246 = concat(axis = concat_246_axis_0, interleave = concat_246_interleave_0, values = (concat_246_values0_0, expand_dims_176, expand_dims_1, expand_dims_178))[name = string("concat_246")];
            tensor<int32, [1]> concat_247_values0_0 = const()[name = string("concat_247_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_247_values1_0 = const()[name = string("concat_247_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_247_values3_0 = const()[name = string("concat_247_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_247_axis_0 = const()[name = string("concat_247_axis_0"), val = int32(0)];
            bool concat_247_interleave_0 = const()[name = string("concat_247_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_247 = concat(axis = concat_247_axis_0, interleave = concat_247_interleave_0, values = (concat_247_values0_0, concat_247_values1_0, expand_dims_179, concat_247_values3_0))[name = string("concat_247")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_12_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_12_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_12_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_246, begin_mask = k_cache1_internal_tensor_assign_12_begin_mask_0, end = concat_247, end_mask = k_cache1_internal_tensor_assign_12_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_12_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_12_stride_0, update = linear_89_cast_fp16, x = coreml_update_state_68)[name = string("k_cache1_internal_tensor_assign_12_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_12_cast_fp16, input = k_cache1)[name = string("coreml_update_state_70_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_70 = read_state(input = k_cache1)[name = string("coreml_update_state_70")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_12_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_12_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_12_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_246, begin_mask = v_cache1_internal_tensor_assign_12_begin_mask_0, end = concat_247, end_mask = v_cache1_internal_tensor_assign_12_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_12_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_12_stride_0, update = linear_90_cast_fp16, x = coreml_update_state_69)[name = string("v_cache1_internal_tensor_assign_12_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_12_cast_fp16, input = v_cache1)[name = string("coreml_update_state_71_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_71 = read_state(input = v_cache1)[name = string("coreml_update_state_71")];
            int32 concat_252_values0_0 = const()[name = string("concat_252_values0_0"), val = int32(1)];
            int32 concat_252_values2_0 = const()[name = string("concat_252_values2_0"), val = int32(1024)];
            int32 concat_252_axis_0 = const()[name = string("concat_252_axis_0"), val = int32(0)];
            bool concat_252_interleave_0 = const()[name = string("concat_252_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_252 = concat(axis = concat_252_axis_0, interleave = concat_252_interleave_0, values = (concat_252_values0_0, end_step_25, concat_252_values2_0))[name = string("concat_252")];
            tensor<int32, [3]> var_2525_begin_0 = const()[name = string("op_2525_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2525_end_mask_0 = const()[name = string("op_2525_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2525_cast_fp16 = slice_by_index(begin = var_2525_begin_0, end = concat_252, end_mask = var_2525_end_mask_0, x = k_cache_45_cast_fp16)[name = string("op_2525_cast_fp16")];
            tensor<int32, [3]> var_2528_begin_0 = const()[name = string("op_2528_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2528_end_mask_0 = const()[name = string("op_2528_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2528_cast_fp16 = slice_by_index(begin = var_2528_begin_0, end = concat_252, end_mask = var_2528_end_mask_0, x = v_cache_45_cast_fp16)[name = string("op_2528_cast_fp16")];
            tensor<int32, [4]> concat_254x = const()[name = string("concat_254x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2538_cast_fp16 = reshape(shape = concat_254x, x = linear_88_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_164_to_fp16 = const()[name = string("const_164_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_91_cast_fp16 = mul(x = var_2538_cast_fp16, y = const_164_to_fp16)[name = string("q_91_cast_fp16")];
            tensor<int32, [4]> concat_255x = const()[name = string("concat_255x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2545_cast_fp16 = reshape(shape = concat_255x, x = var_2525_cast_fp16)[name = string("op_2545_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_165_to_fp16 = const()[name = string("const_165_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_115_cast_fp16 = mul(x = var_2545_cast_fp16, y = const_165_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<int32, [4]> concat_256x = const()[name = string("concat_256x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2552_cast_fp16 = reshape(shape = concat_256x, x = var_2528_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> var_2553 = const()[name = string("op_2553"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_67_transpose_x_0 = const()[name = string("qk_67_transpose_x_0"), val = bool(false)];
            bool qk_67_transpose_y_0 = const()[name = string("qk_67_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_237_perm_0 = const()[name = string("transpose_237_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_238_perm_0 = const()[name = string("transpose_238_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_238 = transpose(perm = transpose_238_perm_0, x = k_115_cast_fp16)[name = string("transpose_390")];
            tensor<fp16, [1, 16, ?, 64]> transpose_237 = transpose(perm = transpose_237_perm_0, x = q_91_cast_fp16)[name = string("transpose_391")];
            tensor<fp16, [1, 16, ?, ?]> qk_67_cast_fp16 = matmul(transpose_x = qk_67_transpose_x_0, transpose_y = qk_67_transpose_y_0, x = transpose_237, y = transpose_238)[name = string("qk_67_cast_fp16")];
            int32 concat_257_values1_0 = const()[name = string("concat_257_values1_0"), val = int32(448)];
            int32 concat_257_axis_0 = const()[name = string("concat_257_axis_0"), val = int32(0)];
            bool concat_257_interleave_0 = const()[name = string("concat_257_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_257 = concat(axis = concat_257_axis_0, interleave = concat_257_interleave_0, values = (gather_134_cast_uint16_to_int32, concat_257_values1_0))[name = string("concat_257")];
            tensor<int32, [2]> var_2556_begin_0 = const()[name = string("op_2556_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2556_end_mask_0 = const()[name = string("op_2556_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_2556_cast_fp16 = slice_by_index(begin = var_2556_begin_0, end = concat_257, end_mask = var_2556_end_mask_0, x = mask_to_fp16)[name = string("op_2556_cast_fp16")];
            int32 concat_258_values0_0 = const()[name = string("concat_258_values0_0"), val = int32(0)];
            int32 concat_258_axis_0 = const()[name = string("concat_258_axis_0"), val = int32(0)];
            bool concat_258_interleave_0 = const()[name = string("concat_258_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_258 = concat(axis = concat_258_axis_0, interleave = concat_258_interleave_0, values = (concat_258_values0_0, gather_134_cast_uint16_to_int32))[name = string("concat_258")];
            tensor<int32, [2]> var_2557_begin_0 = const()[name = string("op_2557_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2557_end_mask_0 = const()[name = string("op_2557_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_2557_cast_fp16 = slice_by_index(begin = var_2557_begin_0, end = concat_258, end_mask = var_2557_end_mask_0, x = var_2556_cast_fp16)[name = string("op_2557_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_69_cast_fp16 = add(x = qk_67_cast_fp16, y = var_2557_cast_fp16)[name = string("qk_69_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_2560_cast_fp16 = softmax(axis = var_2469, x = qk_69_cast_fp16)[name = string("op_2560_cast_fp16")];
            bool var_2562_transpose_x_0 = const()[name = string("op_2562_transpose_x_0"), val = bool(false)];
            bool var_2562_transpose_y_0 = const()[name = string("op_2562_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_115_cast_fp16 = transpose(perm = var_2553, x = var_2552_cast_fp16)[name = string("transpose_392")];
            tensor<fp16, [1, 16, ?, 64]> var_2562_cast_fp16 = matmul(transpose_x = var_2562_transpose_x_0, transpose_y = var_2562_transpose_y_0, x = var_2560_cast_fp16, y = v_115_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_259x = const()[name = string("concat_259x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2564_cast_fp16 = transpose(perm = var_2563, x = var_2562_cast_fp16)[name = string("transpose_389")];
            tensor<fp16, [1, ?, 1024]> x_205_cast_fp16 = reshape(shape = concat_259x, x = var_2564_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2568_to_fp16 = const()[name = string("op_2568_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440249600)))];
            tensor<fp16, [1024]> var_2569_to_fp16 = const()[name = string("op_2569_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442346816)))];
            tensor<fp16, [1, ?, 1024]> linear_91_cast_fp16 = linear(bias = var_2569_to_fp16, weight = var_2568_to_fp16, x = x_205_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_207_cast_fp16 = add(x = x_201_cast_fp16, y = linear_91_cast_fp16)[name = string("x_207_cast_fp16")];
            tensor<int32, [1]> var_2576_axes_0 = const()[name = string("op_2576_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_11_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_11_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442348928)))];
            tensor<fp16, [1024]> blocks_11_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_11_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442351040)))];
            tensor<fp16, [1, ?, 1024]> var_2576_cast_fp16 = layer_norm(axes = var_2576_axes_0, beta = blocks_11_cross_attn_ln_bias_to_fp16, epsilon = var_2475_to_fp16, gamma = blocks_11_cross_attn_ln_weight_to_fp16, x = x_207_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442353152)))];
            tensor<fp16, [1024]> var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444450368)))];
            tensor<fp16, [1, ?, 1024]> linear_92_cast_fp16 = linear(bias = var_2586_to_fp16, weight = var_2585_to_fp16, x = var_2576_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [3]> concat_260 = const()[name = string("concat_260"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_261 = const()[name = string("concat_261"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_117_internal_tensor_assign_1_stride_0 = const()[name = string("k_117_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_117_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_117_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_117_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_117_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_117_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_117_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_117_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_260, begin_mask = k_117_internal_tensor_assign_1_begin_mask_0, end = concat_261, end_mask = k_117_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_117_internal_tensor_assign_1_squeeze_mask_0, stride = k_117_internal_tensor_assign_1_stride_0, update = k_cache_47_cast_fp16, x = k_7_to_fp16)[name = string("k_117_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_262 = const()[name = string("concat_262"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_263 = const()[name = string("concat_263"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_117_internal_tensor_assign_1_stride_0 = const()[name = string("v_117_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_117_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_117_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_117_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_117_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_117_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_117_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_117_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_262, begin_mask = v_117_internal_tensor_assign_1_begin_mask_0, end = concat_263, end_mask = v_117_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_117_internal_tensor_assign_1_squeeze_mask_0, stride = v_117_internal_tensor_assign_1_stride_0, update = v_cache_47_cast_fp16, x = k_7_to_fp16)[name = string("v_117_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_264x = const()[name = string("concat_264x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2606_cast_fp16 = reshape(shape = concat_264x, x = linear_92_cast_fp16)[name = string("op_2606_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_166_to_fp16 = const()[name = string("const_166_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_95_cast_fp16 = mul(x = var_2606_cast_fp16, y = const_166_to_fp16)[name = string("q_95_cast_fp16")];
            tensor<int32, [4]> var_2612 = const()[name = string("op_2612"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2613_cast_fp16 = reshape(shape = var_2612, x = k_117_internal_tensor_assign_1_cast_fp16)[name = string("op_2613_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_167_to_fp16 = const()[name = string("const_167_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_119_cast_fp16 = mul(x = var_2613_cast_fp16, y = const_167_to_fp16)[name = string("k_119_cast_fp16")];
            tensor<int32, [4]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2620_cast_fp16 = reshape(shape = var_2619, x = v_117_internal_tensor_assign_1_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [4]> var_2621 = const()[name = string("op_2621"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_71_transpose_x_0 = const()[name = string("qk_71_transpose_x_0"), val = bool(false)];
            bool qk_71_transpose_y_0 = const()[name = string("qk_71_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_239_perm_0 = const()[name = string("transpose_239_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_240_perm_0 = const()[name = string("transpose_240_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_240 = transpose(perm = transpose_240_perm_0, x = k_119_cast_fp16)[name = string("transpose_386")];
            tensor<fp16, [1, 16, ?, 64]> transpose_239 = transpose(perm = transpose_239_perm_0, x = q_95_cast_fp16)[name = string("transpose_387")];
            tensor<fp16, [1, 16, ?, 1500]> qk_71_cast_fp16 = matmul(transpose_x = qk_71_transpose_x_0, transpose_y = qk_71_transpose_y_0, x = transpose_239, y = transpose_240)[name = string("qk_71_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_2625_cast_fp16 = softmax(axis = var_2469, x = qk_71_cast_fp16)[name = string("op_2625_cast_fp16")];
            bool var_2627_transpose_x_0 = const()[name = string("op_2627_transpose_x_0"), val = bool(false)];
            bool var_2627_transpose_y_0 = const()[name = string("op_2627_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_119_cast_fp16 = transpose(perm = var_2621, x = var_2620_cast_fp16)[name = string("transpose_388")];
            tensor<fp16, [1, 16, ?, 64]> var_2627_cast_fp16 = matmul(transpose_x = var_2627_transpose_x_0, transpose_y = var_2627_transpose_y_0, x = var_2625_cast_fp16, y = v_119_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [4]> var_2628 = const()[name = string("op_2628"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_265x = const()[name = string("concat_265x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2629_cast_fp16 = transpose(perm = var_2628, x = var_2627_cast_fp16)[name = string("transpose_385")];
            tensor<fp16, [1, ?, 1024]> x_211_cast_fp16 = reshape(shape = concat_265x, x = var_2629_cast_fp16)[name = string("x_211_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2633_to_fp16 = const()[name = string("op_2633_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444452480)))];
            tensor<fp16, [1024]> var_2634_to_fp16 = const()[name = string("op_2634_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446549696)))];
            tensor<fp16, [1, ?, 1024]> linear_93_cast_fp16 = linear(bias = var_2634_to_fp16, weight = var_2633_to_fp16, x = x_211_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_213_cast_fp16 = add(x = x_207_cast_fp16, y = linear_93_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [1]> var_2641_axes_0 = const()[name = string("op_2641_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_11_mlp_ln_weight_to_fp16 = const()[name = string("blocks_11_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446551808)))];
            tensor<fp16, [1024]> blocks_11_mlp_ln_bias_to_fp16 = const()[name = string("blocks_11_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446553920)))];
            tensor<fp16, [1, ?, 1024]> var_2641_cast_fp16 = layer_norm(axes = var_2641_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_2475_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_213_cast_fp16)[name = string("op_2641_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_2650_to_fp16 = const()[name = string("op_2650_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446556032)))];
            tensor<fp16, [4096]> var_2651_to_fp16 = const()[name = string("op_2651_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454944704)))];
            tensor<fp16, [1, ?, 4096]> linear_94_cast_fp16 = linear(bias = var_2651_to_fp16, weight = var_2650_to_fp16, x = var_2641_cast_fp16)[name = string("linear_94_cast_fp16")];
            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_217_cast_fp16 = gelu(mode = x_217_mode_0, x = linear_94_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_2656_to_fp16 = const()[name = string("op_2656_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454952960)))];
            tensor<fp16, [1024]> var_2657_to_fp16 = const()[name = string("op_2657_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463341632)))];
            tensor<fp16, [1, ?, 1024]> linear_95_cast_fp16 = linear(bias = var_2657_to_fp16, weight = var_2656_to_fp16, x = x_217_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_219_cast_fp16 = add(x = x_213_cast_fp16, y = linear_95_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<int32, [4]> k_cache_49_begin_0 = const()[name = string("k_cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> k_cache_49_end_0 = const()[name = string("k_cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_49_end_mask_0 = const()[name = string("k_cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_49_squeeze_mask_0 = const()[name = string("k_cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_49_cast_fp16 = slice_by_index(begin = k_cache_49_begin_0, end = k_cache_49_end_0, end_mask = k_cache_49_end_mask_0, squeeze_mask = k_cache_49_squeeze_mask_0, x = coreml_update_state_70)[name = string("k_cache_49_cast_fp16")];
            tensor<int32, [4]> v_cache_49_begin_0 = const()[name = string("v_cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> v_cache_49_end_0 = const()[name = string("v_cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_49_end_mask_0 = const()[name = string("v_cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_49_squeeze_mask_0 = const()[name = string("v_cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_49_cast_fp16 = slice_by_index(begin = v_cache_49_begin_0, end = v_cache_49_end_0, end_mask = v_cache_49_end_mask_0, squeeze_mask = v_cache_49_squeeze_mask_0, x = coreml_update_state_71)[name = string("v_cache_49_cast_fp16")];
            tensor<int32, [4]> k_cache_51_begin_0 = const()[name = string("k_cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> k_cache_51_end_0 = const()[name = string("k_cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_51_end_mask_0 = const()[name = string("k_cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_51_squeeze_mask_0 = const()[name = string("k_cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_51_cast_fp16 = slice_by_index(begin = k_cache_51_begin_0, end = k_cache_51_end_0, end_mask = k_cache_51_end_mask_0, squeeze_mask = k_cache_51_squeeze_mask_0, x = read_state_2)[name = string("k_cache_51_cast_fp16")];
            tensor<int32, [4]> v_cache_51_begin_0 = const()[name = string("v_cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> v_cache_51_end_0 = const()[name = string("v_cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_51_end_mask_0 = const()[name = string("v_cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_51_squeeze_mask_0 = const()[name = string("v_cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_51_cast_fp16 = slice_by_index(begin = v_cache_51_begin_0, end = v_cache_51_end_0, end_mask = v_cache_51_end_mask_0, squeeze_mask = v_cache_51_squeeze_mask_0, x = read_state_3)[name = string("v_cache_51_cast_fp16")];
            int32 var_2680 = const()[name = string("op_2680"), val = int32(-1)];
            tensor<int32, [1]> var_2698_axes_0 = const()[name = string("op_2698_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_12_attn_ln_weight_to_fp16 = const()[name = string("blocks_12_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463343744)))];
            tensor<fp16, [1024]> blocks_12_attn_ln_bias_to_fp16 = const()[name = string("blocks_12_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463345856)))];
            fp16 var_2686_to_fp16 = const()[name = string("op_2686_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_2698_cast_fp16 = layer_norm(axes = var_2698_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_2686_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_219_cast_fp16)[name = string("op_2698_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2709_to_fp16 = const()[name = string("op_2709_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463347968)))];
            tensor<fp16, [1024]> var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465445184)))];
            tensor<fp16, [1, ?, 1024]> linear_96_cast_fp16 = linear(bias = var_2710_to_fp16, weight = var_2709_to_fp16, x = var_2698_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2713_to_fp16 = const()[name = string("op_2713_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465447296)))];
            tensor<fp16, [1, ?, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2713_to_fp16, x = var_2698_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467544512)))];
            tensor<fp16, [1024]> var_2718_to_fp16 = const()[name = string("op_2718_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469641728)))];
            tensor<fp16, [1, ?, 1024]> linear_98_cast_fp16 = linear(bias = var_2718_to_fp16, weight = var_2717_to_fp16, x = var_2698_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [3]> var_2720_shape_cast_fp16 = shape(x = linear_96_cast_fp16)[name = string("op_2720_shape_cast_fp16")];
            int32 gather_146_axis_0 = const()[name = string("gather_146_axis_0"), val = int32(0)];
            int32 gather_146_batch_dims_0 = const()[name = string("gather_146_batch_dims_0"), val = int32(0)];
            bool gather_146_validate_indices_0 = const()[name = string("gather_146_validate_indices_0"), val = bool(false)];
            string var_2720_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2720_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_146_to_uint16 = const()[name = string("select_146_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2720_shape_cast_fp16_to_uint16 = cast(dtype = var_2720_shape_cast_fp16_to_uint16_dtype_0, x = var_2720_shape_cast_fp16)[name = string("cast_270")];
            uint16 gather_146_cast_uint16 = gather(axis = gather_146_axis_0, batch_dims = gather_146_batch_dims_0, indices = select_146_to_uint16, validate_indices = gather_146_validate_indices_0, x = var_2720_shape_cast_fp16_to_uint16)[name = string("gather_146_cast_uint16")];
            string gather_146_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_146_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_146_cast_uint16_to_int32 = cast(dtype = gather_146_cast_uint16_to_int32_dtype_0, x = gather_146_cast_uint16)[name = string("cast_269")];
            int32 end_step_27 = add(x = offset, y = gather_146_cast_uint16_to_int32)[name = string("end_step_27")];
            tensor<int32, [1]> expand_dims_192 = const()[name = string("expand_dims_192"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_194 = const()[name = string("expand_dims_194"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_195_axes_0 = const()[name = string("expand_dims_195_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_195 = expand_dims(axes = expand_dims_195_axes_0, x = end_step_27)[name = string("expand_dims_195")];
            tensor<int32, [1]> concat_268_values0_0 = const()[name = string("concat_268_values0_0"), val = tensor<int32, [1]>([12])];
            int32 concat_268_axis_0 = const()[name = string("concat_268_axis_0"), val = int32(0)];
            bool concat_268_interleave_0 = const()[name = string("concat_268_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_268 = concat(axis = concat_268_axis_0, interleave = concat_268_interleave_0, values = (concat_268_values0_0, expand_dims_192, expand_dims_1, expand_dims_194))[name = string("concat_268")];
            tensor<int32, [1]> concat_269_values0_0 = const()[name = string("concat_269_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_269_values1_0 = const()[name = string("concat_269_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_269_values3_0 = const()[name = string("concat_269_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_269_axis_0 = const()[name = string("concat_269_axis_0"), val = int32(0)];
            bool concat_269_interleave_0 = const()[name = string("concat_269_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_269 = concat(axis = concat_269_axis_0, interleave = concat_269_interleave_0, values = (concat_269_values0_0, concat_269_values1_0, expand_dims_195, concat_269_values3_0))[name = string("concat_269")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_13_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_13_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_13_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_268, begin_mask = k_cache1_internal_tensor_assign_13_begin_mask_0, end = concat_269, end_mask = k_cache1_internal_tensor_assign_13_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_13_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_13_stride_0, update = linear_97_cast_fp16, x = coreml_update_state_70)[name = string("k_cache1_internal_tensor_assign_13_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_13_cast_fp16, input = k_cache1)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_72 = read_state(input = k_cache1)[name = string("coreml_update_state_72")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_13_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_13_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_13_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_268, begin_mask = v_cache1_internal_tensor_assign_13_begin_mask_0, end = concat_269, end_mask = v_cache1_internal_tensor_assign_13_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_13_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_13_stride_0, update = linear_98_cast_fp16, x = coreml_update_state_71)[name = string("v_cache1_internal_tensor_assign_13_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_13_cast_fp16, input = v_cache1)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_73 = read_state(input = v_cache1)[name = string("coreml_update_state_73")];
            int32 concat_274_values0_0 = const()[name = string("concat_274_values0_0"), val = int32(1)];
            int32 concat_274_values2_0 = const()[name = string("concat_274_values2_0"), val = int32(1024)];
            int32 concat_274_axis_0 = const()[name = string("concat_274_axis_0"), val = int32(0)];
            bool concat_274_interleave_0 = const()[name = string("concat_274_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_274 = concat(axis = concat_274_axis_0, interleave = concat_274_interleave_0, values = (concat_274_values0_0, end_step_27, concat_274_values2_0))[name = string("concat_274")];
            tensor<int32, [3]> var_2736_begin_0 = const()[name = string("op_2736_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2736_end_mask_0 = const()[name = string("op_2736_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2736_cast_fp16 = slice_by_index(begin = var_2736_begin_0, end = concat_274, end_mask = var_2736_end_mask_0, x = k_cache_49_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<int32, [3]> var_2739_begin_0 = const()[name = string("op_2739_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2739_end_mask_0 = const()[name = string("op_2739_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2739_cast_fp16 = slice_by_index(begin = var_2739_begin_0, end = concat_274, end_mask = var_2739_end_mask_0, x = v_cache_49_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [4]> concat_276x = const()[name = string("concat_276x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2749_cast_fp16 = reshape(shape = concat_276x, x = linear_96_cast_fp16)[name = string("op_2749_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_168_to_fp16 = const()[name = string("const_168_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_99_cast_fp16 = mul(x = var_2749_cast_fp16, y = const_168_to_fp16)[name = string("q_99_cast_fp16")];
            tensor<int32, [4]> concat_277x = const()[name = string("concat_277x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2756_cast_fp16 = reshape(shape = concat_277x, x = var_2736_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_169_to_fp16 = const()[name = string("const_169_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_125_cast_fp16 = mul(x = var_2756_cast_fp16, y = const_169_to_fp16)[name = string("k_125_cast_fp16")];
            tensor<int32, [4]> concat_278x = const()[name = string("concat_278x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2763_cast_fp16 = reshape(shape = concat_278x, x = var_2739_cast_fp16)[name = string("op_2763_cast_fp16")];
            tensor<int32, [4]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_73_transpose_x_0 = const()[name = string("qk_73_transpose_x_0"), val = bool(false)];
            bool qk_73_transpose_y_0 = const()[name = string("qk_73_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_241_perm_0 = const()[name = string("transpose_241_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_242_perm_0 = const()[name = string("transpose_242_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_242 = transpose(perm = transpose_242_perm_0, x = k_125_cast_fp16)[name = string("transpose_382")];
            tensor<fp16, [1, 16, ?, 64]> transpose_241 = transpose(perm = transpose_241_perm_0, x = q_99_cast_fp16)[name = string("transpose_383")];
            tensor<fp16, [1, 16, ?, ?]> qk_73_cast_fp16 = matmul(transpose_x = qk_73_transpose_x_0, transpose_y = qk_73_transpose_y_0, x = transpose_241, y = transpose_242)[name = string("qk_73_cast_fp16")];
            int32 concat_279_values1_0 = const()[name = string("concat_279_values1_0"), val = int32(448)];
            int32 concat_279_axis_0 = const()[name = string("concat_279_axis_0"), val = int32(0)];
            bool concat_279_interleave_0 = const()[name = string("concat_279_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_279 = concat(axis = concat_279_axis_0, interleave = concat_279_interleave_0, values = (gather_146_cast_uint16_to_int32, concat_279_values1_0))[name = string("concat_279")];
            tensor<int32, [2]> var_2767_begin_0 = const()[name = string("op_2767_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2767_end_mask_0 = const()[name = string("op_2767_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_2767_cast_fp16 = slice_by_index(begin = var_2767_begin_0, end = concat_279, end_mask = var_2767_end_mask_0, x = mask_to_fp16)[name = string("op_2767_cast_fp16")];
            int32 concat_280_values0_0 = const()[name = string("concat_280_values0_0"), val = int32(0)];
            int32 concat_280_axis_0 = const()[name = string("concat_280_axis_0"), val = int32(0)];
            bool concat_280_interleave_0 = const()[name = string("concat_280_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_280 = concat(axis = concat_280_axis_0, interleave = concat_280_interleave_0, values = (concat_280_values0_0, gather_146_cast_uint16_to_int32))[name = string("concat_280")];
            tensor<int32, [2]> var_2768_begin_0 = const()[name = string("op_2768_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2768_end_mask_0 = const()[name = string("op_2768_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_2768_cast_fp16 = slice_by_index(begin = var_2768_begin_0, end = concat_280, end_mask = var_2768_end_mask_0, x = var_2767_cast_fp16)[name = string("op_2768_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_75_cast_fp16 = add(x = qk_73_cast_fp16, y = var_2768_cast_fp16)[name = string("qk_75_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_2771_cast_fp16 = softmax(axis = var_2680, x = qk_75_cast_fp16)[name = string("op_2771_cast_fp16")];
            bool var_2773_transpose_x_0 = const()[name = string("op_2773_transpose_x_0"), val = bool(false)];
            bool var_2773_transpose_y_0 = const()[name = string("op_2773_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_125_cast_fp16 = transpose(perm = var_2764, x = var_2763_cast_fp16)[name = string("transpose_384")];
            tensor<fp16, [1, 16, ?, 64]> var_2773_cast_fp16 = matmul(transpose_x = var_2773_transpose_x_0, transpose_y = var_2773_transpose_y_0, x = var_2771_cast_fp16, y = v_125_cast_fp16)[name = string("op_2773_cast_fp16")];
            tensor<int32, [4]> var_2774 = const()[name = string("op_2774"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_281x = const()[name = string("concat_281x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2775_cast_fp16 = transpose(perm = var_2774, x = var_2773_cast_fp16)[name = string("transpose_381")];
            tensor<fp16, [1, ?, 1024]> x_223_cast_fp16 = reshape(shape = concat_281x, x = var_2775_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2779_to_fp16 = const()[name = string("op_2779_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469643840)))];
            tensor<fp16, [1024]> var_2780_to_fp16 = const()[name = string("op_2780_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471741056)))];
            tensor<fp16, [1, ?, 1024]> linear_99_cast_fp16 = linear(bias = var_2780_to_fp16, weight = var_2779_to_fp16, x = x_223_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_225_cast_fp16 = add(x = x_219_cast_fp16, y = linear_99_cast_fp16)[name = string("x_225_cast_fp16")];
            tensor<int32, [1]> var_2787_axes_0 = const()[name = string("op_2787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_12_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_12_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471743168)))];
            tensor<fp16, [1024]> blocks_12_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_12_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471745280)))];
            tensor<fp16, [1, ?, 1024]> var_2787_cast_fp16 = layer_norm(axes = var_2787_axes_0, beta = blocks_12_cross_attn_ln_bias_to_fp16, epsilon = var_2686_to_fp16, gamma = blocks_12_cross_attn_ln_weight_to_fp16, x = x_225_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2796_to_fp16 = const()[name = string("op_2796_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471747392)))];
            tensor<fp16, [1024]> var_2797_to_fp16 = const()[name = string("op_2797_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473844608)))];
            tensor<fp16, [1, ?, 1024]> linear_100_cast_fp16 = linear(bias = var_2797_to_fp16, weight = var_2796_to_fp16, x = var_2787_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [3]> concat_282 = const()[name = string("concat_282"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_283 = const()[name = string("concat_283"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_127_internal_tensor_assign_1_stride_0 = const()[name = string("k_127_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_127_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_127_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_127_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_127_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_127_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_127_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_127_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_282, begin_mask = k_127_internal_tensor_assign_1_begin_mask_0, end = concat_283, end_mask = k_127_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_127_internal_tensor_assign_1_squeeze_mask_0, stride = k_127_internal_tensor_assign_1_stride_0, update = k_cache_51_cast_fp16, x = k_7_to_fp16)[name = string("k_127_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_284 = const()[name = string("concat_284"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_285 = const()[name = string("concat_285"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_127_internal_tensor_assign_1_stride_0 = const()[name = string("v_127_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_127_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_127_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_127_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_127_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_127_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_127_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_127_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_284, begin_mask = v_127_internal_tensor_assign_1_begin_mask_0, end = concat_285, end_mask = v_127_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_127_internal_tensor_assign_1_squeeze_mask_0, stride = v_127_internal_tensor_assign_1_stride_0, update = v_cache_51_cast_fp16, x = k_7_to_fp16)[name = string("v_127_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_286x = const()[name = string("concat_286x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2817_cast_fp16 = reshape(shape = concat_286x, x = linear_100_cast_fp16)[name = string("op_2817_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_103_cast_fp16 = mul(x = var_2817_cast_fp16, y = const_170_to_fp16)[name = string("q_103_cast_fp16")];
            tensor<int32, [4]> var_2823 = const()[name = string("op_2823"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2824_cast_fp16 = reshape(shape = var_2823, x = k_127_internal_tensor_assign_1_cast_fp16)[name = string("op_2824_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_129_cast_fp16 = mul(x = var_2824_cast_fp16, y = const_171_to_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> var_2830 = const()[name = string("op_2830"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_2831_cast_fp16 = reshape(shape = var_2830, x = v_127_internal_tensor_assign_1_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<int32, [4]> var_2832 = const()[name = string("op_2832"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_77_transpose_x_0 = const()[name = string("qk_77_transpose_x_0"), val = bool(false)];
            bool qk_77_transpose_y_0 = const()[name = string("qk_77_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_243_perm_0 = const()[name = string("transpose_243_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_244_perm_0 = const()[name = string("transpose_244_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_244 = transpose(perm = transpose_244_perm_0, x = k_129_cast_fp16)[name = string("transpose_378")];
            tensor<fp16, [1, 16, ?, 64]> transpose_243 = transpose(perm = transpose_243_perm_0, x = q_103_cast_fp16)[name = string("transpose_379")];
            tensor<fp16, [1, 16, ?, 1500]> qk_77_cast_fp16 = matmul(transpose_x = qk_77_transpose_x_0, transpose_y = qk_77_transpose_y_0, x = transpose_243, y = transpose_244)[name = string("qk_77_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_2836_cast_fp16 = softmax(axis = var_2680, x = qk_77_cast_fp16)[name = string("op_2836_cast_fp16")];
            bool var_2838_transpose_x_0 = const()[name = string("op_2838_transpose_x_0"), val = bool(false)];
            bool var_2838_transpose_y_0 = const()[name = string("op_2838_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_129_cast_fp16 = transpose(perm = var_2832, x = var_2831_cast_fp16)[name = string("transpose_380")];
            tensor<fp16, [1, 16, ?, 64]> var_2838_cast_fp16 = matmul(transpose_x = var_2838_transpose_x_0, transpose_y = var_2838_transpose_y_0, x = var_2836_cast_fp16, y = v_129_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [4]> var_2839 = const()[name = string("op_2839"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_287x = const()[name = string("concat_287x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2840_cast_fp16 = transpose(perm = var_2839, x = var_2838_cast_fp16)[name = string("transpose_377")];
            tensor<fp16, [1, ?, 1024]> x_229_cast_fp16 = reshape(shape = concat_287x, x = var_2840_cast_fp16)[name = string("x_229_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473846720)))];
            tensor<fp16, [1024]> var_2845_to_fp16 = const()[name = string("op_2845_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475943936)))];
            tensor<fp16, [1, ?, 1024]> linear_101_cast_fp16 = linear(bias = var_2845_to_fp16, weight = var_2844_to_fp16, x = x_229_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_231_cast_fp16 = add(x = x_225_cast_fp16, y = linear_101_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_12_mlp_ln_weight_to_fp16 = const()[name = string("blocks_12_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475946048)))];
            tensor<fp16, [1024]> blocks_12_mlp_ln_bias_to_fp16 = const()[name = string("blocks_12_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475948160)))];
            tensor<fp16, [1, ?, 1024]> var_2852_cast_fp16 = layer_norm(axes = var_2852_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_2686_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_231_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475950272)))];
            tensor<fp16, [4096]> var_2862_to_fp16 = const()[name = string("op_2862_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484338944)))];
            tensor<fp16, [1, ?, 4096]> linear_102_cast_fp16 = linear(bias = var_2862_to_fp16, weight = var_2861_to_fp16, x = var_2852_cast_fp16)[name = string("linear_102_cast_fp16")];
            string x_235_mode_0 = const()[name = string("x_235_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_235_cast_fp16 = gelu(mode = x_235_mode_0, x = linear_102_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_2867_to_fp16 = const()[name = string("op_2867_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484347200)))];
            tensor<fp16, [1024]> var_2868_to_fp16 = const()[name = string("op_2868_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492735872)))];
            tensor<fp16, [1, ?, 1024]> linear_103_cast_fp16 = linear(bias = var_2868_to_fp16, weight = var_2867_to_fp16, x = x_235_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_237_cast_fp16 = add(x = x_231_cast_fp16, y = linear_103_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<int32, [4]> k_cache_53_begin_0 = const()[name = string("k_cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> k_cache_53_end_0 = const()[name = string("k_cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_53_end_mask_0 = const()[name = string("k_cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_53_squeeze_mask_0 = const()[name = string("k_cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_53_cast_fp16 = slice_by_index(begin = k_cache_53_begin_0, end = k_cache_53_end_0, end_mask = k_cache_53_end_mask_0, squeeze_mask = k_cache_53_squeeze_mask_0, x = coreml_update_state_72)[name = string("k_cache_53_cast_fp16")];
            tensor<int32, [4]> v_cache_53_begin_0 = const()[name = string("v_cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> v_cache_53_end_0 = const()[name = string("v_cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_53_end_mask_0 = const()[name = string("v_cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_53_squeeze_mask_0 = const()[name = string("v_cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_53_cast_fp16 = slice_by_index(begin = v_cache_53_begin_0, end = v_cache_53_end_0, end_mask = v_cache_53_end_mask_0, squeeze_mask = v_cache_53_squeeze_mask_0, x = coreml_update_state_73)[name = string("v_cache_53_cast_fp16")];
            tensor<int32, [4]> k_cache_55_begin_0 = const()[name = string("k_cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> k_cache_55_end_0 = const()[name = string("k_cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_55_end_mask_0 = const()[name = string("k_cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_55_squeeze_mask_0 = const()[name = string("k_cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_55_cast_fp16 = slice_by_index(begin = k_cache_55_begin_0, end = k_cache_55_end_0, end_mask = k_cache_55_end_mask_0, squeeze_mask = k_cache_55_squeeze_mask_0, x = read_state_2)[name = string("k_cache_55_cast_fp16")];
            tensor<int32, [4]> v_cache_55_begin_0 = const()[name = string("v_cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> v_cache_55_end_0 = const()[name = string("v_cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_55_end_mask_0 = const()[name = string("v_cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_55_squeeze_mask_0 = const()[name = string("v_cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_55_cast_fp16 = slice_by_index(begin = v_cache_55_begin_0, end = v_cache_55_end_0, end_mask = v_cache_55_end_mask_0, squeeze_mask = v_cache_55_squeeze_mask_0, x = read_state_3)[name = string("v_cache_55_cast_fp16")];
            int32 var_2891 = const()[name = string("op_2891"), val = int32(-1)];
            tensor<int32, [1]> var_2909_axes_0 = const()[name = string("op_2909_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_13_attn_ln_weight_to_fp16 = const()[name = string("blocks_13_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492737984)))];
            tensor<fp16, [1024]> blocks_13_attn_ln_bias_to_fp16 = const()[name = string("blocks_13_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492740096)))];
            fp16 var_2897_to_fp16 = const()[name = string("op_2897_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_2909_cast_fp16 = layer_norm(axes = var_2909_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_2897_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_237_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2920_to_fp16 = const()[name = string("op_2920_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492742208)))];
            tensor<fp16, [1024]> var_2921_to_fp16 = const()[name = string("op_2921_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494839424)))];
            tensor<fp16, [1, ?, 1024]> linear_104_cast_fp16 = linear(bias = var_2921_to_fp16, weight = var_2920_to_fp16, x = var_2909_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2924_to_fp16 = const()[name = string("op_2924_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494841536)))];
            tensor<fp16, [1, ?, 1024]> linear_105_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2924_to_fp16, x = var_2909_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2928_to_fp16 = const()[name = string("op_2928_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496938752)))];
            tensor<fp16, [1024]> var_2929_to_fp16 = const()[name = string("op_2929_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499035968)))];
            tensor<fp16, [1, ?, 1024]> linear_106_cast_fp16 = linear(bias = var_2929_to_fp16, weight = var_2928_to_fp16, x = var_2909_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [3]> var_2931_shape_cast_fp16 = shape(x = linear_104_cast_fp16)[name = string("op_2931_shape_cast_fp16")];
            int32 gather_158_axis_0 = const()[name = string("gather_158_axis_0"), val = int32(0)];
            int32 gather_158_batch_dims_0 = const()[name = string("gather_158_batch_dims_0"), val = int32(0)];
            bool gather_158_validate_indices_0 = const()[name = string("gather_158_validate_indices_0"), val = bool(false)];
            string var_2931_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2931_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_158_to_uint16 = const()[name = string("select_158_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2931_shape_cast_fp16_to_uint16 = cast(dtype = var_2931_shape_cast_fp16_to_uint16_dtype_0, x = var_2931_shape_cast_fp16)[name = string("cast_268")];
            uint16 gather_158_cast_uint16 = gather(axis = gather_158_axis_0, batch_dims = gather_158_batch_dims_0, indices = select_158_to_uint16, validate_indices = gather_158_validate_indices_0, x = var_2931_shape_cast_fp16_to_uint16)[name = string("gather_158_cast_uint16")];
            string gather_158_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_158_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_158_cast_uint16_to_int32 = cast(dtype = gather_158_cast_uint16_to_int32_dtype_0, x = gather_158_cast_uint16)[name = string("cast_267")];
            int32 end_step_29 = add(x = offset, y = gather_158_cast_uint16_to_int32)[name = string("end_step_29")];
            tensor<int32, [1]> expand_dims_208 = const()[name = string("expand_dims_208"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_210 = const()[name = string("expand_dims_210"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_211_axes_0 = const()[name = string("expand_dims_211_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_211 = expand_dims(axes = expand_dims_211_axes_0, x = end_step_29)[name = string("expand_dims_211")];
            tensor<int32, [1]> concat_290_values0_0 = const()[name = string("concat_290_values0_0"), val = tensor<int32, [1]>([13])];
            int32 concat_290_axis_0 = const()[name = string("concat_290_axis_0"), val = int32(0)];
            bool concat_290_interleave_0 = const()[name = string("concat_290_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_290 = concat(axis = concat_290_axis_0, interleave = concat_290_interleave_0, values = (concat_290_values0_0, expand_dims_208, expand_dims_1, expand_dims_210))[name = string("concat_290")];
            tensor<int32, [1]> concat_291_values0_0 = const()[name = string("concat_291_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_291_values1_0 = const()[name = string("concat_291_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_291_values3_0 = const()[name = string("concat_291_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_291_axis_0 = const()[name = string("concat_291_axis_0"), val = int32(0)];
            bool concat_291_interleave_0 = const()[name = string("concat_291_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_291 = concat(axis = concat_291_axis_0, interleave = concat_291_interleave_0, values = (concat_291_values0_0, concat_291_values1_0, expand_dims_211, concat_291_values3_0))[name = string("concat_291")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_14_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_14_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_14_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_290, begin_mask = k_cache1_internal_tensor_assign_14_begin_mask_0, end = concat_291, end_mask = k_cache1_internal_tensor_assign_14_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_14_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_14_stride_0, update = linear_105_cast_fp16, x = coreml_update_state_72)[name = string("k_cache1_internal_tensor_assign_14_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_14_cast_fp16, input = k_cache1)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_74 = read_state(input = k_cache1)[name = string("coreml_update_state_74")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_14_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_14_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_14_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_290, begin_mask = v_cache1_internal_tensor_assign_14_begin_mask_0, end = concat_291, end_mask = v_cache1_internal_tensor_assign_14_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_14_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_14_stride_0, update = linear_106_cast_fp16, x = coreml_update_state_73)[name = string("v_cache1_internal_tensor_assign_14_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_14_cast_fp16, input = v_cache1)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_75 = read_state(input = v_cache1)[name = string("coreml_update_state_75")];
            int32 concat_296_values0_0 = const()[name = string("concat_296_values0_0"), val = int32(1)];
            int32 concat_296_values2_0 = const()[name = string("concat_296_values2_0"), val = int32(1024)];
            int32 concat_296_axis_0 = const()[name = string("concat_296_axis_0"), val = int32(0)];
            bool concat_296_interleave_0 = const()[name = string("concat_296_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_296 = concat(axis = concat_296_axis_0, interleave = concat_296_interleave_0, values = (concat_296_values0_0, end_step_29, concat_296_values2_0))[name = string("concat_296")];
            tensor<int32, [3]> var_2947_begin_0 = const()[name = string("op_2947_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2947_end_mask_0 = const()[name = string("op_2947_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2947_cast_fp16 = slice_by_index(begin = var_2947_begin_0, end = concat_296, end_mask = var_2947_end_mask_0, x = k_cache_53_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<int32, [3]> var_2950_begin_0 = const()[name = string("op_2950_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_2950_end_mask_0 = const()[name = string("op_2950_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_2950_cast_fp16 = slice_by_index(begin = var_2950_begin_0, end = concat_296, end_mask = var_2950_end_mask_0, x = v_cache_53_cast_fp16)[name = string("op_2950_cast_fp16")];
            tensor<int32, [4]> concat_298x = const()[name = string("concat_298x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2960_cast_fp16 = reshape(shape = concat_298x, x = linear_104_cast_fp16)[name = string("op_2960_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_107_cast_fp16 = mul(x = var_2960_cast_fp16, y = const_172_to_fp16)[name = string("q_107_cast_fp16")];
            tensor<int32, [4]> concat_299x = const()[name = string("concat_299x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2967_cast_fp16 = reshape(shape = concat_299x, x = var_2947_cast_fp16)[name = string("op_2967_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_173_to_fp16 = const()[name = string("const_173_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_135_cast_fp16 = mul(x = var_2967_cast_fp16, y = const_173_to_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> concat_300x = const()[name = string("concat_300x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_2974_cast_fp16 = reshape(shape = concat_300x, x = var_2950_cast_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [4]> var_2975 = const()[name = string("op_2975"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_79_transpose_x_0 = const()[name = string("qk_79_transpose_x_0"), val = bool(false)];
            bool qk_79_transpose_y_0 = const()[name = string("qk_79_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_245_perm_0 = const()[name = string("transpose_245_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_246_perm_0 = const()[name = string("transpose_246_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_246 = transpose(perm = transpose_246_perm_0, x = k_135_cast_fp16)[name = string("transpose_374")];
            tensor<fp16, [1, 16, ?, 64]> transpose_245 = transpose(perm = transpose_245_perm_0, x = q_107_cast_fp16)[name = string("transpose_375")];
            tensor<fp16, [1, 16, ?, ?]> qk_79_cast_fp16 = matmul(transpose_x = qk_79_transpose_x_0, transpose_y = qk_79_transpose_y_0, x = transpose_245, y = transpose_246)[name = string("qk_79_cast_fp16")];
            int32 concat_301_values1_0 = const()[name = string("concat_301_values1_0"), val = int32(448)];
            int32 concat_301_axis_0 = const()[name = string("concat_301_axis_0"), val = int32(0)];
            bool concat_301_interleave_0 = const()[name = string("concat_301_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_301 = concat(axis = concat_301_axis_0, interleave = concat_301_interleave_0, values = (gather_158_cast_uint16_to_int32, concat_301_values1_0))[name = string("concat_301")];
            tensor<int32, [2]> var_2978_begin_0 = const()[name = string("op_2978_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2978_end_mask_0 = const()[name = string("op_2978_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_2978_cast_fp16 = slice_by_index(begin = var_2978_begin_0, end = concat_301, end_mask = var_2978_end_mask_0, x = mask_to_fp16)[name = string("op_2978_cast_fp16")];
            int32 concat_302_values0_0 = const()[name = string("concat_302_values0_0"), val = int32(0)];
            int32 concat_302_axis_0 = const()[name = string("concat_302_axis_0"), val = int32(0)];
            bool concat_302_interleave_0 = const()[name = string("concat_302_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_302 = concat(axis = concat_302_axis_0, interleave = concat_302_interleave_0, values = (concat_302_values0_0, gather_158_cast_uint16_to_int32))[name = string("concat_302")];
            tensor<int32, [2]> var_2979_begin_0 = const()[name = string("op_2979_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_2979_end_mask_0 = const()[name = string("op_2979_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_2979_cast_fp16 = slice_by_index(begin = var_2979_begin_0, end = concat_302, end_mask = var_2979_end_mask_0, x = var_2978_cast_fp16)[name = string("op_2979_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_81_cast_fp16 = add(x = qk_79_cast_fp16, y = var_2979_cast_fp16)[name = string("qk_81_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_2982_cast_fp16 = softmax(axis = var_2891, x = qk_81_cast_fp16)[name = string("op_2982_cast_fp16")];
            bool var_2984_transpose_x_0 = const()[name = string("op_2984_transpose_x_0"), val = bool(false)];
            bool var_2984_transpose_y_0 = const()[name = string("op_2984_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_135_cast_fp16 = transpose(perm = var_2975, x = var_2974_cast_fp16)[name = string("transpose_376")];
            tensor<fp16, [1, 16, ?, 64]> var_2984_cast_fp16 = matmul(transpose_x = var_2984_transpose_x_0, transpose_y = var_2984_transpose_y_0, x = var_2982_cast_fp16, y = v_135_cast_fp16)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2985 = const()[name = string("op_2985"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_303x = const()[name = string("concat_303x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_2986_cast_fp16 = transpose(perm = var_2985, x = var_2984_cast_fp16)[name = string("transpose_373")];
            tensor<fp16, [1, ?, 1024]> x_241_cast_fp16 = reshape(shape = concat_303x, x = var_2986_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_2990_to_fp16 = const()[name = string("op_2990_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499038080)))];
            tensor<fp16, [1024]> var_2991_to_fp16 = const()[name = string("op_2991_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501135296)))];
            tensor<fp16, [1, ?, 1024]> linear_107_cast_fp16 = linear(bias = var_2991_to_fp16, weight = var_2990_to_fp16, x = x_241_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_243_cast_fp16 = add(x = x_237_cast_fp16, y = linear_107_cast_fp16)[name = string("x_243_cast_fp16")];
            tensor<int32, [1]> var_2998_axes_0 = const()[name = string("op_2998_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_13_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_13_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501137408)))];
            tensor<fp16, [1024]> blocks_13_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_13_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501139520)))];
            tensor<fp16, [1, ?, 1024]> var_2998_cast_fp16 = layer_norm(axes = var_2998_axes_0, beta = blocks_13_cross_attn_ln_bias_to_fp16, epsilon = var_2897_to_fp16, gamma = blocks_13_cross_attn_ln_weight_to_fp16, x = x_243_cast_fp16)[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3007_to_fp16 = const()[name = string("op_3007_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501141632)))];
            tensor<fp16, [1024]> var_3008_to_fp16 = const()[name = string("op_3008_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503238848)))];
            tensor<fp16, [1, ?, 1024]> linear_108_cast_fp16 = linear(bias = var_3008_to_fp16, weight = var_3007_to_fp16, x = var_2998_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [3]> concat_304 = const()[name = string("concat_304"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_305 = const()[name = string("concat_305"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_137_internal_tensor_assign_1_stride_0 = const()[name = string("k_137_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_137_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_137_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_137_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_137_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_137_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_137_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_137_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_304, begin_mask = k_137_internal_tensor_assign_1_begin_mask_0, end = concat_305, end_mask = k_137_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_137_internal_tensor_assign_1_squeeze_mask_0, stride = k_137_internal_tensor_assign_1_stride_0, update = k_cache_55_cast_fp16, x = k_7_to_fp16)[name = string("k_137_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_306 = const()[name = string("concat_306"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_307 = const()[name = string("concat_307"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_137_internal_tensor_assign_1_stride_0 = const()[name = string("v_137_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_137_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_137_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_137_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_137_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_137_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_137_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_137_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_306, begin_mask = v_137_internal_tensor_assign_1_begin_mask_0, end = concat_307, end_mask = v_137_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_137_internal_tensor_assign_1_squeeze_mask_0, stride = v_137_internal_tensor_assign_1_stride_0, update = v_cache_55_cast_fp16, x = k_7_to_fp16)[name = string("v_137_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_308x = const()[name = string("concat_308x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3028_cast_fp16 = reshape(shape = concat_308x, x = linear_108_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_174_to_fp16 = const()[name = string("const_174_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_111_cast_fp16 = mul(x = var_3028_cast_fp16, y = const_174_to_fp16)[name = string("q_111_cast_fp16")];
            tensor<int32, [4]> var_3034 = const()[name = string("op_3034"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3035_cast_fp16 = reshape(shape = var_3034, x = k_137_internal_tensor_assign_1_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_175_to_fp16 = const()[name = string("const_175_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_139_cast_fp16 = mul(x = var_3035_cast_fp16, y = const_175_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<int32, [4]> var_3041 = const()[name = string("op_3041"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3042_cast_fp16 = reshape(shape = var_3041, x = v_137_internal_tensor_assign_1_cast_fp16)[name = string("op_3042_cast_fp16")];
            tensor<int32, [4]> var_3043 = const()[name = string("op_3043"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_83_transpose_x_0 = const()[name = string("qk_83_transpose_x_0"), val = bool(false)];
            bool qk_83_transpose_y_0 = const()[name = string("qk_83_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_247_perm_0 = const()[name = string("transpose_247_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_248_perm_0 = const()[name = string("transpose_248_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_248 = transpose(perm = transpose_248_perm_0, x = k_139_cast_fp16)[name = string("transpose_370")];
            tensor<fp16, [1, 16, ?, 64]> transpose_247 = transpose(perm = transpose_247_perm_0, x = q_111_cast_fp16)[name = string("transpose_371")];
            tensor<fp16, [1, 16, ?, 1500]> qk_83_cast_fp16 = matmul(transpose_x = qk_83_transpose_x_0, transpose_y = qk_83_transpose_y_0, x = transpose_247, y = transpose_248)[name = string("qk_83_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_3047_cast_fp16 = softmax(axis = var_2891, x = qk_83_cast_fp16)[name = string("op_3047_cast_fp16")];
            bool var_3049_transpose_x_0 = const()[name = string("op_3049_transpose_x_0"), val = bool(false)];
            bool var_3049_transpose_y_0 = const()[name = string("op_3049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_139_cast_fp16 = transpose(perm = var_3043, x = var_3042_cast_fp16)[name = string("transpose_372")];
            tensor<fp16, [1, 16, ?, 64]> var_3049_cast_fp16 = matmul(transpose_x = var_3049_transpose_x_0, transpose_y = var_3049_transpose_y_0, x = var_3047_cast_fp16, y = v_139_cast_fp16)[name = string("op_3049_cast_fp16")];
            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_309x = const()[name = string("concat_309x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3051_cast_fp16 = transpose(perm = var_3050, x = var_3049_cast_fp16)[name = string("transpose_369")];
            tensor<fp16, [1, ?, 1024]> x_247_cast_fp16 = reshape(shape = concat_309x, x = var_3051_cast_fp16)[name = string("x_247_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3055_to_fp16 = const()[name = string("op_3055_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503240960)))];
            tensor<fp16, [1024]> var_3056_to_fp16 = const()[name = string("op_3056_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505338176)))];
            tensor<fp16, [1, ?, 1024]> linear_109_cast_fp16 = linear(bias = var_3056_to_fp16, weight = var_3055_to_fp16, x = x_247_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_249_cast_fp16 = add(x = x_243_cast_fp16, y = linear_109_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [1]> var_3063_axes_0 = const()[name = string("op_3063_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_13_mlp_ln_weight_to_fp16 = const()[name = string("blocks_13_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505340288)))];
            tensor<fp16, [1024]> blocks_13_mlp_ln_bias_to_fp16 = const()[name = string("blocks_13_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505342400)))];
            tensor<fp16, [1, ?, 1024]> var_3063_cast_fp16 = layer_norm(axes = var_3063_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_2897_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_249_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505344512)))];
            tensor<fp16, [4096]> var_3073_to_fp16 = const()[name = string("op_3073_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513733184)))];
            tensor<fp16, [1, ?, 4096]> linear_110_cast_fp16 = linear(bias = var_3073_to_fp16, weight = var_3072_to_fp16, x = var_3063_cast_fp16)[name = string("linear_110_cast_fp16")];
            string x_253_mode_0 = const()[name = string("x_253_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_253_cast_fp16 = gelu(mode = x_253_mode_0, x = linear_110_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_3078_to_fp16 = const()[name = string("op_3078_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513741440)))];
            tensor<fp16, [1024]> var_3079_to_fp16 = const()[name = string("op_3079_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522130112)))];
            tensor<fp16, [1, ?, 1024]> linear_111_cast_fp16 = linear(bias = var_3079_to_fp16, weight = var_3078_to_fp16, x = x_253_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_255_cast_fp16 = add(x = x_249_cast_fp16, y = linear_111_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<int32, [4]> k_cache_57_begin_0 = const()[name = string("k_cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> k_cache_57_end_0 = const()[name = string("k_cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_57_end_mask_0 = const()[name = string("k_cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_57_squeeze_mask_0 = const()[name = string("k_cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_57_cast_fp16 = slice_by_index(begin = k_cache_57_begin_0, end = k_cache_57_end_0, end_mask = k_cache_57_end_mask_0, squeeze_mask = k_cache_57_squeeze_mask_0, x = coreml_update_state_74)[name = string("k_cache_57_cast_fp16")];
            tensor<int32, [4]> v_cache_57_begin_0 = const()[name = string("v_cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> v_cache_57_end_0 = const()[name = string("v_cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_57_end_mask_0 = const()[name = string("v_cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_57_squeeze_mask_0 = const()[name = string("v_cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_57_cast_fp16 = slice_by_index(begin = v_cache_57_begin_0, end = v_cache_57_end_0, end_mask = v_cache_57_end_mask_0, squeeze_mask = v_cache_57_squeeze_mask_0, x = coreml_update_state_75)[name = string("v_cache_57_cast_fp16")];
            tensor<int32, [4]> k_cache_59_begin_0 = const()[name = string("k_cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> k_cache_59_end_0 = const()[name = string("k_cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_59_end_mask_0 = const()[name = string("k_cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_59_squeeze_mask_0 = const()[name = string("k_cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_59_cast_fp16 = slice_by_index(begin = k_cache_59_begin_0, end = k_cache_59_end_0, end_mask = k_cache_59_end_mask_0, squeeze_mask = k_cache_59_squeeze_mask_0, x = read_state_2)[name = string("k_cache_59_cast_fp16")];
            tensor<int32, [4]> v_cache_59_begin_0 = const()[name = string("v_cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> v_cache_59_end_0 = const()[name = string("v_cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_59_end_mask_0 = const()[name = string("v_cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_59_squeeze_mask_0 = const()[name = string("v_cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_59_cast_fp16 = slice_by_index(begin = v_cache_59_begin_0, end = v_cache_59_end_0, end_mask = v_cache_59_end_mask_0, squeeze_mask = v_cache_59_squeeze_mask_0, x = read_state_3)[name = string("v_cache_59_cast_fp16")];
            int32 var_3102 = const()[name = string("op_3102"), val = int32(-1)];
            tensor<int32, [1]> var_3120_axes_0 = const()[name = string("op_3120_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_14_attn_ln_weight_to_fp16 = const()[name = string("blocks_14_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522132224)))];
            tensor<fp16, [1024]> blocks_14_attn_ln_bias_to_fp16 = const()[name = string("blocks_14_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522134336)))];
            fp16 var_3108_to_fp16 = const()[name = string("op_3108_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_3120_cast_fp16 = layer_norm(axes = var_3120_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_3108_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_255_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3131_to_fp16 = const()[name = string("op_3131_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522136448)))];
            tensor<fp16, [1024]> var_3132_to_fp16 = const()[name = string("op_3132_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524233664)))];
            tensor<fp16, [1, ?, 1024]> linear_112_cast_fp16 = linear(bias = var_3132_to_fp16, weight = var_3131_to_fp16, x = var_3120_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3135_to_fp16 = const()[name = string("op_3135_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524235776)))];
            tensor<fp16, [1, ?, 1024]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3135_to_fp16, x = var_3120_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3139_to_fp16 = const()[name = string("op_3139_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526332992)))];
            tensor<fp16, [1024]> var_3140_to_fp16 = const()[name = string("op_3140_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528430208)))];
            tensor<fp16, [1, ?, 1024]> linear_114_cast_fp16 = linear(bias = var_3140_to_fp16, weight = var_3139_to_fp16, x = var_3120_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [3]> var_3142_shape_cast_fp16 = shape(x = linear_112_cast_fp16)[name = string("op_3142_shape_cast_fp16")];
            int32 gather_170_axis_0 = const()[name = string("gather_170_axis_0"), val = int32(0)];
            int32 gather_170_batch_dims_0 = const()[name = string("gather_170_batch_dims_0"), val = int32(0)];
            bool gather_170_validate_indices_0 = const()[name = string("gather_170_validate_indices_0"), val = bool(false)];
            string var_3142_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3142_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_170_to_uint16 = const()[name = string("select_170_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3142_shape_cast_fp16_to_uint16 = cast(dtype = var_3142_shape_cast_fp16_to_uint16_dtype_0, x = var_3142_shape_cast_fp16)[name = string("cast_266")];
            uint16 gather_170_cast_uint16 = gather(axis = gather_170_axis_0, batch_dims = gather_170_batch_dims_0, indices = select_170_to_uint16, validate_indices = gather_170_validate_indices_0, x = var_3142_shape_cast_fp16_to_uint16)[name = string("gather_170_cast_uint16")];
            string gather_170_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_170_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_170_cast_uint16_to_int32 = cast(dtype = gather_170_cast_uint16_to_int32_dtype_0, x = gather_170_cast_uint16)[name = string("cast_265")];
            int32 end_step_31 = add(x = offset, y = gather_170_cast_uint16_to_int32)[name = string("end_step_31")];
            tensor<int32, [1]> expand_dims_224 = const()[name = string("expand_dims_224"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_226 = const()[name = string("expand_dims_226"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_227_axes_0 = const()[name = string("expand_dims_227_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_227 = expand_dims(axes = expand_dims_227_axes_0, x = end_step_31)[name = string("expand_dims_227")];
            tensor<int32, [1]> concat_312_values0_0 = const()[name = string("concat_312_values0_0"), val = tensor<int32, [1]>([14])];
            int32 concat_312_axis_0 = const()[name = string("concat_312_axis_0"), val = int32(0)];
            bool concat_312_interleave_0 = const()[name = string("concat_312_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_312 = concat(axis = concat_312_axis_0, interleave = concat_312_interleave_0, values = (concat_312_values0_0, expand_dims_224, expand_dims_1, expand_dims_226))[name = string("concat_312")];
            tensor<int32, [1]> concat_313_values0_0 = const()[name = string("concat_313_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_313_values1_0 = const()[name = string("concat_313_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_313_values3_0 = const()[name = string("concat_313_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_313_axis_0 = const()[name = string("concat_313_axis_0"), val = int32(0)];
            bool concat_313_interleave_0 = const()[name = string("concat_313_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_313 = concat(axis = concat_313_axis_0, interleave = concat_313_interleave_0, values = (concat_313_values0_0, concat_313_values1_0, expand_dims_227, concat_313_values3_0))[name = string("concat_313")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_15_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_15_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_15_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_312, begin_mask = k_cache1_internal_tensor_assign_15_begin_mask_0, end = concat_313, end_mask = k_cache1_internal_tensor_assign_15_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_15_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_15_stride_0, update = linear_113_cast_fp16, x = coreml_update_state_74)[name = string("k_cache1_internal_tensor_assign_15_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_15_cast_fp16, input = k_cache1)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_76 = read_state(input = k_cache1)[name = string("coreml_update_state_76")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_15_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_15_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_15_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_312, begin_mask = v_cache1_internal_tensor_assign_15_begin_mask_0, end = concat_313, end_mask = v_cache1_internal_tensor_assign_15_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_15_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_15_stride_0, update = linear_114_cast_fp16, x = coreml_update_state_75)[name = string("v_cache1_internal_tensor_assign_15_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_15_cast_fp16, input = v_cache1)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_77 = read_state(input = v_cache1)[name = string("coreml_update_state_77")];
            int32 concat_318_values0_0 = const()[name = string("concat_318_values0_0"), val = int32(1)];
            int32 concat_318_values2_0 = const()[name = string("concat_318_values2_0"), val = int32(1024)];
            int32 concat_318_axis_0 = const()[name = string("concat_318_axis_0"), val = int32(0)];
            bool concat_318_interleave_0 = const()[name = string("concat_318_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_318 = concat(axis = concat_318_axis_0, interleave = concat_318_interleave_0, values = (concat_318_values0_0, end_step_31, concat_318_values2_0))[name = string("concat_318")];
            tensor<int32, [3]> var_3158_begin_0 = const()[name = string("op_3158_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3158_end_mask_0 = const()[name = string("op_3158_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3158_cast_fp16 = slice_by_index(begin = var_3158_begin_0, end = concat_318, end_mask = var_3158_end_mask_0, x = k_cache_57_cast_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [3]> var_3161_begin_0 = const()[name = string("op_3161_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3161_end_mask_0 = const()[name = string("op_3161_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3161_cast_fp16 = slice_by_index(begin = var_3161_begin_0, end = concat_318, end_mask = var_3161_end_mask_0, x = v_cache_57_cast_fp16)[name = string("op_3161_cast_fp16")];
            tensor<int32, [4]> concat_320x = const()[name = string("concat_320x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3171_cast_fp16 = reshape(shape = concat_320x, x = linear_112_cast_fp16)[name = string("op_3171_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_176_to_fp16 = const()[name = string("const_176_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_115_cast_fp16 = mul(x = var_3171_cast_fp16, y = const_176_to_fp16)[name = string("q_115_cast_fp16")];
            tensor<int32, [4]> concat_321x = const()[name = string("concat_321x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3178_cast_fp16 = reshape(shape = concat_321x, x = var_3158_cast_fp16)[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_177_to_fp16 = const()[name = string("const_177_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_145_cast_fp16 = mul(x = var_3178_cast_fp16, y = const_177_to_fp16)[name = string("k_145_cast_fp16")];
            tensor<int32, [4]> concat_322x = const()[name = string("concat_322x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3185_cast_fp16 = reshape(shape = concat_322x, x = var_3161_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> var_3186 = const()[name = string("op_3186"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_85_transpose_x_0 = const()[name = string("qk_85_transpose_x_0"), val = bool(false)];
            bool qk_85_transpose_y_0 = const()[name = string("qk_85_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_249_perm_0 = const()[name = string("transpose_249_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_250_perm_0 = const()[name = string("transpose_250_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_250 = transpose(perm = transpose_250_perm_0, x = k_145_cast_fp16)[name = string("transpose_366")];
            tensor<fp16, [1, 16, ?, 64]> transpose_249 = transpose(perm = transpose_249_perm_0, x = q_115_cast_fp16)[name = string("transpose_367")];
            tensor<fp16, [1, 16, ?, ?]> qk_85_cast_fp16 = matmul(transpose_x = qk_85_transpose_x_0, transpose_y = qk_85_transpose_y_0, x = transpose_249, y = transpose_250)[name = string("qk_85_cast_fp16")];
            int32 concat_323_values1_0 = const()[name = string("concat_323_values1_0"), val = int32(448)];
            int32 concat_323_axis_0 = const()[name = string("concat_323_axis_0"), val = int32(0)];
            bool concat_323_interleave_0 = const()[name = string("concat_323_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_323 = concat(axis = concat_323_axis_0, interleave = concat_323_interleave_0, values = (gather_170_cast_uint16_to_int32, concat_323_values1_0))[name = string("concat_323")];
            tensor<int32, [2]> var_3189_begin_0 = const()[name = string("op_3189_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3189_end_mask_0 = const()[name = string("op_3189_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_3189_cast_fp16 = slice_by_index(begin = var_3189_begin_0, end = concat_323, end_mask = var_3189_end_mask_0, x = mask_to_fp16)[name = string("op_3189_cast_fp16")];
            int32 concat_324_values0_0 = const()[name = string("concat_324_values0_0"), val = int32(0)];
            int32 concat_324_axis_0 = const()[name = string("concat_324_axis_0"), val = int32(0)];
            bool concat_324_interleave_0 = const()[name = string("concat_324_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_324 = concat(axis = concat_324_axis_0, interleave = concat_324_interleave_0, values = (concat_324_values0_0, gather_170_cast_uint16_to_int32))[name = string("concat_324")];
            tensor<int32, [2]> var_3190_begin_0 = const()[name = string("op_3190_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3190_end_mask_0 = const()[name = string("op_3190_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_3190_cast_fp16 = slice_by_index(begin = var_3190_begin_0, end = concat_324, end_mask = var_3190_end_mask_0, x = var_3189_cast_fp16)[name = string("op_3190_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_87_cast_fp16 = add(x = qk_85_cast_fp16, y = var_3190_cast_fp16)[name = string("qk_87_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_3193_cast_fp16 = softmax(axis = var_3102, x = qk_87_cast_fp16)[name = string("op_3193_cast_fp16")];
            bool var_3195_transpose_x_0 = const()[name = string("op_3195_transpose_x_0"), val = bool(false)];
            bool var_3195_transpose_y_0 = const()[name = string("op_3195_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_145_cast_fp16 = transpose(perm = var_3186, x = var_3185_cast_fp16)[name = string("transpose_368")];
            tensor<fp16, [1, 16, ?, 64]> var_3195_cast_fp16 = matmul(transpose_x = var_3195_transpose_x_0, transpose_y = var_3195_transpose_y_0, x = var_3193_cast_fp16, y = v_145_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> var_3196 = const()[name = string("op_3196"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_325x = const()[name = string("concat_325x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3197_cast_fp16 = transpose(perm = var_3196, x = var_3195_cast_fp16)[name = string("transpose_365")];
            tensor<fp16, [1, ?, 1024]> x_259_cast_fp16 = reshape(shape = concat_325x, x = var_3197_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3201_to_fp16 = const()[name = string("op_3201_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528432320)))];
            tensor<fp16, [1024]> var_3202_to_fp16 = const()[name = string("op_3202_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530529536)))];
            tensor<fp16, [1, ?, 1024]> linear_115_cast_fp16 = linear(bias = var_3202_to_fp16, weight = var_3201_to_fp16, x = x_259_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_261_cast_fp16 = add(x = x_255_cast_fp16, y = linear_115_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [1]> var_3209_axes_0 = const()[name = string("op_3209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_14_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_14_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530531648)))];
            tensor<fp16, [1024]> blocks_14_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_14_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530533760)))];
            tensor<fp16, [1, ?, 1024]> var_3209_cast_fp16 = layer_norm(axes = var_3209_axes_0, beta = blocks_14_cross_attn_ln_bias_to_fp16, epsilon = var_3108_to_fp16, gamma = blocks_14_cross_attn_ln_weight_to_fp16, x = x_261_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3218_to_fp16 = const()[name = string("op_3218_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530535872)))];
            tensor<fp16, [1024]> var_3219_to_fp16 = const()[name = string("op_3219_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532633088)))];
            tensor<fp16, [1, ?, 1024]> linear_116_cast_fp16 = linear(bias = var_3219_to_fp16, weight = var_3218_to_fp16, x = var_3209_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<int32, [3]> concat_326 = const()[name = string("concat_326"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_327 = const()[name = string("concat_327"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_147_internal_tensor_assign_1_stride_0 = const()[name = string("k_147_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_147_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_147_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_147_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_147_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_147_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_147_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_147_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_326, begin_mask = k_147_internal_tensor_assign_1_begin_mask_0, end = concat_327, end_mask = k_147_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_147_internal_tensor_assign_1_squeeze_mask_0, stride = k_147_internal_tensor_assign_1_stride_0, update = k_cache_59_cast_fp16, x = k_7_to_fp16)[name = string("k_147_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_328 = const()[name = string("concat_328"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_329 = const()[name = string("concat_329"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_147_internal_tensor_assign_1_stride_0 = const()[name = string("v_147_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_147_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_147_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_147_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_147_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_147_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_147_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_147_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_328, begin_mask = v_147_internal_tensor_assign_1_begin_mask_0, end = concat_329, end_mask = v_147_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_147_internal_tensor_assign_1_squeeze_mask_0, stride = v_147_internal_tensor_assign_1_stride_0, update = v_cache_59_cast_fp16, x = k_7_to_fp16)[name = string("v_147_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_330x = const()[name = string("concat_330x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3239_cast_fp16 = reshape(shape = concat_330x, x = linear_116_cast_fp16)[name = string("op_3239_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_178_to_fp16 = const()[name = string("const_178_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_119_cast_fp16 = mul(x = var_3239_cast_fp16, y = const_178_to_fp16)[name = string("q_119_cast_fp16")];
            tensor<int32, [4]> var_3245 = const()[name = string("op_3245"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3246_cast_fp16 = reshape(shape = var_3245, x = k_147_internal_tensor_assign_1_cast_fp16)[name = string("op_3246_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_179_to_fp16 = const()[name = string("const_179_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_149_cast_fp16 = mul(x = var_3246_cast_fp16, y = const_179_to_fp16)[name = string("k_149_cast_fp16")];
            tensor<int32, [4]> var_3252 = const()[name = string("op_3252"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3253_cast_fp16 = reshape(shape = var_3252, x = v_147_internal_tensor_assign_1_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<int32, [4]> var_3254 = const()[name = string("op_3254"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_89_transpose_x_0 = const()[name = string("qk_89_transpose_x_0"), val = bool(false)];
            bool qk_89_transpose_y_0 = const()[name = string("qk_89_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_251_perm_0 = const()[name = string("transpose_251_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_252_perm_0 = const()[name = string("transpose_252_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_252 = transpose(perm = transpose_252_perm_0, x = k_149_cast_fp16)[name = string("transpose_362")];
            tensor<fp16, [1, 16, ?, 64]> transpose_251 = transpose(perm = transpose_251_perm_0, x = q_119_cast_fp16)[name = string("transpose_363")];
            tensor<fp16, [1, 16, ?, 1500]> qk_89_cast_fp16 = matmul(transpose_x = qk_89_transpose_x_0, transpose_y = qk_89_transpose_y_0, x = transpose_251, y = transpose_252)[name = string("qk_89_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_3258_cast_fp16 = softmax(axis = var_3102, x = qk_89_cast_fp16)[name = string("op_3258_cast_fp16")];
            bool var_3260_transpose_x_0 = const()[name = string("op_3260_transpose_x_0"), val = bool(false)];
            bool var_3260_transpose_y_0 = const()[name = string("op_3260_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_149_cast_fp16 = transpose(perm = var_3254, x = var_3253_cast_fp16)[name = string("transpose_364")];
            tensor<fp16, [1, 16, ?, 64]> var_3260_cast_fp16 = matmul(transpose_x = var_3260_transpose_x_0, transpose_y = var_3260_transpose_y_0, x = var_3258_cast_fp16, y = v_149_cast_fp16)[name = string("op_3260_cast_fp16")];
            tensor<int32, [4]> var_3261 = const()[name = string("op_3261"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_331x = const()[name = string("concat_331x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3262_cast_fp16 = transpose(perm = var_3261, x = var_3260_cast_fp16)[name = string("transpose_361")];
            tensor<fp16, [1, ?, 1024]> x_265_cast_fp16 = reshape(shape = concat_331x, x = var_3262_cast_fp16)[name = string("x_265_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3266_to_fp16 = const()[name = string("op_3266_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532635200)))];
            tensor<fp16, [1024]> var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534732416)))];
            tensor<fp16, [1, ?, 1024]> linear_117_cast_fp16 = linear(bias = var_3267_to_fp16, weight = var_3266_to_fp16, x = x_265_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_267_cast_fp16 = add(x = x_261_cast_fp16, y = linear_117_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [1]> var_3274_axes_0 = const()[name = string("op_3274_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_14_mlp_ln_weight_to_fp16 = const()[name = string("blocks_14_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534734528)))];
            tensor<fp16, [1024]> blocks_14_mlp_ln_bias_to_fp16 = const()[name = string("blocks_14_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534736640)))];
            tensor<fp16, [1, ?, 1024]> var_3274_cast_fp16 = layer_norm(axes = var_3274_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_3108_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_267_cast_fp16)[name = string("op_3274_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_3283_to_fp16 = const()[name = string("op_3283_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534738752)))];
            tensor<fp16, [4096]> var_3284_to_fp16 = const()[name = string("op_3284_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543127424)))];
            tensor<fp16, [1, ?, 4096]> linear_118_cast_fp16 = linear(bias = var_3284_to_fp16, weight = var_3283_to_fp16, x = var_3274_cast_fp16)[name = string("linear_118_cast_fp16")];
            string x_271_mode_0 = const()[name = string("x_271_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_271_cast_fp16 = gelu(mode = x_271_mode_0, x = linear_118_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_3289_to_fp16 = const()[name = string("op_3289_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543135680)))];
            tensor<fp16, [1024]> var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551524352)))];
            tensor<fp16, [1, ?, 1024]> linear_119_cast_fp16 = linear(bias = var_3290_to_fp16, weight = var_3289_to_fp16, x = x_271_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_273_cast_fp16 = add(x = x_267_cast_fp16, y = linear_119_cast_fp16)[name = string("x_273_cast_fp16")];
            tensor<int32, [4]> k_cache_61_begin_0 = const()[name = string("k_cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> k_cache_61_end_0 = const()[name = string("k_cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_61_end_mask_0 = const()[name = string("k_cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_61_squeeze_mask_0 = const()[name = string("k_cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_61_cast_fp16 = slice_by_index(begin = k_cache_61_begin_0, end = k_cache_61_end_0, end_mask = k_cache_61_end_mask_0, squeeze_mask = k_cache_61_squeeze_mask_0, x = coreml_update_state_76)[name = string("k_cache_61_cast_fp16")];
            tensor<int32, [4]> v_cache_61_begin_0 = const()[name = string("v_cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> v_cache_61_end_0 = const()[name = string("v_cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_61_end_mask_0 = const()[name = string("v_cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_61_squeeze_mask_0 = const()[name = string("v_cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_61_cast_fp16 = slice_by_index(begin = v_cache_61_begin_0, end = v_cache_61_end_0, end_mask = v_cache_61_end_mask_0, squeeze_mask = v_cache_61_squeeze_mask_0, x = coreml_update_state_77)[name = string("v_cache_61_cast_fp16")];
            tensor<int32, [4]> k_cache_63_begin_0 = const()[name = string("k_cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> k_cache_63_end_0 = const()[name = string("k_cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_63_end_mask_0 = const()[name = string("k_cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_63_squeeze_mask_0 = const()[name = string("k_cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_63_cast_fp16 = slice_by_index(begin = k_cache_63_begin_0, end = k_cache_63_end_0, end_mask = k_cache_63_end_mask_0, squeeze_mask = k_cache_63_squeeze_mask_0, x = read_state_2)[name = string("k_cache_63_cast_fp16")];
            tensor<int32, [4]> v_cache_63_begin_0 = const()[name = string("v_cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> v_cache_63_end_0 = const()[name = string("v_cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_63_end_mask_0 = const()[name = string("v_cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_63_squeeze_mask_0 = const()[name = string("v_cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_63_cast_fp16 = slice_by_index(begin = v_cache_63_begin_0, end = v_cache_63_end_0, end_mask = v_cache_63_end_mask_0, squeeze_mask = v_cache_63_squeeze_mask_0, x = read_state_3)[name = string("v_cache_63_cast_fp16")];
            int32 var_3313 = const()[name = string("op_3313"), val = int32(-1)];
            tensor<int32, [1]> var_3331_axes_0 = const()[name = string("op_3331_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_15_attn_ln_weight_to_fp16 = const()[name = string("blocks_15_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551526464)))];
            tensor<fp16, [1024]> blocks_15_attn_ln_bias_to_fp16 = const()[name = string("blocks_15_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551528576)))];
            fp16 var_3319_to_fp16 = const()[name = string("op_3319_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_3331_cast_fp16 = layer_norm(axes = var_3331_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_273_cast_fp16)[name = string("op_3331_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3342_to_fp16 = const()[name = string("op_3342_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551530688)))];
            tensor<fp16, [1024]> var_3343_to_fp16 = const()[name = string("op_3343_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553627904)))];
            tensor<fp16, [1, ?, 1024]> linear_120_cast_fp16 = linear(bias = var_3343_to_fp16, weight = var_3342_to_fp16, x = var_3331_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3346_to_fp16 = const()[name = string("op_3346_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553630016)))];
            tensor<fp16, [1, ?, 1024]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3346_to_fp16, x = var_3331_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3350_to_fp16 = const()[name = string("op_3350_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555727232)))];
            tensor<fp16, [1024]> var_3351_to_fp16 = const()[name = string("op_3351_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557824448)))];
            tensor<fp16, [1, ?, 1024]> linear_122_cast_fp16 = linear(bias = var_3351_to_fp16, weight = var_3350_to_fp16, x = var_3331_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<int32, [3]> var_3353_shape_cast_fp16 = shape(x = linear_120_cast_fp16)[name = string("op_3353_shape_cast_fp16")];
            int32 gather_182_axis_0 = const()[name = string("gather_182_axis_0"), val = int32(0)];
            int32 gather_182_batch_dims_0 = const()[name = string("gather_182_batch_dims_0"), val = int32(0)];
            bool gather_182_validate_indices_0 = const()[name = string("gather_182_validate_indices_0"), val = bool(false)];
            string var_3353_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3353_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_182_to_uint16 = const()[name = string("select_182_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3353_shape_cast_fp16_to_uint16 = cast(dtype = var_3353_shape_cast_fp16_to_uint16_dtype_0, x = var_3353_shape_cast_fp16)[name = string("cast_264")];
            uint16 gather_182_cast_uint16 = gather(axis = gather_182_axis_0, batch_dims = gather_182_batch_dims_0, indices = select_182_to_uint16, validate_indices = gather_182_validate_indices_0, x = var_3353_shape_cast_fp16_to_uint16)[name = string("gather_182_cast_uint16")];
            string gather_182_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_182_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_182_cast_uint16_to_int32 = cast(dtype = gather_182_cast_uint16_to_int32_dtype_0, x = gather_182_cast_uint16)[name = string("cast_263")];
            int32 end_step_33 = add(x = offset, y = gather_182_cast_uint16_to_int32)[name = string("end_step_33")];
            tensor<int32, [1]> expand_dims_240 = const()[name = string("expand_dims_240"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_242 = const()[name = string("expand_dims_242"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_243_axes_0 = const()[name = string("expand_dims_243_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_243 = expand_dims(axes = expand_dims_243_axes_0, x = end_step_33)[name = string("expand_dims_243")];
            tensor<int32, [1]> concat_334_values0_0 = const()[name = string("concat_334_values0_0"), val = tensor<int32, [1]>([15])];
            int32 concat_334_axis_0 = const()[name = string("concat_334_axis_0"), val = int32(0)];
            bool concat_334_interleave_0 = const()[name = string("concat_334_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_334 = concat(axis = concat_334_axis_0, interleave = concat_334_interleave_0, values = (concat_334_values0_0, expand_dims_240, expand_dims_1, expand_dims_242))[name = string("concat_334")];
            tensor<int32, [1]> concat_335_values0_0 = const()[name = string("concat_335_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_335_values1_0 = const()[name = string("concat_335_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_335_values3_0 = const()[name = string("concat_335_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_335_axis_0 = const()[name = string("concat_335_axis_0"), val = int32(0)];
            bool concat_335_interleave_0 = const()[name = string("concat_335_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_335 = concat(axis = concat_335_axis_0, interleave = concat_335_interleave_0, values = (concat_335_values0_0, concat_335_values1_0, expand_dims_243, concat_335_values3_0))[name = string("concat_335")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_16_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_16_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_16_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_334, begin_mask = k_cache1_internal_tensor_assign_16_begin_mask_0, end = concat_335, end_mask = k_cache1_internal_tensor_assign_16_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_16_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_16_stride_0, update = linear_121_cast_fp16, x = coreml_update_state_76)[name = string("k_cache1_internal_tensor_assign_16_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_16_cast_fp16, input = k_cache1)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_78 = read_state(input = k_cache1)[name = string("coreml_update_state_78")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_16_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_16_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_16_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_334, begin_mask = v_cache1_internal_tensor_assign_16_begin_mask_0, end = concat_335, end_mask = v_cache1_internal_tensor_assign_16_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_16_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_16_stride_0, update = linear_122_cast_fp16, x = coreml_update_state_77)[name = string("v_cache1_internal_tensor_assign_16_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_16_cast_fp16, input = v_cache1)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_79 = read_state(input = v_cache1)[name = string("coreml_update_state_79")];
            int32 concat_340_values0_0 = const()[name = string("concat_340_values0_0"), val = int32(1)];
            int32 concat_340_values2_0 = const()[name = string("concat_340_values2_0"), val = int32(1024)];
            int32 concat_340_axis_0 = const()[name = string("concat_340_axis_0"), val = int32(0)];
            bool concat_340_interleave_0 = const()[name = string("concat_340_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_340 = concat(axis = concat_340_axis_0, interleave = concat_340_interleave_0, values = (concat_340_values0_0, end_step_33, concat_340_values2_0))[name = string("concat_340")];
            tensor<int32, [3]> var_3369_begin_0 = const()[name = string("op_3369_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3369_end_mask_0 = const()[name = string("op_3369_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3369_cast_fp16 = slice_by_index(begin = var_3369_begin_0, end = concat_340, end_mask = var_3369_end_mask_0, x = k_cache_61_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [3]> var_3372_begin_0 = const()[name = string("op_3372_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3372_end_mask_0 = const()[name = string("op_3372_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3372_cast_fp16 = slice_by_index(begin = var_3372_begin_0, end = concat_340, end_mask = var_3372_end_mask_0, x = v_cache_61_cast_fp16)[name = string("op_3372_cast_fp16")];
            tensor<int32, [4]> concat_342x = const()[name = string("concat_342x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3382_cast_fp16 = reshape(shape = concat_342x, x = linear_120_cast_fp16)[name = string("op_3382_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_180_to_fp16 = const()[name = string("const_180_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_123_cast_fp16 = mul(x = var_3382_cast_fp16, y = const_180_to_fp16)[name = string("q_123_cast_fp16")];
            tensor<int32, [4]> concat_343x = const()[name = string("concat_343x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3389_cast_fp16 = reshape(shape = concat_343x, x = var_3369_cast_fp16)[name = string("op_3389_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_181_to_fp16 = const()[name = string("const_181_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_155_cast_fp16 = mul(x = var_3389_cast_fp16, y = const_181_to_fp16)[name = string("k_155_cast_fp16")];
            tensor<int32, [4]> concat_344x = const()[name = string("concat_344x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3396_cast_fp16 = reshape(shape = concat_344x, x = var_3372_cast_fp16)[name = string("op_3396_cast_fp16")];
            tensor<int32, [4]> var_3397 = const()[name = string("op_3397"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_91_transpose_x_0 = const()[name = string("qk_91_transpose_x_0"), val = bool(false)];
            bool qk_91_transpose_y_0 = const()[name = string("qk_91_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_253_perm_0 = const()[name = string("transpose_253_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_254_perm_0 = const()[name = string("transpose_254_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_254 = transpose(perm = transpose_254_perm_0, x = k_155_cast_fp16)[name = string("transpose_358")];
            tensor<fp16, [1, 16, ?, 64]> transpose_253 = transpose(perm = transpose_253_perm_0, x = q_123_cast_fp16)[name = string("transpose_359")];
            tensor<fp16, [1, 16, ?, ?]> qk_91_cast_fp16 = matmul(transpose_x = qk_91_transpose_x_0, transpose_y = qk_91_transpose_y_0, x = transpose_253, y = transpose_254)[name = string("qk_91_cast_fp16")];
            int32 concat_345_values1_0 = const()[name = string("concat_345_values1_0"), val = int32(448)];
            int32 concat_345_axis_0 = const()[name = string("concat_345_axis_0"), val = int32(0)];
            bool concat_345_interleave_0 = const()[name = string("concat_345_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_345 = concat(axis = concat_345_axis_0, interleave = concat_345_interleave_0, values = (gather_182_cast_uint16_to_int32, concat_345_values1_0))[name = string("concat_345")];
            tensor<int32, [2]> var_3400_begin_0 = const()[name = string("op_3400_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3400_end_mask_0 = const()[name = string("op_3400_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_3400_cast_fp16 = slice_by_index(begin = var_3400_begin_0, end = concat_345, end_mask = var_3400_end_mask_0, x = mask_to_fp16)[name = string("op_3400_cast_fp16")];
            int32 concat_346_values0_0 = const()[name = string("concat_346_values0_0"), val = int32(0)];
            int32 concat_346_axis_0 = const()[name = string("concat_346_axis_0"), val = int32(0)];
            bool concat_346_interleave_0 = const()[name = string("concat_346_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_346 = concat(axis = concat_346_axis_0, interleave = concat_346_interleave_0, values = (concat_346_values0_0, gather_182_cast_uint16_to_int32))[name = string("concat_346")];
            tensor<int32, [2]> var_3401_begin_0 = const()[name = string("op_3401_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3401_end_mask_0 = const()[name = string("op_3401_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_3401_cast_fp16 = slice_by_index(begin = var_3401_begin_0, end = concat_346, end_mask = var_3401_end_mask_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_93_cast_fp16 = add(x = qk_91_cast_fp16, y = var_3401_cast_fp16)[name = string("qk_93_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_3404_cast_fp16 = softmax(axis = var_3313, x = qk_93_cast_fp16)[name = string("op_3404_cast_fp16")];
            bool var_3406_transpose_x_0 = const()[name = string("op_3406_transpose_x_0"), val = bool(false)];
            bool var_3406_transpose_y_0 = const()[name = string("op_3406_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_155_cast_fp16 = transpose(perm = var_3397, x = var_3396_cast_fp16)[name = string("transpose_360")];
            tensor<fp16, [1, 16, ?, 64]> var_3406_cast_fp16 = matmul(transpose_x = var_3406_transpose_x_0, transpose_y = var_3406_transpose_y_0, x = var_3404_cast_fp16, y = v_155_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<int32, [4]> var_3407 = const()[name = string("op_3407"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_347x = const()[name = string("concat_347x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3408_cast_fp16 = transpose(perm = var_3407, x = var_3406_cast_fp16)[name = string("transpose_357")];
            tensor<fp16, [1, ?, 1024]> x_277_cast_fp16 = reshape(shape = concat_347x, x = var_3408_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3412_to_fp16 = const()[name = string("op_3412_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557826560)))];
            tensor<fp16, [1024]> var_3413_to_fp16 = const()[name = string("op_3413_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559923776)))];
            tensor<fp16, [1, ?, 1024]> linear_123_cast_fp16 = linear(bias = var_3413_to_fp16, weight = var_3412_to_fp16, x = x_277_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_279_cast_fp16 = add(x = x_273_cast_fp16, y = linear_123_cast_fp16)[name = string("x_279_cast_fp16")];
            tensor<int32, [1]> var_3420_axes_0 = const()[name = string("op_3420_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_15_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_15_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559925888)))];
            tensor<fp16, [1024]> blocks_15_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_15_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559928000)))];
            tensor<fp16, [1, ?, 1024]> var_3420_cast_fp16 = layer_norm(axes = var_3420_axes_0, beta = blocks_15_cross_attn_ln_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = blocks_15_cross_attn_ln_weight_to_fp16, x = x_279_cast_fp16)[name = string("op_3420_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3429_to_fp16 = const()[name = string("op_3429_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559930112)))];
            tensor<fp16, [1024]> var_3430_to_fp16 = const()[name = string("op_3430_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562027328)))];
            tensor<fp16, [1, ?, 1024]> linear_124_cast_fp16 = linear(bias = var_3430_to_fp16, weight = var_3429_to_fp16, x = var_3420_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<int32, [3]> concat_348 = const()[name = string("concat_348"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_349 = const()[name = string("concat_349"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_157_internal_tensor_assign_1_stride_0 = const()[name = string("k_157_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_157_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_157_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_157_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_157_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_157_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_157_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_157_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_348, begin_mask = k_157_internal_tensor_assign_1_begin_mask_0, end = concat_349, end_mask = k_157_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_157_internal_tensor_assign_1_squeeze_mask_0, stride = k_157_internal_tensor_assign_1_stride_0, update = k_cache_63_cast_fp16, x = k_7_to_fp16)[name = string("k_157_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_350 = const()[name = string("concat_350"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_351 = const()[name = string("concat_351"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_157_internal_tensor_assign_1_stride_0 = const()[name = string("v_157_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_157_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_157_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_157_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_157_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_157_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_157_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_157_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_350, begin_mask = v_157_internal_tensor_assign_1_begin_mask_0, end = concat_351, end_mask = v_157_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_157_internal_tensor_assign_1_squeeze_mask_0, stride = v_157_internal_tensor_assign_1_stride_0, update = v_cache_63_cast_fp16, x = k_7_to_fp16)[name = string("v_157_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_352x = const()[name = string("concat_352x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3450_cast_fp16 = reshape(shape = concat_352x, x = linear_124_cast_fp16)[name = string("op_3450_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_182_to_fp16 = const()[name = string("const_182_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_127_cast_fp16 = mul(x = var_3450_cast_fp16, y = const_182_to_fp16)[name = string("q_127_cast_fp16")];
            tensor<int32, [4]> var_3456 = const()[name = string("op_3456"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3457_cast_fp16 = reshape(shape = var_3456, x = k_157_internal_tensor_assign_1_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_159_cast_fp16 = mul(x = var_3457_cast_fp16, y = const_183_to_fp16)[name = string("k_159_cast_fp16")];
            tensor<int32, [4]> var_3463 = const()[name = string("op_3463"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3464_cast_fp16 = reshape(shape = var_3463, x = v_157_internal_tensor_assign_1_cast_fp16)[name = string("op_3464_cast_fp16")];
            tensor<int32, [4]> var_3465 = const()[name = string("op_3465"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_95_transpose_x_0 = const()[name = string("qk_95_transpose_x_0"), val = bool(false)];
            bool qk_95_transpose_y_0 = const()[name = string("qk_95_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_255_perm_0 = const()[name = string("transpose_255_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_256_perm_0 = const()[name = string("transpose_256_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_256 = transpose(perm = transpose_256_perm_0, x = k_159_cast_fp16)[name = string("transpose_354")];
            tensor<fp16, [1, 16, ?, 64]> transpose_255 = transpose(perm = transpose_255_perm_0, x = q_127_cast_fp16)[name = string("transpose_355")];
            tensor<fp16, [1, 16, ?, 1500]> qk_95_cast_fp16 = matmul(transpose_x = qk_95_transpose_x_0, transpose_y = qk_95_transpose_y_0, x = transpose_255, y = transpose_256)[name = string("qk_95_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_3469_cast_fp16 = softmax(axis = var_3313, x = qk_95_cast_fp16)[name = string("op_3469_cast_fp16")];
            bool var_3471_transpose_x_0 = const()[name = string("op_3471_transpose_x_0"), val = bool(false)];
            bool var_3471_transpose_y_0 = const()[name = string("op_3471_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_159_cast_fp16 = transpose(perm = var_3465, x = var_3464_cast_fp16)[name = string("transpose_356")];
            tensor<fp16, [1, 16, ?, 64]> var_3471_cast_fp16 = matmul(transpose_x = var_3471_transpose_x_0, transpose_y = var_3471_transpose_y_0, x = var_3469_cast_fp16, y = v_159_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<int32, [4]> var_3472 = const()[name = string("op_3472"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_353x = const()[name = string("concat_353x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3473_cast_fp16 = transpose(perm = var_3472, x = var_3471_cast_fp16)[name = string("transpose_353")];
            tensor<fp16, [1, ?, 1024]> x_283_cast_fp16 = reshape(shape = concat_353x, x = var_3473_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3477_to_fp16 = const()[name = string("op_3477_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562029440)))];
            tensor<fp16, [1024]> var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564126656)))];
            tensor<fp16, [1, ?, 1024]> linear_125_cast_fp16 = linear(bias = var_3478_to_fp16, weight = var_3477_to_fp16, x = x_283_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_285_cast_fp16 = add(x = x_279_cast_fp16, y = linear_125_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [1]> var_3485_axes_0 = const()[name = string("op_3485_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_15_mlp_ln_weight_to_fp16 = const()[name = string("blocks_15_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564128768)))];
            tensor<fp16, [1024]> blocks_15_mlp_ln_bias_to_fp16 = const()[name = string("blocks_15_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564130880)))];
            tensor<fp16, [1, ?, 1024]> var_3485_cast_fp16 = layer_norm(axes = var_3485_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_285_cast_fp16)[name = string("op_3485_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_3494_to_fp16 = const()[name = string("op_3494_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564132992)))];
            tensor<fp16, [4096]> var_3495_to_fp16 = const()[name = string("op_3495_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572521664)))];
            tensor<fp16, [1, ?, 4096]> linear_126_cast_fp16 = linear(bias = var_3495_to_fp16, weight = var_3494_to_fp16, x = var_3485_cast_fp16)[name = string("linear_126_cast_fp16")];
            string x_289_mode_0 = const()[name = string("x_289_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_289_cast_fp16 = gelu(mode = x_289_mode_0, x = linear_126_cast_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_3500_to_fp16 = const()[name = string("op_3500_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572529920)))];
            tensor<fp16, [1024]> var_3501_to_fp16 = const()[name = string("op_3501_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580918592)))];
            tensor<fp16, [1, ?, 1024]> linear_127_cast_fp16 = linear(bias = var_3501_to_fp16, weight = var_3500_to_fp16, x = x_289_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_291_cast_fp16 = add(x = x_285_cast_fp16, y = linear_127_cast_fp16)[name = string("x_291_cast_fp16")];
            tensor<int32, [4]> k_cache_65_begin_0 = const()[name = string("k_cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> k_cache_65_end_0 = const()[name = string("k_cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_65_end_mask_0 = const()[name = string("k_cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_65_squeeze_mask_0 = const()[name = string("k_cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_65_cast_fp16 = slice_by_index(begin = k_cache_65_begin_0, end = k_cache_65_end_0, end_mask = k_cache_65_end_mask_0, squeeze_mask = k_cache_65_squeeze_mask_0, x = coreml_update_state_78)[name = string("k_cache_65_cast_fp16")];
            tensor<int32, [4]> v_cache_65_begin_0 = const()[name = string("v_cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> v_cache_65_end_0 = const()[name = string("v_cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_65_end_mask_0 = const()[name = string("v_cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_65_squeeze_mask_0 = const()[name = string("v_cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_65_cast_fp16 = slice_by_index(begin = v_cache_65_begin_0, end = v_cache_65_end_0, end_mask = v_cache_65_end_mask_0, squeeze_mask = v_cache_65_squeeze_mask_0, x = coreml_update_state_79)[name = string("v_cache_65_cast_fp16")];
            tensor<int32, [4]> k_cache_67_begin_0 = const()[name = string("k_cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> k_cache_67_end_0 = const()[name = string("k_cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_67_end_mask_0 = const()[name = string("k_cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_67_squeeze_mask_0 = const()[name = string("k_cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_67_cast_fp16 = slice_by_index(begin = k_cache_67_begin_0, end = k_cache_67_end_0, end_mask = k_cache_67_end_mask_0, squeeze_mask = k_cache_67_squeeze_mask_0, x = read_state_2)[name = string("k_cache_67_cast_fp16")];
            tensor<int32, [4]> v_cache_67_begin_0 = const()[name = string("v_cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> v_cache_67_end_0 = const()[name = string("v_cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_67_end_mask_0 = const()[name = string("v_cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_67_squeeze_mask_0 = const()[name = string("v_cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_67_cast_fp16 = slice_by_index(begin = v_cache_67_begin_0, end = v_cache_67_end_0, end_mask = v_cache_67_end_mask_0, squeeze_mask = v_cache_67_squeeze_mask_0, x = read_state_3)[name = string("v_cache_67_cast_fp16")];
            int32 var_3524 = const()[name = string("op_3524"), val = int32(-1)];
            tensor<int32, [1]> var_3542_axes_0 = const()[name = string("op_3542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_16_attn_ln_weight_to_fp16 = const()[name = string("blocks_16_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580920704)))];
            tensor<fp16, [1024]> blocks_16_attn_ln_bias_to_fp16 = const()[name = string("blocks_16_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580922816)))];
            fp16 var_3530_to_fp16 = const()[name = string("op_3530_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_3542_cast_fp16 = layer_norm(axes = var_3542_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_3530_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_291_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3553_to_fp16 = const()[name = string("op_3553_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580924928)))];
            tensor<fp16, [1024]> var_3554_to_fp16 = const()[name = string("op_3554_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583022144)))];
            tensor<fp16, [1, ?, 1024]> linear_128_cast_fp16 = linear(bias = var_3554_to_fp16, weight = var_3553_to_fp16, x = var_3542_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3557_to_fp16 = const()[name = string("op_3557_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583024256)))];
            tensor<fp16, [1, ?, 1024]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3557_to_fp16, x = var_3542_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3561_to_fp16 = const()[name = string("op_3561_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585121472)))];
            tensor<fp16, [1024]> var_3562_to_fp16 = const()[name = string("op_3562_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587218688)))];
            tensor<fp16, [1, ?, 1024]> linear_130_cast_fp16 = linear(bias = var_3562_to_fp16, weight = var_3561_to_fp16, x = var_3542_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<int32, [3]> var_3564_shape_cast_fp16 = shape(x = linear_128_cast_fp16)[name = string("op_3564_shape_cast_fp16")];
            int32 gather_194_axis_0 = const()[name = string("gather_194_axis_0"), val = int32(0)];
            int32 gather_194_batch_dims_0 = const()[name = string("gather_194_batch_dims_0"), val = int32(0)];
            bool gather_194_validate_indices_0 = const()[name = string("gather_194_validate_indices_0"), val = bool(false)];
            string var_3564_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3564_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_194_to_uint16 = const()[name = string("select_194_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3564_shape_cast_fp16_to_uint16 = cast(dtype = var_3564_shape_cast_fp16_to_uint16_dtype_0, x = var_3564_shape_cast_fp16)[name = string("cast_262")];
            uint16 gather_194_cast_uint16 = gather(axis = gather_194_axis_0, batch_dims = gather_194_batch_dims_0, indices = select_194_to_uint16, validate_indices = gather_194_validate_indices_0, x = var_3564_shape_cast_fp16_to_uint16)[name = string("gather_194_cast_uint16")];
            string gather_194_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_194_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_194_cast_uint16_to_int32 = cast(dtype = gather_194_cast_uint16_to_int32_dtype_0, x = gather_194_cast_uint16)[name = string("cast_261")];
            int32 end_step_35 = add(x = offset, y = gather_194_cast_uint16_to_int32)[name = string("end_step_35")];
            tensor<int32, [1]> expand_dims_256 = const()[name = string("expand_dims_256"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_258 = const()[name = string("expand_dims_258"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_259_axes_0 = const()[name = string("expand_dims_259_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_259 = expand_dims(axes = expand_dims_259_axes_0, x = end_step_35)[name = string("expand_dims_259")];
            tensor<int32, [1]> concat_356_values0_0 = const()[name = string("concat_356_values0_0"), val = tensor<int32, [1]>([16])];
            int32 concat_356_axis_0 = const()[name = string("concat_356_axis_0"), val = int32(0)];
            bool concat_356_interleave_0 = const()[name = string("concat_356_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_356 = concat(axis = concat_356_axis_0, interleave = concat_356_interleave_0, values = (concat_356_values0_0, expand_dims_256, expand_dims_1, expand_dims_258))[name = string("concat_356")];
            tensor<int32, [1]> concat_357_values0_0 = const()[name = string("concat_357_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_357_values1_0 = const()[name = string("concat_357_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_357_values3_0 = const()[name = string("concat_357_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_357_axis_0 = const()[name = string("concat_357_axis_0"), val = int32(0)];
            bool concat_357_interleave_0 = const()[name = string("concat_357_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_357 = concat(axis = concat_357_axis_0, interleave = concat_357_interleave_0, values = (concat_357_values0_0, concat_357_values1_0, expand_dims_259, concat_357_values3_0))[name = string("concat_357")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_17_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_17_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_17_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_356, begin_mask = k_cache1_internal_tensor_assign_17_begin_mask_0, end = concat_357, end_mask = k_cache1_internal_tensor_assign_17_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_17_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_17_stride_0, update = linear_129_cast_fp16, x = coreml_update_state_78)[name = string("k_cache1_internal_tensor_assign_17_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_17_cast_fp16, input = k_cache1)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_80 = read_state(input = k_cache1)[name = string("coreml_update_state_80")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_17_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_17_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_17_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_356, begin_mask = v_cache1_internal_tensor_assign_17_begin_mask_0, end = concat_357, end_mask = v_cache1_internal_tensor_assign_17_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_17_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_17_stride_0, update = linear_130_cast_fp16, x = coreml_update_state_79)[name = string("v_cache1_internal_tensor_assign_17_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_17_cast_fp16, input = v_cache1)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_81 = read_state(input = v_cache1)[name = string("coreml_update_state_81")];
            int32 concat_362_values0_0 = const()[name = string("concat_362_values0_0"), val = int32(1)];
            int32 concat_362_values2_0 = const()[name = string("concat_362_values2_0"), val = int32(1024)];
            int32 concat_362_axis_0 = const()[name = string("concat_362_axis_0"), val = int32(0)];
            bool concat_362_interleave_0 = const()[name = string("concat_362_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_362 = concat(axis = concat_362_axis_0, interleave = concat_362_interleave_0, values = (concat_362_values0_0, end_step_35, concat_362_values2_0))[name = string("concat_362")];
            tensor<int32, [3]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = concat_362, end_mask = var_3580_end_mask_0, x = k_cache_65_cast_fp16)[name = string("op_3580_cast_fp16")];
            tensor<int32, [3]> var_3583_begin_0 = const()[name = string("op_3583_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3583_end_mask_0 = const()[name = string("op_3583_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3583_cast_fp16 = slice_by_index(begin = var_3583_begin_0, end = concat_362, end_mask = var_3583_end_mask_0, x = v_cache_65_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [4]> concat_364x = const()[name = string("concat_364x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3593_cast_fp16 = reshape(shape = concat_364x, x = linear_128_cast_fp16)[name = string("op_3593_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_184_to_fp16 = const()[name = string("const_184_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_131_cast_fp16 = mul(x = var_3593_cast_fp16, y = const_184_to_fp16)[name = string("q_131_cast_fp16")];
            tensor<int32, [4]> concat_365x = const()[name = string("concat_365x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3600_cast_fp16 = reshape(shape = concat_365x, x = var_3580_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_185_to_fp16 = const()[name = string("const_185_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_165_cast_fp16 = mul(x = var_3600_cast_fp16, y = const_185_to_fp16)[name = string("k_165_cast_fp16")];
            tensor<int32, [4]> concat_366x = const()[name = string("concat_366x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3607_cast_fp16 = reshape(shape = concat_366x, x = var_3583_cast_fp16)[name = string("op_3607_cast_fp16")];
            tensor<int32, [4]> var_3608 = const()[name = string("op_3608"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_97_transpose_x_0 = const()[name = string("qk_97_transpose_x_0"), val = bool(false)];
            bool qk_97_transpose_y_0 = const()[name = string("qk_97_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_257_perm_0 = const()[name = string("transpose_257_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_258_perm_0 = const()[name = string("transpose_258_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_258 = transpose(perm = transpose_258_perm_0, x = k_165_cast_fp16)[name = string("transpose_350")];
            tensor<fp16, [1, 16, ?, 64]> transpose_257 = transpose(perm = transpose_257_perm_0, x = q_131_cast_fp16)[name = string("transpose_351")];
            tensor<fp16, [1, 16, ?, ?]> qk_97_cast_fp16 = matmul(transpose_x = qk_97_transpose_x_0, transpose_y = qk_97_transpose_y_0, x = transpose_257, y = transpose_258)[name = string("qk_97_cast_fp16")];
            int32 concat_367_values1_0 = const()[name = string("concat_367_values1_0"), val = int32(448)];
            int32 concat_367_axis_0 = const()[name = string("concat_367_axis_0"), val = int32(0)];
            bool concat_367_interleave_0 = const()[name = string("concat_367_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_367 = concat(axis = concat_367_axis_0, interleave = concat_367_interleave_0, values = (gather_194_cast_uint16_to_int32, concat_367_values1_0))[name = string("concat_367")];
            tensor<int32, [2]> var_3611_begin_0 = const()[name = string("op_3611_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3611_end_mask_0 = const()[name = string("op_3611_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_3611_cast_fp16 = slice_by_index(begin = var_3611_begin_0, end = concat_367, end_mask = var_3611_end_mask_0, x = mask_to_fp16)[name = string("op_3611_cast_fp16")];
            int32 concat_368_values0_0 = const()[name = string("concat_368_values0_0"), val = int32(0)];
            int32 concat_368_axis_0 = const()[name = string("concat_368_axis_0"), val = int32(0)];
            bool concat_368_interleave_0 = const()[name = string("concat_368_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_368 = concat(axis = concat_368_axis_0, interleave = concat_368_interleave_0, values = (concat_368_values0_0, gather_194_cast_uint16_to_int32))[name = string("concat_368")];
            tensor<int32, [2]> var_3612_begin_0 = const()[name = string("op_3612_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3612_end_mask_0 = const()[name = string("op_3612_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_3612_cast_fp16 = slice_by_index(begin = var_3612_begin_0, end = concat_368, end_mask = var_3612_end_mask_0, x = var_3611_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_99_cast_fp16 = add(x = qk_97_cast_fp16, y = var_3612_cast_fp16)[name = string("qk_99_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_3615_cast_fp16 = softmax(axis = var_3524, x = qk_99_cast_fp16)[name = string("op_3615_cast_fp16")];
            bool var_3617_transpose_x_0 = const()[name = string("op_3617_transpose_x_0"), val = bool(false)];
            bool var_3617_transpose_y_0 = const()[name = string("op_3617_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_165_cast_fp16 = transpose(perm = var_3608, x = var_3607_cast_fp16)[name = string("transpose_352")];
            tensor<fp16, [1, 16, ?, 64]> var_3617_cast_fp16 = matmul(transpose_x = var_3617_transpose_x_0, transpose_y = var_3617_transpose_y_0, x = var_3615_cast_fp16, y = v_165_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<int32, [4]> var_3618 = const()[name = string("op_3618"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_369x = const()[name = string("concat_369x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3619_cast_fp16 = transpose(perm = var_3618, x = var_3617_cast_fp16)[name = string("transpose_349")];
            tensor<fp16, [1, ?, 1024]> x_295_cast_fp16 = reshape(shape = concat_369x, x = var_3619_cast_fp16)[name = string("x_295_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3623_to_fp16 = const()[name = string("op_3623_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587220800)))];
            tensor<fp16, [1024]> var_3624_to_fp16 = const()[name = string("op_3624_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589318016)))];
            tensor<fp16, [1, ?, 1024]> linear_131_cast_fp16 = linear(bias = var_3624_to_fp16, weight = var_3623_to_fp16, x = x_295_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_297_cast_fp16 = add(x = x_291_cast_fp16, y = linear_131_cast_fp16)[name = string("x_297_cast_fp16")];
            tensor<int32, [1]> var_3631_axes_0 = const()[name = string("op_3631_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_16_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_16_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589320128)))];
            tensor<fp16, [1024]> blocks_16_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_16_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589322240)))];
            tensor<fp16, [1, ?, 1024]> var_3631_cast_fp16 = layer_norm(axes = var_3631_axes_0, beta = blocks_16_cross_attn_ln_bias_to_fp16, epsilon = var_3530_to_fp16, gamma = blocks_16_cross_attn_ln_weight_to_fp16, x = x_297_cast_fp16)[name = string("op_3631_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3640_to_fp16 = const()[name = string("op_3640_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589324352)))];
            tensor<fp16, [1024]> var_3641_to_fp16 = const()[name = string("op_3641_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591421568)))];
            tensor<fp16, [1, ?, 1024]> linear_132_cast_fp16 = linear(bias = var_3641_to_fp16, weight = var_3640_to_fp16, x = var_3631_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<int32, [3]> concat_370 = const()[name = string("concat_370"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_371 = const()[name = string("concat_371"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_167_internal_tensor_assign_1_stride_0 = const()[name = string("k_167_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_167_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_167_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_167_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_167_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_167_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_167_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_167_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_370, begin_mask = k_167_internal_tensor_assign_1_begin_mask_0, end = concat_371, end_mask = k_167_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_167_internal_tensor_assign_1_squeeze_mask_0, stride = k_167_internal_tensor_assign_1_stride_0, update = k_cache_67_cast_fp16, x = k_7_to_fp16)[name = string("k_167_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_372 = const()[name = string("concat_372"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_373 = const()[name = string("concat_373"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_167_internal_tensor_assign_1_stride_0 = const()[name = string("v_167_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_167_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_167_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_167_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_167_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_167_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_167_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_167_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_372, begin_mask = v_167_internal_tensor_assign_1_begin_mask_0, end = concat_373, end_mask = v_167_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_167_internal_tensor_assign_1_squeeze_mask_0, stride = v_167_internal_tensor_assign_1_stride_0, update = v_cache_67_cast_fp16, x = k_7_to_fp16)[name = string("v_167_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_374x = const()[name = string("concat_374x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3661_cast_fp16 = reshape(shape = concat_374x, x = linear_132_cast_fp16)[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_186_to_fp16 = const()[name = string("const_186_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_135_cast_fp16 = mul(x = var_3661_cast_fp16, y = const_186_to_fp16)[name = string("q_135_cast_fp16")];
            tensor<int32, [4]> var_3667 = const()[name = string("op_3667"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3668_cast_fp16 = reshape(shape = var_3667, x = k_167_internal_tensor_assign_1_cast_fp16)[name = string("op_3668_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_187_to_fp16 = const()[name = string("const_187_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_169_cast_fp16 = mul(x = var_3668_cast_fp16, y = const_187_to_fp16)[name = string("k_169_cast_fp16")];
            tensor<int32, [4]> var_3674 = const()[name = string("op_3674"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3675_cast_fp16 = reshape(shape = var_3674, x = v_167_internal_tensor_assign_1_cast_fp16)[name = string("op_3675_cast_fp16")];
            tensor<int32, [4]> var_3676 = const()[name = string("op_3676"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_101_transpose_x_0 = const()[name = string("qk_101_transpose_x_0"), val = bool(false)];
            bool qk_101_transpose_y_0 = const()[name = string("qk_101_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_259_perm_0 = const()[name = string("transpose_259_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_260_perm_0 = const()[name = string("transpose_260_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_260 = transpose(perm = transpose_260_perm_0, x = k_169_cast_fp16)[name = string("transpose_346")];
            tensor<fp16, [1, 16, ?, 64]> transpose_259 = transpose(perm = transpose_259_perm_0, x = q_135_cast_fp16)[name = string("transpose_347")];
            tensor<fp16, [1, 16, ?, 1500]> qk_101_cast_fp16 = matmul(transpose_x = qk_101_transpose_x_0, transpose_y = qk_101_transpose_y_0, x = transpose_259, y = transpose_260)[name = string("qk_101_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_3680_cast_fp16 = softmax(axis = var_3524, x = qk_101_cast_fp16)[name = string("op_3680_cast_fp16")];
            bool var_3682_transpose_x_0 = const()[name = string("op_3682_transpose_x_0"), val = bool(false)];
            bool var_3682_transpose_y_0 = const()[name = string("op_3682_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_169_cast_fp16 = transpose(perm = var_3676, x = var_3675_cast_fp16)[name = string("transpose_348")];
            tensor<fp16, [1, 16, ?, 64]> var_3682_cast_fp16 = matmul(transpose_x = var_3682_transpose_x_0, transpose_y = var_3682_transpose_y_0, x = var_3680_cast_fp16, y = v_169_cast_fp16)[name = string("op_3682_cast_fp16")];
            tensor<int32, [4]> var_3683 = const()[name = string("op_3683"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_375x = const()[name = string("concat_375x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3684_cast_fp16 = transpose(perm = var_3683, x = var_3682_cast_fp16)[name = string("transpose_345")];
            tensor<fp16, [1, ?, 1024]> x_301_cast_fp16 = reshape(shape = concat_375x, x = var_3684_cast_fp16)[name = string("x_301_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3688_to_fp16 = const()[name = string("op_3688_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591423680)))];
            tensor<fp16, [1024]> var_3689_to_fp16 = const()[name = string("op_3689_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593520896)))];
            tensor<fp16, [1, ?, 1024]> linear_133_cast_fp16 = linear(bias = var_3689_to_fp16, weight = var_3688_to_fp16, x = x_301_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_303_cast_fp16 = add(x = x_297_cast_fp16, y = linear_133_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [1]> var_3696_axes_0 = const()[name = string("op_3696_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_16_mlp_ln_weight_to_fp16 = const()[name = string("blocks_16_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593523008)))];
            tensor<fp16, [1024]> blocks_16_mlp_ln_bias_to_fp16 = const()[name = string("blocks_16_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593525120)))];
            tensor<fp16, [1, ?, 1024]> var_3696_cast_fp16 = layer_norm(axes = var_3696_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_3530_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_303_cast_fp16)[name = string("op_3696_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_3705_to_fp16 = const()[name = string("op_3705_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593527232)))];
            tensor<fp16, [4096]> var_3706_to_fp16 = const()[name = string("op_3706_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601915904)))];
            tensor<fp16, [1, ?, 4096]> linear_134_cast_fp16 = linear(bias = var_3706_to_fp16, weight = var_3705_to_fp16, x = var_3696_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_307_mode_0 = const()[name = string("x_307_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_307_cast_fp16 = gelu(mode = x_307_mode_0, x = linear_134_cast_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_3711_to_fp16 = const()[name = string("op_3711_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601924160)))];
            tensor<fp16, [1024]> var_3712_to_fp16 = const()[name = string("op_3712_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610312832)))];
            tensor<fp16, [1, ?, 1024]> linear_135_cast_fp16 = linear(bias = var_3712_to_fp16, weight = var_3711_to_fp16, x = x_307_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_309_cast_fp16 = add(x = x_303_cast_fp16, y = linear_135_cast_fp16)[name = string("x_309_cast_fp16")];
            tensor<int32, [4]> k_cache_69_begin_0 = const()[name = string("k_cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> k_cache_69_end_0 = const()[name = string("k_cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_69_end_mask_0 = const()[name = string("k_cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_69_squeeze_mask_0 = const()[name = string("k_cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_69_cast_fp16 = slice_by_index(begin = k_cache_69_begin_0, end = k_cache_69_end_0, end_mask = k_cache_69_end_mask_0, squeeze_mask = k_cache_69_squeeze_mask_0, x = coreml_update_state_80)[name = string("k_cache_69_cast_fp16")];
            tensor<int32, [4]> v_cache_69_begin_0 = const()[name = string("v_cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> v_cache_69_end_0 = const()[name = string("v_cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_69_end_mask_0 = const()[name = string("v_cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_69_squeeze_mask_0 = const()[name = string("v_cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_69_cast_fp16 = slice_by_index(begin = v_cache_69_begin_0, end = v_cache_69_end_0, end_mask = v_cache_69_end_mask_0, squeeze_mask = v_cache_69_squeeze_mask_0, x = coreml_update_state_81)[name = string("v_cache_69_cast_fp16")];
            tensor<int32, [4]> k_cache_71_begin_0 = const()[name = string("k_cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> k_cache_71_end_0 = const()[name = string("k_cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_71_end_mask_0 = const()[name = string("k_cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_71_squeeze_mask_0 = const()[name = string("k_cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_71_cast_fp16 = slice_by_index(begin = k_cache_71_begin_0, end = k_cache_71_end_0, end_mask = k_cache_71_end_mask_0, squeeze_mask = k_cache_71_squeeze_mask_0, x = read_state_2)[name = string("k_cache_71_cast_fp16")];
            tensor<int32, [4]> v_cache_71_begin_0 = const()[name = string("v_cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> v_cache_71_end_0 = const()[name = string("v_cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_71_end_mask_0 = const()[name = string("v_cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_71_squeeze_mask_0 = const()[name = string("v_cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_71_cast_fp16 = slice_by_index(begin = v_cache_71_begin_0, end = v_cache_71_end_0, end_mask = v_cache_71_end_mask_0, squeeze_mask = v_cache_71_squeeze_mask_0, x = read_state_3)[name = string("v_cache_71_cast_fp16")];
            int32 var_3735 = const()[name = string("op_3735"), val = int32(-1)];
            tensor<int32, [1]> var_3753_axes_0 = const()[name = string("op_3753_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_17_attn_ln_weight_to_fp16 = const()[name = string("blocks_17_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610314944)))];
            tensor<fp16, [1024]> blocks_17_attn_ln_bias_to_fp16 = const()[name = string("blocks_17_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610317056)))];
            fp16 var_3741_to_fp16 = const()[name = string("op_3741_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_3753_cast_fp16 = layer_norm(axes = var_3753_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_3741_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_309_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3764_to_fp16 = const()[name = string("op_3764_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610319168)))];
            tensor<fp16, [1024]> var_3765_to_fp16 = const()[name = string("op_3765_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612416384)))];
            tensor<fp16, [1, ?, 1024]> linear_136_cast_fp16 = linear(bias = var_3765_to_fp16, weight = var_3764_to_fp16, x = var_3753_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3768_to_fp16 = const()[name = string("op_3768_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612418496)))];
            tensor<fp16, [1, ?, 1024]> linear_137_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3768_to_fp16, x = var_3753_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3772_to_fp16 = const()[name = string("op_3772_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614515712)))];
            tensor<fp16, [1024]> var_3773_to_fp16 = const()[name = string("op_3773_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616612928)))];
            tensor<fp16, [1, ?, 1024]> linear_138_cast_fp16 = linear(bias = var_3773_to_fp16, weight = var_3772_to_fp16, x = var_3753_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [3]> var_3775_shape_cast_fp16 = shape(x = linear_136_cast_fp16)[name = string("op_3775_shape_cast_fp16")];
            int32 gather_206_axis_0 = const()[name = string("gather_206_axis_0"), val = int32(0)];
            int32 gather_206_batch_dims_0 = const()[name = string("gather_206_batch_dims_0"), val = int32(0)];
            bool gather_206_validate_indices_0 = const()[name = string("gather_206_validate_indices_0"), val = bool(false)];
            string var_3775_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3775_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_206_to_uint16 = const()[name = string("select_206_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3775_shape_cast_fp16_to_uint16 = cast(dtype = var_3775_shape_cast_fp16_to_uint16_dtype_0, x = var_3775_shape_cast_fp16)[name = string("cast_260")];
            uint16 gather_206_cast_uint16 = gather(axis = gather_206_axis_0, batch_dims = gather_206_batch_dims_0, indices = select_206_to_uint16, validate_indices = gather_206_validate_indices_0, x = var_3775_shape_cast_fp16_to_uint16)[name = string("gather_206_cast_uint16")];
            string gather_206_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_206_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_206_cast_uint16_to_int32 = cast(dtype = gather_206_cast_uint16_to_int32_dtype_0, x = gather_206_cast_uint16)[name = string("cast_259")];
            int32 end_step_37 = add(x = offset, y = gather_206_cast_uint16_to_int32)[name = string("end_step_37")];
            tensor<int32, [1]> expand_dims_272 = const()[name = string("expand_dims_272"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_274 = const()[name = string("expand_dims_274"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_275_axes_0 = const()[name = string("expand_dims_275_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_275 = expand_dims(axes = expand_dims_275_axes_0, x = end_step_37)[name = string("expand_dims_275")];
            tensor<int32, [1]> concat_378_values0_0 = const()[name = string("concat_378_values0_0"), val = tensor<int32, [1]>([17])];
            int32 concat_378_axis_0 = const()[name = string("concat_378_axis_0"), val = int32(0)];
            bool concat_378_interleave_0 = const()[name = string("concat_378_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_378 = concat(axis = concat_378_axis_0, interleave = concat_378_interleave_0, values = (concat_378_values0_0, expand_dims_272, expand_dims_1, expand_dims_274))[name = string("concat_378")];
            tensor<int32, [1]> concat_379_values0_0 = const()[name = string("concat_379_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_379_values1_0 = const()[name = string("concat_379_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_379_values3_0 = const()[name = string("concat_379_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_379_axis_0 = const()[name = string("concat_379_axis_0"), val = int32(0)];
            bool concat_379_interleave_0 = const()[name = string("concat_379_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_379 = concat(axis = concat_379_axis_0, interleave = concat_379_interleave_0, values = (concat_379_values0_0, concat_379_values1_0, expand_dims_275, concat_379_values3_0))[name = string("concat_379")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_18_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_18_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_18_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_378, begin_mask = k_cache1_internal_tensor_assign_18_begin_mask_0, end = concat_379, end_mask = k_cache1_internal_tensor_assign_18_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_18_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_18_stride_0, update = linear_137_cast_fp16, x = coreml_update_state_80)[name = string("k_cache1_internal_tensor_assign_18_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_18_cast_fp16, input = k_cache1)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_82 = read_state(input = k_cache1)[name = string("coreml_update_state_82")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_18_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_18_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_18_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_378, begin_mask = v_cache1_internal_tensor_assign_18_begin_mask_0, end = concat_379, end_mask = v_cache1_internal_tensor_assign_18_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_18_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_18_stride_0, update = linear_138_cast_fp16, x = coreml_update_state_81)[name = string("v_cache1_internal_tensor_assign_18_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_18_cast_fp16, input = v_cache1)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_83 = read_state(input = v_cache1)[name = string("coreml_update_state_83")];
            int32 concat_384_values0_0 = const()[name = string("concat_384_values0_0"), val = int32(1)];
            int32 concat_384_values2_0 = const()[name = string("concat_384_values2_0"), val = int32(1024)];
            int32 concat_384_axis_0 = const()[name = string("concat_384_axis_0"), val = int32(0)];
            bool concat_384_interleave_0 = const()[name = string("concat_384_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_384 = concat(axis = concat_384_axis_0, interleave = concat_384_interleave_0, values = (concat_384_values0_0, end_step_37, concat_384_values2_0))[name = string("concat_384")];
            tensor<int32, [3]> var_3791_begin_0 = const()[name = string("op_3791_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3791_end_mask_0 = const()[name = string("op_3791_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3791_cast_fp16 = slice_by_index(begin = var_3791_begin_0, end = concat_384, end_mask = var_3791_end_mask_0, x = k_cache_69_cast_fp16)[name = string("op_3791_cast_fp16")];
            tensor<int32, [3]> var_3794_begin_0 = const()[name = string("op_3794_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_3794_end_mask_0 = const()[name = string("op_3794_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_3794_cast_fp16 = slice_by_index(begin = var_3794_begin_0, end = concat_384, end_mask = var_3794_end_mask_0, x = v_cache_69_cast_fp16)[name = string("op_3794_cast_fp16")];
            tensor<int32, [4]> concat_386x = const()[name = string("concat_386x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3804_cast_fp16 = reshape(shape = concat_386x, x = linear_136_cast_fp16)[name = string("op_3804_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_188_to_fp16 = const()[name = string("const_188_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_139_cast_fp16 = mul(x = var_3804_cast_fp16, y = const_188_to_fp16)[name = string("q_139_cast_fp16")];
            tensor<int32, [4]> concat_387x = const()[name = string("concat_387x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3811_cast_fp16 = reshape(shape = concat_387x, x = var_3791_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_175_cast_fp16 = mul(x = var_3811_cast_fp16, y = const_189_to_fp16)[name = string("k_175_cast_fp16")];
            tensor<int32, [4]> concat_388x = const()[name = string("concat_388x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3818_cast_fp16 = reshape(shape = concat_388x, x = var_3794_cast_fp16)[name = string("op_3818_cast_fp16")];
            tensor<int32, [4]> var_3819 = const()[name = string("op_3819"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_103_transpose_x_0 = const()[name = string("qk_103_transpose_x_0"), val = bool(false)];
            bool qk_103_transpose_y_0 = const()[name = string("qk_103_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_261_perm_0 = const()[name = string("transpose_261_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_262_perm_0 = const()[name = string("transpose_262_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_262 = transpose(perm = transpose_262_perm_0, x = k_175_cast_fp16)[name = string("transpose_342")];
            tensor<fp16, [1, 16, ?, 64]> transpose_261 = transpose(perm = transpose_261_perm_0, x = q_139_cast_fp16)[name = string("transpose_343")];
            tensor<fp16, [1, 16, ?, ?]> qk_103_cast_fp16 = matmul(transpose_x = qk_103_transpose_x_0, transpose_y = qk_103_transpose_y_0, x = transpose_261, y = transpose_262)[name = string("qk_103_cast_fp16")];
            int32 concat_389_values1_0 = const()[name = string("concat_389_values1_0"), val = int32(448)];
            int32 concat_389_axis_0 = const()[name = string("concat_389_axis_0"), val = int32(0)];
            bool concat_389_interleave_0 = const()[name = string("concat_389_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_389 = concat(axis = concat_389_axis_0, interleave = concat_389_interleave_0, values = (gather_206_cast_uint16_to_int32, concat_389_values1_0))[name = string("concat_389")];
            tensor<int32, [2]> var_3822_begin_0 = const()[name = string("op_3822_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3822_end_mask_0 = const()[name = string("op_3822_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_3822_cast_fp16 = slice_by_index(begin = var_3822_begin_0, end = concat_389, end_mask = var_3822_end_mask_0, x = mask_to_fp16)[name = string("op_3822_cast_fp16")];
            int32 concat_390_values0_0 = const()[name = string("concat_390_values0_0"), val = int32(0)];
            int32 concat_390_axis_0 = const()[name = string("concat_390_axis_0"), val = int32(0)];
            bool concat_390_interleave_0 = const()[name = string("concat_390_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_390 = concat(axis = concat_390_axis_0, interleave = concat_390_interleave_0, values = (concat_390_values0_0, gather_206_cast_uint16_to_int32))[name = string("concat_390")];
            tensor<int32, [2]> var_3823_begin_0 = const()[name = string("op_3823_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_3823_end_mask_0 = const()[name = string("op_3823_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_3823_cast_fp16 = slice_by_index(begin = var_3823_begin_0, end = concat_390, end_mask = var_3823_end_mask_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_105_cast_fp16 = add(x = qk_103_cast_fp16, y = var_3823_cast_fp16)[name = string("qk_105_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_3826_cast_fp16 = softmax(axis = var_3735, x = qk_105_cast_fp16)[name = string("op_3826_cast_fp16")];
            bool var_3828_transpose_x_0 = const()[name = string("op_3828_transpose_x_0"), val = bool(false)];
            bool var_3828_transpose_y_0 = const()[name = string("op_3828_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_175_cast_fp16 = transpose(perm = var_3819, x = var_3818_cast_fp16)[name = string("transpose_344")];
            tensor<fp16, [1, 16, ?, 64]> var_3828_cast_fp16 = matmul(transpose_x = var_3828_transpose_x_0, transpose_y = var_3828_transpose_y_0, x = var_3826_cast_fp16, y = v_175_cast_fp16)[name = string("op_3828_cast_fp16")];
            tensor<int32, [4]> var_3829 = const()[name = string("op_3829"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_391x = const()[name = string("concat_391x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3830_cast_fp16 = transpose(perm = var_3829, x = var_3828_cast_fp16)[name = string("transpose_341")];
            tensor<fp16, [1, ?, 1024]> x_313_cast_fp16 = reshape(shape = concat_391x, x = var_3830_cast_fp16)[name = string("x_313_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616615040)))];
            tensor<fp16, [1024]> var_3835_to_fp16 = const()[name = string("op_3835_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618712256)))];
            tensor<fp16, [1, ?, 1024]> linear_139_cast_fp16 = linear(bias = var_3835_to_fp16, weight = var_3834_to_fp16, x = x_313_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_315_cast_fp16 = add(x = x_309_cast_fp16, y = linear_139_cast_fp16)[name = string("x_315_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_17_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_17_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618714368)))];
            tensor<fp16, [1024]> blocks_17_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_17_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618716480)))];
            tensor<fp16, [1, ?, 1024]> var_3842_cast_fp16 = layer_norm(axes = var_3842_axes_0, beta = blocks_17_cross_attn_ln_bias_to_fp16, epsilon = var_3741_to_fp16, gamma = blocks_17_cross_attn_ln_weight_to_fp16, x = x_315_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3851_to_fp16 = const()[name = string("op_3851_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618718592)))];
            tensor<fp16, [1024]> var_3852_to_fp16 = const()[name = string("op_3852_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620815808)))];
            tensor<fp16, [1, ?, 1024]> linear_140_cast_fp16 = linear(bias = var_3852_to_fp16, weight = var_3851_to_fp16, x = var_3842_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [3]> concat_392 = const()[name = string("concat_392"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_393 = const()[name = string("concat_393"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_177_internal_tensor_assign_1_stride_0 = const()[name = string("k_177_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_177_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_177_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_177_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_177_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_177_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_177_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_177_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_392, begin_mask = k_177_internal_tensor_assign_1_begin_mask_0, end = concat_393, end_mask = k_177_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_177_internal_tensor_assign_1_squeeze_mask_0, stride = k_177_internal_tensor_assign_1_stride_0, update = k_cache_71_cast_fp16, x = k_7_to_fp16)[name = string("k_177_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_394 = const()[name = string("concat_394"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_395 = const()[name = string("concat_395"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_177_internal_tensor_assign_1_stride_0 = const()[name = string("v_177_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_177_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_177_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_177_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_177_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_177_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_177_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_177_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_394, begin_mask = v_177_internal_tensor_assign_1_begin_mask_0, end = concat_395, end_mask = v_177_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_177_internal_tensor_assign_1_squeeze_mask_0, stride = v_177_internal_tensor_assign_1_stride_0, update = v_cache_71_cast_fp16, x = k_7_to_fp16)[name = string("v_177_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_396x = const()[name = string("concat_396x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_3872_cast_fp16 = reshape(shape = concat_396x, x = linear_140_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_143_cast_fp16 = mul(x = var_3872_cast_fp16, y = const_190_to_fp16)[name = string("q_143_cast_fp16")];
            tensor<int32, [4]> var_3878 = const()[name = string("op_3878"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3879_cast_fp16 = reshape(shape = var_3878, x = k_177_internal_tensor_assign_1_cast_fp16)[name = string("op_3879_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_191_to_fp16 = const()[name = string("const_191_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_179_cast_fp16 = mul(x = var_3879_cast_fp16, y = const_191_to_fp16)[name = string("k_179_cast_fp16")];
            tensor<int32, [4]> var_3885 = const()[name = string("op_3885"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_3886_cast_fp16 = reshape(shape = var_3885, x = v_177_internal_tensor_assign_1_cast_fp16)[name = string("op_3886_cast_fp16")];
            tensor<int32, [4]> var_3887 = const()[name = string("op_3887"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_107_transpose_x_0 = const()[name = string("qk_107_transpose_x_0"), val = bool(false)];
            bool qk_107_transpose_y_0 = const()[name = string("qk_107_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_263_perm_0 = const()[name = string("transpose_263_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_264_perm_0 = const()[name = string("transpose_264_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_264 = transpose(perm = transpose_264_perm_0, x = k_179_cast_fp16)[name = string("transpose_338")];
            tensor<fp16, [1, 16, ?, 64]> transpose_263 = transpose(perm = transpose_263_perm_0, x = q_143_cast_fp16)[name = string("transpose_339")];
            tensor<fp16, [1, 16, ?, 1500]> qk_107_cast_fp16 = matmul(transpose_x = qk_107_transpose_x_0, transpose_y = qk_107_transpose_y_0, x = transpose_263, y = transpose_264)[name = string("qk_107_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_3891_cast_fp16 = softmax(axis = var_3735, x = qk_107_cast_fp16)[name = string("op_3891_cast_fp16")];
            bool var_3893_transpose_x_0 = const()[name = string("op_3893_transpose_x_0"), val = bool(false)];
            bool var_3893_transpose_y_0 = const()[name = string("op_3893_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_179_cast_fp16 = transpose(perm = var_3887, x = var_3886_cast_fp16)[name = string("transpose_340")];
            tensor<fp16, [1, 16, ?, 64]> var_3893_cast_fp16 = matmul(transpose_x = var_3893_transpose_x_0, transpose_y = var_3893_transpose_y_0, x = var_3891_cast_fp16, y = v_179_cast_fp16)[name = string("op_3893_cast_fp16")];
            tensor<int32, [4]> var_3894 = const()[name = string("op_3894"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_397x = const()[name = string("concat_397x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_3895_cast_fp16 = transpose(perm = var_3894, x = var_3893_cast_fp16)[name = string("transpose_337")];
            tensor<fp16, [1, ?, 1024]> x_319_cast_fp16 = reshape(shape = concat_397x, x = var_3895_cast_fp16)[name = string("x_319_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3899_to_fp16 = const()[name = string("op_3899_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620817920)))];
            tensor<fp16, [1024]> var_3900_to_fp16 = const()[name = string("op_3900_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622915136)))];
            tensor<fp16, [1, ?, 1024]> linear_141_cast_fp16 = linear(bias = var_3900_to_fp16, weight = var_3899_to_fp16, x = x_319_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_321_cast_fp16 = add(x = x_315_cast_fp16, y = linear_141_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [1]> var_3907_axes_0 = const()[name = string("op_3907_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_17_mlp_ln_weight_to_fp16 = const()[name = string("blocks_17_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622917248)))];
            tensor<fp16, [1024]> blocks_17_mlp_ln_bias_to_fp16 = const()[name = string("blocks_17_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622919360)))];
            tensor<fp16, [1, ?, 1024]> var_3907_cast_fp16 = layer_norm(axes = var_3907_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_3741_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_321_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622921472)))];
            tensor<fp16, [4096]> var_3917_to_fp16 = const()[name = string("op_3917_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631310144)))];
            tensor<fp16, [1, ?, 4096]> linear_142_cast_fp16 = linear(bias = var_3917_to_fp16, weight = var_3916_to_fp16, x = var_3907_cast_fp16)[name = string("linear_142_cast_fp16")];
            string x_325_mode_0 = const()[name = string("x_325_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_325_cast_fp16 = gelu(mode = x_325_mode_0, x = linear_142_cast_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631318400)))];
            tensor<fp16, [1024]> var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639707072)))];
            tensor<fp16, [1, ?, 1024]> linear_143_cast_fp16 = linear(bias = var_3923_to_fp16, weight = var_3922_to_fp16, x = x_325_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_327_cast_fp16 = add(x = x_321_cast_fp16, y = linear_143_cast_fp16)[name = string("x_327_cast_fp16")];
            tensor<int32, [4]> k_cache_73_begin_0 = const()[name = string("k_cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> k_cache_73_end_0 = const()[name = string("k_cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_73_end_mask_0 = const()[name = string("k_cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_73_squeeze_mask_0 = const()[name = string("k_cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_73_cast_fp16 = slice_by_index(begin = k_cache_73_begin_0, end = k_cache_73_end_0, end_mask = k_cache_73_end_mask_0, squeeze_mask = k_cache_73_squeeze_mask_0, x = coreml_update_state_82)[name = string("k_cache_73_cast_fp16")];
            tensor<int32, [4]> v_cache_73_begin_0 = const()[name = string("v_cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> v_cache_73_end_0 = const()[name = string("v_cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_73_end_mask_0 = const()[name = string("v_cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_73_squeeze_mask_0 = const()[name = string("v_cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_73_cast_fp16 = slice_by_index(begin = v_cache_73_begin_0, end = v_cache_73_end_0, end_mask = v_cache_73_end_mask_0, squeeze_mask = v_cache_73_squeeze_mask_0, x = coreml_update_state_83)[name = string("v_cache_73_cast_fp16")];
            tensor<int32, [4]> k_cache_75_begin_0 = const()[name = string("k_cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> k_cache_75_end_0 = const()[name = string("k_cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_75_end_mask_0 = const()[name = string("k_cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_75_squeeze_mask_0 = const()[name = string("k_cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_75_cast_fp16 = slice_by_index(begin = k_cache_75_begin_0, end = k_cache_75_end_0, end_mask = k_cache_75_end_mask_0, squeeze_mask = k_cache_75_squeeze_mask_0, x = read_state_2)[name = string("k_cache_75_cast_fp16")];
            tensor<int32, [4]> v_cache_75_begin_0 = const()[name = string("v_cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> v_cache_75_end_0 = const()[name = string("v_cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_75_end_mask_0 = const()[name = string("v_cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_75_squeeze_mask_0 = const()[name = string("v_cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_75_cast_fp16 = slice_by_index(begin = v_cache_75_begin_0, end = v_cache_75_end_0, end_mask = v_cache_75_end_mask_0, squeeze_mask = v_cache_75_squeeze_mask_0, x = read_state_3)[name = string("v_cache_75_cast_fp16")];
            int32 var_3946 = const()[name = string("op_3946"), val = int32(-1)];
            tensor<int32, [1]> var_3964_axes_0 = const()[name = string("op_3964_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_18_attn_ln_weight_to_fp16 = const()[name = string("blocks_18_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639709184)))];
            tensor<fp16, [1024]> blocks_18_attn_ln_bias_to_fp16 = const()[name = string("blocks_18_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639711296)))];
            fp16 var_3952_to_fp16 = const()[name = string("op_3952_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_3964_cast_fp16 = layer_norm(axes = var_3964_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_3952_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_327_cast_fp16)[name = string("op_3964_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3975_to_fp16 = const()[name = string("op_3975_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639713408)))];
            tensor<fp16, [1024]> var_3976_to_fp16 = const()[name = string("op_3976_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(641810624)))];
            tensor<fp16, [1, ?, 1024]> linear_144_cast_fp16 = linear(bias = var_3976_to_fp16, weight = var_3975_to_fp16, x = var_3964_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3979_to_fp16 = const()[name = string("op_3979_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(641812736)))];
            tensor<fp16, [1, ?, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3979_to_fp16, x = var_3964_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_3983_to_fp16 = const()[name = string("op_3983_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643909952)))];
            tensor<fp16, [1024]> var_3984_to_fp16 = const()[name = string("op_3984_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646007168)))];
            tensor<fp16, [1, ?, 1024]> linear_146_cast_fp16 = linear(bias = var_3984_to_fp16, weight = var_3983_to_fp16, x = var_3964_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [3]> var_3986_shape_cast_fp16 = shape(x = linear_144_cast_fp16)[name = string("op_3986_shape_cast_fp16")];
            int32 gather_218_axis_0 = const()[name = string("gather_218_axis_0"), val = int32(0)];
            int32 gather_218_batch_dims_0 = const()[name = string("gather_218_batch_dims_0"), val = int32(0)];
            bool gather_218_validate_indices_0 = const()[name = string("gather_218_validate_indices_0"), val = bool(false)];
            string var_3986_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3986_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_218_to_uint16 = const()[name = string("select_218_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3986_shape_cast_fp16_to_uint16 = cast(dtype = var_3986_shape_cast_fp16_to_uint16_dtype_0, x = var_3986_shape_cast_fp16)[name = string("cast_258")];
            uint16 gather_218_cast_uint16 = gather(axis = gather_218_axis_0, batch_dims = gather_218_batch_dims_0, indices = select_218_to_uint16, validate_indices = gather_218_validate_indices_0, x = var_3986_shape_cast_fp16_to_uint16)[name = string("gather_218_cast_uint16")];
            string gather_218_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_218_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_218_cast_uint16_to_int32 = cast(dtype = gather_218_cast_uint16_to_int32_dtype_0, x = gather_218_cast_uint16)[name = string("cast_257")];
            int32 end_step_39 = add(x = offset, y = gather_218_cast_uint16_to_int32)[name = string("end_step_39")];
            tensor<int32, [1]> expand_dims_288 = const()[name = string("expand_dims_288"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_290 = const()[name = string("expand_dims_290"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_291_axes_0 = const()[name = string("expand_dims_291_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_291 = expand_dims(axes = expand_dims_291_axes_0, x = end_step_39)[name = string("expand_dims_291")];
            tensor<int32, [1]> concat_400_values0_0 = const()[name = string("concat_400_values0_0"), val = tensor<int32, [1]>([18])];
            int32 concat_400_axis_0 = const()[name = string("concat_400_axis_0"), val = int32(0)];
            bool concat_400_interleave_0 = const()[name = string("concat_400_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_400 = concat(axis = concat_400_axis_0, interleave = concat_400_interleave_0, values = (concat_400_values0_0, expand_dims_288, expand_dims_1, expand_dims_290))[name = string("concat_400")];
            tensor<int32, [1]> concat_401_values0_0 = const()[name = string("concat_401_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_401_values1_0 = const()[name = string("concat_401_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_401_values3_0 = const()[name = string("concat_401_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_401_axis_0 = const()[name = string("concat_401_axis_0"), val = int32(0)];
            bool concat_401_interleave_0 = const()[name = string("concat_401_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_401 = concat(axis = concat_401_axis_0, interleave = concat_401_interleave_0, values = (concat_401_values0_0, concat_401_values1_0, expand_dims_291, concat_401_values3_0))[name = string("concat_401")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_19_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_19_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_19_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_400, begin_mask = k_cache1_internal_tensor_assign_19_begin_mask_0, end = concat_401, end_mask = k_cache1_internal_tensor_assign_19_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_19_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_19_stride_0, update = linear_145_cast_fp16, x = coreml_update_state_82)[name = string("k_cache1_internal_tensor_assign_19_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_19_cast_fp16, input = k_cache1)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_84 = read_state(input = k_cache1)[name = string("coreml_update_state_84")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_19_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_19_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_19_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_400, begin_mask = v_cache1_internal_tensor_assign_19_begin_mask_0, end = concat_401, end_mask = v_cache1_internal_tensor_assign_19_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_19_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_19_stride_0, update = linear_146_cast_fp16, x = coreml_update_state_83)[name = string("v_cache1_internal_tensor_assign_19_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_19_cast_fp16, input = v_cache1)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_85 = read_state(input = v_cache1)[name = string("coreml_update_state_85")];
            int32 concat_406_values0_0 = const()[name = string("concat_406_values0_0"), val = int32(1)];
            int32 concat_406_values2_0 = const()[name = string("concat_406_values2_0"), val = int32(1024)];
            int32 concat_406_axis_0 = const()[name = string("concat_406_axis_0"), val = int32(0)];
            bool concat_406_interleave_0 = const()[name = string("concat_406_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_406 = concat(axis = concat_406_axis_0, interleave = concat_406_interleave_0, values = (concat_406_values0_0, end_step_39, concat_406_values2_0))[name = string("concat_406")];
            tensor<int32, [3]> var_4002_begin_0 = const()[name = string("op_4002_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4002_end_mask_0 = const()[name = string("op_4002_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4002_cast_fp16 = slice_by_index(begin = var_4002_begin_0, end = concat_406, end_mask = var_4002_end_mask_0, x = k_cache_73_cast_fp16)[name = string("op_4002_cast_fp16")];
            tensor<int32, [3]> var_4005_begin_0 = const()[name = string("op_4005_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4005_end_mask_0 = const()[name = string("op_4005_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = concat_406, end_mask = var_4005_end_mask_0, x = v_cache_73_cast_fp16)[name = string("op_4005_cast_fp16")];
            tensor<int32, [4]> concat_408x = const()[name = string("concat_408x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4015_cast_fp16 = reshape(shape = concat_408x, x = linear_144_cast_fp16)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_192_to_fp16 = const()[name = string("const_192_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_147_cast_fp16 = mul(x = var_4015_cast_fp16, y = const_192_to_fp16)[name = string("q_147_cast_fp16")];
            tensor<int32, [4]> concat_409x = const()[name = string("concat_409x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4022_cast_fp16 = reshape(shape = concat_409x, x = var_4002_cast_fp16)[name = string("op_4022_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_193_to_fp16 = const()[name = string("const_193_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_185_cast_fp16 = mul(x = var_4022_cast_fp16, y = const_193_to_fp16)[name = string("k_185_cast_fp16")];
            tensor<int32, [4]> concat_410x = const()[name = string("concat_410x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4029_cast_fp16 = reshape(shape = concat_410x, x = var_4005_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [4]> var_4030 = const()[name = string("op_4030"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_109_transpose_x_0 = const()[name = string("qk_109_transpose_x_0"), val = bool(false)];
            bool qk_109_transpose_y_0 = const()[name = string("qk_109_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_265_perm_0 = const()[name = string("transpose_265_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_266_perm_0 = const()[name = string("transpose_266_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_266 = transpose(perm = transpose_266_perm_0, x = k_185_cast_fp16)[name = string("transpose_334")];
            tensor<fp16, [1, 16, ?, 64]> transpose_265 = transpose(perm = transpose_265_perm_0, x = q_147_cast_fp16)[name = string("transpose_335")];
            tensor<fp16, [1, 16, ?, ?]> qk_109_cast_fp16 = matmul(transpose_x = qk_109_transpose_x_0, transpose_y = qk_109_transpose_y_0, x = transpose_265, y = transpose_266)[name = string("qk_109_cast_fp16")];
            int32 concat_411_values1_0 = const()[name = string("concat_411_values1_0"), val = int32(448)];
            int32 concat_411_axis_0 = const()[name = string("concat_411_axis_0"), val = int32(0)];
            bool concat_411_interleave_0 = const()[name = string("concat_411_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_411 = concat(axis = concat_411_axis_0, interleave = concat_411_interleave_0, values = (gather_218_cast_uint16_to_int32, concat_411_values1_0))[name = string("concat_411")];
            tensor<int32, [2]> var_4033_begin_0 = const()[name = string("op_4033_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4033_end_mask_0 = const()[name = string("op_4033_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_4033_cast_fp16 = slice_by_index(begin = var_4033_begin_0, end = concat_411, end_mask = var_4033_end_mask_0, x = mask_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 concat_412_values0_0 = const()[name = string("concat_412_values0_0"), val = int32(0)];
            int32 concat_412_axis_0 = const()[name = string("concat_412_axis_0"), val = int32(0)];
            bool concat_412_interleave_0 = const()[name = string("concat_412_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_412 = concat(axis = concat_412_axis_0, interleave = concat_412_interleave_0, values = (concat_412_values0_0, gather_218_cast_uint16_to_int32))[name = string("concat_412")];
            tensor<int32, [2]> var_4034_begin_0 = const()[name = string("op_4034_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4034_end_mask_0 = const()[name = string("op_4034_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_4034_cast_fp16 = slice_by_index(begin = var_4034_begin_0, end = concat_412, end_mask = var_4034_end_mask_0, x = var_4033_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_111_cast_fp16 = add(x = qk_109_cast_fp16, y = var_4034_cast_fp16)[name = string("qk_111_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_4037_cast_fp16 = softmax(axis = var_3946, x = qk_111_cast_fp16)[name = string("op_4037_cast_fp16")];
            bool var_4039_transpose_x_0 = const()[name = string("op_4039_transpose_x_0"), val = bool(false)];
            bool var_4039_transpose_y_0 = const()[name = string("op_4039_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_185_cast_fp16 = transpose(perm = var_4030, x = var_4029_cast_fp16)[name = string("transpose_336")];
            tensor<fp16, [1, 16, ?, 64]> var_4039_cast_fp16 = matmul(transpose_x = var_4039_transpose_x_0, transpose_y = var_4039_transpose_y_0, x = var_4037_cast_fp16, y = v_185_cast_fp16)[name = string("op_4039_cast_fp16")];
            tensor<int32, [4]> var_4040 = const()[name = string("op_4040"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_413x = const()[name = string("concat_413x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4041_cast_fp16 = transpose(perm = var_4040, x = var_4039_cast_fp16)[name = string("transpose_333")];
            tensor<fp16, [1, ?, 1024]> x_331_cast_fp16 = reshape(shape = concat_413x, x = var_4041_cast_fp16)[name = string("x_331_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4045_to_fp16 = const()[name = string("op_4045_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646009280)))];
            tensor<fp16, [1024]> var_4046_to_fp16 = const()[name = string("op_4046_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648106496)))];
            tensor<fp16, [1, ?, 1024]> linear_147_cast_fp16 = linear(bias = var_4046_to_fp16, weight = var_4045_to_fp16, x = x_331_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_333_cast_fp16 = add(x = x_327_cast_fp16, y = linear_147_cast_fp16)[name = string("x_333_cast_fp16")];
            tensor<int32, [1]> var_4053_axes_0 = const()[name = string("op_4053_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_18_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_18_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648108608)))];
            tensor<fp16, [1024]> blocks_18_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_18_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648110720)))];
            tensor<fp16, [1, ?, 1024]> var_4053_cast_fp16 = layer_norm(axes = var_4053_axes_0, beta = blocks_18_cross_attn_ln_bias_to_fp16, epsilon = var_3952_to_fp16, gamma = blocks_18_cross_attn_ln_weight_to_fp16, x = x_333_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648112832)))];
            tensor<fp16, [1024]> var_4063_to_fp16 = const()[name = string("op_4063_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650210048)))];
            tensor<fp16, [1, ?, 1024]> linear_148_cast_fp16 = linear(bias = var_4063_to_fp16, weight = var_4062_to_fp16, x = var_4053_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [3]> concat_414 = const()[name = string("concat_414"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_415 = const()[name = string("concat_415"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_187_internal_tensor_assign_1_stride_0 = const()[name = string("k_187_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_187_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_187_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_187_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_187_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_187_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_187_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_187_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_414, begin_mask = k_187_internal_tensor_assign_1_begin_mask_0, end = concat_415, end_mask = k_187_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_187_internal_tensor_assign_1_squeeze_mask_0, stride = k_187_internal_tensor_assign_1_stride_0, update = k_cache_75_cast_fp16, x = k_7_to_fp16)[name = string("k_187_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_416 = const()[name = string("concat_416"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_417 = const()[name = string("concat_417"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_187_internal_tensor_assign_1_stride_0 = const()[name = string("v_187_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_187_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_187_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_187_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_187_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_187_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_187_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_187_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_416, begin_mask = v_187_internal_tensor_assign_1_begin_mask_0, end = concat_417, end_mask = v_187_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_187_internal_tensor_assign_1_squeeze_mask_0, stride = v_187_internal_tensor_assign_1_stride_0, update = v_cache_75_cast_fp16, x = k_7_to_fp16)[name = string("v_187_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_418x = const()[name = string("concat_418x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4083_cast_fp16 = reshape(shape = concat_418x, x = linear_148_cast_fp16)[name = string("op_4083_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_194_to_fp16 = const()[name = string("const_194_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_151_cast_fp16 = mul(x = var_4083_cast_fp16, y = const_194_to_fp16)[name = string("q_151_cast_fp16")];
            tensor<int32, [4]> var_4089 = const()[name = string("op_4089"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4090_cast_fp16 = reshape(shape = var_4089, x = k_187_internal_tensor_assign_1_cast_fp16)[name = string("op_4090_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_195_to_fp16 = const()[name = string("const_195_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_189_cast_fp16 = mul(x = var_4090_cast_fp16, y = const_195_to_fp16)[name = string("k_189_cast_fp16")];
            tensor<int32, [4]> var_4096 = const()[name = string("op_4096"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4097_cast_fp16 = reshape(shape = var_4096, x = v_187_internal_tensor_assign_1_cast_fp16)[name = string("op_4097_cast_fp16")];
            tensor<int32, [4]> var_4098 = const()[name = string("op_4098"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_113_transpose_x_0 = const()[name = string("qk_113_transpose_x_0"), val = bool(false)];
            bool qk_113_transpose_y_0 = const()[name = string("qk_113_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_267_perm_0 = const()[name = string("transpose_267_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_268_perm_0 = const()[name = string("transpose_268_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_268 = transpose(perm = transpose_268_perm_0, x = k_189_cast_fp16)[name = string("transpose_330")];
            tensor<fp16, [1, 16, ?, 64]> transpose_267 = transpose(perm = transpose_267_perm_0, x = q_151_cast_fp16)[name = string("transpose_331")];
            tensor<fp16, [1, 16, ?, 1500]> qk_113_cast_fp16 = matmul(transpose_x = qk_113_transpose_x_0, transpose_y = qk_113_transpose_y_0, x = transpose_267, y = transpose_268)[name = string("qk_113_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_4102_cast_fp16 = softmax(axis = var_3946, x = qk_113_cast_fp16)[name = string("op_4102_cast_fp16")];
            bool var_4104_transpose_x_0 = const()[name = string("op_4104_transpose_x_0"), val = bool(false)];
            bool var_4104_transpose_y_0 = const()[name = string("op_4104_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_189_cast_fp16 = transpose(perm = var_4098, x = var_4097_cast_fp16)[name = string("transpose_332")];
            tensor<fp16, [1, 16, ?, 64]> var_4104_cast_fp16 = matmul(transpose_x = var_4104_transpose_x_0, transpose_y = var_4104_transpose_y_0, x = var_4102_cast_fp16, y = v_189_cast_fp16)[name = string("op_4104_cast_fp16")];
            tensor<int32, [4]> var_4105 = const()[name = string("op_4105"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_419x = const()[name = string("concat_419x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4106_cast_fp16 = transpose(perm = var_4105, x = var_4104_cast_fp16)[name = string("transpose_329")];
            tensor<fp16, [1, ?, 1024]> x_337_cast_fp16 = reshape(shape = concat_419x, x = var_4106_cast_fp16)[name = string("x_337_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4110_to_fp16 = const()[name = string("op_4110_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650212160)))];
            tensor<fp16, [1024]> var_4111_to_fp16 = const()[name = string("op_4111_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652309376)))];
            tensor<fp16, [1, ?, 1024]> linear_149_cast_fp16 = linear(bias = var_4111_to_fp16, weight = var_4110_to_fp16, x = x_337_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_339_cast_fp16 = add(x = x_333_cast_fp16, y = linear_149_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [1]> var_4118_axes_0 = const()[name = string("op_4118_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_18_mlp_ln_weight_to_fp16 = const()[name = string("blocks_18_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652311488)))];
            tensor<fp16, [1024]> blocks_18_mlp_ln_bias_to_fp16 = const()[name = string("blocks_18_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652313600)))];
            tensor<fp16, [1, ?, 1024]> var_4118_cast_fp16 = layer_norm(axes = var_4118_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_3952_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_339_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_4127_to_fp16 = const()[name = string("op_4127_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652315712)))];
            tensor<fp16, [4096]> var_4128_to_fp16 = const()[name = string("op_4128_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660704384)))];
            tensor<fp16, [1, ?, 4096]> linear_150_cast_fp16 = linear(bias = var_4128_to_fp16, weight = var_4127_to_fp16, x = var_4118_cast_fp16)[name = string("linear_150_cast_fp16")];
            string x_343_mode_0 = const()[name = string("x_343_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_343_cast_fp16 = gelu(mode = x_343_mode_0, x = linear_150_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_4133_to_fp16 = const()[name = string("op_4133_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660712640)))];
            tensor<fp16, [1024]> var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669101312)))];
            tensor<fp16, [1, ?, 1024]> linear_151_cast_fp16 = linear(bias = var_4134_to_fp16, weight = var_4133_to_fp16, x = x_343_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_345_cast_fp16 = add(x = x_339_cast_fp16, y = linear_151_cast_fp16)[name = string("x_345_cast_fp16")];
            tensor<int32, [4]> k_cache_77_begin_0 = const()[name = string("k_cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> k_cache_77_end_0 = const()[name = string("k_cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_77_end_mask_0 = const()[name = string("k_cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_77_squeeze_mask_0 = const()[name = string("k_cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_77_cast_fp16 = slice_by_index(begin = k_cache_77_begin_0, end = k_cache_77_end_0, end_mask = k_cache_77_end_mask_0, squeeze_mask = k_cache_77_squeeze_mask_0, x = coreml_update_state_84)[name = string("k_cache_77_cast_fp16")];
            tensor<int32, [4]> v_cache_77_begin_0 = const()[name = string("v_cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> v_cache_77_end_0 = const()[name = string("v_cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_77_end_mask_0 = const()[name = string("v_cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_77_squeeze_mask_0 = const()[name = string("v_cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_77_cast_fp16 = slice_by_index(begin = v_cache_77_begin_0, end = v_cache_77_end_0, end_mask = v_cache_77_end_mask_0, squeeze_mask = v_cache_77_squeeze_mask_0, x = coreml_update_state_85)[name = string("v_cache_77_cast_fp16")];
            tensor<int32, [4]> k_cache_79_begin_0 = const()[name = string("k_cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> k_cache_79_end_0 = const()[name = string("k_cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_79_end_mask_0 = const()[name = string("k_cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_79_squeeze_mask_0 = const()[name = string("k_cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_79_cast_fp16 = slice_by_index(begin = k_cache_79_begin_0, end = k_cache_79_end_0, end_mask = k_cache_79_end_mask_0, squeeze_mask = k_cache_79_squeeze_mask_0, x = read_state_2)[name = string("k_cache_79_cast_fp16")];
            tensor<int32, [4]> v_cache_79_begin_0 = const()[name = string("v_cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> v_cache_79_end_0 = const()[name = string("v_cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_79_end_mask_0 = const()[name = string("v_cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_79_squeeze_mask_0 = const()[name = string("v_cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_79_cast_fp16 = slice_by_index(begin = v_cache_79_begin_0, end = v_cache_79_end_0, end_mask = v_cache_79_end_mask_0, squeeze_mask = v_cache_79_squeeze_mask_0, x = read_state_3)[name = string("v_cache_79_cast_fp16")];
            int32 var_4157 = const()[name = string("op_4157"), val = int32(-1)];
            tensor<int32, [1]> var_4175_axes_0 = const()[name = string("op_4175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_19_attn_ln_weight_to_fp16 = const()[name = string("blocks_19_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669103424)))];
            tensor<fp16, [1024]> blocks_19_attn_ln_bias_to_fp16 = const()[name = string("blocks_19_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669105536)))];
            fp16 var_4163_to_fp16 = const()[name = string("op_4163_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_4175_cast_fp16 = layer_norm(axes = var_4175_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_4163_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_345_cast_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4186_to_fp16 = const()[name = string("op_4186_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669107648)))];
            tensor<fp16, [1024]> var_4187_to_fp16 = const()[name = string("op_4187_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671204864)))];
            tensor<fp16, [1, ?, 1024]> linear_152_cast_fp16 = linear(bias = var_4187_to_fp16, weight = var_4186_to_fp16, x = var_4175_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4190_to_fp16 = const()[name = string("op_4190_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671206976)))];
            tensor<fp16, [1, ?, 1024]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_4190_to_fp16, x = var_4175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4194_to_fp16 = const()[name = string("op_4194_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673304192)))];
            tensor<fp16, [1024]> var_4195_to_fp16 = const()[name = string("op_4195_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675401408)))];
            tensor<fp16, [1, ?, 1024]> linear_154_cast_fp16 = linear(bias = var_4195_to_fp16, weight = var_4194_to_fp16, x = var_4175_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [3]> var_4197_shape_cast_fp16 = shape(x = linear_152_cast_fp16)[name = string("op_4197_shape_cast_fp16")];
            int32 gather_230_axis_0 = const()[name = string("gather_230_axis_0"), val = int32(0)];
            int32 gather_230_batch_dims_0 = const()[name = string("gather_230_batch_dims_0"), val = int32(0)];
            bool gather_230_validate_indices_0 = const()[name = string("gather_230_validate_indices_0"), val = bool(false)];
            string var_4197_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4197_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_230_to_uint16 = const()[name = string("select_230_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4197_shape_cast_fp16_to_uint16 = cast(dtype = var_4197_shape_cast_fp16_to_uint16_dtype_0, x = var_4197_shape_cast_fp16)[name = string("cast_256")];
            uint16 gather_230_cast_uint16 = gather(axis = gather_230_axis_0, batch_dims = gather_230_batch_dims_0, indices = select_230_to_uint16, validate_indices = gather_230_validate_indices_0, x = var_4197_shape_cast_fp16_to_uint16)[name = string("gather_230_cast_uint16")];
            string gather_230_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_230_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_230_cast_uint16_to_int32 = cast(dtype = gather_230_cast_uint16_to_int32_dtype_0, x = gather_230_cast_uint16)[name = string("cast_255")];
            int32 end_step_41 = add(x = offset, y = gather_230_cast_uint16_to_int32)[name = string("end_step_41")];
            tensor<int32, [1]> expand_dims_304 = const()[name = string("expand_dims_304"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_306 = const()[name = string("expand_dims_306"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_307_axes_0 = const()[name = string("expand_dims_307_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_307 = expand_dims(axes = expand_dims_307_axes_0, x = end_step_41)[name = string("expand_dims_307")];
            tensor<int32, [1]> concat_422_values0_0 = const()[name = string("concat_422_values0_0"), val = tensor<int32, [1]>([19])];
            int32 concat_422_axis_0 = const()[name = string("concat_422_axis_0"), val = int32(0)];
            bool concat_422_interleave_0 = const()[name = string("concat_422_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_422 = concat(axis = concat_422_axis_0, interleave = concat_422_interleave_0, values = (concat_422_values0_0, expand_dims_304, expand_dims_1, expand_dims_306))[name = string("concat_422")];
            tensor<int32, [1]> concat_423_values0_0 = const()[name = string("concat_423_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_423_values1_0 = const()[name = string("concat_423_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_423_values3_0 = const()[name = string("concat_423_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_423_axis_0 = const()[name = string("concat_423_axis_0"), val = int32(0)];
            bool concat_423_interleave_0 = const()[name = string("concat_423_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_423 = concat(axis = concat_423_axis_0, interleave = concat_423_interleave_0, values = (concat_423_values0_0, concat_423_values1_0, expand_dims_307, concat_423_values3_0))[name = string("concat_423")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_20_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_20_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_20_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_20_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_422, begin_mask = k_cache1_internal_tensor_assign_20_begin_mask_0, end = concat_423, end_mask = k_cache1_internal_tensor_assign_20_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_20_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_20_stride_0, update = linear_153_cast_fp16, x = coreml_update_state_84)[name = string("k_cache1_internal_tensor_assign_20_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_20_cast_fp16, input = k_cache1)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_86 = read_state(input = k_cache1)[name = string("coreml_update_state_86")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_20_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_20_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_20_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_20_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_422, begin_mask = v_cache1_internal_tensor_assign_20_begin_mask_0, end = concat_423, end_mask = v_cache1_internal_tensor_assign_20_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_20_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_20_stride_0, update = linear_154_cast_fp16, x = coreml_update_state_85)[name = string("v_cache1_internal_tensor_assign_20_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_20_cast_fp16, input = v_cache1)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_87 = read_state(input = v_cache1)[name = string("coreml_update_state_87")];
            int32 concat_428_values0_0 = const()[name = string("concat_428_values0_0"), val = int32(1)];
            int32 concat_428_values2_0 = const()[name = string("concat_428_values2_0"), val = int32(1024)];
            int32 concat_428_axis_0 = const()[name = string("concat_428_axis_0"), val = int32(0)];
            bool concat_428_interleave_0 = const()[name = string("concat_428_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_428 = concat(axis = concat_428_axis_0, interleave = concat_428_interleave_0, values = (concat_428_values0_0, end_step_41, concat_428_values2_0))[name = string("concat_428")];
            tensor<int32, [3]> var_4213_begin_0 = const()[name = string("op_4213_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4213_end_mask_0 = const()[name = string("op_4213_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4213_cast_fp16 = slice_by_index(begin = var_4213_begin_0, end = concat_428, end_mask = var_4213_end_mask_0, x = k_cache_77_cast_fp16)[name = string("op_4213_cast_fp16")];
            tensor<int32, [3]> var_4216_begin_0 = const()[name = string("op_4216_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4216_end_mask_0 = const()[name = string("op_4216_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4216_cast_fp16 = slice_by_index(begin = var_4216_begin_0, end = concat_428, end_mask = var_4216_end_mask_0, x = v_cache_77_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> concat_430x = const()[name = string("concat_430x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4226_cast_fp16 = reshape(shape = concat_430x, x = linear_152_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_155_cast_fp16 = mul(x = var_4226_cast_fp16, y = const_196_to_fp16)[name = string("q_155_cast_fp16")];
            tensor<int32, [4]> concat_431x = const()[name = string("concat_431x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4233_cast_fp16 = reshape(shape = concat_431x, x = var_4213_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_197_to_fp16 = const()[name = string("const_197_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_195_cast_fp16 = mul(x = var_4233_cast_fp16, y = const_197_to_fp16)[name = string("k_195_cast_fp16")];
            tensor<int32, [4]> concat_432x = const()[name = string("concat_432x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4240_cast_fp16 = reshape(shape = concat_432x, x = var_4216_cast_fp16)[name = string("op_4240_cast_fp16")];
            tensor<int32, [4]> var_4241 = const()[name = string("op_4241"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_115_transpose_x_0 = const()[name = string("qk_115_transpose_x_0"), val = bool(false)];
            bool qk_115_transpose_y_0 = const()[name = string("qk_115_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_269_perm_0 = const()[name = string("transpose_269_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_270_perm_0 = const()[name = string("transpose_270_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_270 = transpose(perm = transpose_270_perm_0, x = k_195_cast_fp16)[name = string("transpose_326")];
            tensor<fp16, [1, 16, ?, 64]> transpose_269 = transpose(perm = transpose_269_perm_0, x = q_155_cast_fp16)[name = string("transpose_327")];
            tensor<fp16, [1, 16, ?, ?]> qk_115_cast_fp16 = matmul(transpose_x = qk_115_transpose_x_0, transpose_y = qk_115_transpose_y_0, x = transpose_269, y = transpose_270)[name = string("qk_115_cast_fp16")];
            int32 concat_433_values1_0 = const()[name = string("concat_433_values1_0"), val = int32(448)];
            int32 concat_433_axis_0 = const()[name = string("concat_433_axis_0"), val = int32(0)];
            bool concat_433_interleave_0 = const()[name = string("concat_433_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_433 = concat(axis = concat_433_axis_0, interleave = concat_433_interleave_0, values = (gather_230_cast_uint16_to_int32, concat_433_values1_0))[name = string("concat_433")];
            tensor<int32, [2]> var_4244_begin_0 = const()[name = string("op_4244_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4244_end_mask_0 = const()[name = string("op_4244_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_4244_cast_fp16 = slice_by_index(begin = var_4244_begin_0, end = concat_433, end_mask = var_4244_end_mask_0, x = mask_to_fp16)[name = string("op_4244_cast_fp16")];
            int32 concat_434_values0_0 = const()[name = string("concat_434_values0_0"), val = int32(0)];
            int32 concat_434_axis_0 = const()[name = string("concat_434_axis_0"), val = int32(0)];
            bool concat_434_interleave_0 = const()[name = string("concat_434_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_434 = concat(axis = concat_434_axis_0, interleave = concat_434_interleave_0, values = (concat_434_values0_0, gather_230_cast_uint16_to_int32))[name = string("concat_434")];
            tensor<int32, [2]> var_4245_begin_0 = const()[name = string("op_4245_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4245_end_mask_0 = const()[name = string("op_4245_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = concat_434, end_mask = var_4245_end_mask_0, x = var_4244_cast_fp16)[name = string("op_4245_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_117_cast_fp16 = add(x = qk_115_cast_fp16, y = var_4245_cast_fp16)[name = string("qk_117_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_4248_cast_fp16 = softmax(axis = var_4157, x = qk_117_cast_fp16)[name = string("op_4248_cast_fp16")];
            bool var_4250_transpose_x_0 = const()[name = string("op_4250_transpose_x_0"), val = bool(false)];
            bool var_4250_transpose_y_0 = const()[name = string("op_4250_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_195_cast_fp16 = transpose(perm = var_4241, x = var_4240_cast_fp16)[name = string("transpose_328")];
            tensor<fp16, [1, 16, ?, 64]> var_4250_cast_fp16 = matmul(transpose_x = var_4250_transpose_x_0, transpose_y = var_4250_transpose_y_0, x = var_4248_cast_fp16, y = v_195_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<int32, [4]> var_4251 = const()[name = string("op_4251"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_435x = const()[name = string("concat_435x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4252_cast_fp16 = transpose(perm = var_4251, x = var_4250_cast_fp16)[name = string("transpose_325")];
            tensor<fp16, [1, ?, 1024]> x_349_cast_fp16 = reshape(shape = concat_435x, x = var_4252_cast_fp16)[name = string("x_349_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4256_to_fp16 = const()[name = string("op_4256_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675403520)))];
            tensor<fp16, [1024]> var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677500736)))];
            tensor<fp16, [1, ?, 1024]> linear_155_cast_fp16 = linear(bias = var_4257_to_fp16, weight = var_4256_to_fp16, x = x_349_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_351_cast_fp16 = add(x = x_345_cast_fp16, y = linear_155_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<int32, [1]> var_4264_axes_0 = const()[name = string("op_4264_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_19_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_19_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677502848)))];
            tensor<fp16, [1024]> blocks_19_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_19_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677504960)))];
            tensor<fp16, [1, ?, 1024]> var_4264_cast_fp16 = layer_norm(axes = var_4264_axes_0, beta = blocks_19_cross_attn_ln_bias_to_fp16, epsilon = var_4163_to_fp16, gamma = blocks_19_cross_attn_ln_weight_to_fp16, x = x_351_cast_fp16)[name = string("op_4264_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4273_to_fp16 = const()[name = string("op_4273_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677507072)))];
            tensor<fp16, [1024]> var_4274_to_fp16 = const()[name = string("op_4274_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679604288)))];
            tensor<fp16, [1, ?, 1024]> linear_156_cast_fp16 = linear(bias = var_4274_to_fp16, weight = var_4273_to_fp16, x = var_4264_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [3]> concat_436 = const()[name = string("concat_436"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_437 = const()[name = string("concat_437"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_197_internal_tensor_assign_1_stride_0 = const()[name = string("k_197_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_197_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_197_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_197_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_197_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_197_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_197_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_197_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_436, begin_mask = k_197_internal_tensor_assign_1_begin_mask_0, end = concat_437, end_mask = k_197_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_197_internal_tensor_assign_1_squeeze_mask_0, stride = k_197_internal_tensor_assign_1_stride_0, update = k_cache_79_cast_fp16, x = k_7_to_fp16)[name = string("k_197_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_438 = const()[name = string("concat_438"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_439 = const()[name = string("concat_439"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_197_internal_tensor_assign_1_stride_0 = const()[name = string("v_197_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_197_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_197_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_197_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_197_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_197_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_197_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_197_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_438, begin_mask = v_197_internal_tensor_assign_1_begin_mask_0, end = concat_439, end_mask = v_197_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_197_internal_tensor_assign_1_squeeze_mask_0, stride = v_197_internal_tensor_assign_1_stride_0, update = v_cache_79_cast_fp16, x = k_7_to_fp16)[name = string("v_197_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_440x = const()[name = string("concat_440x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4294_cast_fp16 = reshape(shape = concat_440x, x = linear_156_cast_fp16)[name = string("op_4294_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_198_to_fp16 = const()[name = string("const_198_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_159_cast_fp16 = mul(x = var_4294_cast_fp16, y = const_198_to_fp16)[name = string("q_159_cast_fp16")];
            tensor<int32, [4]> var_4300 = const()[name = string("op_4300"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4301_cast_fp16 = reshape(shape = var_4300, x = k_197_internal_tensor_assign_1_cast_fp16)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_199_to_fp16 = const()[name = string("const_199_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_199_cast_fp16 = mul(x = var_4301_cast_fp16, y = const_199_to_fp16)[name = string("k_199_cast_fp16")];
            tensor<int32, [4]> var_4307 = const()[name = string("op_4307"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4308_cast_fp16 = reshape(shape = var_4307, x = v_197_internal_tensor_assign_1_cast_fp16)[name = string("op_4308_cast_fp16")];
            tensor<int32, [4]> var_4309 = const()[name = string("op_4309"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_119_transpose_x_0 = const()[name = string("qk_119_transpose_x_0"), val = bool(false)];
            bool qk_119_transpose_y_0 = const()[name = string("qk_119_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_271_perm_0 = const()[name = string("transpose_271_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_272_perm_0 = const()[name = string("transpose_272_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_272 = transpose(perm = transpose_272_perm_0, x = k_199_cast_fp16)[name = string("transpose_322")];
            tensor<fp16, [1, 16, ?, 64]> transpose_271 = transpose(perm = transpose_271_perm_0, x = q_159_cast_fp16)[name = string("transpose_323")];
            tensor<fp16, [1, 16, ?, 1500]> qk_119_cast_fp16 = matmul(transpose_x = qk_119_transpose_x_0, transpose_y = qk_119_transpose_y_0, x = transpose_271, y = transpose_272)[name = string("qk_119_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_4313_cast_fp16 = softmax(axis = var_4157, x = qk_119_cast_fp16)[name = string("op_4313_cast_fp16")];
            bool var_4315_transpose_x_0 = const()[name = string("op_4315_transpose_x_0"), val = bool(false)];
            bool var_4315_transpose_y_0 = const()[name = string("op_4315_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_199_cast_fp16 = transpose(perm = var_4309, x = var_4308_cast_fp16)[name = string("transpose_324")];
            tensor<fp16, [1, 16, ?, 64]> var_4315_cast_fp16 = matmul(transpose_x = var_4315_transpose_x_0, transpose_y = var_4315_transpose_y_0, x = var_4313_cast_fp16, y = v_199_cast_fp16)[name = string("op_4315_cast_fp16")];
            tensor<int32, [4]> var_4316 = const()[name = string("op_4316"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_441x = const()[name = string("concat_441x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4317_cast_fp16 = transpose(perm = var_4316, x = var_4315_cast_fp16)[name = string("transpose_321")];
            tensor<fp16, [1, ?, 1024]> x_355_cast_fp16 = reshape(shape = concat_441x, x = var_4317_cast_fp16)[name = string("x_355_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4321_to_fp16 = const()[name = string("op_4321_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679606400)))];
            tensor<fp16, [1024]> var_4322_to_fp16 = const()[name = string("op_4322_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681703616)))];
            tensor<fp16, [1, ?, 1024]> linear_157_cast_fp16 = linear(bias = var_4322_to_fp16, weight = var_4321_to_fp16, x = x_355_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_357_cast_fp16 = add(x = x_351_cast_fp16, y = linear_157_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [1]> var_4329_axes_0 = const()[name = string("op_4329_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_19_mlp_ln_weight_to_fp16 = const()[name = string("blocks_19_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681705728)))];
            tensor<fp16, [1024]> blocks_19_mlp_ln_bias_to_fp16 = const()[name = string("blocks_19_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681707840)))];
            tensor<fp16, [1, ?, 1024]> var_4329_cast_fp16 = layer_norm(axes = var_4329_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_4163_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_357_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_4338_to_fp16 = const()[name = string("op_4338_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681709952)))];
            tensor<fp16, [4096]> var_4339_to_fp16 = const()[name = string("op_4339_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690098624)))];
            tensor<fp16, [1, ?, 4096]> linear_158_cast_fp16 = linear(bias = var_4339_to_fp16, weight = var_4338_to_fp16, x = var_4329_cast_fp16)[name = string("linear_158_cast_fp16")];
            string x_361_mode_0 = const()[name = string("x_361_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_361_cast_fp16 = gelu(mode = x_361_mode_0, x = linear_158_cast_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_4344_to_fp16 = const()[name = string("op_4344_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690106880)))];
            tensor<fp16, [1024]> var_4345_to_fp16 = const()[name = string("op_4345_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698495552)))];
            tensor<fp16, [1, ?, 1024]> linear_159_cast_fp16 = linear(bias = var_4345_to_fp16, weight = var_4344_to_fp16, x = x_361_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_363_cast_fp16 = add(x = x_357_cast_fp16, y = linear_159_cast_fp16)[name = string("x_363_cast_fp16")];
            tensor<int32, [4]> k_cache_81_begin_0 = const()[name = string("k_cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> k_cache_81_end_0 = const()[name = string("k_cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_81_end_mask_0 = const()[name = string("k_cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_81_squeeze_mask_0 = const()[name = string("k_cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_81_cast_fp16 = slice_by_index(begin = k_cache_81_begin_0, end = k_cache_81_end_0, end_mask = k_cache_81_end_mask_0, squeeze_mask = k_cache_81_squeeze_mask_0, x = coreml_update_state_86)[name = string("k_cache_81_cast_fp16")];
            tensor<int32, [4]> v_cache_81_begin_0 = const()[name = string("v_cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> v_cache_81_end_0 = const()[name = string("v_cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_81_end_mask_0 = const()[name = string("v_cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_81_squeeze_mask_0 = const()[name = string("v_cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_81_cast_fp16 = slice_by_index(begin = v_cache_81_begin_0, end = v_cache_81_end_0, end_mask = v_cache_81_end_mask_0, squeeze_mask = v_cache_81_squeeze_mask_0, x = coreml_update_state_87)[name = string("v_cache_81_cast_fp16")];
            tensor<int32, [4]> k_cache_83_begin_0 = const()[name = string("k_cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> k_cache_83_end_0 = const()[name = string("k_cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_83_end_mask_0 = const()[name = string("k_cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_83_squeeze_mask_0 = const()[name = string("k_cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_83_cast_fp16 = slice_by_index(begin = k_cache_83_begin_0, end = k_cache_83_end_0, end_mask = k_cache_83_end_mask_0, squeeze_mask = k_cache_83_squeeze_mask_0, x = read_state_2)[name = string("k_cache_83_cast_fp16")];
            tensor<int32, [4]> v_cache_83_begin_0 = const()[name = string("v_cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> v_cache_83_end_0 = const()[name = string("v_cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_83_end_mask_0 = const()[name = string("v_cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_83_squeeze_mask_0 = const()[name = string("v_cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_83_cast_fp16 = slice_by_index(begin = v_cache_83_begin_0, end = v_cache_83_end_0, end_mask = v_cache_83_end_mask_0, squeeze_mask = v_cache_83_squeeze_mask_0, x = read_state_3)[name = string("v_cache_83_cast_fp16")];
            int32 var_4368 = const()[name = string("op_4368"), val = int32(-1)];
            tensor<int32, [1]> var_4386_axes_0 = const()[name = string("op_4386_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_20_attn_ln_weight_to_fp16 = const()[name = string("blocks_20_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698497664)))];
            tensor<fp16, [1024]> blocks_20_attn_ln_bias_to_fp16 = const()[name = string("blocks_20_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698499776)))];
            fp16 var_4374_to_fp16 = const()[name = string("op_4374_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_4386_cast_fp16 = layer_norm(axes = var_4386_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_4374_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_363_cast_fp16)[name = string("op_4386_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4397_to_fp16 = const()[name = string("op_4397_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698501888)))];
            tensor<fp16, [1024]> var_4398_to_fp16 = const()[name = string("op_4398_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700599104)))];
            tensor<fp16, [1, ?, 1024]> linear_160_cast_fp16 = linear(bias = var_4398_to_fp16, weight = var_4397_to_fp16, x = var_4386_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4401_to_fp16 = const()[name = string("op_4401_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700601216)))];
            tensor<fp16, [1, ?, 1024]> linear_161_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_4401_to_fp16, x = var_4386_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4405_to_fp16 = const()[name = string("op_4405_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702698432)))];
            tensor<fp16, [1024]> var_4406_to_fp16 = const()[name = string("op_4406_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704795648)))];
            tensor<fp16, [1, ?, 1024]> linear_162_cast_fp16 = linear(bias = var_4406_to_fp16, weight = var_4405_to_fp16, x = var_4386_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [3]> var_4408_shape_cast_fp16 = shape(x = linear_160_cast_fp16)[name = string("op_4408_shape_cast_fp16")];
            int32 gather_242_axis_0 = const()[name = string("gather_242_axis_0"), val = int32(0)];
            int32 gather_242_batch_dims_0 = const()[name = string("gather_242_batch_dims_0"), val = int32(0)];
            bool gather_242_validate_indices_0 = const()[name = string("gather_242_validate_indices_0"), val = bool(false)];
            string var_4408_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4408_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_242_to_uint16 = const()[name = string("select_242_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4408_shape_cast_fp16_to_uint16 = cast(dtype = var_4408_shape_cast_fp16_to_uint16_dtype_0, x = var_4408_shape_cast_fp16)[name = string("cast_254")];
            uint16 gather_242_cast_uint16 = gather(axis = gather_242_axis_0, batch_dims = gather_242_batch_dims_0, indices = select_242_to_uint16, validate_indices = gather_242_validate_indices_0, x = var_4408_shape_cast_fp16_to_uint16)[name = string("gather_242_cast_uint16")];
            string gather_242_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_242_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_242_cast_uint16_to_int32 = cast(dtype = gather_242_cast_uint16_to_int32_dtype_0, x = gather_242_cast_uint16)[name = string("cast_253")];
            int32 end_step_43 = add(x = offset, y = gather_242_cast_uint16_to_int32)[name = string("end_step_43")];
            tensor<int32, [1]> expand_dims_320 = const()[name = string("expand_dims_320"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_322 = const()[name = string("expand_dims_322"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_323_axes_0 = const()[name = string("expand_dims_323_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_323 = expand_dims(axes = expand_dims_323_axes_0, x = end_step_43)[name = string("expand_dims_323")];
            tensor<int32, [1]> concat_444_values0_0 = const()[name = string("concat_444_values0_0"), val = tensor<int32, [1]>([20])];
            int32 concat_444_axis_0 = const()[name = string("concat_444_axis_0"), val = int32(0)];
            bool concat_444_interleave_0 = const()[name = string("concat_444_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_444 = concat(axis = concat_444_axis_0, interleave = concat_444_interleave_0, values = (concat_444_values0_0, expand_dims_320, expand_dims_1, expand_dims_322))[name = string("concat_444")];
            tensor<int32, [1]> concat_445_values0_0 = const()[name = string("concat_445_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_445_values1_0 = const()[name = string("concat_445_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_445_values3_0 = const()[name = string("concat_445_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_445_axis_0 = const()[name = string("concat_445_axis_0"), val = int32(0)];
            bool concat_445_interleave_0 = const()[name = string("concat_445_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_445 = concat(axis = concat_445_axis_0, interleave = concat_445_interleave_0, values = (concat_445_values0_0, concat_445_values1_0, expand_dims_323, concat_445_values3_0))[name = string("concat_445")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_21_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_21_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_21_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_444, begin_mask = k_cache1_internal_tensor_assign_21_begin_mask_0, end = concat_445, end_mask = k_cache1_internal_tensor_assign_21_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_21_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_21_stride_0, update = linear_161_cast_fp16, x = coreml_update_state_86)[name = string("k_cache1_internal_tensor_assign_21_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_21_cast_fp16, input = k_cache1)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_88 = read_state(input = k_cache1)[name = string("coreml_update_state_88")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_21_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_21_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_21_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_444, begin_mask = v_cache1_internal_tensor_assign_21_begin_mask_0, end = concat_445, end_mask = v_cache1_internal_tensor_assign_21_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_21_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_21_stride_0, update = linear_162_cast_fp16, x = coreml_update_state_87)[name = string("v_cache1_internal_tensor_assign_21_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_21_cast_fp16, input = v_cache1)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_89 = read_state(input = v_cache1)[name = string("coreml_update_state_89")];
            int32 concat_450_values0_0 = const()[name = string("concat_450_values0_0"), val = int32(1)];
            int32 concat_450_values2_0 = const()[name = string("concat_450_values2_0"), val = int32(1024)];
            int32 concat_450_axis_0 = const()[name = string("concat_450_axis_0"), val = int32(0)];
            bool concat_450_interleave_0 = const()[name = string("concat_450_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_450 = concat(axis = concat_450_axis_0, interleave = concat_450_interleave_0, values = (concat_450_values0_0, end_step_43, concat_450_values2_0))[name = string("concat_450")];
            tensor<int32, [3]> var_4424_begin_0 = const()[name = string("op_4424_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4424_end_mask_0 = const()[name = string("op_4424_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = concat_450, end_mask = var_4424_end_mask_0, x = k_cache_81_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<int32, [3]> var_4427_begin_0 = const()[name = string("op_4427_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4427_end_mask_0 = const()[name = string("op_4427_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4427_cast_fp16 = slice_by_index(begin = var_4427_begin_0, end = concat_450, end_mask = var_4427_end_mask_0, x = v_cache_81_cast_fp16)[name = string("op_4427_cast_fp16")];
            tensor<int32, [4]> concat_452x = const()[name = string("concat_452x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4437_cast_fp16 = reshape(shape = concat_452x, x = linear_160_cast_fp16)[name = string("op_4437_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_200_to_fp16 = const()[name = string("const_200_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_163_cast_fp16 = mul(x = var_4437_cast_fp16, y = const_200_to_fp16)[name = string("q_163_cast_fp16")];
            tensor<int32, [4]> concat_453x = const()[name = string("concat_453x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4444_cast_fp16 = reshape(shape = concat_453x, x = var_4424_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_201_to_fp16 = const()[name = string("const_201_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_205_cast_fp16 = mul(x = var_4444_cast_fp16, y = const_201_to_fp16)[name = string("k_205_cast_fp16")];
            tensor<int32, [4]> concat_454x = const()[name = string("concat_454x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4451_cast_fp16 = reshape(shape = concat_454x, x = var_4427_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<int32, [4]> var_4452 = const()[name = string("op_4452"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_121_transpose_x_0 = const()[name = string("qk_121_transpose_x_0"), val = bool(false)];
            bool qk_121_transpose_y_0 = const()[name = string("qk_121_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_273_perm_0 = const()[name = string("transpose_273_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_274_perm_0 = const()[name = string("transpose_274_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_274 = transpose(perm = transpose_274_perm_0, x = k_205_cast_fp16)[name = string("transpose_318")];
            tensor<fp16, [1, 16, ?, 64]> transpose_273 = transpose(perm = transpose_273_perm_0, x = q_163_cast_fp16)[name = string("transpose_319")];
            tensor<fp16, [1, 16, ?, ?]> qk_121_cast_fp16 = matmul(transpose_x = qk_121_transpose_x_0, transpose_y = qk_121_transpose_y_0, x = transpose_273, y = transpose_274)[name = string("qk_121_cast_fp16")];
            int32 concat_455_values1_0 = const()[name = string("concat_455_values1_0"), val = int32(448)];
            int32 concat_455_axis_0 = const()[name = string("concat_455_axis_0"), val = int32(0)];
            bool concat_455_interleave_0 = const()[name = string("concat_455_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_455 = concat(axis = concat_455_axis_0, interleave = concat_455_interleave_0, values = (gather_242_cast_uint16_to_int32, concat_455_values1_0))[name = string("concat_455")];
            tensor<int32, [2]> var_4455_begin_0 = const()[name = string("op_4455_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4455_end_mask_0 = const()[name = string("op_4455_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_4455_cast_fp16 = slice_by_index(begin = var_4455_begin_0, end = concat_455, end_mask = var_4455_end_mask_0, x = mask_to_fp16)[name = string("op_4455_cast_fp16")];
            int32 concat_456_values0_0 = const()[name = string("concat_456_values0_0"), val = int32(0)];
            int32 concat_456_axis_0 = const()[name = string("concat_456_axis_0"), val = int32(0)];
            bool concat_456_interleave_0 = const()[name = string("concat_456_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_456 = concat(axis = concat_456_axis_0, interleave = concat_456_interleave_0, values = (concat_456_values0_0, gather_242_cast_uint16_to_int32))[name = string("concat_456")];
            tensor<int32, [2]> var_4456_begin_0 = const()[name = string("op_4456_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4456_end_mask_0 = const()[name = string("op_4456_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_4456_cast_fp16 = slice_by_index(begin = var_4456_begin_0, end = concat_456, end_mask = var_4456_end_mask_0, x = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_123_cast_fp16 = add(x = qk_121_cast_fp16, y = var_4456_cast_fp16)[name = string("qk_123_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_4459_cast_fp16 = softmax(axis = var_4368, x = qk_123_cast_fp16)[name = string("op_4459_cast_fp16")];
            bool var_4461_transpose_x_0 = const()[name = string("op_4461_transpose_x_0"), val = bool(false)];
            bool var_4461_transpose_y_0 = const()[name = string("op_4461_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_205_cast_fp16 = transpose(perm = var_4452, x = var_4451_cast_fp16)[name = string("transpose_320")];
            tensor<fp16, [1, 16, ?, 64]> var_4461_cast_fp16 = matmul(transpose_x = var_4461_transpose_x_0, transpose_y = var_4461_transpose_y_0, x = var_4459_cast_fp16, y = v_205_cast_fp16)[name = string("op_4461_cast_fp16")];
            tensor<int32, [4]> var_4462 = const()[name = string("op_4462"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_457x = const()[name = string("concat_457x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4463_cast_fp16 = transpose(perm = var_4462, x = var_4461_cast_fp16)[name = string("transpose_317")];
            tensor<fp16, [1, ?, 1024]> x_367_cast_fp16 = reshape(shape = concat_457x, x = var_4463_cast_fp16)[name = string("x_367_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704797760)))];
            tensor<fp16, [1024]> var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706894976)))];
            tensor<fp16, [1, ?, 1024]> linear_163_cast_fp16 = linear(bias = var_4468_to_fp16, weight = var_4467_to_fp16, x = x_367_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_369_cast_fp16 = add(x = x_363_cast_fp16, y = linear_163_cast_fp16)[name = string("x_369_cast_fp16")];
            tensor<int32, [1]> var_4475_axes_0 = const()[name = string("op_4475_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_20_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_20_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706897088)))];
            tensor<fp16, [1024]> blocks_20_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_20_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706899200)))];
            tensor<fp16, [1, ?, 1024]> var_4475_cast_fp16 = layer_norm(axes = var_4475_axes_0, beta = blocks_20_cross_attn_ln_bias_to_fp16, epsilon = var_4374_to_fp16, gamma = blocks_20_cross_attn_ln_weight_to_fp16, x = x_369_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4484_to_fp16 = const()[name = string("op_4484_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706901312)))];
            tensor<fp16, [1024]> var_4485_to_fp16 = const()[name = string("op_4485_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708998528)))];
            tensor<fp16, [1, ?, 1024]> linear_164_cast_fp16 = linear(bias = var_4485_to_fp16, weight = var_4484_to_fp16, x = var_4475_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [3]> concat_458 = const()[name = string("concat_458"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_459 = const()[name = string("concat_459"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_207_internal_tensor_assign_1_stride_0 = const()[name = string("k_207_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_207_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_207_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_207_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_207_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_207_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_207_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_207_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_458, begin_mask = k_207_internal_tensor_assign_1_begin_mask_0, end = concat_459, end_mask = k_207_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_207_internal_tensor_assign_1_squeeze_mask_0, stride = k_207_internal_tensor_assign_1_stride_0, update = k_cache_83_cast_fp16, x = k_7_to_fp16)[name = string("k_207_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_460 = const()[name = string("concat_460"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_461 = const()[name = string("concat_461"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_207_internal_tensor_assign_1_stride_0 = const()[name = string("v_207_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_207_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_207_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_207_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_207_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_207_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_207_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_207_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_460, begin_mask = v_207_internal_tensor_assign_1_begin_mask_0, end = concat_461, end_mask = v_207_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_207_internal_tensor_assign_1_squeeze_mask_0, stride = v_207_internal_tensor_assign_1_stride_0, update = v_cache_83_cast_fp16, x = k_7_to_fp16)[name = string("v_207_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_462x = const()[name = string("concat_462x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4505_cast_fp16 = reshape(shape = concat_462x, x = linear_164_cast_fp16)[name = string("op_4505_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_202_to_fp16 = const()[name = string("const_202_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_167_cast_fp16 = mul(x = var_4505_cast_fp16, y = const_202_to_fp16)[name = string("q_167_cast_fp16")];
            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4512_cast_fp16 = reshape(shape = var_4511, x = k_207_internal_tensor_assign_1_cast_fp16)[name = string("op_4512_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_203_to_fp16 = const()[name = string("const_203_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_209_cast_fp16 = mul(x = var_4512_cast_fp16, y = const_203_to_fp16)[name = string("k_209_cast_fp16")];
            tensor<int32, [4]> var_4518 = const()[name = string("op_4518"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4519_cast_fp16 = reshape(shape = var_4518, x = v_207_internal_tensor_assign_1_cast_fp16)[name = string("op_4519_cast_fp16")];
            tensor<int32, [4]> var_4520 = const()[name = string("op_4520"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_125_transpose_x_0 = const()[name = string("qk_125_transpose_x_0"), val = bool(false)];
            bool qk_125_transpose_y_0 = const()[name = string("qk_125_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_275_perm_0 = const()[name = string("transpose_275_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_276_perm_0 = const()[name = string("transpose_276_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_276 = transpose(perm = transpose_276_perm_0, x = k_209_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 16, ?, 64]> transpose_275 = transpose(perm = transpose_275_perm_0, x = q_167_cast_fp16)[name = string("transpose_315")];
            tensor<fp16, [1, 16, ?, 1500]> qk_125_cast_fp16 = matmul(transpose_x = qk_125_transpose_x_0, transpose_y = qk_125_transpose_y_0, x = transpose_275, y = transpose_276)[name = string("qk_125_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_4524_cast_fp16 = softmax(axis = var_4368, x = qk_125_cast_fp16)[name = string("op_4524_cast_fp16")];
            bool var_4526_transpose_x_0 = const()[name = string("op_4526_transpose_x_0"), val = bool(false)];
            bool var_4526_transpose_y_0 = const()[name = string("op_4526_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_209_cast_fp16 = transpose(perm = var_4520, x = var_4519_cast_fp16)[name = string("transpose_316")];
            tensor<fp16, [1, 16, ?, 64]> var_4526_cast_fp16 = matmul(transpose_x = var_4526_transpose_x_0, transpose_y = var_4526_transpose_y_0, x = var_4524_cast_fp16, y = v_209_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<int32, [4]> var_4527 = const()[name = string("op_4527"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_463x = const()[name = string("concat_463x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4528_cast_fp16 = transpose(perm = var_4527, x = var_4526_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, ?, 1024]> x_373_cast_fp16 = reshape(shape = concat_463x, x = var_4528_cast_fp16)[name = string("x_373_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4532_to_fp16 = const()[name = string("op_4532_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709000640)))];
            tensor<fp16, [1024]> var_4533_to_fp16 = const()[name = string("op_4533_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711097856)))];
            tensor<fp16, [1, ?, 1024]> linear_165_cast_fp16 = linear(bias = var_4533_to_fp16, weight = var_4532_to_fp16, x = x_373_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_375_cast_fp16 = add(x = x_369_cast_fp16, y = linear_165_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [1]> var_4540_axes_0 = const()[name = string("op_4540_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_20_mlp_ln_weight_to_fp16 = const()[name = string("blocks_20_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711099968)))];
            tensor<fp16, [1024]> blocks_20_mlp_ln_bias_to_fp16 = const()[name = string("blocks_20_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711102080)))];
            tensor<fp16, [1, ?, 1024]> var_4540_cast_fp16 = layer_norm(axes = var_4540_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_4374_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_375_cast_fp16)[name = string("op_4540_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_4549_to_fp16 = const()[name = string("op_4549_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711104192)))];
            tensor<fp16, [4096]> var_4550_to_fp16 = const()[name = string("op_4550_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719492864)))];
            tensor<fp16, [1, ?, 4096]> linear_166_cast_fp16 = linear(bias = var_4550_to_fp16, weight = var_4549_to_fp16, x = var_4540_cast_fp16)[name = string("linear_166_cast_fp16")];
            string x_379_mode_0 = const()[name = string("x_379_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_379_cast_fp16 = gelu(mode = x_379_mode_0, x = linear_166_cast_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_4555_to_fp16 = const()[name = string("op_4555_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719501120)))];
            tensor<fp16, [1024]> var_4556_to_fp16 = const()[name = string("op_4556_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727889792)))];
            tensor<fp16, [1, ?, 1024]> linear_167_cast_fp16 = linear(bias = var_4556_to_fp16, weight = var_4555_to_fp16, x = x_379_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_381_cast_fp16 = add(x = x_375_cast_fp16, y = linear_167_cast_fp16)[name = string("x_381_cast_fp16")];
            tensor<int32, [4]> k_cache_85_begin_0 = const()[name = string("k_cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> k_cache_85_end_0 = const()[name = string("k_cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_85_end_mask_0 = const()[name = string("k_cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_85_squeeze_mask_0 = const()[name = string("k_cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_85_cast_fp16 = slice_by_index(begin = k_cache_85_begin_0, end = k_cache_85_end_0, end_mask = k_cache_85_end_mask_0, squeeze_mask = k_cache_85_squeeze_mask_0, x = coreml_update_state_88)[name = string("k_cache_85_cast_fp16")];
            tensor<int32, [4]> v_cache_85_begin_0 = const()[name = string("v_cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> v_cache_85_end_0 = const()[name = string("v_cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_85_end_mask_0 = const()[name = string("v_cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_85_squeeze_mask_0 = const()[name = string("v_cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_85_cast_fp16 = slice_by_index(begin = v_cache_85_begin_0, end = v_cache_85_end_0, end_mask = v_cache_85_end_mask_0, squeeze_mask = v_cache_85_squeeze_mask_0, x = coreml_update_state_89)[name = string("v_cache_85_cast_fp16")];
            tensor<int32, [4]> k_cache_87_begin_0 = const()[name = string("k_cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> k_cache_87_end_0 = const()[name = string("k_cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_87_end_mask_0 = const()[name = string("k_cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_87_squeeze_mask_0 = const()[name = string("k_cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_87_cast_fp16 = slice_by_index(begin = k_cache_87_begin_0, end = k_cache_87_end_0, end_mask = k_cache_87_end_mask_0, squeeze_mask = k_cache_87_squeeze_mask_0, x = read_state_2)[name = string("k_cache_87_cast_fp16")];
            tensor<int32, [4]> v_cache_87_begin_0 = const()[name = string("v_cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> v_cache_87_end_0 = const()[name = string("v_cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_87_end_mask_0 = const()[name = string("v_cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_87_squeeze_mask_0 = const()[name = string("v_cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_87_cast_fp16 = slice_by_index(begin = v_cache_87_begin_0, end = v_cache_87_end_0, end_mask = v_cache_87_end_mask_0, squeeze_mask = v_cache_87_squeeze_mask_0, x = read_state_3)[name = string("v_cache_87_cast_fp16")];
            int32 var_4579 = const()[name = string("op_4579"), val = int32(-1)];
            tensor<int32, [1]> var_4597_axes_0 = const()[name = string("op_4597_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_21_attn_ln_weight_to_fp16 = const()[name = string("blocks_21_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727891904)))];
            tensor<fp16, [1024]> blocks_21_attn_ln_bias_to_fp16 = const()[name = string("blocks_21_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727894016)))];
            fp16 var_4585_to_fp16 = const()[name = string("op_4585_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_4597_cast_fp16 = layer_norm(axes = var_4597_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_4585_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_381_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4608_to_fp16 = const()[name = string("op_4608_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727896128)))];
            tensor<fp16, [1024]> var_4609_to_fp16 = const()[name = string("op_4609_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729993344)))];
            tensor<fp16, [1, ?, 1024]> linear_168_cast_fp16 = linear(bias = var_4609_to_fp16, weight = var_4608_to_fp16, x = var_4597_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4612_to_fp16 = const()[name = string("op_4612_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729995456)))];
            tensor<fp16, [1, ?, 1024]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_4612_to_fp16, x = var_4597_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4616_to_fp16 = const()[name = string("op_4616_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(732092672)))];
            tensor<fp16, [1024]> var_4617_to_fp16 = const()[name = string("op_4617_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734189888)))];
            tensor<fp16, [1, ?, 1024]> linear_170_cast_fp16 = linear(bias = var_4617_to_fp16, weight = var_4616_to_fp16, x = var_4597_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [3]> var_4619_shape_cast_fp16 = shape(x = linear_168_cast_fp16)[name = string("op_4619_shape_cast_fp16")];
            int32 gather_254_axis_0 = const()[name = string("gather_254_axis_0"), val = int32(0)];
            int32 gather_254_batch_dims_0 = const()[name = string("gather_254_batch_dims_0"), val = int32(0)];
            bool gather_254_validate_indices_0 = const()[name = string("gather_254_validate_indices_0"), val = bool(false)];
            string var_4619_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4619_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_254_to_uint16 = const()[name = string("select_254_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4619_shape_cast_fp16_to_uint16 = cast(dtype = var_4619_shape_cast_fp16_to_uint16_dtype_0, x = var_4619_shape_cast_fp16)[name = string("cast_252")];
            uint16 gather_254_cast_uint16 = gather(axis = gather_254_axis_0, batch_dims = gather_254_batch_dims_0, indices = select_254_to_uint16, validate_indices = gather_254_validate_indices_0, x = var_4619_shape_cast_fp16_to_uint16)[name = string("gather_254_cast_uint16")];
            string gather_254_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_254_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_254_cast_uint16_to_int32 = cast(dtype = gather_254_cast_uint16_to_int32_dtype_0, x = gather_254_cast_uint16)[name = string("cast_251")];
            int32 end_step_45 = add(x = offset, y = gather_254_cast_uint16_to_int32)[name = string("end_step_45")];
            tensor<int32, [1]> expand_dims_336 = const()[name = string("expand_dims_336"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_338 = const()[name = string("expand_dims_338"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_339_axes_0 = const()[name = string("expand_dims_339_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_339 = expand_dims(axes = expand_dims_339_axes_0, x = end_step_45)[name = string("expand_dims_339")];
            tensor<int32, [1]> concat_466_values0_0 = const()[name = string("concat_466_values0_0"), val = tensor<int32, [1]>([21])];
            int32 concat_466_axis_0 = const()[name = string("concat_466_axis_0"), val = int32(0)];
            bool concat_466_interleave_0 = const()[name = string("concat_466_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_466 = concat(axis = concat_466_axis_0, interleave = concat_466_interleave_0, values = (concat_466_values0_0, expand_dims_336, expand_dims_1, expand_dims_338))[name = string("concat_466")];
            tensor<int32, [1]> concat_467_values0_0 = const()[name = string("concat_467_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_467_values1_0 = const()[name = string("concat_467_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_467_values3_0 = const()[name = string("concat_467_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_467_axis_0 = const()[name = string("concat_467_axis_0"), val = int32(0)];
            bool concat_467_interleave_0 = const()[name = string("concat_467_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_467 = concat(axis = concat_467_axis_0, interleave = concat_467_interleave_0, values = (concat_467_values0_0, concat_467_values1_0, expand_dims_339, concat_467_values3_0))[name = string("concat_467")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_22_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_22_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_22_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_22_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_466, begin_mask = k_cache1_internal_tensor_assign_22_begin_mask_0, end = concat_467, end_mask = k_cache1_internal_tensor_assign_22_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_22_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_22_stride_0, update = linear_169_cast_fp16, x = coreml_update_state_88)[name = string("k_cache1_internal_tensor_assign_22_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_22_cast_fp16, input = k_cache1)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_90 = read_state(input = k_cache1)[name = string("coreml_update_state_90")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_22_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_22_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_22_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_22_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_466, begin_mask = v_cache1_internal_tensor_assign_22_begin_mask_0, end = concat_467, end_mask = v_cache1_internal_tensor_assign_22_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_22_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_22_stride_0, update = linear_170_cast_fp16, x = coreml_update_state_89)[name = string("v_cache1_internal_tensor_assign_22_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_22_cast_fp16, input = v_cache1)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_91 = read_state(input = v_cache1)[name = string("coreml_update_state_91")];
            int32 concat_472_values0_0 = const()[name = string("concat_472_values0_0"), val = int32(1)];
            int32 concat_472_values2_0 = const()[name = string("concat_472_values2_0"), val = int32(1024)];
            int32 concat_472_axis_0 = const()[name = string("concat_472_axis_0"), val = int32(0)];
            bool concat_472_interleave_0 = const()[name = string("concat_472_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_472 = concat(axis = concat_472_axis_0, interleave = concat_472_interleave_0, values = (concat_472_values0_0, end_step_45, concat_472_values2_0))[name = string("concat_472")];
            tensor<int32, [3]> var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = concat_472, end_mask = var_4635_end_mask_0, x = k_cache_85_cast_fp16)[name = string("op_4635_cast_fp16")];
            tensor<int32, [3]> var_4638_begin_0 = const()[name = string("op_4638_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4638_end_mask_0 = const()[name = string("op_4638_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4638_cast_fp16 = slice_by_index(begin = var_4638_begin_0, end = concat_472, end_mask = var_4638_end_mask_0, x = v_cache_85_cast_fp16)[name = string("op_4638_cast_fp16")];
            tensor<int32, [4]> concat_474x = const()[name = string("concat_474x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4648_cast_fp16 = reshape(shape = concat_474x, x = linear_168_cast_fp16)[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_204_to_fp16 = const()[name = string("const_204_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_171_cast_fp16 = mul(x = var_4648_cast_fp16, y = const_204_to_fp16)[name = string("q_171_cast_fp16")];
            tensor<int32, [4]> concat_475x = const()[name = string("concat_475x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4655_cast_fp16 = reshape(shape = concat_475x, x = var_4635_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_205_to_fp16 = const()[name = string("const_205_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_215_cast_fp16 = mul(x = var_4655_cast_fp16, y = const_205_to_fp16)[name = string("k_215_cast_fp16")];
            tensor<int32, [4]> concat_476x = const()[name = string("concat_476x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4662_cast_fp16 = reshape(shape = concat_476x, x = var_4638_cast_fp16)[name = string("op_4662_cast_fp16")];
            tensor<int32, [4]> var_4663 = const()[name = string("op_4663"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_127_transpose_x_0 = const()[name = string("qk_127_transpose_x_0"), val = bool(false)];
            bool qk_127_transpose_y_0 = const()[name = string("qk_127_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_277_perm_0 = const()[name = string("transpose_277_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_278_perm_0 = const()[name = string("transpose_278_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_278 = transpose(perm = transpose_278_perm_0, x = k_215_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [1, 16, ?, 64]> transpose_277 = transpose(perm = transpose_277_perm_0, x = q_171_cast_fp16)[name = string("transpose_311")];
            tensor<fp16, [1, 16, ?, ?]> qk_127_cast_fp16 = matmul(transpose_x = qk_127_transpose_x_0, transpose_y = qk_127_transpose_y_0, x = transpose_277, y = transpose_278)[name = string("qk_127_cast_fp16")];
            int32 concat_477_values1_0 = const()[name = string("concat_477_values1_0"), val = int32(448)];
            int32 concat_477_axis_0 = const()[name = string("concat_477_axis_0"), val = int32(0)];
            bool concat_477_interleave_0 = const()[name = string("concat_477_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_477 = concat(axis = concat_477_axis_0, interleave = concat_477_interleave_0, values = (gather_254_cast_uint16_to_int32, concat_477_values1_0))[name = string("concat_477")];
            tensor<int32, [2]> var_4666_begin_0 = const()[name = string("op_4666_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4666_end_mask_0 = const()[name = string("op_4666_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_4666_cast_fp16 = slice_by_index(begin = var_4666_begin_0, end = concat_477, end_mask = var_4666_end_mask_0, x = mask_to_fp16)[name = string("op_4666_cast_fp16")];
            int32 concat_478_values0_0 = const()[name = string("concat_478_values0_0"), val = int32(0)];
            int32 concat_478_axis_0 = const()[name = string("concat_478_axis_0"), val = int32(0)];
            bool concat_478_interleave_0 = const()[name = string("concat_478_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_478 = concat(axis = concat_478_axis_0, interleave = concat_478_interleave_0, values = (concat_478_values0_0, gather_254_cast_uint16_to_int32))[name = string("concat_478")];
            tensor<int32, [2]> var_4667_begin_0 = const()[name = string("op_4667_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4667_end_mask_0 = const()[name = string("op_4667_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_4667_cast_fp16 = slice_by_index(begin = var_4667_begin_0, end = concat_478, end_mask = var_4667_end_mask_0, x = var_4666_cast_fp16)[name = string("op_4667_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_129_cast_fp16 = add(x = qk_127_cast_fp16, y = var_4667_cast_fp16)[name = string("qk_129_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_4670_cast_fp16 = softmax(axis = var_4579, x = qk_129_cast_fp16)[name = string("op_4670_cast_fp16")];
            bool var_4672_transpose_x_0 = const()[name = string("op_4672_transpose_x_0"), val = bool(false)];
            bool var_4672_transpose_y_0 = const()[name = string("op_4672_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_215_cast_fp16 = transpose(perm = var_4663, x = var_4662_cast_fp16)[name = string("transpose_312")];
            tensor<fp16, [1, 16, ?, 64]> var_4672_cast_fp16 = matmul(transpose_x = var_4672_transpose_x_0, transpose_y = var_4672_transpose_y_0, x = var_4670_cast_fp16, y = v_215_cast_fp16)[name = string("op_4672_cast_fp16")];
            tensor<int32, [4]> var_4673 = const()[name = string("op_4673"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_479x = const()[name = string("concat_479x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4674_cast_fp16 = transpose(perm = var_4673, x = var_4672_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, ?, 1024]> x_385_cast_fp16 = reshape(shape = concat_479x, x = var_4674_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734192000)))];
            tensor<fp16, [1024]> var_4679_to_fp16 = const()[name = string("op_4679_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736289216)))];
            tensor<fp16, [1, ?, 1024]> linear_171_cast_fp16 = linear(bias = var_4679_to_fp16, weight = var_4678_to_fp16, x = x_385_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_387_cast_fp16 = add(x = x_381_cast_fp16, y = linear_171_cast_fp16)[name = string("x_387_cast_fp16")];
            tensor<int32, [1]> var_4686_axes_0 = const()[name = string("op_4686_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_21_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_21_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736291328)))];
            tensor<fp16, [1024]> blocks_21_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_21_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736293440)))];
            tensor<fp16, [1, ?, 1024]> var_4686_cast_fp16 = layer_norm(axes = var_4686_axes_0, beta = blocks_21_cross_attn_ln_bias_to_fp16, epsilon = var_4585_to_fp16, gamma = blocks_21_cross_attn_ln_weight_to_fp16, x = x_387_cast_fp16)[name = string("op_4686_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4695_to_fp16 = const()[name = string("op_4695_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736295552)))];
            tensor<fp16, [1024]> var_4696_to_fp16 = const()[name = string("op_4696_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738392768)))];
            tensor<fp16, [1, ?, 1024]> linear_172_cast_fp16 = linear(bias = var_4696_to_fp16, weight = var_4695_to_fp16, x = var_4686_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [3]> concat_480 = const()[name = string("concat_480"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_481 = const()[name = string("concat_481"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_217_internal_tensor_assign_1_stride_0 = const()[name = string("k_217_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_217_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_217_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_217_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_217_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_217_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_217_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_217_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_480, begin_mask = k_217_internal_tensor_assign_1_begin_mask_0, end = concat_481, end_mask = k_217_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_217_internal_tensor_assign_1_squeeze_mask_0, stride = k_217_internal_tensor_assign_1_stride_0, update = k_cache_87_cast_fp16, x = k_7_to_fp16)[name = string("k_217_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_482 = const()[name = string("concat_482"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_483 = const()[name = string("concat_483"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_217_internal_tensor_assign_1_stride_0 = const()[name = string("v_217_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_217_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_217_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_217_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_217_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_217_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_217_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_217_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_482, begin_mask = v_217_internal_tensor_assign_1_begin_mask_0, end = concat_483, end_mask = v_217_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_217_internal_tensor_assign_1_squeeze_mask_0, stride = v_217_internal_tensor_assign_1_stride_0, update = v_cache_87_cast_fp16, x = k_7_to_fp16)[name = string("v_217_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_484x = const()[name = string("concat_484x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4716_cast_fp16 = reshape(shape = concat_484x, x = linear_172_cast_fp16)[name = string("op_4716_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_206_to_fp16 = const()[name = string("const_206_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_175_cast_fp16 = mul(x = var_4716_cast_fp16, y = const_206_to_fp16)[name = string("q_175_cast_fp16")];
            tensor<int32, [4]> var_4722 = const()[name = string("op_4722"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4723_cast_fp16 = reshape(shape = var_4722, x = k_217_internal_tensor_assign_1_cast_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_219_cast_fp16 = mul(x = var_4723_cast_fp16, y = const_207_to_fp16)[name = string("k_219_cast_fp16")];
            tensor<int32, [4]> var_4729 = const()[name = string("op_4729"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4730_cast_fp16 = reshape(shape = var_4729, x = v_217_internal_tensor_assign_1_cast_fp16)[name = string("op_4730_cast_fp16")];
            tensor<int32, [4]> var_4731 = const()[name = string("op_4731"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_131_transpose_x_0 = const()[name = string("qk_131_transpose_x_0"), val = bool(false)];
            bool qk_131_transpose_y_0 = const()[name = string("qk_131_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_279_perm_0 = const()[name = string("transpose_279_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_280_perm_0 = const()[name = string("transpose_280_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_280 = transpose(perm = transpose_280_perm_0, x = k_219_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 16, ?, 64]> transpose_279 = transpose(perm = transpose_279_perm_0, x = q_175_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [1, 16, ?, 1500]> qk_131_cast_fp16 = matmul(transpose_x = qk_131_transpose_x_0, transpose_y = qk_131_transpose_y_0, x = transpose_279, y = transpose_280)[name = string("qk_131_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_4735_cast_fp16 = softmax(axis = var_4579, x = qk_131_cast_fp16)[name = string("op_4735_cast_fp16")];
            bool var_4737_transpose_x_0 = const()[name = string("op_4737_transpose_x_0"), val = bool(false)];
            bool var_4737_transpose_y_0 = const()[name = string("op_4737_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_219_cast_fp16 = transpose(perm = var_4731, x = var_4730_cast_fp16)[name = string("transpose_308")];
            tensor<fp16, [1, 16, ?, 64]> var_4737_cast_fp16 = matmul(transpose_x = var_4737_transpose_x_0, transpose_y = var_4737_transpose_y_0, x = var_4735_cast_fp16, y = v_219_cast_fp16)[name = string("op_4737_cast_fp16")];
            tensor<int32, [4]> var_4738 = const()[name = string("op_4738"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_485x = const()[name = string("concat_485x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4739_cast_fp16 = transpose(perm = var_4738, x = var_4737_cast_fp16)[name = string("transpose_305")];
            tensor<fp16, [1, ?, 1024]> x_391_cast_fp16 = reshape(shape = concat_485x, x = var_4739_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4743_to_fp16 = const()[name = string("op_4743_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738394880)))];
            tensor<fp16, [1024]> var_4744_to_fp16 = const()[name = string("op_4744_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740492096)))];
            tensor<fp16, [1, ?, 1024]> linear_173_cast_fp16 = linear(bias = var_4744_to_fp16, weight = var_4743_to_fp16, x = x_391_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_393_cast_fp16 = add(x = x_387_cast_fp16, y = linear_173_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [1]> var_4751_axes_0 = const()[name = string("op_4751_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_21_mlp_ln_weight_to_fp16 = const()[name = string("blocks_21_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740494208)))];
            tensor<fp16, [1024]> blocks_21_mlp_ln_bias_to_fp16 = const()[name = string("blocks_21_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740496320)))];
            tensor<fp16, [1, ?, 1024]> var_4751_cast_fp16 = layer_norm(axes = var_4751_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_4585_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_393_cast_fp16)[name = string("op_4751_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_4760_to_fp16 = const()[name = string("op_4760_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740498432)))];
            tensor<fp16, [4096]> var_4761_to_fp16 = const()[name = string("op_4761_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748887104)))];
            tensor<fp16, [1, ?, 4096]> linear_174_cast_fp16 = linear(bias = var_4761_to_fp16, weight = var_4760_to_fp16, x = var_4751_cast_fp16)[name = string("linear_174_cast_fp16")];
            string x_397_mode_0 = const()[name = string("x_397_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_397_cast_fp16 = gelu(mode = x_397_mode_0, x = linear_174_cast_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_4766_to_fp16 = const()[name = string("op_4766_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748895360)))];
            tensor<fp16, [1024]> var_4767_to_fp16 = const()[name = string("op_4767_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757284032)))];
            tensor<fp16, [1, ?, 1024]> linear_175_cast_fp16 = linear(bias = var_4767_to_fp16, weight = var_4766_to_fp16, x = x_397_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_399_cast_fp16 = add(x = x_393_cast_fp16, y = linear_175_cast_fp16)[name = string("x_399_cast_fp16")];
            tensor<int32, [4]> k_cache_89_begin_0 = const()[name = string("k_cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> k_cache_89_end_0 = const()[name = string("k_cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_89_end_mask_0 = const()[name = string("k_cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_89_squeeze_mask_0 = const()[name = string("k_cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_89_cast_fp16 = slice_by_index(begin = k_cache_89_begin_0, end = k_cache_89_end_0, end_mask = k_cache_89_end_mask_0, squeeze_mask = k_cache_89_squeeze_mask_0, x = coreml_update_state_90)[name = string("k_cache_89_cast_fp16")];
            tensor<int32, [4]> v_cache_89_begin_0 = const()[name = string("v_cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> v_cache_89_end_0 = const()[name = string("v_cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_89_end_mask_0 = const()[name = string("v_cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_89_squeeze_mask_0 = const()[name = string("v_cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_89_cast_fp16 = slice_by_index(begin = v_cache_89_begin_0, end = v_cache_89_end_0, end_mask = v_cache_89_end_mask_0, squeeze_mask = v_cache_89_squeeze_mask_0, x = coreml_update_state_91)[name = string("v_cache_89_cast_fp16")];
            tensor<int32, [4]> k_cache_91_begin_0 = const()[name = string("k_cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> k_cache_91_end_0 = const()[name = string("k_cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_91_end_mask_0 = const()[name = string("k_cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_91_squeeze_mask_0 = const()[name = string("k_cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_91_cast_fp16 = slice_by_index(begin = k_cache_91_begin_0, end = k_cache_91_end_0, end_mask = k_cache_91_end_mask_0, squeeze_mask = k_cache_91_squeeze_mask_0, x = read_state_2)[name = string("k_cache_91_cast_fp16")];
            tensor<int32, [4]> v_cache_91_begin_0 = const()[name = string("v_cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> v_cache_91_end_0 = const()[name = string("v_cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_91_end_mask_0 = const()[name = string("v_cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_91_squeeze_mask_0 = const()[name = string("v_cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_91_cast_fp16 = slice_by_index(begin = v_cache_91_begin_0, end = v_cache_91_end_0, end_mask = v_cache_91_end_mask_0, squeeze_mask = v_cache_91_squeeze_mask_0, x = read_state_3)[name = string("v_cache_91_cast_fp16")];
            int32 var_4790 = const()[name = string("op_4790"), val = int32(-1)];
            tensor<int32, [1]> var_4808_axes_0 = const()[name = string("op_4808_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_22_attn_ln_weight_to_fp16 = const()[name = string("blocks_22_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757286144)))];
            tensor<fp16, [1024]> blocks_22_attn_ln_bias_to_fp16 = const()[name = string("blocks_22_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757288256)))];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_4808_cast_fp16 = layer_norm(axes = var_4808_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_399_cast_fp16)[name = string("op_4808_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4819_to_fp16 = const()[name = string("op_4819_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757290368)))];
            tensor<fp16, [1024]> var_4820_to_fp16 = const()[name = string("op_4820_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759387584)))];
            tensor<fp16, [1, ?, 1024]> linear_176_cast_fp16 = linear(bias = var_4820_to_fp16, weight = var_4819_to_fp16, x = var_4808_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4823_to_fp16 = const()[name = string("op_4823_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759389696)))];
            tensor<fp16, [1, ?, 1024]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_4823_to_fp16, x = var_4808_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4827_to_fp16 = const()[name = string("op_4827_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761486912)))];
            tensor<fp16, [1024]> var_4828_to_fp16 = const()[name = string("op_4828_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763584128)))];
            tensor<fp16, [1, ?, 1024]> linear_178_cast_fp16 = linear(bias = var_4828_to_fp16, weight = var_4827_to_fp16, x = var_4808_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<int32, [3]> var_4830_shape_cast_fp16 = shape(x = linear_176_cast_fp16)[name = string("op_4830_shape_cast_fp16")];
            int32 gather_266_axis_0 = const()[name = string("gather_266_axis_0"), val = int32(0)];
            int32 gather_266_batch_dims_0 = const()[name = string("gather_266_batch_dims_0"), val = int32(0)];
            bool gather_266_validate_indices_0 = const()[name = string("gather_266_validate_indices_0"), val = bool(false)];
            string var_4830_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4830_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_266_to_uint16 = const()[name = string("select_266_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4830_shape_cast_fp16_to_uint16 = cast(dtype = var_4830_shape_cast_fp16_to_uint16_dtype_0, x = var_4830_shape_cast_fp16)[name = string("cast_250")];
            uint16 gather_266_cast_uint16 = gather(axis = gather_266_axis_0, batch_dims = gather_266_batch_dims_0, indices = select_266_to_uint16, validate_indices = gather_266_validate_indices_0, x = var_4830_shape_cast_fp16_to_uint16)[name = string("gather_266_cast_uint16")];
            string gather_266_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_266_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_266_cast_uint16_to_int32 = cast(dtype = gather_266_cast_uint16_to_int32_dtype_0, x = gather_266_cast_uint16)[name = string("cast_249")];
            int32 end_step_47 = add(x = offset, y = gather_266_cast_uint16_to_int32)[name = string("end_step_47")];
            tensor<int32, [1]> expand_dims_352 = const()[name = string("expand_dims_352"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_354 = const()[name = string("expand_dims_354"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_355_axes_0 = const()[name = string("expand_dims_355_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_355 = expand_dims(axes = expand_dims_355_axes_0, x = end_step_47)[name = string("expand_dims_355")];
            tensor<int32, [1]> concat_488_values0_0 = const()[name = string("concat_488_values0_0"), val = tensor<int32, [1]>([22])];
            int32 concat_488_axis_0 = const()[name = string("concat_488_axis_0"), val = int32(0)];
            bool concat_488_interleave_0 = const()[name = string("concat_488_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_488 = concat(axis = concat_488_axis_0, interleave = concat_488_interleave_0, values = (concat_488_values0_0, expand_dims_352, expand_dims_1, expand_dims_354))[name = string("concat_488")];
            tensor<int32, [1]> concat_489_values0_0 = const()[name = string("concat_489_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_489_values1_0 = const()[name = string("concat_489_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_489_values3_0 = const()[name = string("concat_489_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_489_axis_0 = const()[name = string("concat_489_axis_0"), val = int32(0)];
            bool concat_489_interleave_0 = const()[name = string("concat_489_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_489 = concat(axis = concat_489_axis_0, interleave = concat_489_interleave_0, values = (concat_489_values0_0, concat_489_values1_0, expand_dims_355, concat_489_values3_0))[name = string("concat_489")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_23_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_23_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_23_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_488, begin_mask = k_cache1_internal_tensor_assign_23_begin_mask_0, end = concat_489, end_mask = k_cache1_internal_tensor_assign_23_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_23_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_23_stride_0, update = linear_177_cast_fp16, x = coreml_update_state_90)[name = string("k_cache1_internal_tensor_assign_23_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_23_cast_fp16, input = k_cache1)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_92 = read_state(input = k_cache1)[name = string("coreml_update_state_92")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_23_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_23_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_23_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_488, begin_mask = v_cache1_internal_tensor_assign_23_begin_mask_0, end = concat_489, end_mask = v_cache1_internal_tensor_assign_23_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_23_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_23_stride_0, update = linear_178_cast_fp16, x = coreml_update_state_91)[name = string("v_cache1_internal_tensor_assign_23_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_23_cast_fp16, input = v_cache1)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [24, 1, 448, 1024]> coreml_update_state_93 = read_state(input = v_cache1)[name = string("coreml_update_state_93")];
            int32 concat_494_values0_0 = const()[name = string("concat_494_values0_0"), val = int32(1)];
            int32 concat_494_values2_0 = const()[name = string("concat_494_values2_0"), val = int32(1024)];
            int32 concat_494_axis_0 = const()[name = string("concat_494_axis_0"), val = int32(0)];
            bool concat_494_interleave_0 = const()[name = string("concat_494_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_494 = concat(axis = concat_494_axis_0, interleave = concat_494_interleave_0, values = (concat_494_values0_0, end_step_47, concat_494_values2_0))[name = string("concat_494")];
            tensor<int32, [3]> var_4846_begin_0 = const()[name = string("op_4846_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4846_end_mask_0 = const()[name = string("op_4846_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4846_cast_fp16 = slice_by_index(begin = var_4846_begin_0, end = concat_494, end_mask = var_4846_end_mask_0, x = k_cache_89_cast_fp16)[name = string("op_4846_cast_fp16")];
            tensor<int32, [3]> var_4849_begin_0 = const()[name = string("op_4849_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_4849_end_mask_0 = const()[name = string("op_4849_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_4849_cast_fp16 = slice_by_index(begin = var_4849_begin_0, end = concat_494, end_mask = var_4849_end_mask_0, x = v_cache_89_cast_fp16)[name = string("op_4849_cast_fp16")];
            tensor<int32, [4]> concat_496x = const()[name = string("concat_496x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4859_cast_fp16 = reshape(shape = concat_496x, x = linear_176_cast_fp16)[name = string("op_4859_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_179_cast_fp16 = mul(x = var_4859_cast_fp16, y = const_208_to_fp16)[name = string("q_179_cast_fp16")];
            tensor<int32, [4]> concat_497x = const()[name = string("concat_497x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4866_cast_fp16 = reshape(shape = concat_497x, x = var_4846_cast_fp16)[name = string("op_4866_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_225_cast_fp16 = mul(x = var_4866_cast_fp16, y = const_209_to_fp16)[name = string("k_225_cast_fp16")];
            tensor<int32, [4]> concat_498x = const()[name = string("concat_498x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4873_cast_fp16 = reshape(shape = concat_498x, x = var_4849_cast_fp16)[name = string("op_4873_cast_fp16")];
            tensor<int32, [4]> var_4874 = const()[name = string("op_4874"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_133_transpose_x_0 = const()[name = string("qk_133_transpose_x_0"), val = bool(false)];
            bool qk_133_transpose_y_0 = const()[name = string("qk_133_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_281_perm_0 = const()[name = string("transpose_281_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_282_perm_0 = const()[name = string("transpose_282_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_282 = transpose(perm = transpose_282_perm_0, x = k_225_cast_fp16)[name = string("transpose_302")];
            tensor<fp16, [1, 16, ?, 64]> transpose_281 = transpose(perm = transpose_281_perm_0, x = q_179_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 16, ?, ?]> qk_133_cast_fp16 = matmul(transpose_x = qk_133_transpose_x_0, transpose_y = qk_133_transpose_y_0, x = transpose_281, y = transpose_282)[name = string("qk_133_cast_fp16")];
            int32 concat_499_values1_0 = const()[name = string("concat_499_values1_0"), val = int32(448)];
            int32 concat_499_axis_0 = const()[name = string("concat_499_axis_0"), val = int32(0)];
            bool concat_499_interleave_0 = const()[name = string("concat_499_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_499 = concat(axis = concat_499_axis_0, interleave = concat_499_interleave_0, values = (gather_266_cast_uint16_to_int32, concat_499_values1_0))[name = string("concat_499")];
            tensor<int32, [2]> var_4877_begin_0 = const()[name = string("op_4877_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4877_end_mask_0 = const()[name = string("op_4877_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_4877_cast_fp16 = slice_by_index(begin = var_4877_begin_0, end = concat_499, end_mask = var_4877_end_mask_0, x = mask_to_fp16)[name = string("op_4877_cast_fp16")];
            int32 concat_500_values0_0 = const()[name = string("concat_500_values0_0"), val = int32(0)];
            int32 concat_500_axis_0 = const()[name = string("concat_500_axis_0"), val = int32(0)];
            bool concat_500_interleave_0 = const()[name = string("concat_500_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_500 = concat(axis = concat_500_axis_0, interleave = concat_500_interleave_0, values = (concat_500_values0_0, gather_266_cast_uint16_to_int32))[name = string("concat_500")];
            tensor<int32, [2]> var_4878_begin_0 = const()[name = string("op_4878_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_4878_end_mask_0 = const()[name = string("op_4878_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_4878_cast_fp16 = slice_by_index(begin = var_4878_begin_0, end = concat_500, end_mask = var_4878_end_mask_0, x = var_4877_cast_fp16)[name = string("op_4878_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_135_cast_fp16 = add(x = qk_133_cast_fp16, y = var_4878_cast_fp16)[name = string("qk_135_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_4881_cast_fp16 = softmax(axis = var_4790, x = qk_135_cast_fp16)[name = string("op_4881_cast_fp16")];
            bool var_4883_transpose_x_0 = const()[name = string("op_4883_transpose_x_0"), val = bool(false)];
            bool var_4883_transpose_y_0 = const()[name = string("op_4883_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_225_cast_fp16 = transpose(perm = var_4874, x = var_4873_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [1, 16, ?, 64]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_0, transpose_y = var_4883_transpose_y_0, x = var_4881_cast_fp16, y = v_225_cast_fp16)[name = string("op_4883_cast_fp16")];
            tensor<int32, [4]> var_4884 = const()[name = string("op_4884"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_501x = const()[name = string("concat_501x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4885_cast_fp16 = transpose(perm = var_4884, x = var_4883_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [1, ?, 1024]> x_403_cast_fp16 = reshape(shape = concat_501x, x = var_4885_cast_fp16)[name = string("x_403_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4889_to_fp16 = const()[name = string("op_4889_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763586240)))];
            tensor<fp16, [1024]> var_4890_to_fp16 = const()[name = string("op_4890_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765683456)))];
            tensor<fp16, [1, ?, 1024]> linear_179_cast_fp16 = linear(bias = var_4890_to_fp16, weight = var_4889_to_fp16, x = x_403_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_405_cast_fp16 = add(x = x_399_cast_fp16, y = linear_179_cast_fp16)[name = string("x_405_cast_fp16")];
            tensor<int32, [1]> var_4897_axes_0 = const()[name = string("op_4897_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_22_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_22_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765685568)))];
            tensor<fp16, [1024]> blocks_22_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_22_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765687680)))];
            tensor<fp16, [1, ?, 1024]> var_4897_cast_fp16 = layer_norm(axes = var_4897_axes_0, beta = blocks_22_cross_attn_ln_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = blocks_22_cross_attn_ln_weight_to_fp16, x = x_405_cast_fp16)[name = string("op_4897_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765689792)))];
            tensor<fp16, [1024]> var_4907_to_fp16 = const()[name = string("op_4907_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767787008)))];
            tensor<fp16, [1, ?, 1024]> linear_180_cast_fp16 = linear(bias = var_4907_to_fp16, weight = var_4906_to_fp16, x = var_4897_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<int32, [3]> concat_502 = const()[name = string("concat_502"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_503 = const()[name = string("concat_503"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_227_internal_tensor_assign_1_stride_0 = const()[name = string("k_227_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_227_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_227_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_227_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_227_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_227_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_227_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_227_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_502, begin_mask = k_227_internal_tensor_assign_1_begin_mask_0, end = concat_503, end_mask = k_227_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_227_internal_tensor_assign_1_squeeze_mask_0, stride = k_227_internal_tensor_assign_1_stride_0, update = k_cache_91_cast_fp16, x = k_7_to_fp16)[name = string("k_227_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_504 = const()[name = string("concat_504"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_505 = const()[name = string("concat_505"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_227_internal_tensor_assign_1_stride_0 = const()[name = string("v_227_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_227_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_227_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_227_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_227_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_227_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_227_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_227_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_504, begin_mask = v_227_internal_tensor_assign_1_begin_mask_0, end = concat_505, end_mask = v_227_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_227_internal_tensor_assign_1_squeeze_mask_0, stride = v_227_internal_tensor_assign_1_stride_0, update = v_cache_91_cast_fp16, x = k_7_to_fp16)[name = string("v_227_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_506x = const()[name = string("concat_506x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_4927_cast_fp16 = reshape(shape = concat_506x, x = linear_180_cast_fp16)[name = string("op_4927_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_210_to_fp16 = const()[name = string("const_210_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_183_cast_fp16 = mul(x = var_4927_cast_fp16, y = const_210_to_fp16)[name = string("q_183_cast_fp16")];
            tensor<int32, [4]> var_4933 = const()[name = string("op_4933"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4934_cast_fp16 = reshape(shape = var_4933, x = k_227_internal_tensor_assign_1_cast_fp16)[name = string("op_4934_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_211_to_fp16 = const()[name = string("const_211_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_229_cast_fp16 = mul(x = var_4934_cast_fp16, y = const_211_to_fp16)[name = string("k_229_cast_fp16")];
            tensor<int32, [4]> var_4940 = const()[name = string("op_4940"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_4941_cast_fp16 = reshape(shape = var_4940, x = v_227_internal_tensor_assign_1_cast_fp16)[name = string("op_4941_cast_fp16")];
            tensor<int32, [4]> var_4942 = const()[name = string("op_4942"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_137_transpose_x_0 = const()[name = string("qk_137_transpose_x_0"), val = bool(false)];
            bool qk_137_transpose_y_0 = const()[name = string("qk_137_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_283_perm_0 = const()[name = string("transpose_283_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_284_perm_0 = const()[name = string("transpose_284_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_284 = transpose(perm = transpose_284_perm_0, x = k_229_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [1, 16, ?, 64]> transpose_283 = transpose(perm = transpose_283_perm_0, x = q_183_cast_fp16)[name = string("transpose_299")];
            tensor<fp16, [1, 16, ?, 1500]> qk_137_cast_fp16 = matmul(transpose_x = qk_137_transpose_x_0, transpose_y = qk_137_transpose_y_0, x = transpose_283, y = transpose_284)[name = string("qk_137_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_4946_cast_fp16 = softmax(axis = var_4790, x = qk_137_cast_fp16)[name = string("op_4946_cast_fp16")];
            bool var_4948_transpose_x_0 = const()[name = string("op_4948_transpose_x_0"), val = bool(false)];
            bool var_4948_transpose_y_0 = const()[name = string("op_4948_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_229_cast_fp16 = transpose(perm = var_4942, x = var_4941_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 16, ?, 64]> var_4948_cast_fp16 = matmul(transpose_x = var_4948_transpose_x_0, transpose_y = var_4948_transpose_y_0, x = var_4946_cast_fp16, y = v_229_cast_fp16)[name = string("op_4948_cast_fp16")];
            tensor<int32, [4]> var_4949 = const()[name = string("op_4949"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_507x = const()[name = string("concat_507x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_4950_cast_fp16 = transpose(perm = var_4949, x = var_4948_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, ?, 1024]> x_409_cast_fp16 = reshape(shape = concat_507x, x = var_4950_cast_fp16)[name = string("x_409_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_4954_to_fp16 = const()[name = string("op_4954_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767789120)))];
            tensor<fp16, [1024]> var_4955_to_fp16 = const()[name = string("op_4955_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(769886336)))];
            tensor<fp16, [1, ?, 1024]> linear_181_cast_fp16 = linear(bias = var_4955_to_fp16, weight = var_4954_to_fp16, x = x_409_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_411_cast_fp16 = add(x = x_405_cast_fp16, y = linear_181_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [1]> var_4962_axes_0 = const()[name = string("op_4962_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_22_mlp_ln_weight_to_fp16 = const()[name = string("blocks_22_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(769888448)))];
            tensor<fp16, [1024]> blocks_22_mlp_ln_bias_to_fp16 = const()[name = string("blocks_22_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(769890560)))];
            tensor<fp16, [1, ?, 1024]> var_4962_cast_fp16 = layer_norm(axes = var_4962_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_411_cast_fp16)[name = string("op_4962_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_4971_to_fp16 = const()[name = string("op_4971_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(769892672)))];
            tensor<fp16, [4096]> var_4972_to_fp16 = const()[name = string("op_4972_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(778281344)))];
            tensor<fp16, [1, ?, 4096]> linear_182_cast_fp16 = linear(bias = var_4972_to_fp16, weight = var_4971_to_fp16, x = var_4962_cast_fp16)[name = string("linear_182_cast_fp16")];
            string x_415_mode_0 = const()[name = string("x_415_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_415_cast_fp16 = gelu(mode = x_415_mode_0, x = linear_182_cast_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_4977_to_fp16 = const()[name = string("op_4977_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(778289600)))];
            tensor<fp16, [1024]> var_4978_to_fp16 = const()[name = string("op_4978_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(786678272)))];
            tensor<fp16, [1, ?, 1024]> linear_183_cast_fp16 = linear(bias = var_4978_to_fp16, weight = var_4977_to_fp16, x = x_415_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_417_cast_fp16 = add(x = x_411_cast_fp16, y = linear_183_cast_fp16)[name = string("x_417_cast_fp16")];
            tensor<int32, [4]> k_cache_93_begin_0 = const()[name = string("k_cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> k_cache_93_end_0 = const()[name = string("k_cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 448, 1024])];
            tensor<bool, [4]> k_cache_93_end_mask_0 = const()[name = string("k_cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_93_squeeze_mask_0 = const()[name = string("k_cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> k_cache_93_cast_fp16 = slice_by_index(begin = k_cache_93_begin_0, end = k_cache_93_end_0, end_mask = k_cache_93_end_mask_0, squeeze_mask = k_cache_93_squeeze_mask_0, x = coreml_update_state_92)[name = string("k_cache_93_cast_fp16")];
            tensor<int32, [4]> v_cache_93_begin_0 = const()[name = string("v_cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> v_cache_93_end_0 = const()[name = string("v_cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 448, 1024])];
            tensor<bool, [4]> v_cache_93_end_mask_0 = const()[name = string("v_cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_93_squeeze_mask_0 = const()[name = string("v_cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 448, 1024]> v_cache_93_cast_fp16 = slice_by_index(begin = v_cache_93_begin_0, end = v_cache_93_end_0, end_mask = v_cache_93_end_mask_0, squeeze_mask = v_cache_93_squeeze_mask_0, x = coreml_update_state_93)[name = string("v_cache_93_cast_fp16")];
            tensor<int32, [4]> k_cache_begin_0 = const()[name = string("k_cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> k_cache_end_0 = const()[name = string("k_cache_end_0"), val = tensor<int32, [4]>([24, 1, 1500, 1024])];
            tensor<bool, [4]> k_cache_end_mask_0 = const()[name = string("k_cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_squeeze_mask_0 = const()[name = string("k_cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_cache_cast_fp16 = slice_by_index(begin = k_cache_begin_0, end = k_cache_end_0, end_mask = k_cache_end_mask_0, squeeze_mask = k_cache_squeeze_mask_0, x = read_state_2)[name = string("k_cache_cast_fp16")];
            tensor<int32, [4]> v_cache_begin_0 = const()[name = string("v_cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> v_cache_end_0 = const()[name = string("v_cache_end_0"), val = tensor<int32, [4]>([24, 1, 1500, 1024])];
            tensor<bool, [4]> v_cache_end_mask_0 = const()[name = string("v_cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_squeeze_mask_0 = const()[name = string("v_cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_cache_cast_fp16 = slice_by_index(begin = v_cache_begin_0, end = v_cache_end_0, end_mask = v_cache_end_mask_0, squeeze_mask = v_cache_squeeze_mask_0, x = read_state_3)[name = string("v_cache_cast_fp16")];
            int32 var_5001 = const()[name = string("op_5001"), val = int32(-1)];
            tensor<int32, [1]> var_5019_axes_0 = const()[name = string("op_5019_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_23_attn_ln_weight_to_fp16 = const()[name = string("blocks_23_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(786680384)))];
            tensor<fp16, [1024]> blocks_23_attn_ln_bias_to_fp16 = const()[name = string("blocks_23_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(786682496)))];
            fp16 var_5007_to_fp16 = const()[name = string("op_5007_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_5019_cast_fp16 = layer_norm(axes = var_5019_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_5007_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_417_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5030_to_fp16 = const()[name = string("op_5030_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(786684608)))];
            tensor<fp16, [1024]> var_5031_to_fp16 = const()[name = string("op_5031_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788781824)))];
            tensor<fp16, [1, ?, 1024]> linear_184_cast_fp16 = linear(bias = var_5031_to_fp16, weight = var_5030_to_fp16, x = var_5019_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5034_to_fp16 = const()[name = string("op_5034_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788783936)))];
            tensor<fp16, [1, ?, 1024]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_5034_to_fp16, x = var_5019_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5038_to_fp16 = const()[name = string("op_5038_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790881152)))];
            tensor<fp16, [1024]> var_5039_to_fp16 = const()[name = string("op_5039_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792978368)))];
            tensor<fp16, [1, ?, 1024]> linear_186_cast_fp16 = linear(bias = var_5039_to_fp16, weight = var_5038_to_fp16, x = var_5019_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<int32, [3]> var_5041_shape_cast_fp16 = shape(x = linear_184_cast_fp16)[name = string("op_5041_shape_cast_fp16")];
            int32 gather_278_axis_0 = const()[name = string("gather_278_axis_0"), val = int32(0)];
            int32 gather_278_batch_dims_0 = const()[name = string("gather_278_batch_dims_0"), val = int32(0)];
            bool gather_278_validate_indices_0 = const()[name = string("gather_278_validate_indices_0"), val = bool(false)];
            string var_5041_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5041_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_278_to_uint16 = const()[name = string("select_278_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_5041_shape_cast_fp16_to_uint16 = cast(dtype = var_5041_shape_cast_fp16_to_uint16_dtype_0, x = var_5041_shape_cast_fp16)[name = string("cast_248")];
            uint16 gather_278_cast_uint16 = gather(axis = gather_278_axis_0, batch_dims = gather_278_batch_dims_0, indices = select_278_to_uint16, validate_indices = gather_278_validate_indices_0, x = var_5041_shape_cast_fp16_to_uint16)[name = string("gather_278_cast_uint16")];
            string gather_278_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_278_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_278_cast_uint16_to_int32 = cast(dtype = gather_278_cast_uint16_to_int32_dtype_0, x = gather_278_cast_uint16)[name = string("cast_247")];
            int32 end_step = add(x = offset, y = gather_278_cast_uint16_to_int32)[name = string("end_step")];
            tensor<int32, [1]> expand_dims_368 = const()[name = string("expand_dims_368"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_370 = const()[name = string("expand_dims_370"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_371_axes_0 = const()[name = string("expand_dims_371_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_371 = expand_dims(axes = expand_dims_371_axes_0, x = end_step)[name = string("expand_dims_371")];
            tensor<int32, [1]> concat_510_values0_0 = const()[name = string("concat_510_values0_0"), val = tensor<int32, [1]>([23])];
            int32 concat_510_axis_0 = const()[name = string("concat_510_axis_0"), val = int32(0)];
            bool concat_510_interleave_0 = const()[name = string("concat_510_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_510 = concat(axis = concat_510_axis_0, interleave = concat_510_interleave_0, values = (concat_510_values0_0, expand_dims_368, expand_dims_1, expand_dims_370))[name = string("concat_510")];
            tensor<int32, [1]> concat_511_values0_0 = const()[name = string("concat_511_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_511_values1_0 = const()[name = string("concat_511_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_511_values3_0 = const()[name = string("concat_511_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_511_axis_0 = const()[name = string("concat_511_axis_0"), val = int32(0)];
            bool concat_511_interleave_0 = const()[name = string("concat_511_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_511 = concat(axis = concat_511_axis_0, interleave = concat_511_interleave_0, values = (concat_511_values0_0, concat_511_values1_0, expand_dims_371, concat_511_values3_0))[name = string("concat_511")];
            tensor<int32, [4]> k_cache1_internal_tensor_assign_24_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_24_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_24_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_24_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> k_cache1_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> k_cache1_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_510, begin_mask = k_cache1_internal_tensor_assign_24_begin_mask_0, end = concat_511, end_mask = k_cache1_internal_tensor_assign_24_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_24_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_24_stride_0, update = linear_185_cast_fp16, x = coreml_update_state_92)[name = string("k_cache1_internal_tensor_assign_24_cast_fp16")];
            write_state(data = k_cache1_internal_tensor_assign_24_cast_fp16, input = k_cache1)[name = string("coreml_update_state_94_write_state")];
            tensor<int32, [4]> v_cache1_internal_tensor_assign_24_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_24_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_24_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_24_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> v_cache1_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [24, 1, 448, 1024]> v_cache1_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_510, begin_mask = v_cache1_internal_tensor_assign_24_begin_mask_0, end = concat_511, end_mask = v_cache1_internal_tensor_assign_24_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_24_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_24_stride_0, update = linear_186_cast_fp16, x = coreml_update_state_93)[name = string("v_cache1_internal_tensor_assign_24_cast_fp16")];
            write_state(data = v_cache1_internal_tensor_assign_24_cast_fp16, input = v_cache1)[name = string("coreml_update_state_95_write_state")];
            int32 concat_516_values0_0 = const()[name = string("concat_516_values0_0"), val = int32(1)];
            int32 concat_516_values2_0 = const()[name = string("concat_516_values2_0"), val = int32(1024)];
            int32 concat_516_axis_0 = const()[name = string("concat_516_axis_0"), val = int32(0)];
            bool concat_516_interleave_0 = const()[name = string("concat_516_interleave_0"), val = bool(false)];
            tensor<int32, [3]> concat_516 = concat(axis = concat_516_axis_0, interleave = concat_516_interleave_0, values = (concat_516_values0_0, end_step, concat_516_values2_0))[name = string("concat_516")];
            tensor<int32, [3]> var_5057_begin_0 = const()[name = string("op_5057_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_5057_end_mask_0 = const()[name = string("op_5057_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_5057_cast_fp16 = slice_by_index(begin = var_5057_begin_0, end = concat_516, end_mask = var_5057_end_mask_0, x = k_cache_93_cast_fp16)[name = string("op_5057_cast_fp16")];
            tensor<int32, [3]> var_5060_begin_0 = const()[name = string("op_5060_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<bool, [3]> var_5060_end_mask_0 = const()[name = string("op_5060_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, ?, 1024]> var_5060_cast_fp16 = slice_by_index(begin = var_5060_begin_0, end = concat_516, end_mask = var_5060_end_mask_0, x = v_cache_93_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<int32, [4]> concat_518x = const()[name = string("concat_518x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_5070_cast_fp16 = reshape(shape = concat_518x, x = linear_184_cast_fp16)[name = string("op_5070_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_212_to_fp16 = const()[name = string("const_212_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_187_cast_fp16 = mul(x = var_5070_cast_fp16, y = const_212_to_fp16)[name = string("q_187_cast_fp16")];
            tensor<int32, [4]> concat_519x = const()[name = string("concat_519x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_5077_cast_fp16 = reshape(shape = concat_519x, x = var_5057_cast_fp16)[name = string("op_5077_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_213_to_fp16 = const()[name = string("const_213_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> k_235_cast_fp16 = mul(x = var_5077_cast_fp16, y = const_213_to_fp16)[name = string("k_235_cast_fp16")];
            tensor<int32, [4]> concat_520x = const()[name = string("concat_520x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_5084_cast_fp16 = reshape(shape = concat_520x, x = var_5060_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<int32, [4]> var_5085 = const()[name = string("op_5085"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_139_transpose_x_0 = const()[name = string("qk_139_transpose_x_0"), val = bool(false)];
            bool qk_139_transpose_y_0 = const()[name = string("qk_139_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_285_perm_0 = const()[name = string("transpose_285_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_286_perm_0 = const()[name = string("transpose_286_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, ?]> transpose_286 = transpose(perm = transpose_286_perm_0, x = k_235_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 16, ?, 64]> transpose_285 = transpose(perm = transpose_285_perm_0, x = q_187_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [1, 16, ?, ?]> qk_139_cast_fp16 = matmul(transpose_x = qk_139_transpose_x_0, transpose_y = qk_139_transpose_y_0, x = transpose_285, y = transpose_286)[name = string("qk_139_cast_fp16")];
            int32 concat_521_values1_0 = const()[name = string("concat_521_values1_0"), val = int32(448)];
            int32 concat_521_axis_0 = const()[name = string("concat_521_axis_0"), val = int32(0)];
            bool concat_521_interleave_0 = const()[name = string("concat_521_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_521 = concat(axis = concat_521_axis_0, interleave = concat_521_interleave_0, values = (gather_278_cast_uint16_to_int32, concat_521_values1_0))[name = string("concat_521")];
            tensor<int32, [2]> var_5088_begin_0 = const()[name = string("op_5088_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_5088_end_mask_0 = const()[name = string("op_5088_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [?, 448]> var_5088_cast_fp16 = slice_by_index(begin = var_5088_begin_0, end = concat_521, end_mask = var_5088_end_mask_0, x = mask_to_fp16)[name = string("op_5088_cast_fp16")];
            int32 concat_522_values0_0 = const()[name = string("concat_522_values0_0"), val = int32(0)];
            int32 concat_522_axis_0 = const()[name = string("concat_522_axis_0"), val = int32(0)];
            bool concat_522_interleave_0 = const()[name = string("concat_522_interleave_0"), val = bool(false)];
            tensor<int32, [2]> concat_522 = concat(axis = concat_522_axis_0, interleave = concat_522_interleave_0, values = (concat_522_values0_0, gather_278_cast_uint16_to_int32))[name = string("concat_522")];
            tensor<int32, [2]> var_5089_begin_0 = const()[name = string("op_5089_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<bool, [2]> var_5089_end_mask_0 = const()[name = string("op_5089_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [?, ?]> var_5089_cast_fp16 = slice_by_index(begin = var_5089_begin_0, end = concat_522, end_mask = var_5089_end_mask_0, x = var_5088_cast_fp16)[name = string("op_5089_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> qk_141_cast_fp16 = add(x = qk_139_cast_fp16, y = var_5089_cast_fp16)[name = string("qk_141_cast_fp16")];
            tensor<fp16, [1, 16, ?, ?]> var_5092_cast_fp16 = softmax(axis = var_5001, x = qk_141_cast_fp16)[name = string("op_5092_cast_fp16")];
            bool var_5094_transpose_x_0 = const()[name = string("op_5094_transpose_x_0"), val = bool(false)];
            bool var_5094_transpose_y_0 = const()[name = string("op_5094_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 64]> v_235_cast_fp16 = transpose(perm = var_5085, x = var_5084_cast_fp16)[name = string("transpose_296")];
            tensor<fp16, [1, 16, ?, 64]> var_5094_cast_fp16 = matmul(transpose_x = var_5094_transpose_x_0, transpose_y = var_5094_transpose_y_0, x = var_5092_cast_fp16, y = v_235_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [4]> var_5095 = const()[name = string("op_5095"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_523x = const()[name = string("concat_523x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_5096_cast_fp16 = transpose(perm = var_5095, x = var_5094_cast_fp16)[name = string("transpose_293")];
            tensor<fp16, [1, ?, 1024]> x_421_cast_fp16 = reshape(shape = concat_523x, x = var_5096_cast_fp16)[name = string("x_421_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5100_to_fp16 = const()[name = string("op_5100_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792980480)))];
            tensor<fp16, [1024]> var_5101_to_fp16 = const()[name = string("op_5101_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(795077696)))];
            tensor<fp16, [1, ?, 1024]> linear_187_cast_fp16 = linear(bias = var_5101_to_fp16, weight = var_5100_to_fp16, x = x_421_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_423_cast_fp16 = add(x = x_417_cast_fp16, y = linear_187_cast_fp16)[name = string("x_423_cast_fp16")];
            tensor<int32, [1]> var_5108_axes_0 = const()[name = string("op_5108_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_23_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_23_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(795079808)))];
            tensor<fp16, [1024]> blocks_23_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_23_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(795081920)))];
            tensor<fp16, [1, ?, 1024]> var_5108_cast_fp16 = layer_norm(axes = var_5108_axes_0, beta = blocks_23_cross_attn_ln_bias_to_fp16, epsilon = var_5007_to_fp16, gamma = blocks_23_cross_attn_ln_weight_to_fp16, x = x_423_cast_fp16)[name = string("op_5108_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5117_to_fp16 = const()[name = string("op_5117_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(795084032)))];
            tensor<fp16, [1024]> var_5118_to_fp16 = const()[name = string("op_5118_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(797181248)))];
            tensor<fp16, [1, ?, 1024]> linear_188_cast_fp16 = linear(bias = var_5118_to_fp16, weight = var_5117_to_fp16, x = var_5108_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<int32, [3]> concat_524 = const()[name = string("concat_524"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_525 = const()[name = string("concat_525"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> k_237_internal_tensor_assign_1_stride_0 = const()[name = string("k_237_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> k_237_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_237_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> k_237_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_237_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> k_237_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_237_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> k_237_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_524, begin_mask = k_237_internal_tensor_assign_1_begin_mask_0, end = concat_525, end_mask = k_237_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_237_internal_tensor_assign_1_squeeze_mask_0, stride = k_237_internal_tensor_assign_1_stride_0, update = k_cache_cast_fp16, x = k_7_to_fp16)[name = string("k_237_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [3]> concat_526 = const()[name = string("concat_526"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> concat_527 = const()[name = string("concat_527"), val = tensor<int32, [3]>([0, 1500, 0])];
            tensor<int32, [3]> v_237_internal_tensor_assign_1_stride_0 = const()[name = string("v_237_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<bool, [3]> v_237_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_237_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<bool, [3]> v_237_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_237_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> v_237_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_237_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])];
            tensor<fp16, [1, 1500, 1024]> v_237_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_526, begin_mask = v_237_internal_tensor_assign_1_begin_mask_0, end = concat_527, end_mask = v_237_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_237_internal_tensor_assign_1_squeeze_mask_0, stride = v_237_internal_tensor_assign_1_stride_0, update = v_cache_cast_fp16, x = k_7_to_fp16)[name = string("v_237_internal_tensor_assign_1_cast_fp16")];
            tensor<int32, [4]> concat_528x = const()[name = string("concat_528x"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_5138_cast_fp16 = reshape(shape = concat_528x, x = linear_188_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_214_to_fp16 = const()[name = string("const_214_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, ?, 16, 64]> q_cast_fp16 = mul(x = var_5138_cast_fp16, y = const_214_to_fp16)[name = string("q_cast_fp16")];
            tensor<int32, [4]> var_5144 = const()[name = string("op_5144"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_5145_cast_fp16 = reshape(shape = var_5144, x = k_237_internal_tensor_assign_1_cast_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 1, 1, 1]> const_215_to_fp16 = const()[name = string("const_215_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
            tensor<fp16, [1, 1500, 16, 64]> k_cast_fp16 = mul(x = var_5145_cast_fp16, y = const_215_to_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_5151 = const()[name = string("op_5151"), val = tensor<int32, [4]>([1, 1500, 16, -1])];
            tensor<fp16, [1, 1500, 16, 64]> var_5152_cast_fp16 = reshape(shape = var_5151, x = v_237_internal_tensor_assign_1_cast_fp16)[name = string("op_5152_cast_fp16")];
            tensor<int32, [4]> var_5153 = const()[name = string("op_5153"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool qk_transpose_x_0 = const()[name = string("qk_transpose_x_0"), val = bool(false)];
            bool qk_transpose_y_0 = const()[name = string("qk_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_287_perm_0 = const()[name = string("transpose_287_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_288_perm_0 = const()[name = string("transpose_288_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 64, 1500]> transpose_288 = transpose(perm = transpose_288_perm_0, x = k_cast_fp16)[name = string("transpose_290")];
            tensor<fp16, [1, 16, ?, 64]> transpose_287 = transpose(perm = transpose_287_perm_0, x = q_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 16, ?, 1500]> qk_cast_fp16 = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_287, y = transpose_288)[name = string("qk_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1500]> var_5157_cast_fp16 = softmax(axis = var_5001, x = qk_cast_fp16)[name = string("op_5157_cast_fp16")];
            bool var_5159_transpose_x_0 = const()[name = string("op_5159_transpose_x_0"), val = bool(false)];
            bool var_5159_transpose_y_0 = const()[name = string("op_5159_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1500, 64]> v_cast_fp16 = transpose(perm = var_5153, x = var_5152_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [1, 16, ?, 64]> var_5159_cast_fp16 = matmul(transpose_x = var_5159_transpose_x_0, transpose_y = var_5159_transpose_y_0, x = var_5157_cast_fp16, y = v_cast_fp16)[name = string("op_5159_cast_fp16")];
            tensor<int32, [4]> var_5160 = const()[name = string("op_5160"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_529x = const()[name = string("concat_529x"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_5161_cast_fp16 = transpose(perm = var_5160, x = var_5159_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [1, ?, 1024]> x_427_cast_fp16 = reshape(shape = concat_529x, x = var_5161_cast_fp16)[name = string("x_427_cast_fp16")];
            tensor<fp16, [1024, 1024]> var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(797183360)))];
            tensor<fp16, [1024]> var_5166_to_fp16 = const()[name = string("op_5166_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799280576)))];
            tensor<fp16, [1, ?, 1024]> linear_189_cast_fp16 = linear(bias = var_5166_to_fp16, weight = var_5165_to_fp16, x = x_427_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_429_cast_fp16 = add(x = x_423_cast_fp16, y = linear_189_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [1]> var_5173_axes_0 = const()[name = string("op_5173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> blocks_23_mlp_ln_weight_to_fp16 = const()[name = string("blocks_23_mlp_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799282688)))];
            tensor<fp16, [1024]> blocks_23_mlp_ln_bias_to_fp16 = const()[name = string("blocks_23_mlp_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799284800)))];
            tensor<fp16, [1, ?, 1024]> var_5173_cast_fp16 = layer_norm(axes = var_5173_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_5007_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_429_cast_fp16)[name = string("op_5173_cast_fp16")];
            tensor<fp16, [4096, 1024]> var_5182_to_fp16 = const()[name = string("op_5182_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799286912)))];
            tensor<fp16, [4096]> var_5183_to_fp16 = const()[name = string("op_5183_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(807675584)))];
            tensor<fp16, [1, ?, 4096]> linear_190_cast_fp16 = linear(bias = var_5183_to_fp16, weight = var_5182_to_fp16, x = var_5173_cast_fp16)[name = string("linear_190_cast_fp16")];
            string x_433_mode_0 = const()[name = string("x_433_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, ?, 4096]> x_433_cast_fp16 = gelu(mode = x_433_mode_0, x = linear_190_cast_fp16)[name = string("x_433_cast_fp16")];
            tensor<fp16, [1024, 4096]> var_5188_to_fp16 = const()[name = string("op_5188_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(807683840)))];
            tensor<fp16, [1024]> var_5189_to_fp16 = const()[name = string("op_5189_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816072512)))];
            tensor<fp16, [1, ?, 1024]> linear_191_cast_fp16 = linear(bias = var_5189_to_fp16, weight = var_5188_to_fp16, x = x_433_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_435_cast_fp16 = add(x = x_429_cast_fp16, y = linear_191_cast_fp16)[name = string("x_435_cast_fp16")];
            tensor<int32, [1]> var_5202_axes_0 = const()[name = string("op_5202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> ln_weight_to_fp16 = const()[name = string("ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816074624)))];
            tensor<fp16, [1024]> ln_bias_to_fp16 = const()[name = string("ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816076736)))];
            fp16 var_5193_to_fp16 = const()[name = string("op_5193_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1024]> var_5202_cast_fp16 = layer_norm(axes = var_5202_axes_0, beta = ln_bias_to_fp16, epsilon = var_5193_to_fp16, gamma = ln_weight_to_fp16, x = x_435_cast_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [51865]> var_5212_bias_0_to_fp16 = const()[name = string("op_5212_bias_0_to_fp16"), val = tensor<fp16, [51865]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816078848)))];
            tensor<fp16, [1, ?, 51865]> logits = linear(bias = var_5212_bias_0_to_fp16, weight = token_embedding_weight_to_fp16, x = var_5202_cast_fp16)[name = string("op_5212_cast_fp16")];
        } -> (logits);
}