program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
{
    func main<ios17>(tensor<fp32, [1, 1, 1, 108]> attention_mask, tensor<fp32, [1, 1, 1, 438]> cross_attention_mask, tensor<fp32, [1, 438, 1024]> encoder_hidden_states, tensor<int32, [1, 1]> input_id, tensor<fp32, [1, 8, 108, 128]> k_cache_0, tensor<fp32, [1, 8, 108, 128]> k_cache_1, tensor<fp32, [1, 8, 108, 128]> k_cache_2, tensor<fp32, [1, 8, 108, 128]> k_cache_3, tensor<fp32, [1, 8, 108, 128]> k_cache_4, tensor<fp32, [1, 8, 108, 128]> k_cache_5, tensor<fp32, [1, 8, 108, 128]> k_cache_6, tensor<fp32, [1, 8, 108, 128]> k_cache_7, tensor<int32, [1, 1]> position_id, tensor<fp32, [1, 8, 108, 128]> v_cache_0, tensor<fp32, [1, 8, 108, 128]> v_cache_1, tensor<fp32, [1, 8, 108, 128]> v_cache_2, tensor<fp32, [1, 8, 108, 128]> v_cache_3, tensor<fp32, [1, 8, 108, 128]> v_cache_4, tensor<fp32, [1, 8, 108, 128]> v_cache_5, tensor<fp32, [1, 8, 108, 128]> v_cache_6, tensor<fp32, [1, 8, 108, 128]> v_cache_7) {
            tensor<int32, [4]> var_282 = const()[name = tensor<string, []>("op_282"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [1, 1, 1, 1]> var_283 = reshape(shape = var_282, x = position_id)[name = tensor<string, []>("op_283")];
            tensor<int32, [4]> pos_idx_reps_0 = const()[name = tensor<string, []>("pos_idx_reps_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<int32, [1, 8, 1, 128]> pos_idx = tile(reps = pos_idx_reps_0, x = var_283)[name = tensor<string, []>("pos_idx")];
            tensor<int32, []> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_303_batch_dims_0 = const()[name = tensor<string, []>("op_303_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_303_validate_indices_0 = const()[name = tensor<string, []>("op_303_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [16384, 1024]> embedding_token_embedding_weight_to_fp16 = const()[name = tensor<string, []>("embedding_token_embedding_weight_to_fp16"), val = tensor<fp16, [16384, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<string, []> input_id_to_int16_dtype_0 = const()[name = tensor<string, []>("input_id_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<string, []> cast_68_dtype_0 = const()[name = tensor<string, []>("cast_68_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
            tensor<int16, [1, 1]> input_id_to_int16 = cast(dtype = input_id_to_int16_dtype_0, x = input_id)[name = tensor<string, []>("cast_108")];
            tensor<int32, [1, 1]> cast_68 = cast(dtype = cast_68_dtype_0, x = input_id_to_int16)[name = tensor<string, []>("cast_107")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_68, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(16384)];
            tensor<int32, [1, 1]> add_16 = add(x = cast_68, y = slice_by_index_0)[name = tensor<string, []>("add_16")];
            tensor<int32, [1, 1]> select_0 = select(a = cast_68, b = add_16, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<int32, []> var_303_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_303_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_106")];
            tensor<fp16, [1, 1, 1024]> var_303_cast_fp16_cast_uint16_cast_uint16 = gather(axis = var_303_cast_fp16_cast_uint16_axis_0, batch_dims = var_303_batch_dims_0, indices = select_0_to_int16, validate_indices = var_303_validate_indices_0, x = embedding_token_embedding_weight_to_fp16)[name = tensor<string, []>("op_303_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, [1]> var_305 = const()[name = tensor<string, []>("op_305"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1]> var_306 = reshape(shape = var_305, x = position_id)[name = tensor<string, []>("op_306")];
            tensor<int32, []> var_307_batch_dims_0 = const()[name = tensor<string, []>("op_307_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_307_validate_indices_0 = const()[name = tensor<string, []>("op_307_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 1024]> embedding_position_embedding_pos_enc_to_fp16 = const()[name = tensor<string, []>("embedding_position_embedding_pos_enc_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33554560)))];
            tensor<string, []> var_306_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_306_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1]> var_306_to_uint16 = cast(dtype = var_306_to_uint16_dtype_0, x = var_306)[name = tensor<string, []>("cast_105")];
            tensor<fp16, [1, 1024]> var_307_cast_fp16_cast_uint16 = gather(axis = var_295, batch_dims = var_307_batch_dims_0, indices = var_306_to_uint16, validate_indices = var_307_validate_indices_0, x = embedding_position_embedding_pos_enc_to_fp16)[name = tensor<string, []>("op_307_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_310 = const()[name = tensor<string, []>("op_310"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<fp16, [1, 1, 1024]> var_311_cast_fp16 = reshape(shape = var_310, x = var_307_cast_fp16_cast_uint16)[name = tensor<string, []>("op_311_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = add(x = var_303_cast_fp16_cast_uint16_cast_uint16, y = var_311_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> embedding_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("embedding_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35651776)))];
            tensor<fp16, [1024]> embedding_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("embedding_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35653888)))];
            tensor<fp16, []> var_292_to_fp16 = const()[name = tensor<string, []>("op_292_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = embedding_layer_norm_bias_to_fp16, epsilon = var_292_to_fp16, gamma = embedding_layer_norm_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35656000)))];
            tensor<fp16, [1024]> layers_0_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35658112)))];
            tensor<fp16, []> var_318_to_fp16 = const()[name = tensor<string, []>("op_318_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = layers_0_layer_norm_1_bias_to_fp16, epsilon = var_318_to_fp16, gamma = layers_0_layer_norm_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35660224)))];
            tensor<fp16, [1024]> layers_0_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37757440)))];
            tensor<fp16, [1, 1, 1024]> linear_0_cast_fp16 = linear(bias = layers_0_first_sub_layer_query_net_bias_to_fp16, weight = layers_0_first_sub_layer_query_net_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37759552)))];
            tensor<fp16, [1024]> layers_0_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39856768)))];
            tensor<fp16, [1, 1, 1024]> linear_1_cast_fp16 = linear(bias = layers_0_first_sub_layer_key_net_bias_to_fp16, weight = layers_0_first_sub_layer_key_net_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39858880)))];
            tensor<fp16, [1024]> layers_0_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41956096)))];
            tensor<fp16, [1, 1, 1024]> linear_2_cast_fp16 = linear(bias = layers_0_first_sub_layer_value_net_bias_to_fp16, weight = layers_0_first_sub_layer_value_net_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [4]> var_343 = const()[name = tensor<string, []>("op_343"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_344_cast_fp16 = reshape(shape = var_343, x = linear_0_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")];
            tensor<int32, [4]> query_1_perm_0 = const()[name = tensor<string, []>("query_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_358 = const()[name = tensor<string, []>("op_358"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_359_cast_fp16 = reshape(shape = var_358, x = linear_1_cast_fp16)[name = tensor<string, []>("op_359_cast_fp16")];
            tensor<int32, [4]> key_1_perm_0 = const()[name = tensor<string, []>("key_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_374_cast_fp16 = reshape(shape = var_373, x = linear_2_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")];
            tensor<int32, [4]> value_1_perm_0 = const()[name = tensor<string, []>("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_1_axis_0 = const()[name = tensor<string, []>("k_cache_new_1_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_1_mode_0 = const()[name = tensor<string, []>("k_cache_new_1_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_1_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_0_to_fp16 = cast(dtype = k_cache_0_to_fp16_dtype_0, x = k_cache_0)[name = tensor<string, []>("cast_104")];
            tensor<fp16, [1, 8, 1, 128]> key_1_cast_fp16 = transpose(perm = key_1_perm_0, x = var_359_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_1_cast_fp16 = scatter_along_axis(axis = k_cache_new_1_axis_0, data = k_cache_0_to_fp16, indices = pos_idx, mode = k_cache_new_1_mode_0, updates = key_1_cast_fp16, validate_indices = k_cache_new_1_validate_indices_0)[name = tensor<string, []>("k_cache_new_1_cast_fp16")];
            tensor<string, []> k_cache_new_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_1_axis_0 = const()[name = tensor<string, []>("v_cache_new_1_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_1_mode_0 = const()[name = tensor<string, []>("v_cache_new_1_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_1_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_0_to_fp16 = cast(dtype = v_cache_0_to_fp16_dtype_0, x = v_cache_0)[name = tensor<string, []>("cast_102")];
            tensor<fp16, [1, 8, 1, 128]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = var_374_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_1_cast_fp16 = scatter_along_axis(axis = v_cache_new_1_axis_0, data = v_cache_0_to_fp16, indices = pos_idx, mode = v_cache_new_1_mode_0, updates = value_1_cast_fp16, validate_indices = v_cache_new_1_validate_indices_0)[name = tensor<string, []>("v_cache_new_1_cast_fp16")];
            tensor<string, []> v_cache_new_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_384_to_fp16 = const()[name = tensor<string, []>("op_384_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_1_cast_fp16 = transpose(perm = query_1_perm_0, x = var_344_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 8, 1, 128]> mul_0_cast_fp16 = mul(x = query_1_cast_fp16, y = var_384_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_0_cast_fp16, y = k_cache_new_1_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
            tensor<string, []> attention_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("attention_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1, 108]> attention_mask_to_fp16 = cast(dtype = attention_mask_to_fp16_dtype_0, x = attention_mask)[name = tensor<string, []>("cast_100")];
            tensor<fp16, [1, 8, 1, 108]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")];
            tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_cache_new_1_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_389_perm_0 = const()[name = tensor<string, []>("op_389_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = tensor<string, []>("op_395"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_389_cast_fp16 = transpose(perm = var_389_perm_0, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = reshape(shape = var_395, x = var_389_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41958208)))];
            tensor<fp16, [1024]> layers_0_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44055424)))];
            tensor<fp16, [1, 1, 1024]> linear_3_cast_fp16 = linear(bias = layers_0_first_sub_layer_out_projection_bias_to_fp16, weight = layers_0_first_sub_layer_out_projection_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_9_cast_fp16 = add(x = input_3_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = tensor<string, []>("input_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44057536)))];
            tensor<fp16, [1024]> layers_0_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44059648)))];
            tensor<fp16, []> var_403_to_fp16 = const()[name = tensor<string, []>("op_403_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = layers_0_layer_norm_2_bias_to_fp16, epsilon = var_403_to_fp16, gamma = layers_0_layer_norm_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44061760)))];
            tensor<fp16, [1024]> layers_0_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46158976)))];
            tensor<fp16, [1, 1, 1024]> linear_4_cast_fp16 = linear(bias = layers_0_second_sub_layer_query_net_bias_to_fp16, weight = layers_0_second_sub_layer_query_net_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<int32, [4]> var_427 = const()[name = tensor<string, []>("op_427"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_428_cast_fp16 = reshape(shape = var_427, x = linear_4_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")];
            tensor<string, []> encoder_hidden_states_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_hidden_states_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46161088)))];
            tensor<fp16, [1024]> layers_0_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48258304)))];
            tensor<fp16, [1, 438, 1024]> encoder_hidden_states_to_fp16 = cast(dtype = encoder_hidden_states_to_fp16_dtype_0, x = encoder_hidden_states)[name = tensor<string, []>("cast_99")];
            tensor<fp16, [1, 438, 1024]> linear_5_cast_fp16 = linear(bias = layers_0_second_sub_layer_key_net_bias_to_fp16, weight = layers_0_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [4]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_436_cast_fp16 = reshape(shape = var_435, x = linear_5_cast_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48260416)))];
            tensor<fp16, [1024]> layers_0_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50357632)))];
            tensor<fp16, [1, 438, 1024]> linear_6_cast_fp16 = linear(bias = layers_0_second_sub_layer_value_net_bias_to_fp16, weight = layers_0_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<int32, [4]> var_443 = const()[name = tensor<string, []>("op_443"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_444_cast_fp16 = reshape(shape = var_443, x = linear_6_cast_fp16)[name = tensor<string, []>("op_444_cast_fp16")];
            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_410_to_fp16 = const()[name = tensor<string, []>("op_410_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_1_cast_fp16 = mul(x = var_428_cast_fp16, y = var_410_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")];
            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_33 = transpose(perm = transpose_33_perm_0, x = var_436_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [1, 8, 1, 128]> transpose_32 = transpose(perm = transpose_32_perm_0, x = mul_1_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [1, 8, 1, 438]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_32, y = transpose_33)[name = tensor<string, []>("matmul_1_cast_fp16")];
            tensor<string, []> cross_attention_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("cross_attention_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1, 438]> cross_attention_mask_to_fp16 = cast(dtype = cross_attention_mask_to_fp16_dtype_0, x = cross_attention_mask)[name = tensor<string, []>("cast_98")];
            tensor<fp16, [1, 8, 1, 438]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")];
            tensor<bool, []> attn_output_5_transpose_x_0 = const()[name = tensor<string, []>("attn_output_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_5_transpose_y_0 = const()[name = tensor<string, []>("attn_output_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = var_444_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_447_perm_0 = const()[name = tensor<string, []>("op_447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_447_cast_fp16 = transpose(perm = var_447_perm_0, x = attn_output_5_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [1, 1, 1024]> input_13_cast_fp16 = reshape(shape = var_451, x = var_447_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50359744)))];
            tensor<fp16, [1024]> layers_0_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52456960)))];
            tensor<fp16, [1, 1, 1024]> linear_7_cast_fp16 = linear(bias = layers_0_second_sub_layer_out_projection_bias_to_fp16, weight = layers_0_second_sub_layer_out_projection_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52459072)))];
            tensor<fp16, [1024]> layers_0_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52461184)))];
            tensor<fp16, []> var_459_to_fp16 = const()[name = tensor<string, []>("op_459_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = layers_0_layer_norm_3_bias_to_fp16, epsilon = var_459_to_fp16, gamma = layers_0_layer_norm_3_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_0_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52463296)))];
            tensor<fp16, [4096]> layers_0_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60851968)))];
            tensor<fp16, [1, 1, 4096]> linear_8_cast_fp16 = linear(bias = layers_0_third_sub_layer_dense_in_bias_to_fp16, weight = layers_0_third_sub_layer_dense_in_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_21_cast_fp16 = relu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_0_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60860224)))];
            tensor<fp16, [1024]> layers_0_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69248896)))];
            tensor<fp16, [1, 1, 1024]> linear_9_cast_fp16 = linear(bias = layers_0_third_sub_layer_dense_out_bias_to_fp16, weight = layers_0_third_sub_layer_dense_out_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_23_cast_fp16 = add(x = input_15_cast_fp16, y = linear_9_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [1]> input_25_axes_0 = const()[name = tensor<string, []>("input_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69251008)))];
            tensor<fp16, [1024]> layers_1_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69253120)))];
            tensor<fp16, []> var_477_to_fp16 = const()[name = tensor<string, []>("op_477_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = layers_1_layer_norm_1_bias_to_fp16, epsilon = var_477_to_fp16, gamma = layers_1_layer_norm_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69255232)))];
            tensor<fp16, [1024]> layers_1_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71352448)))];
            tensor<fp16, [1, 1, 1024]> linear_10_cast_fp16 = linear(bias = layers_1_first_sub_layer_query_net_bias_to_fp16, weight = layers_1_first_sub_layer_query_net_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71354560)))];
            tensor<fp16, [1024]> layers_1_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73451776)))];
            tensor<fp16, [1, 1, 1024]> linear_11_cast_fp16 = linear(bias = layers_1_first_sub_layer_key_net_bias_to_fp16, weight = layers_1_first_sub_layer_key_net_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73453888)))];
            tensor<fp16, [1024]> layers_1_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75551104)))];
            tensor<fp16, [1, 1, 1024]> linear_12_cast_fp16 = linear(bias = layers_1_first_sub_layer_value_net_bias_to_fp16, weight = layers_1_first_sub_layer_value_net_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [4]> var_502 = const()[name = tensor<string, []>("op_502"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_503_cast_fp16 = reshape(shape = var_502, x = linear_10_cast_fp16)[name = tensor<string, []>("op_503_cast_fp16")];
            tensor<int32, [4]> query_5_perm_0 = const()[name = tensor<string, []>("query_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_517 = const()[name = tensor<string, []>("op_517"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_518_cast_fp16 = reshape(shape = var_517, x = linear_11_cast_fp16)[name = tensor<string, []>("op_518_cast_fp16")];
            tensor<int32, [4]> key_5_perm_0 = const()[name = tensor<string, []>("key_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_532 = const()[name = tensor<string, []>("op_532"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_533_cast_fp16 = reshape(shape = var_532, x = linear_12_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_3_axis_0 = const()[name = tensor<string, []>("k_cache_new_3_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_3_mode_0 = const()[name = tensor<string, []>("k_cache_new_3_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_3_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_3_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_1_to_fp16 = cast(dtype = k_cache_1_to_fp16_dtype_0, x = k_cache_1)[name = tensor<string, []>("cast_97")];
            tensor<fp16, [1, 8, 1, 128]> key_5_cast_fp16 = transpose(perm = key_5_perm_0, x = var_518_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_3_cast_fp16 = scatter_along_axis(axis = k_cache_new_3_axis_0, data = k_cache_1_to_fp16, indices = pos_idx, mode = k_cache_new_3_mode_0, updates = key_5_cast_fp16, validate_indices = k_cache_new_3_validate_indices_0)[name = tensor<string, []>("k_cache_new_3_cast_fp16")];
            tensor<string, []> k_cache_new_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_3_axis_0 = const()[name = tensor<string, []>("v_cache_new_3_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_3_mode_0 = const()[name = tensor<string, []>("v_cache_new_3_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_3_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_3_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_1_to_fp16 = cast(dtype = v_cache_1_to_fp16_dtype_0, x = v_cache_1)[name = tensor<string, []>("cast_95")];
            tensor<fp16, [1, 8, 1, 128]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = var_533_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_3_cast_fp16 = scatter_along_axis(axis = v_cache_new_3_axis_0, data = v_cache_1_to_fp16, indices = pos_idx, mode = v_cache_new_3_mode_0, updates = value_5_cast_fp16, validate_indices = v_cache_new_3_validate_indices_0)[name = tensor<string, []>("v_cache_new_3_cast_fp16")];
            tensor<string, []> v_cache_new_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_543_to_fp16 = const()[name = tensor<string, []>("op_543_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_5_cast_fp16 = transpose(perm = query_5_perm_0, x = var_503_cast_fp16)[name = tensor<string, []>("transpose_103")];
            tensor<fp16, [1, 8, 1, 128]> mul_2_cast_fp16 = mul(x = query_5_cast_fp16, y = var_543_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")];
            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_2_cast_fp16, y = k_cache_new_3_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
            tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")];
            tensor<bool, []> attn_output_7_transpose_x_0 = const()[name = tensor<string, []>("attn_output_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_7_transpose_y_0 = const()[name = tensor<string, []>("attn_output_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = softmax_2_cast_fp16, y = v_cache_new_3_cast_fp16)[name = tensor<string, []>("attn_output_7_cast_fp16")];
            tensor<int32, [4]> var_548_perm_0 = const()[name = tensor<string, []>("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_554 = const()[name = tensor<string, []>("op_554"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = attn_output_7_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [1, 1, 1024]> input_27_cast_fp16 = reshape(shape = var_554, x = var_548_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75553216)))];
            tensor<fp16, [1024]> layers_1_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77650432)))];
            tensor<fp16, [1, 1, 1024]> linear_13_cast_fp16 = linear(bias = layers_1_first_sub_layer_out_projection_bias_to_fp16, weight = layers_1_first_sub_layer_out_projection_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_29_cast_fp16 = add(x = input_23_cast_fp16, y = linear_13_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<int32, [1]> input_31_axes_0 = const()[name = tensor<string, []>("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77652544)))];
            tensor<fp16, [1024]> layers_1_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77654656)))];
            tensor<fp16, []> var_562_to_fp16 = const()[name = tensor<string, []>("op_562_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = layers_1_layer_norm_2_bias_to_fp16, epsilon = var_562_to_fp16, gamma = layers_1_layer_norm_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77656768)))];
            tensor<fp16, [1024]> layers_1_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79753984)))];
            tensor<fp16, [1, 1, 1024]> linear_14_cast_fp16 = linear(bias = layers_1_second_sub_layer_query_net_bias_to_fp16, weight = layers_1_second_sub_layer_query_net_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_586 = const()[name = tensor<string, []>("op_586"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_587_cast_fp16 = reshape(shape = var_586, x = linear_14_cast_fp16)[name = tensor<string, []>("op_587_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79756096)))];
            tensor<fp16, [1024]> layers_1_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81853312)))];
            tensor<fp16, [1, 438, 1024]> linear_15_cast_fp16 = linear(bias = layers_1_second_sub_layer_key_net_bias_to_fp16, weight = layers_1_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<int32, [4]> var_594 = const()[name = tensor<string, []>("op_594"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_595_cast_fp16 = reshape(shape = var_594, x = linear_15_cast_fp16)[name = tensor<string, []>("op_595_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81855424)))];
            tensor<fp16, [1024]> layers_1_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83952640)))];
            tensor<fp16, [1, 438, 1024]> linear_16_cast_fp16 = linear(bias = layers_1_second_sub_layer_value_net_bias_to_fp16, weight = layers_1_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<int32, [4]> var_602 = const()[name = tensor<string, []>("op_602"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_603_cast_fp16 = reshape(shape = var_602, x = linear_16_cast_fp16)[name = tensor<string, []>("op_603_cast_fp16")];
            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_569_to_fp16 = const()[name = tensor<string, []>("op_569_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_3_cast_fp16 = mul(x = var_587_cast_fp16, y = var_569_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")];
            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_35 = transpose(perm = transpose_35_perm_0, x = var_595_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [1, 8, 1, 128]> transpose_34 = transpose(perm = transpose_34_perm_0, x = mul_3_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 8, 1, 438]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_34, y = transpose_35)[name = tensor<string, []>("matmul_3_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_3_cast_fp16")];
            tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")];
            tensor<bool, []> attn_output_11_transpose_x_0 = const()[name = tensor<string, []>("attn_output_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_11_transpose_y_0 = const()[name = tensor<string, []>("attn_output_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = var_603_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_606_perm_0 = const()[name = tensor<string, []>("op_606_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_610 = const()[name = tensor<string, []>("op_610"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_606_cast_fp16 = transpose(perm = var_606_perm_0, x = attn_output_11_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [1, 1, 1024]> input_33_cast_fp16 = reshape(shape = var_610, x = var_606_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83954752)))];
            tensor<fp16, [1024]> layers_1_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86051968)))];
            tensor<fp16, [1, 1, 1024]> linear_17_cast_fp16 = linear(bias = layers_1_second_sub_layer_out_projection_bias_to_fp16, weight = layers_1_second_sub_layer_out_projection_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_35_cast_fp16 = add(x = input_29_cast_fp16, y = linear_17_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [1]> input_37_axes_0 = const()[name = tensor<string, []>("input_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86054080)))];
            tensor<fp16, [1024]> layers_1_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86056192)))];
            tensor<fp16, []> var_618_to_fp16 = const()[name = tensor<string, []>("op_618_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = layers_1_layer_norm_3_bias_to_fp16, epsilon = var_618_to_fp16, gamma = layers_1_layer_norm_3_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_1_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86058304)))];
            tensor<fp16, [4096]> layers_1_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94446976)))];
            tensor<fp16, [1, 1, 4096]> linear_18_cast_fp16 = linear(bias = layers_1_third_sub_layer_dense_in_bias_to_fp16, weight = layers_1_third_sub_layer_dense_in_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_41_cast_fp16 = relu(x = linear_18_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_1_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94455232)))];
            tensor<fp16, [1024]> layers_1_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102843904)))];
            tensor<fp16, [1, 1, 1024]> linear_19_cast_fp16 = linear(bias = layers_1_third_sub_layer_dense_out_bias_to_fp16, weight = layers_1_third_sub_layer_dense_out_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_43_cast_fp16 = add(x = input_35_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<int32, [1]> input_45_axes_0 = const()[name = tensor<string, []>("input_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102846016)))];
            tensor<fp16, [1024]> layers_2_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102848128)))];
            tensor<fp16, []> var_636_to_fp16 = const()[name = tensor<string, []>("op_636_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = layers_2_layer_norm_1_bias_to_fp16, epsilon = var_636_to_fp16, gamma = layers_2_layer_norm_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102850240)))];
            tensor<fp16, [1024]> layers_2_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104947456)))];
            tensor<fp16, [1, 1, 1024]> linear_20_cast_fp16 = linear(bias = layers_2_first_sub_layer_query_net_bias_to_fp16, weight = layers_2_first_sub_layer_query_net_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104949568)))];
            tensor<fp16, [1024]> layers_2_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107046784)))];
            tensor<fp16, [1, 1, 1024]> linear_21_cast_fp16 = linear(bias = layers_2_first_sub_layer_key_net_bias_to_fp16, weight = layers_2_first_sub_layer_key_net_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107048896)))];
            tensor<fp16, [1024]> layers_2_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109146112)))];
            tensor<fp16, [1, 1, 1024]> linear_22_cast_fp16 = linear(bias = layers_2_first_sub_layer_value_net_bias_to_fp16, weight = layers_2_first_sub_layer_value_net_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_662_cast_fp16 = reshape(shape = var_661, x = linear_20_cast_fp16)[name = tensor<string, []>("op_662_cast_fp16")];
            tensor<int32, [4]> query_9_perm_0 = const()[name = tensor<string, []>("query_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_677_cast_fp16 = reshape(shape = var_676, x = linear_21_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
            tensor<int32, [4]> key_9_perm_0 = const()[name = tensor<string, []>("key_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_691 = const()[name = tensor<string, []>("op_691"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_692_cast_fp16 = reshape(shape = var_691, x = linear_22_cast_fp16)[name = tensor<string, []>("op_692_cast_fp16")];
            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_5_axis_0 = const()[name = tensor<string, []>("k_cache_new_5_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_5_mode_0 = const()[name = tensor<string, []>("k_cache_new_5_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_5_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_5_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_2_to_fp16 = cast(dtype = k_cache_2_to_fp16_dtype_0, x = k_cache_2)[name = tensor<string, []>("cast_93")];
            tensor<fp16, [1, 8, 1, 128]> key_9_cast_fp16 = transpose(perm = key_9_perm_0, x = var_677_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_5_cast_fp16 = scatter_along_axis(axis = k_cache_new_5_axis_0, data = k_cache_2_to_fp16, indices = pos_idx, mode = k_cache_new_5_mode_0, updates = key_9_cast_fp16, validate_indices = k_cache_new_5_validate_indices_0)[name = tensor<string, []>("k_cache_new_5_cast_fp16")];
            tensor<string, []> k_cache_new_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_5_axis_0 = const()[name = tensor<string, []>("v_cache_new_5_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_5_mode_0 = const()[name = tensor<string, []>("v_cache_new_5_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_5_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_5_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_2_to_fp16 = cast(dtype = v_cache_2_to_fp16_dtype_0, x = v_cache_2)[name = tensor<string, []>("cast_91")];
            tensor<fp16, [1, 8, 1, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = var_692_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_5_cast_fp16 = scatter_along_axis(axis = v_cache_new_5_axis_0, data = v_cache_2_to_fp16, indices = pos_idx, mode = v_cache_new_5_mode_0, updates = value_9_cast_fp16, validate_indices = v_cache_new_5_validate_indices_0)[name = tensor<string, []>("v_cache_new_5_cast_fp16")];
            tensor<string, []> v_cache_new_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_702_to_fp16 = const()[name = tensor<string, []>("op_702_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_9_cast_fp16 = transpose(perm = query_9_perm_0, x = var_662_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [1, 8, 1, 128]> mul_4_cast_fp16 = mul(x = query_9_cast_fp16, y = var_702_to_fp16)[name = tensor<string, []>("mul_4_cast_fp16")];
            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_4_cast_fp16, y = k_cache_new_5_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
            tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")];
            tensor<bool, []> attn_output_13_transpose_x_0 = const()[name = tensor<string, []>("attn_output_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_13_transpose_y_0 = const()[name = tensor<string, []>("attn_output_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_4_cast_fp16, y = v_cache_new_5_cast_fp16)[name = tensor<string, []>("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_707_perm_0 = const()[name = tensor<string, []>("op_707_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_707_cast_fp16 = transpose(perm = var_707_perm_0, x = attn_output_13_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [1, 1, 1024]> input_47_cast_fp16 = reshape(shape = var_713, x = var_707_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109148224)))];
            tensor<fp16, [1024]> layers_2_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111245440)))];
            tensor<fp16, [1, 1, 1024]> linear_23_cast_fp16 = linear(bias = layers_2_first_sub_layer_out_projection_bias_to_fp16, weight = layers_2_first_sub_layer_out_projection_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<int32, [1]> input_51_axes_0 = const()[name = tensor<string, []>("input_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111247552)))];
            tensor<fp16, [1024]> layers_2_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111249664)))];
            tensor<fp16, []> var_721_to_fp16 = const()[name = tensor<string, []>("op_721_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = layers_2_layer_norm_2_bias_to_fp16, epsilon = var_721_to_fp16, gamma = layers_2_layer_norm_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111251776)))];
            tensor<fp16, [1024]> layers_2_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113348992)))];
            tensor<fp16, [1, 1, 1024]> linear_24_cast_fp16 = linear(bias = layers_2_second_sub_layer_query_net_bias_to_fp16, weight = layers_2_second_sub_layer_query_net_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<int32, [4]> var_745 = const()[name = tensor<string, []>("op_745"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_746_cast_fp16 = reshape(shape = var_745, x = linear_24_cast_fp16)[name = tensor<string, []>("op_746_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113351104)))];
            tensor<fp16, [1024]> layers_2_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115448320)))];
            tensor<fp16, [1, 438, 1024]> linear_25_cast_fp16 = linear(bias = layers_2_second_sub_layer_key_net_bias_to_fp16, weight = layers_2_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<int32, [4]> var_753 = const()[name = tensor<string, []>("op_753"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_754_cast_fp16 = reshape(shape = var_753, x = linear_25_cast_fp16)[name = tensor<string, []>("op_754_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115450432)))];
            tensor<fp16, [1024]> layers_2_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117547648)))];
            tensor<fp16, [1, 438, 1024]> linear_26_cast_fp16 = linear(bias = layers_2_second_sub_layer_value_net_bias_to_fp16, weight = layers_2_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<int32, [4]> var_761 = const()[name = tensor<string, []>("op_761"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_762_cast_fp16 = reshape(shape = var_761, x = linear_26_cast_fp16)[name = tensor<string, []>("op_762_cast_fp16")];
            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_728_to_fp16 = const()[name = tensor<string, []>("op_728_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_5_cast_fp16 = mul(x = var_746_cast_fp16, y = var_728_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")];
            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = tensor<string, []>("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_37_perm_0 = const()[name = tensor<string, []>("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_37 = transpose(perm = transpose_37_perm_0, x = var_754_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [1, 8, 1, 128]> transpose_36 = transpose(perm = transpose_36_perm_0, x = mul_5_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 8, 1, 438]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_36, y = transpose_37)[name = tensor<string, []>("matmul_5_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_5_cast_fp16")];
            tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")];
            tensor<bool, []> attn_output_17_transpose_x_0 = const()[name = tensor<string, []>("attn_output_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_17_transpose_y_0 = const()[name = tensor<string, []>("attn_output_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = var_762_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_765_perm_0 = const()[name = tensor<string, []>("op_765_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_769 = const()[name = tensor<string, []>("op_769"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_765_cast_fp16 = transpose(perm = var_765_perm_0, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [1, 1, 1024]> input_53_cast_fp16 = reshape(shape = var_769, x = var_765_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117549760)))];
            tensor<fp16, [1024]> layers_2_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119646976)))];
            tensor<fp16, [1, 1, 1024]> linear_27_cast_fp16 = linear(bias = layers_2_second_sub_layer_out_projection_bias_to_fp16, weight = layers_2_second_sub_layer_out_projection_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [1]> input_57_axes_0 = const()[name = tensor<string, []>("input_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119649088)))];
            tensor<fp16, [1024]> layers_2_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119651200)))];
            tensor<fp16, []> var_777_to_fp16 = const()[name = tensor<string, []>("op_777_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = layers_2_layer_norm_3_bias_to_fp16, epsilon = var_777_to_fp16, gamma = layers_2_layer_norm_3_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_2_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119653312)))];
            tensor<fp16, [4096]> layers_2_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128041984)))];
            tensor<fp16, [1, 1, 4096]> linear_28_cast_fp16 = linear(bias = layers_2_third_sub_layer_dense_in_bias_to_fp16, weight = layers_2_third_sub_layer_dense_in_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_61_cast_fp16 = relu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_2_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128050240)))];
            tensor<fp16, [1024]> layers_2_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136438912)))];
            tensor<fp16, [1, 1, 1024]> linear_29_cast_fp16 = linear(bias = layers_2_third_sub_layer_dense_out_bias_to_fp16, weight = layers_2_third_sub_layer_dense_out_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_63_cast_fp16 = add(x = input_55_cast_fp16, y = linear_29_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136441024)))];
            tensor<fp16, [1024]> layers_3_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136443136)))];
            tensor<fp16, []> var_795_to_fp16 = const()[name = tensor<string, []>("op_795_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = layers_3_layer_norm_1_bias_to_fp16, epsilon = var_795_to_fp16, gamma = layers_3_layer_norm_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136445248)))];
            tensor<fp16, [1024]> layers_3_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138542464)))];
            tensor<fp16, [1, 1, 1024]> linear_30_cast_fp16 = linear(bias = layers_3_first_sub_layer_query_net_bias_to_fp16, weight = layers_3_first_sub_layer_query_net_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138544576)))];
            tensor<fp16, [1024]> layers_3_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140641792)))];
            tensor<fp16, [1, 1, 1024]> linear_31_cast_fp16 = linear(bias = layers_3_first_sub_layer_key_net_bias_to_fp16, weight = layers_3_first_sub_layer_key_net_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140643904)))];
            tensor<fp16, [1024]> layers_3_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142741120)))];
            tensor<fp16, [1, 1, 1024]> linear_32_cast_fp16 = linear(bias = layers_3_first_sub_layer_value_net_bias_to_fp16, weight = layers_3_first_sub_layer_value_net_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_820 = const()[name = tensor<string, []>("op_820"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_821_cast_fp16 = reshape(shape = var_820, x = linear_30_cast_fp16)[name = tensor<string, []>("op_821_cast_fp16")];
            tensor<int32, [4]> query_13_perm_0 = const()[name = tensor<string, []>("query_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_836_cast_fp16 = reshape(shape = var_835, x = linear_31_cast_fp16)[name = tensor<string, []>("op_836_cast_fp16")];
            tensor<int32, [4]> key_13_perm_0 = const()[name = tensor<string, []>("key_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_850 = const()[name = tensor<string, []>("op_850"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_851_cast_fp16 = reshape(shape = var_850, x = linear_32_cast_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_7_axis_0 = const()[name = tensor<string, []>("k_cache_new_7_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_7_mode_0 = const()[name = tensor<string, []>("k_cache_new_7_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_7_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_7_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_3_to_fp16 = cast(dtype = k_cache_3_to_fp16_dtype_0, x = k_cache_3)[name = tensor<string, []>("cast_89")];
            tensor<fp16, [1, 8, 1, 128]> key_13_cast_fp16 = transpose(perm = key_13_perm_0, x = var_836_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_7_cast_fp16 = scatter_along_axis(axis = k_cache_new_7_axis_0, data = k_cache_3_to_fp16, indices = pos_idx, mode = k_cache_new_7_mode_0, updates = key_13_cast_fp16, validate_indices = k_cache_new_7_validate_indices_0)[name = tensor<string, []>("k_cache_new_7_cast_fp16")];
            tensor<string, []> k_cache_new_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_7_axis_0 = const()[name = tensor<string, []>("v_cache_new_7_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_7_mode_0 = const()[name = tensor<string, []>("v_cache_new_7_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_7_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_7_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_3_to_fp16 = cast(dtype = v_cache_3_to_fp16_dtype_0, x = v_cache_3)[name = tensor<string, []>("cast_87")];
            tensor<fp16, [1, 8, 1, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = var_851_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_7_cast_fp16 = scatter_along_axis(axis = v_cache_new_7_axis_0, data = v_cache_3_to_fp16, indices = pos_idx, mode = v_cache_new_7_mode_0, updates = value_13_cast_fp16, validate_indices = v_cache_new_7_validate_indices_0)[name = tensor<string, []>("v_cache_new_7_cast_fp16")];
            tensor<string, []> v_cache_new_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_861_to_fp16 = const()[name = tensor<string, []>("op_861_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_13_cast_fp16 = transpose(perm = query_13_perm_0, x = var_821_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [1, 8, 1, 128]> mul_6_cast_fp16 = mul(x = query_13_cast_fp16, y = var_861_to_fp16)[name = tensor<string, []>("mul_6_cast_fp16")];
            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_6_cast_fp16, y = k_cache_new_7_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_6_cast_fp16")];
            tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = tensor<string, []>("softmax_6_cast_fp16")];
            tensor<bool, []> attn_output_19_transpose_x_0 = const()[name = tensor<string, []>("attn_output_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_19_transpose_y_0 = const()[name = tensor<string, []>("attn_output_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = softmax_6_cast_fp16, y = v_cache_new_7_cast_fp16)[name = tensor<string, []>("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_866_perm_0 = const()[name = tensor<string, []>("op_866_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_872 = const()[name = tensor<string, []>("op_872"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_866_cast_fp16 = transpose(perm = var_866_perm_0, x = attn_output_19_cast_fp16)[name = tensor<string, []>("transpose_84")];
            tensor<fp16, [1, 1, 1024]> input_67_cast_fp16 = reshape(shape = var_872, x = var_866_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142743232)))];
            tensor<fp16, [1024]> layers_3_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144840448)))];
            tensor<fp16, [1, 1, 1024]> linear_33_cast_fp16 = linear(bias = layers_3_first_sub_layer_out_projection_bias_to_fp16, weight = layers_3_first_sub_layer_out_projection_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_33_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = tensor<string, []>("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144842560)))];
            tensor<fp16, [1024]> layers_3_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144844672)))];
            tensor<fp16, []> var_880_to_fp16 = const()[name = tensor<string, []>("op_880_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = layers_3_layer_norm_2_bias_to_fp16, epsilon = var_880_to_fp16, gamma = layers_3_layer_norm_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144846784)))];
            tensor<fp16, [1024]> layers_3_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944000)))];
            tensor<fp16, [1, 1, 1024]> linear_34_cast_fp16 = linear(bias = layers_3_second_sub_layer_query_net_bias_to_fp16, weight = layers_3_second_sub_layer_query_net_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<int32, [4]> var_904 = const()[name = tensor<string, []>("op_904"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_905_cast_fp16 = reshape(shape = var_904, x = linear_34_cast_fp16)[name = tensor<string, []>("op_905_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146946112)))];
            tensor<fp16, [1024]> layers_3_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149043328)))];
            tensor<fp16, [1, 438, 1024]> linear_35_cast_fp16 = linear(bias = layers_3_second_sub_layer_key_net_bias_to_fp16, weight = layers_3_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<int32, [4]> var_912 = const()[name = tensor<string, []>("op_912"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_913_cast_fp16 = reshape(shape = var_912, x = linear_35_cast_fp16)[name = tensor<string, []>("op_913_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149045440)))];
            tensor<fp16, [1024]> layers_3_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151142656)))];
            tensor<fp16, [1, 438, 1024]> linear_36_cast_fp16 = linear(bias = layers_3_second_sub_layer_value_net_bias_to_fp16, weight = layers_3_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<int32, [4]> var_920 = const()[name = tensor<string, []>("op_920"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_921_cast_fp16 = reshape(shape = var_920, x = linear_36_cast_fp16)[name = tensor<string, []>("op_921_cast_fp16")];
            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_887_to_fp16 = const()[name = tensor<string, []>("op_887_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_7_cast_fp16 = mul(x = var_905_cast_fp16, y = var_887_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")];
            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = tensor<string, []>("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_39_perm_0 = const()[name = tensor<string, []>("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_39 = transpose(perm = transpose_39_perm_0, x = var_913_cast_fp16)[name = tensor<string, []>("transpose_81")];
            tensor<fp16, [1, 8, 1, 128]> transpose_38 = transpose(perm = transpose_38_perm_0, x = mul_7_cast_fp16)[name = tensor<string, []>("transpose_82")];
            tensor<fp16, [1, 8, 1, 438]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_38, y = transpose_39)[name = tensor<string, []>("matmul_7_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_7_cast_fp16")];
            tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = tensor<string, []>("softmax_7_cast_fp16")];
            tensor<bool, []> attn_output_23_transpose_x_0 = const()[name = tensor<string, []>("attn_output_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_23_transpose_y_0 = const()[name = tensor<string, []>("attn_output_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = var_921_cast_fp16)[name = tensor<string, []>("transpose_83")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("attn_output_23_cast_fp16")];
            tensor<int32, [4]> var_924_perm_0 = const()[name = tensor<string, []>("op_924_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_928 = const()[name = tensor<string, []>("op_928"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_924_cast_fp16 = transpose(perm = var_924_perm_0, x = attn_output_23_cast_fp16)[name = tensor<string, []>("transpose_80")];
            tensor<fp16, [1, 1, 1024]> input_73_cast_fp16 = reshape(shape = var_928, x = var_924_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151144768)))];
            tensor<fp16, [1024]> layers_3_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153241984)))];
            tensor<fp16, [1, 1, 1024]> linear_37_cast_fp16 = linear(bias = layers_3_second_sub_layer_out_projection_bias_to_fp16, weight = layers_3_second_sub_layer_out_projection_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_37_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153244096)))];
            tensor<fp16, [1024]> layers_3_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153246208)))];
            tensor<fp16, []> var_936_to_fp16 = const()[name = tensor<string, []>("op_936_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = layers_3_layer_norm_3_bias_to_fp16, epsilon = var_936_to_fp16, gamma = layers_3_layer_norm_3_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_3_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153248320)))];
            tensor<fp16, [4096]> layers_3_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161636992)))];
            tensor<fp16, [1, 1, 4096]> linear_38_cast_fp16 = linear(bias = layers_3_third_sub_layer_dense_in_bias_to_fp16, weight = layers_3_third_sub_layer_dense_in_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_81_cast_fp16 = relu(x = linear_38_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_3_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161645248)))];
            tensor<fp16, [1024]> layers_3_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170033920)))];
            tensor<fp16, [1, 1, 1024]> linear_39_cast_fp16 = linear(bias = layers_3_third_sub_layer_dense_out_bias_to_fp16, weight = layers_3_third_sub_layer_dense_out_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_83_cast_fp16 = add(x = input_75_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = tensor<string, []>("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170036032)))];
            tensor<fp16, [1024]> layers_4_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170038144)))];
            tensor<fp16, []> var_954_to_fp16 = const()[name = tensor<string, []>("op_954_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, beta = layers_4_layer_norm_1_bias_to_fp16, epsilon = var_954_to_fp16, gamma = layers_4_layer_norm_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170040256)))];
            tensor<fp16, [1024]> layers_4_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172137472)))];
            tensor<fp16, [1, 1, 1024]> linear_40_cast_fp16 = linear(bias = layers_4_first_sub_layer_query_net_bias_to_fp16, weight = layers_4_first_sub_layer_query_net_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172139584)))];
            tensor<fp16, [1024]> layers_4_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174236800)))];
            tensor<fp16, [1, 1, 1024]> linear_41_cast_fp16 = linear(bias = layers_4_first_sub_layer_key_net_bias_to_fp16, weight = layers_4_first_sub_layer_key_net_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174238912)))];
            tensor<fp16, [1024]> layers_4_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176336128)))];
            tensor<fp16, [1, 1, 1024]> linear_42_cast_fp16 = linear(bias = layers_4_first_sub_layer_value_net_bias_to_fp16, weight = layers_4_first_sub_layer_value_net_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<int32, [4]> var_979 = const()[name = tensor<string, []>("op_979"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_980_cast_fp16 = reshape(shape = var_979, x = linear_40_cast_fp16)[name = tensor<string, []>("op_980_cast_fp16")];
            tensor<int32, [4]> query_17_perm_0 = const()[name = tensor<string, []>("query_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_994 = const()[name = tensor<string, []>("op_994"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_995_cast_fp16 = reshape(shape = var_994, x = linear_41_cast_fp16)[name = tensor<string, []>("op_995_cast_fp16")];
            tensor<int32, [4]> key_17_perm_0 = const()[name = tensor<string, []>("key_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1009 = const()[name = tensor<string, []>("op_1009"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1010_cast_fp16 = reshape(shape = var_1009, x = linear_42_cast_fp16)[name = tensor<string, []>("op_1010_cast_fp16")];
            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_9_axis_0 = const()[name = tensor<string, []>("k_cache_new_9_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_9_mode_0 = const()[name = tensor<string, []>("k_cache_new_9_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_9_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_9_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_4_to_fp16 = cast(dtype = k_cache_4_to_fp16_dtype_0, x = k_cache_4)[name = tensor<string, []>("cast_85")];
            tensor<fp16, [1, 8, 1, 128]> key_17_cast_fp16 = transpose(perm = key_17_perm_0, x = var_995_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_9_cast_fp16 = scatter_along_axis(axis = k_cache_new_9_axis_0, data = k_cache_4_to_fp16, indices = pos_idx, mode = k_cache_new_9_mode_0, updates = key_17_cast_fp16, validate_indices = k_cache_new_9_validate_indices_0)[name = tensor<string, []>("k_cache_new_9_cast_fp16")];
            tensor<string, []> k_cache_new_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_9_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_9_axis_0 = const()[name = tensor<string, []>("v_cache_new_9_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_9_mode_0 = const()[name = tensor<string, []>("v_cache_new_9_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_9_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_9_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_4_to_fp16 = cast(dtype = v_cache_4_to_fp16_dtype_0, x = v_cache_4)[name = tensor<string, []>("cast_83")];
            tensor<fp16, [1, 8, 1, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = var_1010_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_9_cast_fp16 = scatter_along_axis(axis = v_cache_new_9_axis_0, data = v_cache_4_to_fp16, indices = pos_idx, mode = v_cache_new_9_mode_0, updates = value_17_cast_fp16, validate_indices = v_cache_new_9_validate_indices_0)[name = tensor<string, []>("v_cache_new_9_cast_fp16")];
            tensor<string, []> v_cache_new_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_9_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_1020_to_fp16 = const()[name = tensor<string, []>("op_1020_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_17_cast_fp16 = transpose(perm = query_17_perm_0, x = var_980_cast_fp16)[name = tensor<string, []>("transpose_79")];
            tensor<fp16, [1, 8, 1, 128]> mul_8_cast_fp16 = mul(x = query_17_cast_fp16, y = var_1020_to_fp16)[name = tensor<string, []>("mul_8_cast_fp16")];
            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_8_cast_fp16, y = k_cache_new_9_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_8_cast_fp16")];
            tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = tensor<string, []>("softmax_8_cast_fp16")];
            tensor<bool, []> attn_output_25_transpose_x_0 = const()[name = tensor<string, []>("attn_output_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_25_transpose_y_0 = const()[name = tensor<string, []>("attn_output_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_8_cast_fp16, y = v_cache_new_9_cast_fp16)[name = tensor<string, []>("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1025_perm_0 = const()[name = tensor<string, []>("op_1025_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1031 = const()[name = tensor<string, []>("op_1031"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1025_cast_fp16 = transpose(perm = var_1025_perm_0, x = attn_output_25_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [1, 1, 1024]> input_87_cast_fp16 = reshape(shape = var_1031, x = var_1025_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176338240)))];
            tensor<fp16, [1024]> layers_4_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178435456)))];
            tensor<fp16, [1, 1, 1024]> linear_43_cast_fp16 = linear(bias = layers_4_first_sub_layer_out_projection_bias_to_fp16, weight = layers_4_first_sub_layer_out_projection_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_89_cast_fp16 = add(x = input_83_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<int32, [1]> input_91_axes_0 = const()[name = tensor<string, []>("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178437568)))];
            tensor<fp16, [1024]> layers_4_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178439680)))];
            tensor<fp16, []> var_1039_to_fp16 = const()[name = tensor<string, []>("op_1039_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = layers_4_layer_norm_2_bias_to_fp16, epsilon = var_1039_to_fp16, gamma = layers_4_layer_norm_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178441792)))];
            tensor<fp16, [1024]> layers_4_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180539008)))];
            tensor<fp16, [1, 1, 1024]> linear_44_cast_fp16 = linear(bias = layers_4_second_sub_layer_query_net_bias_to_fp16, weight = layers_4_second_sub_layer_query_net_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1063 = const()[name = tensor<string, []>("op_1063"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1064_cast_fp16 = reshape(shape = var_1063, x = linear_44_cast_fp16)[name = tensor<string, []>("op_1064_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180541120)))];
            tensor<fp16, [1024]> layers_4_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182638336)))];
            tensor<fp16, [1, 438, 1024]> linear_45_cast_fp16 = linear(bias = layers_4_second_sub_layer_key_net_bias_to_fp16, weight = layers_4_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<int32, [4]> var_1071 = const()[name = tensor<string, []>("op_1071"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1072_cast_fp16 = reshape(shape = var_1071, x = linear_45_cast_fp16)[name = tensor<string, []>("op_1072_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182640448)))];
            tensor<fp16, [1024]> layers_4_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184737664)))];
            tensor<fp16, [1, 438, 1024]> linear_46_cast_fp16 = linear(bias = layers_4_second_sub_layer_value_net_bias_to_fp16, weight = layers_4_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<int32, [4]> var_1079 = const()[name = tensor<string, []>("op_1079"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1080_cast_fp16 = reshape(shape = var_1079, x = linear_46_cast_fp16)[name = tensor<string, []>("op_1080_cast_fp16")];
            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_1046_to_fp16 = const()[name = tensor<string, []>("op_1046_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_9_cast_fp16 = mul(x = var_1064_cast_fp16, y = var_1046_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")];
            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = tensor<string, []>("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_41_perm_0 = const()[name = tensor<string, []>("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_41 = transpose(perm = transpose_41_perm_0, x = var_1072_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [1, 8, 1, 128]> transpose_40 = transpose(perm = transpose_40_perm_0, x = mul_9_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [1, 8, 1, 438]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_40, y = transpose_41)[name = tensor<string, []>("matmul_9_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_9_cast_fp16")];
            tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = tensor<string, []>("softmax_9_cast_fp16")];
            tensor<bool, []> attn_output_29_transpose_x_0 = const()[name = tensor<string, []>("attn_output_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_29_transpose_y_0 = const()[name = tensor<string, []>("attn_output_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = var_1080_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1083_perm_0 = const()[name = tensor<string, []>("op_1083_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1087 = const()[name = tensor<string, []>("op_1087"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1083_cast_fp16 = transpose(perm = var_1083_perm_0, x = attn_output_29_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 1, 1024]> input_93_cast_fp16 = reshape(shape = var_1087, x = var_1083_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184739776)))];
            tensor<fp16, [1024]> layers_4_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186836992)))];
            tensor<fp16, [1, 1, 1024]> linear_47_cast_fp16 = linear(bias = layers_4_second_sub_layer_out_projection_bias_to_fp16, weight = layers_4_second_sub_layer_out_projection_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_95_cast_fp16 = add(x = input_89_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<int32, [1]> input_97_axes_0 = const()[name = tensor<string, []>("input_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186839104)))];
            tensor<fp16, [1024]> layers_4_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186841216)))];
            tensor<fp16, []> var_1095_to_fp16 = const()[name = tensor<string, []>("op_1095_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = layers_4_layer_norm_3_bias_to_fp16, epsilon = var_1095_to_fp16, gamma = layers_4_layer_norm_3_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_4_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186843328)))];
            tensor<fp16, [4096]> layers_4_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195232000)))];
            tensor<fp16, [1, 1, 4096]> linear_48_cast_fp16 = linear(bias = layers_4_third_sub_layer_dense_in_bias_to_fp16, weight = layers_4_third_sub_layer_dense_in_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_101_cast_fp16 = relu(x = linear_48_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_4_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195240256)))];
            tensor<fp16, [1024]> layers_4_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203628928)))];
            tensor<fp16, [1, 1, 1024]> linear_49_cast_fp16 = linear(bias = layers_4_third_sub_layer_dense_out_bias_to_fp16, weight = layers_4_third_sub_layer_dense_out_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_103_cast_fp16 = add(x = input_95_cast_fp16, y = linear_49_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<int32, [1]> input_105_axes_0 = const()[name = tensor<string, []>("input_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203631040)))];
            tensor<fp16, [1024]> layers_5_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203633152)))];
            tensor<fp16, []> var_1113_to_fp16 = const()[name = tensor<string, []>("op_1113_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = layers_5_layer_norm_1_bias_to_fp16, epsilon = var_1113_to_fp16, gamma = layers_5_layer_norm_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203635264)))];
            tensor<fp16, [1024]> layers_5_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205732480)))];
            tensor<fp16, [1, 1, 1024]> linear_50_cast_fp16 = linear(bias = layers_5_first_sub_layer_query_net_bias_to_fp16, weight = layers_5_first_sub_layer_query_net_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205734592)))];
            tensor<fp16, [1024]> layers_5_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207831808)))];
            tensor<fp16, [1, 1, 1024]> linear_51_cast_fp16 = linear(bias = layers_5_first_sub_layer_key_net_bias_to_fp16, weight = layers_5_first_sub_layer_key_net_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207833920)))];
            tensor<fp16, [1024]> layers_5_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209931136)))];
            tensor<fp16, [1, 1, 1024]> linear_52_cast_fp16 = linear(bias = layers_5_first_sub_layer_value_net_bias_to_fp16, weight = layers_5_first_sub_layer_value_net_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<int32, [4]> var_1138 = const()[name = tensor<string, []>("op_1138"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1139_cast_fp16 = reshape(shape = var_1138, x = linear_50_cast_fp16)[name = tensor<string, []>("op_1139_cast_fp16")];
            tensor<int32, [4]> query_21_perm_0 = const()[name = tensor<string, []>("query_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1153 = const()[name = tensor<string, []>("op_1153"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1154_cast_fp16 = reshape(shape = var_1153, x = linear_51_cast_fp16)[name = tensor<string, []>("op_1154_cast_fp16")];
            tensor<int32, [4]> key_21_perm_0 = const()[name = tensor<string, []>("key_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1168 = const()[name = tensor<string, []>("op_1168"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1169_cast_fp16 = reshape(shape = var_1168, x = linear_52_cast_fp16)[name = tensor<string, []>("op_1169_cast_fp16")];
            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_11_axis_0 = const()[name = tensor<string, []>("k_cache_new_11_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_11_mode_0 = const()[name = tensor<string, []>("k_cache_new_11_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_11_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_11_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_5_to_fp16 = cast(dtype = k_cache_5_to_fp16_dtype_0, x = k_cache_5)[name = tensor<string, []>("cast_81")];
            tensor<fp16, [1, 8, 1, 128]> key_21_cast_fp16 = transpose(perm = key_21_perm_0, x = var_1154_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_11_cast_fp16 = scatter_along_axis(axis = k_cache_new_11_axis_0, data = k_cache_5_to_fp16, indices = pos_idx, mode = k_cache_new_11_mode_0, updates = key_21_cast_fp16, validate_indices = k_cache_new_11_validate_indices_0)[name = tensor<string, []>("k_cache_new_11_cast_fp16")];
            tensor<string, []> k_cache_new_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_11_axis_0 = const()[name = tensor<string, []>("v_cache_new_11_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_11_mode_0 = const()[name = tensor<string, []>("v_cache_new_11_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_11_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_11_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_5_to_fp16 = cast(dtype = v_cache_5_to_fp16_dtype_0, x = v_cache_5)[name = tensor<string, []>("cast_79")];
            tensor<fp16, [1, 8, 1, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = var_1169_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_11_cast_fp16 = scatter_along_axis(axis = v_cache_new_11_axis_0, data = v_cache_5_to_fp16, indices = pos_idx, mode = v_cache_new_11_mode_0, updates = value_21_cast_fp16, validate_indices = v_cache_new_11_validate_indices_0)[name = tensor<string, []>("v_cache_new_11_cast_fp16")];
            tensor<string, []> v_cache_new_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_1179_to_fp16 = const()[name = tensor<string, []>("op_1179_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_21_cast_fp16 = transpose(perm = query_21_perm_0, x = var_1139_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [1, 8, 1, 128]> mul_10_cast_fp16 = mul(x = query_21_cast_fp16, y = var_1179_to_fp16)[name = tensor<string, []>("mul_10_cast_fp16")];
            tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_10_cast_fp16, y = k_cache_new_11_cast_fp16)[name = tensor<string, []>("matmul_10_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_10_cast_fp16")];
            tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = tensor<string, []>("softmax_10_cast_fp16")];
            tensor<bool, []> attn_output_31_transpose_x_0 = const()[name = tensor<string, []>("attn_output_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_31_transpose_y_0 = const()[name = tensor<string, []>("attn_output_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = softmax_10_cast_fp16, y = v_cache_new_11_cast_fp16)[name = tensor<string, []>("attn_output_31_cast_fp16")];
            tensor<int32, [4]> var_1184_perm_0 = const()[name = tensor<string, []>("op_1184_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1190 = const()[name = tensor<string, []>("op_1190"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1184_cast_fp16 = transpose(perm = var_1184_perm_0, x = attn_output_31_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 1, 1024]> input_107_cast_fp16 = reshape(shape = var_1190, x = var_1184_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209933248)))];
            tensor<fp16, [1024]> layers_5_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212030464)))];
            tensor<fp16, [1, 1, 1024]> linear_53_cast_fp16 = linear(bias = layers_5_first_sub_layer_out_projection_bias_to_fp16, weight = layers_5_first_sub_layer_out_projection_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_53_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<int32, [1]> input_111_axes_0 = const()[name = tensor<string, []>("input_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212032576)))];
            tensor<fp16, [1024]> layers_5_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212034688)))];
            tensor<fp16, []> var_1198_to_fp16 = const()[name = tensor<string, []>("op_1198_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = layers_5_layer_norm_2_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = layers_5_layer_norm_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212036800)))];
            tensor<fp16, [1024]> layers_5_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214134016)))];
            tensor<fp16, [1, 1, 1024]> linear_54_cast_fp16 = linear(bias = layers_5_second_sub_layer_query_net_bias_to_fp16, weight = layers_5_second_sub_layer_query_net_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<int32, [4]> var_1222 = const()[name = tensor<string, []>("op_1222"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1223_cast_fp16 = reshape(shape = var_1222, x = linear_54_cast_fp16)[name = tensor<string, []>("op_1223_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214136128)))];
            tensor<fp16, [1024]> layers_5_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216233344)))];
            tensor<fp16, [1, 438, 1024]> linear_55_cast_fp16 = linear(bias = layers_5_second_sub_layer_key_net_bias_to_fp16, weight = layers_5_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<int32, [4]> var_1230 = const()[name = tensor<string, []>("op_1230"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1231_cast_fp16 = reshape(shape = var_1230, x = linear_55_cast_fp16)[name = tensor<string, []>("op_1231_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216235456)))];
            tensor<fp16, [1024]> layers_5_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218332672)))];
            tensor<fp16, [1, 438, 1024]> linear_56_cast_fp16 = linear(bias = layers_5_second_sub_layer_value_net_bias_to_fp16, weight = layers_5_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1238 = const()[name = tensor<string, []>("op_1238"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1239_cast_fp16 = reshape(shape = var_1238, x = linear_56_cast_fp16)[name = tensor<string, []>("op_1239_cast_fp16")];
            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_1205_to_fp16 = const()[name = tensor<string, []>("op_1205_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_11_cast_fp16 = mul(x = var_1223_cast_fp16, y = var_1205_to_fp16)[name = tensor<string, []>("mul_11_cast_fp16")];
            tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = tensor<string, []>("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_43_perm_0 = const()[name = tensor<string, []>("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_43 = transpose(perm = transpose_43_perm_0, x = var_1231_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [1, 8, 1, 128]> transpose_42 = transpose(perm = transpose_42_perm_0, x = mul_11_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [1, 8, 1, 438]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_42, y = transpose_43)[name = tensor<string, []>("matmul_11_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_11_cast_fp16")];
            tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = tensor<string, []>("softmax_11_cast_fp16")];
            tensor<bool, []> attn_output_35_transpose_x_0 = const()[name = tensor<string, []>("attn_output_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_35_transpose_y_0 = const()[name = tensor<string, []>("attn_output_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = var_1239_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = tensor<string, []>("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1242_perm_0 = const()[name = tensor<string, []>("op_1242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1246 = const()[name = tensor<string, []>("op_1246"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1242_cast_fp16 = transpose(perm = var_1242_perm_0, x = attn_output_35_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 1, 1024]> input_113_cast_fp16 = reshape(shape = var_1246, x = var_1242_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218334784)))];
            tensor<fp16, [1024]> layers_5_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220432000)))];
            tensor<fp16, [1, 1, 1024]> linear_57_cast_fp16 = linear(bias = layers_5_second_sub_layer_out_projection_bias_to_fp16, weight = layers_5_second_sub_layer_out_projection_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_57_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220434112)))];
            tensor<fp16, [1024]> layers_5_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220436224)))];
            tensor<fp16, []> var_1254_to_fp16 = const()[name = tensor<string, []>("op_1254_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = layers_5_layer_norm_3_bias_to_fp16, epsilon = var_1254_to_fp16, gamma = layers_5_layer_norm_3_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_5_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220438336)))];
            tensor<fp16, [4096]> layers_5_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228827008)))];
            tensor<fp16, [1, 1, 4096]> linear_58_cast_fp16 = linear(bias = layers_5_third_sub_layer_dense_in_bias_to_fp16, weight = layers_5_third_sub_layer_dense_in_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_121_cast_fp16 = relu(x = linear_58_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_5_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228835264)))];
            tensor<fp16, [1024]> layers_5_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237223936)))];
            tensor<fp16, [1, 1, 1024]> linear_59_cast_fp16 = linear(bias = layers_5_third_sub_layer_dense_out_bias_to_fp16, weight = layers_5_third_sub_layer_dense_out_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_123_cast_fp16 = add(x = input_115_cast_fp16, y = linear_59_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<int32, [1]> input_125_axes_0 = const()[name = tensor<string, []>("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237226048)))];
            tensor<fp16, [1024]> layers_6_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237228160)))];
            tensor<fp16, []> var_1272_to_fp16 = const()[name = tensor<string, []>("op_1272_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = layers_6_layer_norm_1_bias_to_fp16, epsilon = var_1272_to_fp16, gamma = layers_6_layer_norm_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237230272)))];
            tensor<fp16, [1024]> layers_6_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239327488)))];
            tensor<fp16, [1, 1, 1024]> linear_60_cast_fp16 = linear(bias = layers_6_first_sub_layer_query_net_bias_to_fp16, weight = layers_6_first_sub_layer_query_net_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239329600)))];
            tensor<fp16, [1024]> layers_6_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241426816)))];
            tensor<fp16, [1, 1, 1024]> linear_61_cast_fp16 = linear(bias = layers_6_first_sub_layer_key_net_bias_to_fp16, weight = layers_6_first_sub_layer_key_net_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241428928)))];
            tensor<fp16, [1024]> layers_6_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243526144)))];
            tensor<fp16, [1, 1, 1024]> linear_62_cast_fp16 = linear(bias = layers_6_first_sub_layer_value_net_bias_to_fp16, weight = layers_6_first_sub_layer_value_net_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<int32, [4]> var_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1298_cast_fp16 = reshape(shape = var_1297, x = linear_60_cast_fp16)[name = tensor<string, []>("op_1298_cast_fp16")];
            tensor<int32, [4]> query_25_perm_0 = const()[name = tensor<string, []>("query_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1312 = const()[name = tensor<string, []>("op_1312"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1313_cast_fp16 = reshape(shape = var_1312, x = linear_61_cast_fp16)[name = tensor<string, []>("op_1313_cast_fp16")];
            tensor<int32, [4]> key_25_perm_0 = const()[name = tensor<string, []>("key_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1327 = const()[name = tensor<string, []>("op_1327"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1328_cast_fp16 = reshape(shape = var_1327, x = linear_62_cast_fp16)[name = tensor<string, []>("op_1328_cast_fp16")];
            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_13_axis_0 = const()[name = tensor<string, []>("k_cache_new_13_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_13_mode_0 = const()[name = tensor<string, []>("k_cache_new_13_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_13_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_13_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_6_to_fp16 = cast(dtype = k_cache_6_to_fp16_dtype_0, x = k_cache_6)[name = tensor<string, []>("cast_77")];
            tensor<fp16, [1, 8, 1, 128]> key_25_cast_fp16 = transpose(perm = key_25_perm_0, x = var_1313_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_13_cast_fp16 = scatter_along_axis(axis = k_cache_new_13_axis_0, data = k_cache_6_to_fp16, indices = pos_idx, mode = k_cache_new_13_mode_0, updates = key_25_cast_fp16, validate_indices = k_cache_new_13_validate_indices_0)[name = tensor<string, []>("k_cache_new_13_cast_fp16")];
            tensor<string, []> k_cache_new_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_13_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_13_axis_0 = const()[name = tensor<string, []>("v_cache_new_13_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_13_mode_0 = const()[name = tensor<string, []>("v_cache_new_13_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_13_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_13_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_6_to_fp16 = cast(dtype = v_cache_6_to_fp16_dtype_0, x = v_cache_6)[name = tensor<string, []>("cast_75")];
            tensor<fp16, [1, 8, 1, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = var_1328_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_13_cast_fp16 = scatter_along_axis(axis = v_cache_new_13_axis_0, data = v_cache_6_to_fp16, indices = pos_idx, mode = v_cache_new_13_mode_0, updates = value_25_cast_fp16, validate_indices = v_cache_new_13_validate_indices_0)[name = tensor<string, []>("v_cache_new_13_cast_fp16")];
            tensor<string, []> v_cache_new_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_13_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_1338_to_fp16 = const()[name = tensor<string, []>("op_1338_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_25_cast_fp16 = transpose(perm = query_25_perm_0, x = var_1298_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [1, 8, 1, 128]> mul_12_cast_fp16 = mul(x = query_25_cast_fp16, y = var_1338_to_fp16)[name = tensor<string, []>("mul_12_cast_fp16")];
            tensor<bool, []> matmul_12_transpose_y_0 = const()[name = tensor<string, []>("matmul_12_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_12_transpose_x_0 = const()[name = tensor<string, []>("matmul_12_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_12_cast_fp16, y = k_cache_new_13_cast_fp16)[name = tensor<string, []>("matmul_12_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_12_cast_fp16")];
            tensor<int32, []> softmax_12_axis_0 = const()[name = tensor<string, []>("softmax_12_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = tensor<string, []>("softmax_12_cast_fp16")];
            tensor<bool, []> attn_output_37_transpose_x_0 = const()[name = tensor<string, []>("attn_output_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_37_transpose_y_0 = const()[name = tensor<string, []>("attn_output_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_12_cast_fp16, y = v_cache_new_13_cast_fp16)[name = tensor<string, []>("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1343_perm_0 = const()[name = tensor<string, []>("op_1343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1349 = const()[name = tensor<string, []>("op_1349"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1343_cast_fp16 = transpose(perm = var_1343_perm_0, x = attn_output_37_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [1, 1, 1024]> input_127_cast_fp16 = reshape(shape = var_1349, x = var_1343_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243528256)))];
            tensor<fp16, [1024]> layers_6_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245625472)))];
            tensor<fp16, [1, 1, 1024]> linear_63_cast_fp16 = linear(bias = layers_6_first_sub_layer_out_projection_bias_to_fp16, weight = layers_6_first_sub_layer_out_projection_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_63_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245627584)))];
            tensor<fp16, [1024]> layers_6_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245629696)))];
            tensor<fp16, []> var_1357_to_fp16 = const()[name = tensor<string, []>("op_1357_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = layers_6_layer_norm_2_bias_to_fp16, epsilon = var_1357_to_fp16, gamma = layers_6_layer_norm_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245631808)))];
            tensor<fp16, [1024]> layers_6_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247729024)))];
            tensor<fp16, [1, 1, 1024]> linear_64_cast_fp16 = linear(bias = layers_6_second_sub_layer_query_net_bias_to_fp16, weight = layers_6_second_sub_layer_query_net_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1381 = const()[name = tensor<string, []>("op_1381"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1382_cast_fp16 = reshape(shape = var_1381, x = linear_64_cast_fp16)[name = tensor<string, []>("op_1382_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247731136)))];
            tensor<fp16, [1024]> layers_6_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249828352)))];
            tensor<fp16, [1, 438, 1024]> linear_65_cast_fp16 = linear(bias = layers_6_second_sub_layer_key_net_bias_to_fp16, weight = layers_6_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1389 = const()[name = tensor<string, []>("op_1389"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1390_cast_fp16 = reshape(shape = var_1389, x = linear_65_cast_fp16)[name = tensor<string, []>("op_1390_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249830464)))];
            tensor<fp16, [1024]> layers_6_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251927680)))];
            tensor<fp16, [1, 438, 1024]> linear_66_cast_fp16 = linear(bias = layers_6_second_sub_layer_value_net_bias_to_fp16, weight = layers_6_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<int32, [4]> var_1397 = const()[name = tensor<string, []>("op_1397"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1398_cast_fp16 = reshape(shape = var_1397, x = linear_66_cast_fp16)[name = tensor<string, []>("op_1398_cast_fp16")];
            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_1364_to_fp16 = const()[name = tensor<string, []>("op_1364_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_13_cast_fp16 = mul(x = var_1382_cast_fp16, y = var_1364_to_fp16)[name = tensor<string, []>("mul_13_cast_fp16")];
            tensor<bool, []> matmul_13_transpose_y_0 = const()[name = tensor<string, []>("matmul_13_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_13_transpose_x_0 = const()[name = tensor<string, []>("matmul_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = tensor<string, []>("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_45_perm_0 = const()[name = tensor<string, []>("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_45 = transpose(perm = transpose_45_perm_0, x = var_1390_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [1, 8, 1, 128]> transpose_44 = transpose(perm = transpose_44_perm_0, x = mul_13_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [1, 8, 1, 438]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = transpose_44, y = transpose_45)[name = tensor<string, []>("matmul_13_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_13_cast_fp16")];
            tensor<int32, []> softmax_13_axis_0 = const()[name = tensor<string, []>("softmax_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = tensor<string, []>("softmax_13_cast_fp16")];
            tensor<bool, []> attn_output_41_transpose_x_0 = const()[name = tensor<string, []>("attn_output_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_41_transpose_y_0 = const()[name = tensor<string, []>("attn_output_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = var_1398_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = tensor<string, []>("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1401_perm_0 = const()[name = tensor<string, []>("op_1401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1405 = const()[name = tensor<string, []>("op_1405"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1401_cast_fp16 = transpose(perm = var_1401_perm_0, x = attn_output_41_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [1, 1, 1024]> input_133_cast_fp16 = reshape(shape = var_1405, x = var_1401_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251929792)))];
            tensor<fp16, [1024]> layers_6_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254027008)))];
            tensor<fp16, [1, 1, 1024]> linear_67_cast_fp16 = linear(bias = layers_6_second_sub_layer_out_projection_bias_to_fp16, weight = layers_6_second_sub_layer_out_projection_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_67_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [1]> input_137_axes_0 = const()[name = tensor<string, []>("input_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254029120)))];
            tensor<fp16, [1024]> layers_6_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254031232)))];
            tensor<fp16, []> var_1413_to_fp16 = const()[name = tensor<string, []>("op_1413_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = layers_6_layer_norm_3_bias_to_fp16, epsilon = var_1413_to_fp16, gamma = layers_6_layer_norm_3_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_6_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254033344)))];
            tensor<fp16, [4096]> layers_6_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262422016)))];
            tensor<fp16, [1, 1, 4096]> linear_68_cast_fp16 = linear(bias = layers_6_third_sub_layer_dense_in_bias_to_fp16, weight = layers_6_third_sub_layer_dense_in_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_141_cast_fp16 = relu(x = linear_68_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_6_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262430272)))];
            tensor<fp16, [1024]> layers_6_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270818944)))];
            tensor<fp16, [1, 1, 1024]> linear_69_cast_fp16 = linear(bias = layers_6_third_sub_layer_dense_out_bias_to_fp16, weight = layers_6_third_sub_layer_dense_out_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_143_cast_fp16 = add(x = input_135_cast_fp16, y = linear_69_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [1]> input_145_axes_0 = const()[name = tensor<string, []>("input_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270821056)))];
            tensor<fp16, [1024]> layers_7_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270823168)))];
            tensor<fp16, []> var_1431_to_fp16 = const()[name = tensor<string, []>("op_1431_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_145_cast_fp16 = layer_norm(axes = input_145_axes_0, beta = layers_7_layer_norm_1_bias_to_fp16, epsilon = var_1431_to_fp16, gamma = layers_7_layer_norm_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_first_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270825280)))];
            tensor<fp16, [1024]> layers_7_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272922496)))];
            tensor<fp16, [1, 1, 1024]> linear_70_cast_fp16 = linear(bias = layers_7_first_sub_layer_query_net_bias_to_fp16, weight = layers_7_first_sub_layer_query_net_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_first_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272924608)))];
            tensor<fp16, [1024]> layers_7_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275021824)))];
            tensor<fp16, [1, 1, 1024]> linear_71_cast_fp16 = linear(bias = layers_7_first_sub_layer_key_net_bias_to_fp16, weight = layers_7_first_sub_layer_key_net_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_first_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275023936)))];
            tensor<fp16, [1024]> layers_7_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277121152)))];
            tensor<fp16, [1, 1, 1024]> linear_72_cast_fp16 = linear(bias = layers_7_first_sub_layer_value_net_bias_to_fp16, weight = layers_7_first_sub_layer_value_net_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1456 = const()[name = tensor<string, []>("op_1456"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1457_cast_fp16 = reshape(shape = var_1456, x = linear_70_cast_fp16)[name = tensor<string, []>("op_1457_cast_fp16")];
            tensor<int32, [4]> query_29_perm_0 = const()[name = tensor<string, []>("query_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1471 = const()[name = tensor<string, []>("op_1471"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1472_cast_fp16 = reshape(shape = var_1471, x = linear_71_cast_fp16)[name = tensor<string, []>("op_1472_cast_fp16")];
            tensor<int32, [4]> key_29_perm_0 = const()[name = tensor<string, []>("key_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1486 = const()[name = tensor<string, []>("op_1486"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1487_cast_fp16 = reshape(shape = var_1486, x = linear_72_cast_fp16)[name = tensor<string, []>("op_1487_cast_fp16")];
            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, []> k_cache_new_axis_0 = const()[name = tensor<string, []>("k_cache_new_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> k_cache_new_mode_0 = const()[name = tensor<string, []>("k_cache_new_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> k_cache_new_validate_indices_0 = const()[name = tensor<string, []>("k_cache_new_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> k_cache_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("k_cache_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_7_to_fp16 = cast(dtype = k_cache_7_to_fp16_dtype_0, x = k_cache_7)[name = tensor<string, []>("cast_73")];
            tensor<fp16, [1, 8, 1, 128]> key_29_cast_fp16 = transpose(perm = key_29_perm_0, x = var_1472_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [1, 8, 108, 128]> k_cache_new_cast_fp16 = scatter_along_axis(axis = k_cache_new_axis_0, data = k_cache_7_to_fp16, indices = pos_idx, mode = k_cache_new_mode_0, updates = key_29_cast_fp16, validate_indices = k_cache_new_validate_indices_0)[name = tensor<string, []>("k_cache_new_cast_fp16")];
            tensor<string, []> k_cache_new_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("k_cache_new_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> v_cache_new_axis_0 = const()[name = tensor<string, []>("v_cache_new_axis_0"), val = tensor<int32, []>(2)];
            tensor<string, []> v_cache_new_mode_0 = const()[name = tensor<string, []>("v_cache_new_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> v_cache_new_validate_indices_0 = const()[name = tensor<string, []>("v_cache_new_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> v_cache_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("v_cache_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_7_to_fp16 = cast(dtype = v_cache_7_to_fp16_dtype_0, x = v_cache_7)[name = tensor<string, []>("cast_71")];
            tensor<fp16, [1, 8, 1, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = var_1487_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 8, 108, 128]> v_cache_new_cast_fp16 = scatter_along_axis(axis = v_cache_new_axis_0, data = v_cache_7_to_fp16, indices = pos_idx, mode = v_cache_new_mode_0, updates = value_29_cast_fp16, validate_indices = v_cache_new_validate_indices_0)[name = tensor<string, []>("v_cache_new_cast_fp16")];
            tensor<string, []> v_cache_new_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("v_cache_new_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, []> var_1497_to_fp16 = const()[name = tensor<string, []>("op_1497_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 8, 1, 128]> query_29_cast_fp16 = transpose(perm = query_29_perm_0, x = var_1457_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 8, 1, 128]> mul_14_cast_fp16 = mul(x = query_29_cast_fp16, y = var_1497_to_fp16)[name = tensor<string, []>("mul_14_cast_fp16")];
            tensor<bool, []> matmul_14_transpose_y_0 = const()[name = tensor<string, []>("matmul_14_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_14_transpose_x_0 = const()[name = tensor<string, []>("matmul_14_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 108]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_14_cast_fp16, y = k_cache_new_cast_fp16)[name = tensor<string, []>("matmul_14_cast_fp16")];
            tensor<fp16, [1, 8, 1, 108]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = attention_mask_to_fp16)[name = tensor<string, []>("add_14_cast_fp16")];
            tensor<int32, []> softmax_14_axis_0 = const()[name = tensor<string, []>("softmax_14_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 108]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = tensor<string, []>("softmax_14_cast_fp16")];
            tensor<bool, []> attn_output_43_transpose_x_0 = const()[name = tensor<string, []>("attn_output_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_43_transpose_y_0 = const()[name = tensor<string, []>("attn_output_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = softmax_14_cast_fp16, y = v_cache_new_cast_fp16)[name = tensor<string, []>("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_1502_perm_0 = const()[name = tensor<string, []>("op_1502_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1508 = const()[name = tensor<string, []>("op_1508"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1502_cast_fp16 = transpose(perm = var_1502_perm_0, x = attn_output_43_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 1, 1024]> input_147_cast_fp16 = reshape(shape = var_1508, x = var_1502_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_first_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277123264)))];
            tensor<fp16, [1024]> layers_7_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279220480)))];
            tensor<fp16, [1, 1, 1024]> linear_73_cast_fp16 = linear(bias = layers_7_first_sub_layer_out_projection_bias_to_fp16, weight = layers_7_first_sub_layer_out_projection_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_149_cast_fp16 = add(x = input_143_cast_fp16, y = linear_73_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<int32, [1]> input_151_axes_0 = const()[name = tensor<string, []>("input_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279222592)))];
            tensor<fp16, [1024]> layers_7_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279224704)))];
            tensor<fp16, []> var_1516_to_fp16 = const()[name = tensor<string, []>("op_1516_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_151_cast_fp16 = layer_norm(axes = input_151_axes_0, beta = layers_7_layer_norm_2_bias_to_fp16, epsilon = var_1516_to_fp16, gamma = layers_7_layer_norm_2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_query_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_query_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279226816)))];
            tensor<fp16, [1024]> layers_7_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281324032)))];
            tensor<fp16, [1, 1, 1024]> linear_74_cast_fp16 = linear(bias = layers_7_second_sub_layer_query_net_bias_to_fp16, weight = layers_7_second_sub_layer_query_net_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
            tensor<int32, [4]> var_1540 = const()[name = tensor<string, []>("op_1540"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> var_1541_cast_fp16 = reshape(shape = var_1540, x = linear_74_cast_fp16)[name = tensor<string, []>("op_1541_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_key_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281326144)))];
            tensor<fp16, [1024]> layers_7_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283423360)))];
            tensor<fp16, [1, 438, 1024]> linear_75_cast_fp16 = linear(bias = layers_7_second_sub_layer_key_net_bias_to_fp16, weight = layers_7_second_sub_layer_key_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<int32, [4]> var_1548 = const()[name = tensor<string, []>("op_1548"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1549_cast_fp16 = reshape(shape = var_1548, x = linear_75_cast_fp16)[name = tensor<string, []>("op_1549_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_value_net_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283425472)))];
            tensor<fp16, [1024]> layers_7_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285522688)))];
            tensor<fp16, [1, 438, 1024]> linear_76_cast_fp16 = linear(bias = layers_7_second_sub_layer_value_net_bias_to_fp16, weight = layers_7_second_sub_layer_value_net_weight_to_fp16, x = encoder_hidden_states_to_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<int32, [4]> var_1556 = const()[name = tensor<string, []>("op_1556"), val = tensor<int32, [4]>([1, 438, 8, 128])];
            tensor<fp16, [1, 438, 8, 128]> var_1557_cast_fp16 = reshape(shape = var_1556, x = linear_76_cast_fp16)[name = tensor<string, []>("op_1557_cast_fp16")];
            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, []> var_1523_to_fp16 = const()[name = tensor<string, []>("op_1523_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 1, 8, 128]> mul_15_cast_fp16 = mul(x = var_1541_cast_fp16, y = var_1523_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")];
            tensor<bool, []> matmul_15_transpose_y_0 = const()[name = tensor<string, []>("matmul_15_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_15_transpose_x_0 = const()[name = tensor<string, []>("matmul_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = tensor<string, []>("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_47_perm_0 = const()[name = tensor<string, []>("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 438, 128]> transpose_47 = transpose(perm = transpose_47_perm_0, x = var_1549_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 8, 1, 128]> transpose_46 = transpose(perm = transpose_46_perm_0, x = mul_15_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 8, 1, 438]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = transpose_46, y = transpose_47)[name = tensor<string, []>("matmul_15_cast_fp16")];
            tensor<fp16, [1, 8, 1, 438]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = cross_attention_mask_to_fp16)[name = tensor<string, []>("add_15_cast_fp16")];
            tensor<int32, []> softmax_15_axis_0 = const()[name = tensor<string, []>("softmax_15_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 8, 1, 438]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = tensor<string, []>("softmax_15_cast_fp16")];
            tensor<bool, []> attn_output_transpose_x_0 = const()[name = tensor<string, []>("attn_output_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_transpose_y_0 = const()[name = tensor<string, []>("attn_output_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 438, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = var_1557_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 8, 1, 128]> attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = softmax_15_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("attn_output_cast_fp16")];
            tensor<int32, [4]> var_1560_perm_0 = const()[name = tensor<string, []>("op_1560_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1564 = const()[name = tensor<string, []>("op_1564"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_1560_cast_fp16 = transpose(perm = var_1560_perm_0, x = attn_output_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 1, 1024]> input_153_cast_fp16 = reshape(shape = var_1564, x = var_1560_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_out_projection_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_out_projection_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285524800)))];
            tensor<fp16, [1024]> layers_7_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287622016)))];
            tensor<fp16, [1, 1, 1024]> linear_77_cast_fp16 = linear(bias = layers_7_second_sub_layer_out_projection_bias_to_fp16, weight = layers_7_second_sub_layer_out_projection_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_155_cast_fp16 = add(x = input_149_cast_fp16, y = linear_77_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<int32, [1]> input_157_axes_0 = const()[name = tensor<string, []>("input_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287624128)))];
            tensor<fp16, [1024]> layers_7_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287626240)))];
            tensor<fp16, []> var_1572_to_fp16 = const()[name = tensor<string, []>("op_1572_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = layers_7_layer_norm_3_bias_to_fp16, epsilon = var_1572_to_fp16, gamma = layers_7_layer_norm_3_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_7_third_sub_layer_dense_in_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_third_sub_layer_dense_in_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287628352)))];
            tensor<fp16, [4096]> layers_7_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296017024)))];
            tensor<fp16, [1, 1, 4096]> linear_78_cast_fp16 = linear(bias = layers_7_third_sub_layer_dense_in_bias_to_fp16, weight = layers_7_third_sub_layer_dense_in_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_161_cast_fp16 = relu(x = linear_78_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_7_third_sub_layer_dense_out_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_third_sub_layer_dense_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296025280)))];
            tensor<fp16, [1024]> layers_7_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304413952)))];
            tensor<fp16, [1, 1, 1024]> linear_79_cast_fp16 = linear(bias = layers_7_third_sub_layer_dense_out_bias_to_fp16, weight = layers_7_third_sub_layer_dense_out_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_163_cast_fp16 = add(x = input_155_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> final_norm_weight_to_fp16 = const()[name = tensor<string, []>("final_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304416064)))];
            tensor<fp16, [1024]> final_norm_bias_to_fp16 = const()[name = tensor<string, []>("final_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304418176)))];
            tensor<fp16, []> var_1590_to_fp16 = const()[name = tensor<string, []>("op_1590_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_1590_to_fp16, gamma = final_norm_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<fp16, [16384]> lm_head_bias_to_fp16 = const()[name = tensor<string, []>("lm_head_bias_to_fp16"), val = tensor<fp16, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304420288)))];
            tensor<fp16, [1, 1, 16384]> linear_80_cast_fp16 = linear(bias = lm_head_bias_to_fp16, weight = embedding_token_embedding_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<int32, [1]> var_1600_axes_0 = const()[name = tensor<string, []>("op_1600_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 16384]> var_1600_cast_fp16 = squeeze(axes = var_1600_axes_0, x = linear_80_cast_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
            tensor<string, []> var_1600_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1600_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 16384]> logits = cast(dtype = var_1600_cast_fp16_to_fp32_dtype_0, x = var_1600_cast_fp16)[name = tensor<string, []>("cast_69")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_7_out = cast(dtype = v_cache_new_cast_fp16_to_fp32_dtype_0, x = v_cache_new_cast_fp16)[name = tensor<string, []>("cast_70")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_7_out = cast(dtype = k_cache_new_cast_fp16_to_fp32_dtype_0, x = k_cache_new_cast_fp16)[name = tensor<string, []>("cast_72")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_6_out = cast(dtype = v_cache_new_13_cast_fp16_to_fp32_dtype_0, x = v_cache_new_13_cast_fp16)[name = tensor<string, []>("cast_74")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_6_out = cast(dtype = k_cache_new_13_cast_fp16_to_fp32_dtype_0, x = k_cache_new_13_cast_fp16)[name = tensor<string, []>("cast_76")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_5_out = cast(dtype = v_cache_new_11_cast_fp16_to_fp32_dtype_0, x = v_cache_new_11_cast_fp16)[name = tensor<string, []>("cast_78")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_5_out = cast(dtype = k_cache_new_11_cast_fp16_to_fp32_dtype_0, x = k_cache_new_11_cast_fp16)[name = tensor<string, []>("cast_80")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_4_out = cast(dtype = v_cache_new_9_cast_fp16_to_fp32_dtype_0, x = v_cache_new_9_cast_fp16)[name = tensor<string, []>("cast_82")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_4_out = cast(dtype = k_cache_new_9_cast_fp16_to_fp32_dtype_0, x = k_cache_new_9_cast_fp16)[name = tensor<string, []>("cast_84")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_3_out = cast(dtype = v_cache_new_7_cast_fp16_to_fp32_dtype_0, x = v_cache_new_7_cast_fp16)[name = tensor<string, []>("cast_86")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_3_out = cast(dtype = k_cache_new_7_cast_fp16_to_fp32_dtype_0, x = k_cache_new_7_cast_fp16)[name = tensor<string, []>("cast_88")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_2_out = cast(dtype = v_cache_new_5_cast_fp16_to_fp32_dtype_0, x = v_cache_new_5_cast_fp16)[name = tensor<string, []>("cast_90")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_2_out = cast(dtype = k_cache_new_5_cast_fp16_to_fp32_dtype_0, x = k_cache_new_5_cast_fp16)[name = tensor<string, []>("cast_92")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_1_out = cast(dtype = v_cache_new_3_cast_fp16_to_fp32_dtype_0, x = v_cache_new_3_cast_fp16)[name = tensor<string, []>("cast_94")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_1_out = cast(dtype = k_cache_new_3_cast_fp16_to_fp32_dtype_0, x = k_cache_new_3_cast_fp16)[name = tensor<string, []>("cast_96")];
            tensor<fp32, [1, 8, 108, 128]> v_cache_0_out = cast(dtype = v_cache_new_1_cast_fp16_to_fp32_dtype_0, x = v_cache_new_1_cast_fp16)[name = tensor<string, []>("cast_101")];
            tensor<fp32, [1, 8, 108, 128]> k_cache_0_out = cast(dtype = k_cache_new_1_cast_fp16_to_fp32_dtype_0, x = k_cache_new_1_cast_fp16)[name = tensor<string, []>("cast_103")];
        } -> (logits, k_cache_0_out, v_cache_0_out, k_cache_1_out, v_cache_1_out, k_cache_2_out, v_cache_2_out, k_cache_3_out, v_cache_3_out, k_cache_4_out, v_cache_4_out, k_cache_5_out, v_cache_5_out, k_cache_6_out, v_cache_6_out, k_cache_7_out, v_cache_7_out);
}