program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})]
{
    func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 1024, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 24576, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 24576, 1, 448]> value_cache) {
            tensor<int32, []> var_64_axis_0 = const()[name = tensor<string, []>("op_64_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_64_batch_dims_0 = const()[name = tensor<string, []>("op_64_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [51864, 1024]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51864, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 1024]> var_64_cast_fp16 = gather(axis = var_64_axis_0, batch_dims = var_64_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_64_cast_fp16")];
            tensor<int32, []> var_68_axis_0 = const()[name = tensor<string, []>("op_68_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_68_batch_dims_0 = const()[name = tensor<string, []>("op_68_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [448, 1024]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106217600)))];
            tensor<fp16, [1, 1024]> var_68_cast_fp16 = gather(axis = var_68_axis_0, batch_dims = var_68_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_68_cast_fp16")];
            tensor<fp16, [1, 1024]> hidden_states_1_cast_fp16 = add(x = var_64_cast_fp16, y = var_68_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
            tensor<int32, [1]> var_82_axes_0 = const()[name = tensor<string, []>("op_82_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1]> var_82_cast_fp16 = expand_dims(axes = var_82_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_82_cast_fp16")];
            tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1024, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_82_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
            tensor<int32, [24]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [24]>([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])];
            tensor<int32, []> var_87_axis_0 = const()[name = tensor<string, []>("op_87_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_0, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_1, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_2, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_3, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_4, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_5, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_6, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_7, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_8, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_9, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_10, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_11, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_12, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_13, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_14, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_15, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_16, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_17, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_18, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_19, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_20, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_21, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_22, tensor<fp16, [1, 1024, 1, 448]> var_87_cast_fp16_23 = split(axis = var_87_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_87_cast_fp16")];
            tensor<int32, [24]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [24]>([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])];
            tensor<int32, []> var_114_axis_0 = const()[name = tensor<string, []>("op_114_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_0, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_1, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_2, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_3, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_4, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_5, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_6, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_7, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_8, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_9, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_10, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_11, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_12, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_13, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_14, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_15, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_16, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_17, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_18, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_19, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_20, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_21, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_22, tensor<fp16, [1, 1024, 1, 448]> var_114_cast_fp16_23 = split(axis = var_114_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_114_cast_fp16")];
            tensor<int32, []> var_144 = const()[name = tensor<string, []>("op_144"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_169_to_fp16 = const()[name = tensor<string, []>("op_169_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_169_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<fp16, [1024]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107135168)))];
            tensor<fp16, [1024]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107137280)))];
            tensor<fp16, [1024]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107139392)))];
            tensor<fp16, [1024]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107141504)))];
            tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
            tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107143616)))];
            tensor<fp16, [1024]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109240832)))];
            tensor<fp16, [1, 1024, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
            tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_1_strides_0 = const()[name = tensor<string, []>("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_1_dilations_0 = const()[name = tensor<string, []>("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_1_groups_0 = const()[name = tensor<string, []>("current_key_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109242944)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")];
            tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_1_strides_0 = const()[name = tensor<string, []>("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_1_dilations_0 = const()[name = tensor<string, []>("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_1_groups_0 = const()[name = tensor<string, []>("current_value_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111340160)))];
            tensor<fp16, [1024]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113437376)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")];
            tensor<int32, [1]> var_204_axes_0 = const()[name = tensor<string, []>("op_204_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_204_cast_fp16 = expand_dims(axes = var_204_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_204_cast_fp16")];
            tensor<int32, [1]> var_205_axes_0 = const()[name = tensor<string, []>("op_205_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_205_cast_fp16 = expand_dims(axes = var_205_axes_0, x = var_204_cast_fp16)[name = tensor<string, []>("op_205_cast_fp16")];
            tensor<fp16, []> var_145_to_fp16 = const()[name = tensor<string, []>("op_145_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 1, 1, 448]> var_207_cast_fp16 = sub(x = var_145_to_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_207_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_208_cast_fp16 = mul(x = var_87_cast_fp16_0, y = var_207_cast_fp16)[name = tensor<string, []>("op_208_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_209_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_209_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_1_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_212_cast_fp16 = mul(x = var_114_cast_fp16_0, y = var_207_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_213_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_213_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_1_cast_fp16 = add(x = var_212_cast_fp16, y = var_213_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
            tensor<int32, [4]> var_217 = const()[name = tensor<string, []>("op_217"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_217, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
            tensor<fp16, []> var_219_to_fp16 = const()[name = tensor<string, []>("op_219_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_220_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_219_to_fp16)[name = tensor<string, []>("op_220_cast_fp16")];
            tensor<int32, [4]> var_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_224_cast_fp16 = reshape(shape = var_223, x = key_1_cast_fp16)[name = tensor<string, []>("op_224_cast_fp16")];
            tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_220_cast_fp16, y = var_224_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
            tensor<int32, [1]> var_228_axes_0 = const()[name = tensor<string, []>("op_228_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 448]> var_228_cast_fp16 = expand_dims(axes = var_228_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_228_cast_fp16")];
            tensor<int32, [1]> var_229_axes_0 = const()[name = tensor<string, []>("op_229_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 448]> var_229_cast_fp16 = expand_dims(axes = var_229_axes_0, x = var_228_cast_fp16)[name = tensor<string, []>("op_229_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_232_cast_fp16 = softmax(axis = var_144, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_232_cast_fp16")];
            tensor<int32, [4]> var_233 = const()[name = tensor<string, []>("op_233"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_234_cast_fp16 = reshape(shape = var_233, x = value_1_cast_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
            tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_234_cast_fp16, y = var_232_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<int32, [4]> var_237 = const()[name = tensor<string, []>("op_237"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_1_cast_fp16 = reshape(shape = var_237, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113439488)))];
            tensor<fp16, [1024]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115536704)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_259_to_fp16 = const()[name = tensor<string, []>("op_259_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_259_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<fp16, [1024]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115538816)))];
            tensor<fp16, [1024]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115540928)))];
            tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
            tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115543040)))];
            tensor<fp16, [1024]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117640256)))];
            tensor<fp16, [1, 1024, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
            tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117642368)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")];
            tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119739584)))];
            tensor<fp16, [1024]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121836800)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")];
            tensor<int32, [4]> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_295, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
            tensor<fp16, []> var_297_to_fp16 = const()[name = tensor<string, []>("op_297_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_298_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_297_to_fp16)[name = tensor<string, []>("op_298_cast_fp16")];
            tensor<int32, [4]> var_301 = const()[name = tensor<string, []>("op_301"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_302_cast_fp16 = reshape(shape = var_301, x = key_3_cast_fp16)[name = tensor<string, []>("op_302_cast_fp16")];
            tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_298_cast_fp16, y = var_302_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_144, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
            tensor<int32, [4]> var_306 = const()[name = tensor<string, []>("op_306"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_307_cast_fp16 = reshape(shape = var_306, x = value_3_cast_fp16)[name = tensor<string, []>("op_307_cast_fp16")];
            tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_307_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
            tensor<int32, [4]> var_310 = const()[name = tensor<string, []>("op_310"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_3_cast_fp16 = reshape(shape = var_310, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121838912)))];
            tensor<fp16, [1024]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123936128)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_328_to_fp16 = const()[name = tensor<string, []>("op_328_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_328_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<fp16, [1024]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123938240)))];
            tensor<fp16, [1024]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123940352)))];
            tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123942464)))];
            tensor<fp16, [4096]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132331136)))];
            tensor<fp16, [1, 4096, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = tensor<string, []>("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = tensor<string, []>("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_3_groups_0 = const()[name = tensor<string, []>("hidden_states_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132339392)))];
            tensor<fp16, [1024]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140728064)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
            tensor<int32, []> var_363 = const()[name = tensor<string, []>("op_363"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_388_to_fp16 = const()[name = tensor<string, []>("op_388_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_388_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<fp16, [1024]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140730176)))];
            tensor<fp16, [1024]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140732288)))];
            tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
            tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140734400)))];
            tensor<fp16, [1024]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142831616)))];
            tensor<fp16, [1, 1024, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
            tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_3_strides_0 = const()[name = tensor<string, []>("current_key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_3_dilations_0 = const()[name = tensor<string, []>("current_key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_3_groups_0 = const()[name = tensor<string, []>("current_key_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142833728)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
            tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_3_strides_0 = const()[name = tensor<string, []>("current_value_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_3_dilations_0 = const()[name = tensor<string, []>("current_value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_3_groups_0 = const()[name = tensor<string, []>("current_value_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144930944)))];
            tensor<fp16, [1024]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147028160)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_427_cast_fp16 = mul(x = var_87_cast_fp16_1, y = var_207_cast_fp16)[name = tensor<string, []>("op_427_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_428_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_5_cast_fp16 = add(x = var_427_cast_fp16, y = var_428_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_431_cast_fp16 = mul(x = var_114_cast_fp16_1, y = var_207_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_432_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_432_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_5_cast_fp16 = add(x = var_431_cast_fp16, y = var_432_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
            tensor<int32, [4]> var_436 = const()[name = tensor<string, []>("op_436"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_436, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
            tensor<fp16, []> var_438_to_fp16 = const()[name = tensor<string, []>("op_438_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_439_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_438_to_fp16)[name = tensor<string, []>("op_439_cast_fp16")];
            tensor<int32, [4]> var_442 = const()[name = tensor<string, []>("op_442"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_443_cast_fp16 = reshape(shape = var_442, x = key_5_cast_fp16)[name = tensor<string, []>("op_443_cast_fp16")];
            tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_439_cast_fp16, y = var_443_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_451_cast_fp16 = softmax(axis = var_363, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_451_cast_fp16")];
            tensor<int32, [4]> var_452 = const()[name = tensor<string, []>("op_452"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_453_cast_fp16 = reshape(shape = var_452, x = value_5_cast_fp16)[name = tensor<string, []>("op_453_cast_fp16")];
            tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_453_cast_fp16, y = var_451_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<int32, [4]> var_456 = const()[name = tensor<string, []>("op_456"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_11_cast_fp16 = reshape(shape = var_456, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_21_strides_0 = const()[name = tensor<string, []>("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_21_dilations_0 = const()[name = tensor<string, []>("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_21_groups_0 = const()[name = tensor<string, []>("obj_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147030272)))];
            tensor<fp16, [1024]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149127488)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_478_to_fp16 = const()[name = tensor<string, []>("op_478_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_478_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<fp16, [1024]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149129600)))];
            tensor<fp16, [1024]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149131712)))];
            tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
            tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_7_strides_0 = const()[name = tensor<string, []>("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_7_dilations_0 = const()[name = tensor<string, []>("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_7_groups_0 = const()[name = tensor<string, []>("query_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149133824)))];
            tensor<fp16, [1024]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151231040)))];
            tensor<fp16, [1, 1024, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
            tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_7_dilations_0 = const()[name = tensor<string, []>("key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_7_groups_0 = const()[name = tensor<string, []>("key_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151233152)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_7_cast_fp16")];
            tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_7_strides_0 = const()[name = tensor<string, []>("value_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153330368)))];
            tensor<fp16, [1024]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155427584)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_7_cast_fp16")];
            tensor<int32, [4]> var_514 = const()[name = tensor<string, []>("op_514"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_514, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
            tensor<fp16, []> var_516_to_fp16 = const()[name = tensor<string, []>("op_516_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_517_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_516_to_fp16)[name = tensor<string, []>("op_517_cast_fp16")];
            tensor<int32, [4]> var_520 = const()[name = tensor<string, []>("op_520"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_521_cast_fp16 = reshape(shape = var_520, x = key_7_cast_fp16)[name = tensor<string, []>("op_521_cast_fp16")];
            tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_517_cast_fp16, y = var_521_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_363, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
            tensor<int32, [4]> var_525 = const()[name = tensor<string, []>("op_525"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_526_cast_fp16 = reshape(shape = var_525, x = value_7_cast_fp16)[name = tensor<string, []>("op_526_cast_fp16")];
            tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_526_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
            tensor<int32, [4]> var_529 = const()[name = tensor<string, []>("op_529"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_13_cast_fp16 = reshape(shape = var_529, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_25_strides_0 = const()[name = tensor<string, []>("obj_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_25_dilations_0 = const()[name = tensor<string, []>("obj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_25_groups_0 = const()[name = tensor<string, []>("obj_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155429696)))];
            tensor<fp16, [1024]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157526912)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_547_to_fp16 = const()[name = tensor<string, []>("op_547_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_547_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<fp16, [1024]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157529024)))];
            tensor<fp16, [1024]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157531136)))];
            tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157533248)))];
            tensor<fp16, [4096]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165921920)))];
            tensor<fp16, [1, 4096, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165930176)))];
            tensor<fp16, [1024]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174318848)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
            tensor<int32, []> var_582 = const()[name = tensor<string, []>("op_582"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_607_to_fp16 = const()[name = tensor<string, []>("op_607_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_607_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<fp16, [1024]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174320960)))];
            tensor<fp16, [1024]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174323072)))];
            tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
            tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174325184)))];
            tensor<fp16, [1024]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176422400)))];
            tensor<fp16, [1, 1024, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
            tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_5_strides_0 = const()[name = tensor<string, []>("current_key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_5_dilations_0 = const()[name = tensor<string, []>("current_key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_5_groups_0 = const()[name = tensor<string, []>("current_key_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176424512)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
            tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_5_strides_0 = const()[name = tensor<string, []>("current_value_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_5_dilations_0 = const()[name = tensor<string, []>("current_value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_5_groups_0 = const()[name = tensor<string, []>("current_value_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178521728)))];
            tensor<fp16, [1024]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180618944)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_646_cast_fp16 = mul(x = var_87_cast_fp16_2, y = var_207_cast_fp16)[name = tensor<string, []>("op_646_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_647_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_647_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_9_cast_fp16 = add(x = var_646_cast_fp16, y = var_647_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_650_cast_fp16 = mul(x = var_114_cast_fp16_2, y = var_207_cast_fp16)[name = tensor<string, []>("op_650_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_651_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_651_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_9_cast_fp16 = add(x = var_650_cast_fp16, y = var_651_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
            tensor<int32, [4]> var_655 = const()[name = tensor<string, []>("op_655"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_655, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
            tensor<fp16, []> var_657_to_fp16 = const()[name = tensor<string, []>("op_657_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_658_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_657_to_fp16)[name = tensor<string, []>("op_658_cast_fp16")];
            tensor<int32, [4]> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_662_cast_fp16 = reshape(shape = var_661, x = key_9_cast_fp16)[name = tensor<string, []>("op_662_cast_fp16")];
            tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_658_cast_fp16, y = var_662_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_670_cast_fp16 = softmax(axis = var_582, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_670_cast_fp16")];
            tensor<int32, [4]> var_671 = const()[name = tensor<string, []>("op_671"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_672_cast_fp16 = reshape(shape = var_671, x = value_9_cast_fp16)[name = tensor<string, []>("op_672_cast_fp16")];
            tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_672_cast_fp16, y = var_670_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = tensor<string, []>("op_675"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_21_cast_fp16 = reshape(shape = var_675, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_35_strides_0 = const()[name = tensor<string, []>("obj_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_35_dilations_0 = const()[name = tensor<string, []>("obj_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_35_groups_0 = const()[name = tensor<string, []>("obj_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180621056)))];
            tensor<fp16, [1024]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182718272)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_697_to_fp16 = const()[name = tensor<string, []>("op_697_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_697_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<fp16, [1024]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182720384)))];
            tensor<fp16, [1024]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182722496)))];
            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
            tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_11_strides_0 = const()[name = tensor<string, []>("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_11_dilations_0 = const()[name = tensor<string, []>("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_11_groups_0 = const()[name = tensor<string, []>("query_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182724608)))];
            tensor<fp16, [1024]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184821824)))];
            tensor<fp16, [1, 1024, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
            tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_11_dilations_0 = const()[name = tensor<string, []>("key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_11_groups_0 = const()[name = tensor<string, []>("key_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184823936)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")];
            tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_11_strides_0 = const()[name = tensor<string, []>("value_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186921152)))];
            tensor<fp16, [1024]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189018368)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")];
            tensor<int32, [4]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_733, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
            tensor<fp16, []> var_735_to_fp16 = const()[name = tensor<string, []>("op_735_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_736_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_735_to_fp16)[name = tensor<string, []>("op_736_cast_fp16")];
            tensor<int32, [4]> var_739 = const()[name = tensor<string, []>("op_739"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_740_cast_fp16 = reshape(shape = var_739, x = key_11_cast_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
            tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_736_cast_fp16, y = var_740_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_582, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
            tensor<int32, [4]> var_744 = const()[name = tensor<string, []>("op_744"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_745_cast_fp16 = reshape(shape = var_744, x = value_11_cast_fp16)[name = tensor<string, []>("op_745_cast_fp16")];
            tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_745_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
            tensor<int32, [4]> var_748 = const()[name = tensor<string, []>("op_748"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_23_cast_fp16 = reshape(shape = var_748, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_39_strides_0 = const()[name = tensor<string, []>("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_39_dilations_0 = const()[name = tensor<string, []>("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_39_groups_0 = const()[name = tensor<string, []>("obj_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189020480)))];
            tensor<fp16, [1024]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191117696)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_766_to_fp16 = const()[name = tensor<string, []>("op_766_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_766_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<fp16, [1024]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191119808)))];
            tensor<fp16, [1024]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191121920)))];
            tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_27_strides_0 = const()[name = tensor<string, []>("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_27_dilations_0 = const()[name = tensor<string, []>("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_27_groups_0 = const()[name = tensor<string, []>("input_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191124032)))];
            tensor<fp16, [4096]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199512704)))];
            tensor<fp16, [1, 4096, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199520960)))];
            tensor<fp16, [1024]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207909632)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
            tensor<int32, []> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_826_to_fp16 = const()[name = tensor<string, []>("op_826_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_826_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<fp16, [1024]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207911744)))];
            tensor<fp16, [1024]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207913856)))];
            tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
            tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207915968)))];
            tensor<fp16, [1024]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210013184)))];
            tensor<fp16, [1, 1024, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
            tensor<string, []> current_key_7_pad_type_0 = const()[name = tensor<string, []>("current_key_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_7_strides_0 = const()[name = tensor<string, []>("current_key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_7_pad_0 = const()[name = tensor<string, []>("current_key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_7_dilations_0 = const()[name = tensor<string, []>("current_key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_7_groups_0 = const()[name = tensor<string, []>("current_key_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210015296)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_key_7_cast_fp16")];
            tensor<string, []> current_value_7_pad_type_0 = const()[name = tensor<string, []>("current_value_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_7_strides_0 = const()[name = tensor<string, []>("current_value_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_7_pad_0 = const()[name = tensor<string, []>("current_value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_7_dilations_0 = const()[name = tensor<string, []>("current_value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_7_groups_0 = const()[name = tensor<string, []>("current_value_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212112512)))];
            tensor<fp16, [1024]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214209728)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_value_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_865_cast_fp16 = mul(x = var_87_cast_fp16_3, y = var_207_cast_fp16)[name = tensor<string, []>("op_865_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_866_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_866_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_13_cast_fp16 = add(x = var_865_cast_fp16, y = var_866_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_869_cast_fp16 = mul(x = var_114_cast_fp16_3, y = var_207_cast_fp16)[name = tensor<string, []>("op_869_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_870_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_870_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_13_cast_fp16 = add(x = var_869_cast_fp16, y = var_870_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
            tensor<int32, [4]> var_874 = const()[name = tensor<string, []>("op_874"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_874, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
            tensor<fp16, []> var_876_to_fp16 = const()[name = tensor<string, []>("op_876_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_877_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_876_to_fp16)[name = tensor<string, []>("op_877_cast_fp16")];
            tensor<int32, [4]> var_880 = const()[name = tensor<string, []>("op_880"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_881_cast_fp16 = reshape(shape = var_880, x = key_13_cast_fp16)[name = tensor<string, []>("op_881_cast_fp16")];
            tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_877_cast_fp16, y = var_881_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_889_cast_fp16 = softmax(axis = var_801, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_889_cast_fp16")];
            tensor<int32, [4]> var_890 = const()[name = tensor<string, []>("op_890"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_891_cast_fp16 = reshape(shape = var_890, x = value_13_cast_fp16)[name = tensor<string, []>("op_891_cast_fp16")];
            tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_891_cast_fp16, y = var_889_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<int32, [4]> var_894 = const()[name = tensor<string, []>("op_894"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_31_cast_fp16 = reshape(shape = var_894, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_49_strides_0 = const()[name = tensor<string, []>("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_49_dilations_0 = const()[name = tensor<string, []>("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_49_groups_0 = const()[name = tensor<string, []>("obj_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214211840)))];
            tensor<fp16, [1024]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216309056)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_916_to_fp16 = const()[name = tensor<string, []>("op_916_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_916_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<fp16, [1024]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216311168)))];
            tensor<fp16, [1024]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216313280)))];
            tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
            tensor<string, []> query_15_pad_type_0 = const()[name = tensor<string, []>("query_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_15_strides_0 = const()[name = tensor<string, []>("query_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_15_pad_0 = const()[name = tensor<string, []>("query_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_15_dilations_0 = const()[name = tensor<string, []>("query_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_15_groups_0 = const()[name = tensor<string, []>("query_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216315392)))];
            tensor<fp16, [1024]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218412608)))];
            tensor<fp16, [1, 1024, 1, 1]> query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
            tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_15_strides_0 = const()[name = tensor<string, []>("key_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_15_dilations_0 = const()[name = tensor<string, []>("key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_15_groups_0 = const()[name = tensor<string, []>("key_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218414720)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_15_cast_fp16")];
            tensor<string, []> value_15_pad_type_0 = const()[name = tensor<string, []>("value_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_15_strides_0 = const()[name = tensor<string, []>("value_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_15_pad_0 = const()[name = tensor<string, []>("value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_15_dilations_0 = const()[name = tensor<string, []>("value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_15_groups_0 = const()[name = tensor<string, []>("value_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220511936)))];
            tensor<fp16, [1024]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222609152)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_15_cast_fp16")];
            tensor<int32, [4]> var_952 = const()[name = tensor<string, []>("op_952"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_15_cast_fp16 = reshape(shape = var_952, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")];
            tensor<fp16, []> var_954_to_fp16 = const()[name = tensor<string, []>("op_954_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_955_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_954_to_fp16)[name = tensor<string, []>("op_955_cast_fp16")];
            tensor<int32, [4]> var_958 = const()[name = tensor<string, []>("op_958"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_959_cast_fp16 = reshape(shape = var_958, x = key_15_cast_fp16)[name = tensor<string, []>("op_959_cast_fp16")];
            tensor<bool, []> mh_w_23_transpose_x_0 = const()[name = tensor<string, []>("mh_w_23_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_23_transpose_y_0 = const()[name = tensor<string, []>("mh_w_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_955_cast_fp16, y = var_959_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_801, x = mh_w_23_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
            tensor<int32, [4]> var_963 = const()[name = tensor<string, []>("op_963"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_964_cast_fp16 = reshape(shape = var_963, x = value_15_cast_fp16)[name = tensor<string, []>("op_964_cast_fp16")];
            tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_964_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
            tensor<int32, [4]> var_967 = const()[name = tensor<string, []>("op_967"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_33_cast_fp16 = reshape(shape = var_967, x = attn_15_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_53_strides_0 = const()[name = tensor<string, []>("obj_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_53_dilations_0 = const()[name = tensor<string, []>("obj_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_53_groups_0 = const()[name = tensor<string, []>("obj_53_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222611264)))];
            tensor<fp16, [1024]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224708480)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_985_to_fp16 = const()[name = tensor<string, []>("op_985_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_985_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<fp16, [1024]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224710592)))];
            tensor<fp16, [1024]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224712704)))];
            tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224714816)))];
            tensor<fp16, [4096]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233103488)))];
            tensor<fp16, [1, 4096, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233111744)))];
            tensor<fp16, [1024]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241500416)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
            tensor<int32, []> var_1020 = const()[name = tensor<string, []>("op_1020"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_25_axes_0 = const()[name = tensor<string, []>("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1045_to_fp16 = const()[name = tensor<string, []>("op_1045_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1045_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<fp16, [1024]> obj_57_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_57_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241502528)))];
            tensor<fp16, [1024]> obj_57_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_57_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241504640)))];
            tensor<fp16, []> obj_57_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_57_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
            tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_17_strides_0 = const()[name = tensor<string, []>("query_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_17_dilations_0 = const()[name = tensor<string, []>("query_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_17_groups_0 = const()[name = tensor<string, []>("query_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241506752)))];
            tensor<fp16, [1024]> layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243603968)))];
            tensor<fp16, [1, 1024, 1, 1]> query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
            tensor<string, []> current_key_9_pad_type_0 = const()[name = tensor<string, []>("current_key_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_9_strides_0 = const()[name = tensor<string, []>("current_key_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_9_pad_0 = const()[name = tensor<string, []>("current_key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_9_dilations_0 = const()[name = tensor<string, []>("current_key_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_9_groups_0 = const()[name = tensor<string, []>("current_key_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243606080)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("current_key_9_cast_fp16")];
            tensor<string, []> current_value_9_pad_type_0 = const()[name = tensor<string, []>("current_value_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_9_strides_0 = const()[name = tensor<string, []>("current_value_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_9_pad_0 = const()[name = tensor<string, []>("current_value_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_9_dilations_0 = const()[name = tensor<string, []>("current_value_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_9_groups_0 = const()[name = tensor<string, []>("current_value_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245703296)))];
            tensor<fp16, [1024]> layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247800512)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor<string, []>("current_value_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1084_cast_fp16 = mul(x = var_87_cast_fp16_4, y = var_207_cast_fp16)[name = tensor<string, []>("op_1084_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1085_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1085_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_17_cast_fp16 = add(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1088_cast_fp16 = mul(x = var_114_cast_fp16_4, y = var_207_cast_fp16)[name = tensor<string, []>("op_1088_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1089_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1089_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_17_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
            tensor<int32, [4]> var_1093 = const()[name = tensor<string, []>("op_1093"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_17_cast_fp16 = reshape(shape = var_1093, x = query_17_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")];
            tensor<fp16, []> var_1095_to_fp16 = const()[name = tensor<string, []>("op_1095_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1096_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1095_to_fp16)[name = tensor<string, []>("op_1096_cast_fp16")];
            tensor<int32, [4]> var_1099 = const()[name = tensor<string, []>("op_1099"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1100_cast_fp16 = reshape(shape = var_1099, x = key_17_cast_fp16)[name = tensor<string, []>("op_1100_cast_fp16")];
            tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1096_cast_fp16, y = var_1100_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_1108_cast_fp16 = softmax(axis = var_1020, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1108_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1110_cast_fp16 = reshape(shape = var_1109, x = value_17_cast_fp16)[name = tensor<string, []>("op_1110_cast_fp16")];
            tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1110_cast_fp16, y = var_1108_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<int32, [4]> var_1113 = const()[name = tensor<string, []>("op_1113"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_41_cast_fp16 = reshape(shape = var_1113, x = attn_17_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<string, []> obj_63_pad_type_0 = const()[name = tensor<string, []>("obj_63_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_63_strides_0 = const()[name = tensor<string, []>("obj_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_63_pad_0 = const()[name = tensor<string, []>("obj_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_63_dilations_0 = const()[name = tensor<string, []>("obj_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_63_groups_0 = const()[name = tensor<string, []>("obj_63_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247802624)))];
            tensor<fp16, [1024]> layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249899840)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1135_to_fp16 = const()[name = tensor<string, []>("op_1135_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1135_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<fp16, [1024]> obj_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_65_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249901952)))];
            tensor<fp16, [1024]> obj_65_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_65_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249904064)))];
            tensor<fp16, []> obj_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")];
            tensor<string, []> query_19_pad_type_0 = const()[name = tensor<string, []>("query_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_19_strides_0 = const()[name = tensor<string, []>("query_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_19_pad_0 = const()[name = tensor<string, []>("query_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_19_dilations_0 = const()[name = tensor<string, []>("query_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_19_groups_0 = const()[name = tensor<string, []>("query_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249906176)))];
            tensor<fp16, [1024]> layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252003392)))];
            tensor<fp16, [1, 1024, 1, 1]> query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
            tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_19_strides_0 = const()[name = tensor<string, []>("key_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_19_dilations_0 = const()[name = tensor<string, []>("key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_19_groups_0 = const()[name = tensor<string, []>("key_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252005504)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_19_cast_fp16")];
            tensor<string, []> value_19_pad_type_0 = const()[name = tensor<string, []>("value_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_19_strides_0 = const()[name = tensor<string, []>("value_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_19_pad_0 = const()[name = tensor<string, []>("value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_19_dilations_0 = const()[name = tensor<string, []>("value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_19_groups_0 = const()[name = tensor<string, []>("value_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254102720)))];
            tensor<fp16, [1024]> layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256199936)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_19_cast_fp16")];
            tensor<int32, [4]> var_1171 = const()[name = tensor<string, []>("op_1171"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_19_cast_fp16 = reshape(shape = var_1171, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")];
            tensor<fp16, []> var_1173_to_fp16 = const()[name = tensor<string, []>("op_1173_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1174_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1173_to_fp16)[name = tensor<string, []>("op_1174_cast_fp16")];
            tensor<int32, [4]> var_1177 = const()[name = tensor<string, []>("op_1177"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1178_cast_fp16 = reshape(shape = var_1177, x = key_19_cast_fp16)[name = tensor<string, []>("op_1178_cast_fp16")];
            tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1174_cast_fp16, y = var_1178_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_69_cast_fp16 = softmax(axis = var_1020, x = mh_w_29_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
            tensor<int32, [4]> var_1182 = const()[name = tensor<string, []>("op_1182"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1183_cast_fp16 = reshape(shape = var_1182, x = value_19_cast_fp16)[name = tensor<string, []>("op_1183_cast_fp16")];
            tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1183_cast_fp16, y = obj_69_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
            tensor<int32, [4]> var_1186 = const()[name = tensor<string, []>("op_1186"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_43_cast_fp16 = reshape(shape = var_1186, x = attn_19_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<string, []> obj_67_pad_type_0 = const()[name = tensor<string, []>("obj_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_67_strides_0 = const()[name = tensor<string, []>("obj_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_67_pad_0 = const()[name = tensor<string, []>("obj_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_67_dilations_0 = const()[name = tensor<string, []>("obj_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_67_groups_0 = const()[name = tensor<string, []>("obj_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256202048)))];
            tensor<fp16, [1024]> layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258299264)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = tensor<string, []>("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1204_to_fp16 = const()[name = tensor<string, []>("op_1204_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1204_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<fp16, [1024]> input_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_45_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258301376)))];
            tensor<fp16, [1024]> input_45_beta_0_to_fp16 = const()[name = tensor<string, []>("input_45_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258303488)))];
            tensor<fp16, []> input_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<string, []> input_47_pad_type_0 = const()[name = tensor<string, []>("input_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = tensor<string, []>("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = tensor<string, []>("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = tensor<string, []>("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_47_groups_0 = const()[name = tensor<string, []>("input_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258305600)))];
            tensor<fp16, [4096]> layers_4_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(266694272)))];
            tensor<fp16, [1, 4096, 1, 1]> input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<string, []> input_49_mode_0 = const()[name = tensor<string, []>("input_49_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<string, []> hidden_states_11_pad_type_0 = const()[name = tensor<string, []>("hidden_states_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_11_strides_0 = const()[name = tensor<string, []>("hidden_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_11_pad_0 = const()[name = tensor<string, []>("hidden_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_11_dilations_0 = const()[name = tensor<string, []>("hidden_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_11_groups_0 = const()[name = tensor<string, []>("hidden_states_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_4_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(266702528)))];
            tensor<fp16, [1024]> layers_4_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275091200)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("hidden_states_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
            tensor<int32, []> var_1239 = const()[name = tensor<string, []>("op_1239"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_31_axes_0 = const()[name = tensor<string, []>("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1264_to_fp16 = const()[name = tensor<string, []>("op_1264_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1264_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<fp16, [1024]> obj_71_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_71_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275093312)))];
            tensor<fp16, [1024]> obj_71_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_71_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275095424)))];
            tensor<fp16, []> obj_71_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_71_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
            tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_21_strides_0 = const()[name = tensor<string, []>("query_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_21_dilations_0 = const()[name = tensor<string, []>("query_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_21_groups_0 = const()[name = tensor<string, []>("query_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275097536)))];
            tensor<fp16, [1024]> layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277194752)))];
            tensor<fp16, [1, 1024, 1, 1]> query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
            tensor<string, []> current_key_11_pad_type_0 = const()[name = tensor<string, []>("current_key_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_11_strides_0 = const()[name = tensor<string, []>("current_key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_11_pad_0 = const()[name = tensor<string, []>("current_key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_11_dilations_0 = const()[name = tensor<string, []>("current_key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_11_groups_0 = const()[name = tensor<string, []>("current_key_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277196864)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("current_key_11_cast_fp16")];
            tensor<string, []> current_value_11_pad_type_0 = const()[name = tensor<string, []>("current_value_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_11_strides_0 = const()[name = tensor<string, []>("current_value_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_11_pad_0 = const()[name = tensor<string, []>("current_value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_11_dilations_0 = const()[name = tensor<string, []>("current_value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_11_groups_0 = const()[name = tensor<string, []>("current_value_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279294080)))];
            tensor<fp16, [1024]> layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281391296)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("current_value_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1303_cast_fp16 = mul(x = var_87_cast_fp16_5, y = var_207_cast_fp16)[name = tensor<string, []>("op_1303_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1304_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_21_cast_fp16 = add(x = var_1303_cast_fp16, y = var_1304_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1307_cast_fp16 = mul(x = var_114_cast_fp16_5, y = var_207_cast_fp16)[name = tensor<string, []>("op_1307_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1308_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1308_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_21_cast_fp16 = add(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
            tensor<int32, [4]> var_1312 = const()[name = tensor<string, []>("op_1312"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_21_cast_fp16 = reshape(shape = var_1312, x = query_21_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")];
            tensor<fp16, []> var_1314_to_fp16 = const()[name = tensor<string, []>("op_1314_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1315_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1314_to_fp16)[name = tensor<string, []>("op_1315_cast_fp16")];
            tensor<int32, [4]> var_1318 = const()[name = tensor<string, []>("op_1318"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1319_cast_fp16 = reshape(shape = var_1318, x = key_21_cast_fp16)[name = tensor<string, []>("op_1319_cast_fp16")];
            tensor<bool, []> mh_w_31_transpose_x_0 = const()[name = tensor<string, []>("mh_w_31_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_31_transpose_y_0 = const()[name = tensor<string, []>("mh_w_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1315_cast_fp16, y = var_1319_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_1327_cast_fp16 = softmax(axis = var_1239, x = mh_w_33_cast_fp16)[name = tensor<string, []>("op_1327_cast_fp16")];
            tensor<int32, [4]> var_1328 = const()[name = tensor<string, []>("op_1328"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1329_cast_fp16 = reshape(shape = var_1328, x = value_21_cast_fp16)[name = tensor<string, []>("op_1329_cast_fp16")];
            tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1329_cast_fp16, y = var_1327_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<int32, [4]> var_1332 = const()[name = tensor<string, []>("op_1332"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_51_cast_fp16 = reshape(shape = var_1332, x = attn_21_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<string, []> obj_77_pad_type_0 = const()[name = tensor<string, []>("obj_77_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_77_strides_0 = const()[name = tensor<string, []>("obj_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_77_pad_0 = const()[name = tensor<string, []>("obj_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_77_dilations_0 = const()[name = tensor<string, []>("obj_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_77_groups_0 = const()[name = tensor<string, []>("obj_77_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281393408)))];
            tensor<fp16, [1024]> layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283490624)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1354_to_fp16 = const()[name = tensor<string, []>("op_1354_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1354_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<fp16, [1024]> obj_79_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_79_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283492736)))];
            tensor<fp16, [1024]> obj_79_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_79_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283494848)))];
            tensor<fp16, []> obj_79_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_79_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
            tensor<string, []> query_23_pad_type_0 = const()[name = tensor<string, []>("query_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_23_strides_0 = const()[name = tensor<string, []>("query_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_23_pad_0 = const()[name = tensor<string, []>("query_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_23_dilations_0 = const()[name = tensor<string, []>("query_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_23_groups_0 = const()[name = tensor<string, []>("query_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283496960)))];
            tensor<fp16, [1024]> layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285594176)))];
            tensor<fp16, [1, 1024, 1, 1]> query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
            tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_23_strides_0 = const()[name = tensor<string, []>("key_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_23_dilations_0 = const()[name = tensor<string, []>("key_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_23_groups_0 = const()[name = tensor<string, []>("key_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285596288)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_23_cast_fp16")];
            tensor<string, []> value_23_pad_type_0 = const()[name = tensor<string, []>("value_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_23_strides_0 = const()[name = tensor<string, []>("value_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_23_pad_0 = const()[name = tensor<string, []>("value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_23_dilations_0 = const()[name = tensor<string, []>("value_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_23_groups_0 = const()[name = tensor<string, []>("value_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287693504)))];
            tensor<fp16, [1024]> layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289790720)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_23_cast_fp16")];
            tensor<int32, [4]> var_1390 = const()[name = tensor<string, []>("op_1390"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_23_cast_fp16 = reshape(shape = var_1390, x = query_23_cast_fp16)[name = tensor<string, []>("mh_q_23_cast_fp16")];
            tensor<fp16, []> var_1392_to_fp16 = const()[name = tensor<string, []>("op_1392_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1393_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1392_to_fp16)[name = tensor<string, []>("op_1393_cast_fp16")];
            tensor<int32, [4]> var_1396 = const()[name = tensor<string, []>("op_1396"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1397_cast_fp16 = reshape(shape = var_1396, x = key_23_cast_fp16)[name = tensor<string, []>("op_1397_cast_fp16")];
            tensor<bool, []> mh_w_35_transpose_x_0 = const()[name = tensor<string, []>("mh_w_35_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_35_transpose_y_0 = const()[name = tensor<string, []>("mh_w_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1393_cast_fp16, y = var_1397_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_83_cast_fp16 = softmax(axis = var_1239, x = mh_w_35_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
            tensor<int32, [4]> var_1401 = const()[name = tensor<string, []>("op_1401"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1402_cast_fp16 = reshape(shape = var_1401, x = value_23_cast_fp16)[name = tensor<string, []>("op_1402_cast_fp16")];
            tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1402_cast_fp16, y = obj_83_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
            tensor<int32, [4]> var_1405 = const()[name = tensor<string, []>("op_1405"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_53_cast_fp16 = reshape(shape = var_1405, x = attn_23_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<string, []> obj_81_pad_type_0 = const()[name = tensor<string, []>("obj_81_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_81_strides_0 = const()[name = tensor<string, []>("obj_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_81_pad_0 = const()[name = tensor<string, []>("obj_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_81_dilations_0 = const()[name = tensor<string, []>("obj_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_81_groups_0 = const()[name = tensor<string, []>("obj_81_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289792832)))];
            tensor<fp16, [1024]> layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291890048)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = tensor<string, []>("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1423_to_fp16 = const()[name = tensor<string, []>("op_1423_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1423_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<fp16, [1024]> input_55_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_55_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291892160)))];
            tensor<fp16, [1024]> input_55_beta_0_to_fp16 = const()[name = tensor<string, []>("input_55_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291894272)))];
            tensor<fp16, []> input_55_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_55_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<string, []> input_57_pad_type_0 = const()[name = tensor<string, []>("input_57_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_57_strides_0 = const()[name = tensor<string, []>("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_57_pad_0 = const()[name = tensor<string, []>("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_57_dilations_0 = const()[name = tensor<string, []>("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_57_groups_0 = const()[name = tensor<string, []>("input_57_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291896384)))];
            tensor<fp16, [4096]> layers_5_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300285056)))];
            tensor<fp16, [1, 4096, 1, 1]> input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<string, []> input_59_mode_0 = const()[name = tensor<string, []>("input_59_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> hidden_states_13_pad_type_0 = const()[name = tensor<string, []>("hidden_states_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = tensor<string, []>("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = tensor<string, []>("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = tensor<string, []>("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_13_groups_0 = const()[name = tensor<string, []>("hidden_states_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_5_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300293312)))];
            tensor<fp16, [1024]> layers_5_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308681984)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
            tensor<int32, []> var_1458 = const()[name = tensor<string, []>("op_1458"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_37_axes_0 = const()[name = tensor<string, []>("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1483_to_fp16 = const()[name = tensor<string, []>("op_1483_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1483_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<fp16, [1024]> obj_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_85_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308684096)))];
            tensor<fp16, [1024]> obj_85_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_85_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308686208)))];
            tensor<fp16, []> obj_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
            tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_25_strides_0 = const()[name = tensor<string, []>("query_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_25_dilations_0 = const()[name = tensor<string, []>("query_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_25_groups_0 = const()[name = tensor<string, []>("query_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308688320)))];
            tensor<fp16, [1024]> layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310785536)))];
            tensor<fp16, [1, 1024, 1, 1]> query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
            tensor<string, []> current_key_13_pad_type_0 = const()[name = tensor<string, []>("current_key_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_13_strides_0 = const()[name = tensor<string, []>("current_key_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_13_pad_0 = const()[name = tensor<string, []>("current_key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_13_dilations_0 = const()[name = tensor<string, []>("current_key_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_13_groups_0 = const()[name = tensor<string, []>("current_key_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310787648)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("current_key_13_cast_fp16")];
            tensor<string, []> current_value_13_pad_type_0 = const()[name = tensor<string, []>("current_value_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_13_strides_0 = const()[name = tensor<string, []>("current_value_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_13_pad_0 = const()[name = tensor<string, []>("current_value_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_13_dilations_0 = const()[name = tensor<string, []>("current_value_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_13_groups_0 = const()[name = tensor<string, []>("current_value_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312884864)))];
            tensor<fp16, [1024]> layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314982080)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor<string, []>("current_value_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1522_cast_fp16 = mul(x = var_87_cast_fp16_6, y = var_207_cast_fp16)[name = tensor<string, []>("op_1522_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1523_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1523_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_25_cast_fp16 = add(x = var_1522_cast_fp16, y = var_1523_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1526_cast_fp16 = mul(x = var_114_cast_fp16_6, y = var_207_cast_fp16)[name = tensor<string, []>("op_1526_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1527_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1527_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_25_cast_fp16 = add(x = var_1526_cast_fp16, y = var_1527_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
            tensor<int32, [4]> var_1531 = const()[name = tensor<string, []>("op_1531"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_25_cast_fp16 = reshape(shape = var_1531, x = query_25_cast_fp16)[name = tensor<string, []>("mh_q_25_cast_fp16")];
            tensor<fp16, []> var_1533_to_fp16 = const()[name = tensor<string, []>("op_1533_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1534_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1533_to_fp16)[name = tensor<string, []>("op_1534_cast_fp16")];
            tensor<int32, [4]> var_1537 = const()[name = tensor<string, []>("op_1537"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1538_cast_fp16 = reshape(shape = var_1537, x = key_25_cast_fp16)[name = tensor<string, []>("op_1538_cast_fp16")];
            tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1534_cast_fp16, y = var_1538_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_1546_cast_fp16 = softmax(axis = var_1458, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_1546_cast_fp16")];
            tensor<int32, [4]> var_1547 = const()[name = tensor<string, []>("op_1547"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1548_cast_fp16 = reshape(shape = var_1547, x = value_25_cast_fp16)[name = tensor<string, []>("op_1548_cast_fp16")];
            tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1548_cast_fp16, y = var_1546_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<int32, [4]> var_1551 = const()[name = tensor<string, []>("op_1551"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_61_cast_fp16 = reshape(shape = var_1551, x = attn_25_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<string, []> obj_91_pad_type_0 = const()[name = tensor<string, []>("obj_91_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_91_strides_0 = const()[name = tensor<string, []>("obj_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_91_pad_0 = const()[name = tensor<string, []>("obj_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_91_dilations_0 = const()[name = tensor<string, []>("obj_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_91_groups_0 = const()[name = tensor<string, []>("obj_91_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314984192)))];
            tensor<fp16, [1024]> layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317081408)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = tensor<string, []>("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1573_to_fp16 = const()[name = tensor<string, []>("op_1573_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1573_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<fp16, [1024]> obj_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_93_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317083520)))];
            tensor<fp16, [1024]> obj_93_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_93_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317085632)))];
            tensor<fp16, []> obj_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
            tensor<string, []> query_27_pad_type_0 = const()[name = tensor<string, []>("query_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_27_strides_0 = const()[name = tensor<string, []>("query_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_27_pad_0 = const()[name = tensor<string, []>("query_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_27_dilations_0 = const()[name = tensor<string, []>("query_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_27_groups_0 = const()[name = tensor<string, []>("query_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317087744)))];
            tensor<fp16, [1024]> layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319184960)))];
            tensor<fp16, [1, 1024, 1, 1]> query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
            tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_27_strides_0 = const()[name = tensor<string, []>("key_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_27_dilations_0 = const()[name = tensor<string, []>("key_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_27_groups_0 = const()[name = tensor<string, []>("key_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319187072)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_27_cast_fp16")];
            tensor<string, []> value_27_pad_type_0 = const()[name = tensor<string, []>("value_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_27_strides_0 = const()[name = tensor<string, []>("value_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_27_pad_0 = const()[name = tensor<string, []>("value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_27_dilations_0 = const()[name = tensor<string, []>("value_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_27_groups_0 = const()[name = tensor<string, []>("value_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321284288)))];
            tensor<fp16, [1024]> layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323381504)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_27_cast_fp16")];
            tensor<int32, [4]> var_1609 = const()[name = tensor<string, []>("op_1609"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_27_cast_fp16 = reshape(shape = var_1609, x = query_27_cast_fp16)[name = tensor<string, []>("mh_q_27_cast_fp16")];
            tensor<fp16, []> var_1611_to_fp16 = const()[name = tensor<string, []>("op_1611_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1612_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1611_to_fp16)[name = tensor<string, []>("op_1612_cast_fp16")];
            tensor<int32, [4]> var_1615 = const()[name = tensor<string, []>("op_1615"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1616_cast_fp16 = reshape(shape = var_1615, x = key_27_cast_fp16)[name = tensor<string, []>("op_1616_cast_fp16")];
            tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1612_cast_fp16, y = var_1616_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_97_cast_fp16 = softmax(axis = var_1458, x = mh_w_41_cast_fp16)[name = tensor<string, []>("obj_97_cast_fp16")];
            tensor<int32, [4]> var_1620 = const()[name = tensor<string, []>("op_1620"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1621_cast_fp16 = reshape(shape = var_1620, x = value_27_cast_fp16)[name = tensor<string, []>("op_1621_cast_fp16")];
            tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1621_cast_fp16, y = obj_97_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = tensor<string, []>("op_1624"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_63_cast_fp16 = reshape(shape = var_1624, x = attn_27_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<string, []> obj_95_pad_type_0 = const()[name = tensor<string, []>("obj_95_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_95_strides_0 = const()[name = tensor<string, []>("obj_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_95_pad_0 = const()[name = tensor<string, []>("obj_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_95_dilations_0 = const()[name = tensor<string, []>("obj_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_95_groups_0 = const()[name = tensor<string, []>("obj_95_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323383616)))];
            tensor<fp16, [1024]> layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325480832)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = tensor<string, []>("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1642_to_fp16 = const()[name = tensor<string, []>("op_1642_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1642_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<fp16, [1024]> input_65_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_65_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325482944)))];
            tensor<fp16, [1024]> input_65_beta_0_to_fp16 = const()[name = tensor<string, []>("input_65_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325485056)))];
            tensor<fp16, []> input_65_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_65_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<string, []> input_67_pad_type_0 = const()[name = tensor<string, []>("input_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = tensor<string, []>("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = tensor<string, []>("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = tensor<string, []>("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_67_groups_0 = const()[name = tensor<string, []>("input_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325487168)))];
            tensor<fp16, [4096]> layers_6_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333875840)))];
            tensor<fp16, [1, 4096, 1, 1]> input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> input_69_mode_0 = const()[name = tensor<string, []>("input_69_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<string, []> hidden_states_15_pad_type_0 = const()[name = tensor<string, []>("hidden_states_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = tensor<string, []>("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = tensor<string, []>("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = tensor<string, []>("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_15_groups_0 = const()[name = tensor<string, []>("hidden_states_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_6_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333884096)))];
            tensor<fp16, [1024]> layers_6_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342272768)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("hidden_states_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
            tensor<int32, []> var_1677 = const()[name = tensor<string, []>("op_1677"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_43_axes_0 = const()[name = tensor<string, []>("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1702_to_fp16 = const()[name = tensor<string, []>("op_1702_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1702_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<fp16, [1024]> obj_99_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_99_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342274880)))];
            tensor<fp16, [1024]> obj_99_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_99_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342276992)))];
            tensor<fp16, []> obj_99_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_99_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("obj_99_cast_fp16")];
            tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_29_strides_0 = const()[name = tensor<string, []>("query_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_29_dilations_0 = const()[name = tensor<string, []>("query_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_29_groups_0 = const()[name = tensor<string, []>("query_29_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342279104)))];
            tensor<fp16, [1024]> layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344376320)))];
            tensor<fp16, [1, 1024, 1, 1]> query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
            tensor<string, []> current_key_15_pad_type_0 = const()[name = tensor<string, []>("current_key_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_15_strides_0 = const()[name = tensor<string, []>("current_key_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_15_pad_0 = const()[name = tensor<string, []>("current_key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_15_dilations_0 = const()[name = tensor<string, []>("current_key_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_15_groups_0 = const()[name = tensor<string, []>("current_key_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344378432)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("current_key_15_cast_fp16")];
            tensor<string, []> current_value_15_pad_type_0 = const()[name = tensor<string, []>("current_value_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_15_strides_0 = const()[name = tensor<string, []>("current_value_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_15_pad_0 = const()[name = tensor<string, []>("current_value_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_15_dilations_0 = const()[name = tensor<string, []>("current_value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_15_groups_0 = const()[name = tensor<string, []>("current_value_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346475648)))];
            tensor<fp16, [1024]> layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348572864)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor<string, []>("current_value_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1741_cast_fp16 = mul(x = var_87_cast_fp16_7, y = var_207_cast_fp16)[name = tensor<string, []>("op_1741_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1742_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1742_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_29_cast_fp16 = add(x = var_1741_cast_fp16, y = var_1742_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1745_cast_fp16 = mul(x = var_114_cast_fp16_7, y = var_207_cast_fp16)[name = tensor<string, []>("op_1745_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1746_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1746_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_29_cast_fp16 = add(x = var_1745_cast_fp16, y = var_1746_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
            tensor<int32, [4]> var_1750 = const()[name = tensor<string, []>("op_1750"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_29_cast_fp16 = reshape(shape = var_1750, x = query_29_cast_fp16)[name = tensor<string, []>("mh_q_29_cast_fp16")];
            tensor<fp16, []> var_1752_to_fp16 = const()[name = tensor<string, []>("op_1752_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1753_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1752_to_fp16)[name = tensor<string, []>("op_1753_cast_fp16")];
            tensor<int32, [4]> var_1756 = const()[name = tensor<string, []>("op_1756"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1757_cast_fp16 = reshape(shape = var_1756, x = key_29_cast_fp16)[name = tensor<string, []>("op_1757_cast_fp16")];
            tensor<bool, []> mh_w_43_transpose_x_0 = const()[name = tensor<string, []>("mh_w_43_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_43_transpose_y_0 = const()[name = tensor<string, []>("mh_w_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1753_cast_fp16, y = var_1757_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_1765_cast_fp16 = softmax(axis = var_1677, x = mh_w_45_cast_fp16)[name = tensor<string, []>("op_1765_cast_fp16")];
            tensor<int32, [4]> var_1766 = const()[name = tensor<string, []>("op_1766"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1767_cast_fp16 = reshape(shape = var_1766, x = value_29_cast_fp16)[name = tensor<string, []>("op_1767_cast_fp16")];
            tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1767_cast_fp16, y = var_1765_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<int32, [4]> var_1770 = const()[name = tensor<string, []>("op_1770"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_71_cast_fp16 = reshape(shape = var_1770, x = attn_29_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<string, []> obj_105_pad_type_0 = const()[name = tensor<string, []>("obj_105_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_105_strides_0 = const()[name = tensor<string, []>("obj_105_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_105_pad_0 = const()[name = tensor<string, []>("obj_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_105_dilations_0 = const()[name = tensor<string, []>("obj_105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_105_groups_0 = const()[name = tensor<string, []>("obj_105_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348574976)))];
            tensor<fp16, [1024]> layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350672192)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("obj_105_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = tensor<string, []>("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1792_to_fp16 = const()[name = tensor<string, []>("op_1792_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1792_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<fp16, [1024]> obj_107_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_107_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350674304)))];
            tensor<fp16, [1024]> obj_107_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_107_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350676416)))];
            tensor<fp16, []> obj_107_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_107_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("obj_107_cast_fp16")];
            tensor<string, []> query_31_pad_type_0 = const()[name = tensor<string, []>("query_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_31_strides_0 = const()[name = tensor<string, []>("query_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_31_pad_0 = const()[name = tensor<string, []>("query_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_31_dilations_0 = const()[name = tensor<string, []>("query_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_31_groups_0 = const()[name = tensor<string, []>("query_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350678528)))];
            tensor<fp16, [1024]> layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352775744)))];
            tensor<fp16, [1, 1024, 1, 1]> query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
            tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_31_strides_0 = const()[name = tensor<string, []>("key_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_31_dilations_0 = const()[name = tensor<string, []>("key_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_31_groups_0 = const()[name = tensor<string, []>("key_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352777856)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_31_cast_fp16")];
            tensor<string, []> value_31_pad_type_0 = const()[name = tensor<string, []>("value_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_31_strides_0 = const()[name = tensor<string, []>("value_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_31_pad_0 = const()[name = tensor<string, []>("value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_31_dilations_0 = const()[name = tensor<string, []>("value_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_31_groups_0 = const()[name = tensor<string, []>("value_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354875072)))];
            tensor<fp16, [1024]> layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356972288)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_31_cast_fp16")];
            tensor<int32, [4]> var_1828 = const()[name = tensor<string, []>("op_1828"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_31_cast_fp16 = reshape(shape = var_1828, x = query_31_cast_fp16)[name = tensor<string, []>("mh_q_31_cast_fp16")];
            tensor<fp16, []> var_1830_to_fp16 = const()[name = tensor<string, []>("op_1830_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1831_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1830_to_fp16)[name = tensor<string, []>("op_1831_cast_fp16")];
            tensor<int32, [4]> var_1834 = const()[name = tensor<string, []>("op_1834"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1835_cast_fp16 = reshape(shape = var_1834, x = key_31_cast_fp16)[name = tensor<string, []>("op_1835_cast_fp16")];
            tensor<bool, []> mh_w_47_transpose_x_0 = const()[name = tensor<string, []>("mh_w_47_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_47_transpose_y_0 = const()[name = tensor<string, []>("mh_w_47_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1831_cast_fp16, y = var_1835_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_111_cast_fp16 = softmax(axis = var_1677, x = mh_w_47_cast_fp16)[name = tensor<string, []>("obj_111_cast_fp16")];
            tensor<int32, [4]> var_1839 = const()[name = tensor<string, []>("op_1839"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_1840_cast_fp16 = reshape(shape = var_1839, x = value_31_cast_fp16)[name = tensor<string, []>("op_1840_cast_fp16")];
            tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1840_cast_fp16, y = obj_111_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
            tensor<int32, [4]> var_1843 = const()[name = tensor<string, []>("op_1843"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_73_cast_fp16 = reshape(shape = var_1843, x = attn_31_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<string, []> obj_109_pad_type_0 = const()[name = tensor<string, []>("obj_109_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_109_strides_0 = const()[name = tensor<string, []>("obj_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_109_pad_0 = const()[name = tensor<string, []>("obj_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_109_dilations_0 = const()[name = tensor<string, []>("obj_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_109_groups_0 = const()[name = tensor<string, []>("obj_109_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356974400)))];
            tensor<fp16, [1024]> layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359071616)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("obj_109_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = tensor<string, []>("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1861_to_fp16 = const()[name = tensor<string, []>("op_1861_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1861_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
            tensor<fp16, [1024]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359073728)))];
            tensor<fp16, [1024]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359075840)))];
            tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_77_strides_0 = const()[name = tensor<string, []>("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_77_dilations_0 = const()[name = tensor<string, []>("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_77_groups_0 = const()[name = tensor<string, []>("input_77_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359077952)))];
            tensor<fp16, [4096]> layers_7_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367466624)))];
            tensor<fp16, [1, 4096, 1, 1]> input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> hidden_states_17_pad_type_0 = const()[name = tensor<string, []>("hidden_states_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = tensor<string, []>("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = tensor<string, []>("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = tensor<string, []>("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_17_groups_0 = const()[name = tensor<string, []>("hidden_states_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_7_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367474880)))];
            tensor<fp16, [1024]> layers_7_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375863552)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
            tensor<int32, []> var_1896 = const()[name = tensor<string, []>("op_1896"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_49_axes_0 = const()[name = tensor<string, []>("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_1921_to_fp16 = const()[name = tensor<string, []>("op_1921_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1921_to_fp16, x = inputs_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
            tensor<fp16, [1024]> obj_113_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_113_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375865664)))];
            tensor<fp16, [1024]> obj_113_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_113_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375867776)))];
            tensor<fp16, []> obj_113_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_113_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("obj_113_cast_fp16")];
            tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_33_strides_0 = const()[name = tensor<string, []>("query_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_33_dilations_0 = const()[name = tensor<string, []>("query_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_33_groups_0 = const()[name = tensor<string, []>("query_33_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375869888)))];
            tensor<fp16, [1024]> layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377967104)))];
            tensor<fp16, [1, 1024, 1, 1]> query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
            tensor<string, []> current_key_17_pad_type_0 = const()[name = tensor<string, []>("current_key_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_17_strides_0 = const()[name = tensor<string, []>("current_key_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_17_pad_0 = const()[name = tensor<string, []>("current_key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_17_dilations_0 = const()[name = tensor<string, []>("current_key_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_17_groups_0 = const()[name = tensor<string, []>("current_key_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377969216)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("current_key_17_cast_fp16")];
            tensor<string, []> current_value_17_pad_type_0 = const()[name = tensor<string, []>("current_value_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_17_strides_0 = const()[name = tensor<string, []>("current_value_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_17_pad_0 = const()[name = tensor<string, []>("current_value_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_17_dilations_0 = const()[name = tensor<string, []>("current_value_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_17_groups_0 = const()[name = tensor<string, []>("current_value_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380066432)))];
            tensor<fp16, [1024]> layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382163648)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor<string, []>("current_value_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1960_cast_fp16 = mul(x = var_87_cast_fp16_8, y = var_207_cast_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1961_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1961_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_33_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1961_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1964_cast_fp16 = mul(x = var_114_cast_fp16_8, y = var_207_cast_fp16)[name = tensor<string, []>("op_1964_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_1965_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_1965_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_33_cast_fp16 = add(x = var_1964_cast_fp16, y = var_1965_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
            tensor<int32, [4]> var_1969 = const()[name = tensor<string, []>("op_1969"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_33_cast_fp16 = reshape(shape = var_1969, x = query_33_cast_fp16)[name = tensor<string, []>("mh_q_33_cast_fp16")];
            tensor<fp16, []> var_1971_to_fp16 = const()[name = tensor<string, []>("op_1971_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_1972_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1971_to_fp16)[name = tensor<string, []>("op_1972_cast_fp16")];
            tensor<int32, [4]> var_1975 = const()[name = tensor<string, []>("op_1975"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1976_cast_fp16 = reshape(shape = var_1975, x = key_33_cast_fp16)[name = tensor<string, []>("op_1976_cast_fp16")];
            tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1972_cast_fp16, y = var_1976_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_1984_cast_fp16 = softmax(axis = var_1896, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_1984_cast_fp16")];
            tensor<int32, [4]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_1986_cast_fp16 = reshape(shape = var_1985, x = value_33_cast_fp16)[name = tensor<string, []>("op_1986_cast_fp16")];
            tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1986_cast_fp16, y = var_1984_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<int32, [4]> var_1989 = const()[name = tensor<string, []>("op_1989"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_81_cast_fp16 = reshape(shape = var_1989, x = attn_33_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<string, []> obj_119_pad_type_0 = const()[name = tensor<string, []>("obj_119_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_119_strides_0 = const()[name = tensor<string, []>("obj_119_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_119_pad_0 = const()[name = tensor<string, []>("obj_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_119_dilations_0 = const()[name = tensor<string, []>("obj_119_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_119_groups_0 = const()[name = tensor<string, []>("obj_119_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382165760)))];
            tensor<fp16, [1024]> layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384262976)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("obj_119_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = tensor<string, []>("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2011_to_fp16 = const()[name = tensor<string, []>("op_2011_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_2011_to_fp16, x = inputs_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
            tensor<fp16, [1024]> obj_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_121_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384265088)))];
            tensor<fp16, [1024]> obj_121_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_121_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384267200)))];
            tensor<fp16, []> obj_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("obj_121_cast_fp16")];
            tensor<string, []> query_35_pad_type_0 = const()[name = tensor<string, []>("query_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_35_strides_0 = const()[name = tensor<string, []>("query_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_35_pad_0 = const()[name = tensor<string, []>("query_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_35_dilations_0 = const()[name = tensor<string, []>("query_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_35_groups_0 = const()[name = tensor<string, []>("query_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384269312)))];
            tensor<fp16, [1024]> layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386366528)))];
            tensor<fp16, [1, 1024, 1, 1]> query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
            tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_35_strides_0 = const()[name = tensor<string, []>("key_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_35_dilations_0 = const()[name = tensor<string, []>("key_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_35_groups_0 = const()[name = tensor<string, []>("key_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386368640)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_35_cast_fp16")];
            tensor<string, []> value_35_pad_type_0 = const()[name = tensor<string, []>("value_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_35_strides_0 = const()[name = tensor<string, []>("value_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_35_pad_0 = const()[name = tensor<string, []>("value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_35_dilations_0 = const()[name = tensor<string, []>("value_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_35_groups_0 = const()[name = tensor<string, []>("value_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388465856)))];
            tensor<fp16, [1024]> layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390563072)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_35_cast_fp16")];
            tensor<int32, [4]> var_2047 = const()[name = tensor<string, []>("op_2047"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_35_cast_fp16 = reshape(shape = var_2047, x = query_35_cast_fp16)[name = tensor<string, []>("mh_q_35_cast_fp16")];
            tensor<fp16, []> var_2049_to_fp16 = const()[name = tensor<string, []>("op_2049_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2050_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2049_to_fp16)[name = tensor<string, []>("op_2050_cast_fp16")];
            tensor<int32, [4]> var_2053 = const()[name = tensor<string, []>("op_2053"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2054_cast_fp16 = reshape(shape = var_2053, x = key_35_cast_fp16)[name = tensor<string, []>("op_2054_cast_fp16")];
            tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_2050_cast_fp16, y = var_2054_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_125_cast_fp16 = softmax(axis = var_1896, x = mh_w_53_cast_fp16)[name = tensor<string, []>("obj_125_cast_fp16")];
            tensor<int32, [4]> var_2058 = const()[name = tensor<string, []>("op_2058"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2059_cast_fp16 = reshape(shape = var_2058, x = value_35_cast_fp16)[name = tensor<string, []>("op_2059_cast_fp16")];
            tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2059_cast_fp16, y = obj_125_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
            tensor<int32, [4]> var_2062 = const()[name = tensor<string, []>("op_2062"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_83_cast_fp16 = reshape(shape = var_2062, x = attn_35_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<string, []> obj_123_pad_type_0 = const()[name = tensor<string, []>("obj_123_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_123_strides_0 = const()[name = tensor<string, []>("obj_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_123_pad_0 = const()[name = tensor<string, []>("obj_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_123_dilations_0 = const()[name = tensor<string, []>("obj_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_123_groups_0 = const()[name = tensor<string, []>("obj_123_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390565184)))];
            tensor<fp16, [1024]> layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392662400)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("obj_123_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = tensor<string, []>("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2080_to_fp16 = const()[name = tensor<string, []>("op_2080_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_2080_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
            tensor<fp16, [1024]> input_85_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_85_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392664512)))];
            tensor<fp16, [1024]> input_85_beta_0_to_fp16 = const()[name = tensor<string, []>("input_85_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392666624)))];
            tensor<fp16, []> input_85_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_85_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<string, []> input_87_pad_type_0 = const()[name = tensor<string, []>("input_87_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_87_strides_0 = const()[name = tensor<string, []>("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_87_pad_0 = const()[name = tensor<string, []>("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_87_dilations_0 = const()[name = tensor<string, []>("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_87_groups_0 = const()[name = tensor<string, []>("input_87_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392668736)))];
            tensor<fp16, [4096]> layers_8_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401057408)))];
            tensor<fp16, [1, 4096, 1, 1]> input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<string, []> input_89_mode_0 = const()[name = tensor<string, []>("input_89_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<string, []> hidden_states_19_pad_type_0 = const()[name = tensor<string, []>("hidden_states_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_19_strides_0 = const()[name = tensor<string, []>("hidden_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = tensor<string, []>("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_19_dilations_0 = const()[name = tensor<string, []>("hidden_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_19_groups_0 = const()[name = tensor<string, []>("hidden_states_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_8_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401065664)))];
            tensor<fp16, [1024]> layers_8_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409454336)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")];
            tensor<int32, []> var_2115 = const()[name = tensor<string, []>("op_2115"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_55_axes_0 = const()[name = tensor<string, []>("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2140_to_fp16 = const()[name = tensor<string, []>("op_2140_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2140_to_fp16, x = inputs_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
            tensor<fp16, [1024]> obj_127_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_127_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409456448)))];
            tensor<fp16, [1024]> obj_127_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_127_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409458560)))];
            tensor<fp16, []> obj_127_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_127_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("obj_127_cast_fp16")];
            tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_37_strides_0 = const()[name = tensor<string, []>("query_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_37_dilations_0 = const()[name = tensor<string, []>("query_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_37_groups_0 = const()[name = tensor<string, []>("query_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409460672)))];
            tensor<fp16, [1024]> layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411557888)))];
            tensor<fp16, [1, 1024, 1, 1]> query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
            tensor<string, []> current_key_19_pad_type_0 = const()[name = tensor<string, []>("current_key_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_19_strides_0 = const()[name = tensor<string, []>("current_key_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_19_pad_0 = const()[name = tensor<string, []>("current_key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_19_dilations_0 = const()[name = tensor<string, []>("current_key_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_19_groups_0 = const()[name = tensor<string, []>("current_key_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411560000)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("current_key_19_cast_fp16")];
            tensor<string, []> current_value_19_pad_type_0 = const()[name = tensor<string, []>("current_value_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_19_strides_0 = const()[name = tensor<string, []>("current_value_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_19_pad_0 = const()[name = tensor<string, []>("current_value_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_19_dilations_0 = const()[name = tensor<string, []>("current_value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_19_groups_0 = const()[name = tensor<string, []>("current_value_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413657216)))];
            tensor<fp16, [1024]> layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415754432)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor<string, []>("current_value_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2179_cast_fp16 = mul(x = var_87_cast_fp16_9, y = var_207_cast_fp16)[name = tensor<string, []>("op_2179_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2180_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2180_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_37_cast_fp16 = add(x = var_2179_cast_fp16, y = var_2180_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2183_cast_fp16 = mul(x = var_114_cast_fp16_9, y = var_207_cast_fp16)[name = tensor<string, []>("op_2183_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2184_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2184_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_37_cast_fp16 = add(x = var_2183_cast_fp16, y = var_2184_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
            tensor<int32, [4]> var_2188 = const()[name = tensor<string, []>("op_2188"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_37_cast_fp16 = reshape(shape = var_2188, x = query_37_cast_fp16)[name = tensor<string, []>("mh_q_37_cast_fp16")];
            tensor<fp16, []> var_2190_to_fp16 = const()[name = tensor<string, []>("op_2190_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2191_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2190_to_fp16)[name = tensor<string, []>("op_2191_cast_fp16")];
            tensor<int32, [4]> var_2194 = const()[name = tensor<string, []>("op_2194"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2195_cast_fp16 = reshape(shape = var_2194, x = key_37_cast_fp16)[name = tensor<string, []>("op_2195_cast_fp16")];
            tensor<bool, []> mh_w_55_transpose_x_0 = const()[name = tensor<string, []>("mh_w_55_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_55_transpose_y_0 = const()[name = tensor<string, []>("mh_w_55_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2191_cast_fp16, y = var_2195_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_2203_cast_fp16 = softmax(axis = var_2115, x = mh_w_57_cast_fp16)[name = tensor<string, []>("op_2203_cast_fp16")];
            tensor<int32, [4]> var_2204 = const()[name = tensor<string, []>("op_2204"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2205_cast_fp16 = reshape(shape = var_2204, x = value_37_cast_fp16)[name = tensor<string, []>("op_2205_cast_fp16")];
            tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2205_cast_fp16, y = var_2203_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<int32, [4]> var_2208 = const()[name = tensor<string, []>("op_2208"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_91_cast_fp16 = reshape(shape = var_2208, x = attn_37_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<string, []> obj_133_pad_type_0 = const()[name = tensor<string, []>("obj_133_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_133_strides_0 = const()[name = tensor<string, []>("obj_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_133_pad_0 = const()[name = tensor<string, []>("obj_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_133_dilations_0 = const()[name = tensor<string, []>("obj_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_133_groups_0 = const()[name = tensor<string, []>("obj_133_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415756544)))];
            tensor<fp16, [1024]> layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417853760)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("obj_133_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
            tensor<int32, [1]> out_57_axes_0 = const()[name = tensor<string, []>("out_57_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2230_to_fp16 = const()[name = tensor<string, []>("op_2230_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2230_to_fp16, x = inputs_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
            tensor<fp16, [1024]> obj_135_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_135_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417855872)))];
            tensor<fp16, [1024]> obj_135_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_135_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417857984)))];
            tensor<fp16, []> obj_135_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_135_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("obj_135_cast_fp16")];
            tensor<string, []> query_39_pad_type_0 = const()[name = tensor<string, []>("query_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_39_strides_0 = const()[name = tensor<string, []>("query_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_39_pad_0 = const()[name = tensor<string, []>("query_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_39_dilations_0 = const()[name = tensor<string, []>("query_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_39_groups_0 = const()[name = tensor<string, []>("query_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417860096)))];
            tensor<fp16, [1024]> layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419957312)))];
            tensor<fp16, [1, 1024, 1, 1]> query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
            tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_39_strides_0 = const()[name = tensor<string, []>("key_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_39_dilations_0 = const()[name = tensor<string, []>("key_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_39_groups_0 = const()[name = tensor<string, []>("key_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419959424)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_39_cast_fp16")];
            tensor<string, []> value_39_pad_type_0 = const()[name = tensor<string, []>("value_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_39_strides_0 = const()[name = tensor<string, []>("value_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_39_pad_0 = const()[name = tensor<string, []>("value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_39_dilations_0 = const()[name = tensor<string, []>("value_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_39_groups_0 = const()[name = tensor<string, []>("value_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422056640)))];
            tensor<fp16, [1024]> layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424153856)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_39_cast_fp16")];
            tensor<int32, [4]> var_2266 = const()[name = tensor<string, []>("op_2266"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_39_cast_fp16 = reshape(shape = var_2266, x = query_39_cast_fp16)[name = tensor<string, []>("mh_q_39_cast_fp16")];
            tensor<fp16, []> var_2268_to_fp16 = const()[name = tensor<string, []>("op_2268_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2269_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2268_to_fp16)[name = tensor<string, []>("op_2269_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = tensor<string, []>("op_2272"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2273_cast_fp16 = reshape(shape = var_2272, x = key_39_cast_fp16)[name = tensor<string, []>("op_2273_cast_fp16")];
            tensor<bool, []> mh_w_59_transpose_x_0 = const()[name = tensor<string, []>("mh_w_59_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_59_transpose_y_0 = const()[name = tensor<string, []>("mh_w_59_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2269_cast_fp16, y = var_2273_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_139_cast_fp16 = softmax(axis = var_2115, x = mh_w_59_cast_fp16)[name = tensor<string, []>("obj_139_cast_fp16")];
            tensor<int32, [4]> var_2277 = const()[name = tensor<string, []>("op_2277"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2278_cast_fp16 = reshape(shape = var_2277, x = value_39_cast_fp16)[name = tensor<string, []>("op_2278_cast_fp16")];
            tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2278_cast_fp16, y = obj_139_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
            tensor<int32, [4]> var_2281 = const()[name = tensor<string, []>("op_2281"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_93_cast_fp16 = reshape(shape = var_2281, x = attn_39_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<string, []> obj_137_pad_type_0 = const()[name = tensor<string, []>("obj_137_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_137_strides_0 = const()[name = tensor<string, []>("obj_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_137_pad_0 = const()[name = tensor<string, []>("obj_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_137_dilations_0 = const()[name = tensor<string, []>("obj_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_137_groups_0 = const()[name = tensor<string, []>("obj_137_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424155968)))];
            tensor<fp16, [1024]> layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426253184)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("obj_137_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
            tensor<int32, [1]> out_59_axes_0 = const()[name = tensor<string, []>("out_59_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2299_to_fp16 = const()[name = tensor<string, []>("op_2299_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2299_to_fp16, x = inputs_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
            tensor<fp16, [1024]> input_95_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_95_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426255296)))];
            tensor<fp16, [1024]> input_95_beta_0_to_fp16 = const()[name = tensor<string, []>("input_95_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426257408)))];
            tensor<fp16, []> input_95_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_95_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<string, []> input_97_pad_type_0 = const()[name = tensor<string, []>("input_97_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_97_strides_0 = const()[name = tensor<string, []>("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_97_pad_0 = const()[name = tensor<string, []>("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_97_dilations_0 = const()[name = tensor<string, []>("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_97_groups_0 = const()[name = tensor<string, []>("input_97_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426259520)))];
            tensor<fp16, [4096]> layers_9_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434648192)))];
            tensor<fp16, [1, 4096, 1, 1]> input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<string, []> hidden_states_21_pad_type_0 = const()[name = tensor<string, []>("hidden_states_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_21_strides_0 = const()[name = tensor<string, []>("hidden_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_21_pad_0 = const()[name = tensor<string, []>("hidden_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_21_dilations_0 = const()[name = tensor<string, []>("hidden_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_21_groups_0 = const()[name = tensor<string, []>("hidden_states_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_9_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434656448)))];
            tensor<fp16, [1024]> layers_9_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443045120)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("hidden_states_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
            tensor<int32, []> var_2334 = const()[name = tensor<string, []>("op_2334"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_61_axes_0 = const()[name = tensor<string, []>("out_61_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2359_to_fp16 = const()[name = tensor<string, []>("op_2359_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2359_to_fp16, x = inputs_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
            tensor<fp16, [1024]> obj_141_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_141_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443047232)))];
            tensor<fp16, [1024]> obj_141_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_141_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443049344)))];
            tensor<fp16, []> obj_141_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_141_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("obj_141_cast_fp16")];
            tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_41_strides_0 = const()[name = tensor<string, []>("query_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_41_dilations_0 = const()[name = tensor<string, []>("query_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_41_groups_0 = const()[name = tensor<string, []>("query_41_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443051456)))];
            tensor<fp16, [1024]> layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445148672)))];
            tensor<fp16, [1, 1024, 1, 1]> query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
            tensor<string, []> current_key_21_pad_type_0 = const()[name = tensor<string, []>("current_key_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_21_strides_0 = const()[name = tensor<string, []>("current_key_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_21_pad_0 = const()[name = tensor<string, []>("current_key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_21_dilations_0 = const()[name = tensor<string, []>("current_key_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_21_groups_0 = const()[name = tensor<string, []>("current_key_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445150784)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("current_key_21_cast_fp16")];
            tensor<string, []> current_value_21_pad_type_0 = const()[name = tensor<string, []>("current_value_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_21_strides_0 = const()[name = tensor<string, []>("current_value_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_21_pad_0 = const()[name = tensor<string, []>("current_value_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_21_dilations_0 = const()[name = tensor<string, []>("current_value_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_21_groups_0 = const()[name = tensor<string, []>("current_value_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447248000)))];
            tensor<fp16, [1024]> layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449345216)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor<string, []>("current_value_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2398_cast_fp16 = mul(x = var_87_cast_fp16_10, y = var_207_cast_fp16)[name = tensor<string, []>("op_2398_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2399_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2399_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_41_cast_fp16 = add(x = var_2398_cast_fp16, y = var_2399_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2402_cast_fp16 = mul(x = var_114_cast_fp16_10, y = var_207_cast_fp16)[name = tensor<string, []>("op_2402_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2403_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2403_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_41_cast_fp16 = add(x = var_2402_cast_fp16, y = var_2403_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
            tensor<int32, [4]> var_2407 = const()[name = tensor<string, []>("op_2407"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_41_cast_fp16 = reshape(shape = var_2407, x = query_41_cast_fp16)[name = tensor<string, []>("mh_q_41_cast_fp16")];
            tensor<fp16, []> var_2409_to_fp16 = const()[name = tensor<string, []>("op_2409_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2410_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2409_to_fp16)[name = tensor<string, []>("op_2410_cast_fp16")];
            tensor<int32, [4]> var_2413 = const()[name = tensor<string, []>("op_2413"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2414_cast_fp16 = reshape(shape = var_2413, x = key_41_cast_fp16)[name = tensor<string, []>("op_2414_cast_fp16")];
            tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2410_cast_fp16, y = var_2414_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_63_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_2422_cast_fp16 = softmax(axis = var_2334, x = mh_w_63_cast_fp16)[name = tensor<string, []>("op_2422_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = tensor<string, []>("op_2423"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2424_cast_fp16 = reshape(shape = var_2423, x = value_41_cast_fp16)[name = tensor<string, []>("op_2424_cast_fp16")];
            tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2424_cast_fp16, y = var_2422_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<int32, [4]> var_2427 = const()[name = tensor<string, []>("op_2427"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_101_cast_fp16 = reshape(shape = var_2427, x = attn_41_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<string, []> obj_147_pad_type_0 = const()[name = tensor<string, []>("obj_147_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_147_strides_0 = const()[name = tensor<string, []>("obj_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_147_pad_0 = const()[name = tensor<string, []>("obj_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_147_dilations_0 = const()[name = tensor<string, []>("obj_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_147_groups_0 = const()[name = tensor<string, []>("obj_147_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449347328)))];
            tensor<fp16, [1024]> layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451444544)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("obj_147_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
            tensor<int32, [1]> out_63_axes_0 = const()[name = tensor<string, []>("out_63_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2449_to_fp16 = const()[name = tensor<string, []>("op_2449_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2449_to_fp16, x = inputs_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
            tensor<fp16, [1024]> obj_149_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_149_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451446656)))];
            tensor<fp16, [1024]> obj_149_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_149_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451448768)))];
            tensor<fp16, []> obj_149_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_149_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("obj_149_cast_fp16")];
            tensor<string, []> query_43_pad_type_0 = const()[name = tensor<string, []>("query_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_43_strides_0 = const()[name = tensor<string, []>("query_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_43_pad_0 = const()[name = tensor<string, []>("query_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_43_dilations_0 = const()[name = tensor<string, []>("query_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_43_groups_0 = const()[name = tensor<string, []>("query_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451450880)))];
            tensor<fp16, [1024]> layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453548096)))];
            tensor<fp16, [1, 1024, 1, 1]> query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
            tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_43_strides_0 = const()[name = tensor<string, []>("key_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_43_dilations_0 = const()[name = tensor<string, []>("key_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_43_groups_0 = const()[name = tensor<string, []>("key_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453550208)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_43_cast_fp16")];
            tensor<string, []> value_43_pad_type_0 = const()[name = tensor<string, []>("value_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_43_strides_0 = const()[name = tensor<string, []>("value_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_43_pad_0 = const()[name = tensor<string, []>("value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_43_dilations_0 = const()[name = tensor<string, []>("value_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_43_groups_0 = const()[name = tensor<string, []>("value_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455647424)))];
            tensor<fp16, [1024]> layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457744640)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_43_cast_fp16")];
            tensor<int32, [4]> var_2485 = const()[name = tensor<string, []>("op_2485"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_43_cast_fp16 = reshape(shape = var_2485, x = query_43_cast_fp16)[name = tensor<string, []>("mh_q_43_cast_fp16")];
            tensor<fp16, []> var_2487_to_fp16 = const()[name = tensor<string, []>("op_2487_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2488_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2487_to_fp16)[name = tensor<string, []>("op_2488_cast_fp16")];
            tensor<int32, [4]> var_2491 = const()[name = tensor<string, []>("op_2491"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2492_cast_fp16 = reshape(shape = var_2491, x = key_43_cast_fp16)[name = tensor<string, []>("op_2492_cast_fp16")];
            tensor<bool, []> mh_w_65_transpose_x_0 = const()[name = tensor<string, []>("mh_w_65_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_65_transpose_y_0 = const()[name = tensor<string, []>("mh_w_65_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2488_cast_fp16, y = var_2492_cast_fp16)[name = tensor<string, []>("mh_w_65_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_153_cast_fp16 = softmax(axis = var_2334, x = mh_w_65_cast_fp16)[name = tensor<string, []>("obj_153_cast_fp16")];
            tensor<int32, [4]> var_2496 = const()[name = tensor<string, []>("op_2496"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2497_cast_fp16 = reshape(shape = var_2496, x = value_43_cast_fp16)[name = tensor<string, []>("op_2497_cast_fp16")];
            tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2497_cast_fp16, y = obj_153_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
            tensor<int32, [4]> var_2500 = const()[name = tensor<string, []>("op_2500"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_103_cast_fp16 = reshape(shape = var_2500, x = attn_43_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<string, []> obj_151_pad_type_0 = const()[name = tensor<string, []>("obj_151_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_151_strides_0 = const()[name = tensor<string, []>("obj_151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_151_pad_0 = const()[name = tensor<string, []>("obj_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_151_dilations_0 = const()[name = tensor<string, []>("obj_151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_151_groups_0 = const()[name = tensor<string, []>("obj_151_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457746752)))];
            tensor<fp16, [1024]> layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459843968)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("obj_151_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")];
            tensor<int32, [1]> out_65_axes_0 = const()[name = tensor<string, []>("out_65_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2518_to_fp16 = const()[name = tensor<string, []>("op_2518_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2518_to_fp16, x = inputs_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
            tensor<fp16, [1024]> input_105_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_105_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459846080)))];
            tensor<fp16, [1024]> input_105_beta_0_to_fp16 = const()[name = tensor<string, []>("input_105_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459848192)))];
            tensor<fp16, []> input_105_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_105_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<string, []> input_107_pad_type_0 = const()[name = tensor<string, []>("input_107_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_107_strides_0 = const()[name = tensor<string, []>("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_107_pad_0 = const()[name = tensor<string, []>("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_107_dilations_0 = const()[name = tensor<string, []>("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_107_groups_0 = const()[name = tensor<string, []>("input_107_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459850304)))];
            tensor<fp16, [4096]> layers_10_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468238976)))];
            tensor<fp16, [1, 4096, 1, 1]> input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<string, []> input_109_mode_0 = const()[name = tensor<string, []>("input_109_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<string, []> hidden_states_23_pad_type_0 = const()[name = tensor<string, []>("hidden_states_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = tensor<string, []>("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = tensor<string, []>("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = tensor<string, []>("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_23_groups_0 = const()[name = tensor<string, []>("hidden_states_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_10_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468247232)))];
            tensor<fp16, [1024]> layers_10_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476635904)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")];
            tensor<int32, []> var_2553 = const()[name = tensor<string, []>("op_2553"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_67_axes_0 = const()[name = tensor<string, []>("out_67_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2578_to_fp16 = const()[name = tensor<string, []>("op_2578_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2578_to_fp16, x = inputs_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
            tensor<fp16, [1024]> obj_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_155_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476638016)))];
            tensor<fp16, [1024]> obj_155_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_155_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476640128)))];
            tensor<fp16, []> obj_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("obj_155_cast_fp16")];
            tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_45_strides_0 = const()[name = tensor<string, []>("query_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_45_dilations_0 = const()[name = tensor<string, []>("query_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_45_groups_0 = const()[name = tensor<string, []>("query_45_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476642240)))];
            tensor<fp16, [1024]> layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478739456)))];
            tensor<fp16, [1, 1024, 1, 1]> query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
            tensor<string, []> current_key_23_pad_type_0 = const()[name = tensor<string, []>("current_key_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_23_strides_0 = const()[name = tensor<string, []>("current_key_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_23_pad_0 = const()[name = tensor<string, []>("current_key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_23_dilations_0 = const()[name = tensor<string, []>("current_key_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_23_groups_0 = const()[name = tensor<string, []>("current_key_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478741568)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_23_cast_fp16 = conv(dilations = current_key_23_dilations_0, groups = current_key_23_groups_0, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = current_key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("current_key_23_cast_fp16")];
            tensor<string, []> current_value_23_pad_type_0 = const()[name = tensor<string, []>("current_value_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_23_strides_0 = const()[name = tensor<string, []>("current_value_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_23_pad_0 = const()[name = tensor<string, []>("current_value_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_23_dilations_0 = const()[name = tensor<string, []>("current_value_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_23_groups_0 = const()[name = tensor<string, []>("current_value_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480838784)))];
            tensor<fp16, [1024]> layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(482936000)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_23_dilations_0, groups = current_value_23_groups_0, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = current_value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor<string, []>("current_value_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2617_cast_fp16 = mul(x = var_87_cast_fp16_11, y = var_207_cast_fp16)[name = tensor<string, []>("op_2617_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2618_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2618_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_45_cast_fp16 = add(x = var_2617_cast_fp16, y = var_2618_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2621_cast_fp16 = mul(x = var_114_cast_fp16_11, y = var_207_cast_fp16)[name = tensor<string, []>("op_2621_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2622_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2622_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_45_cast_fp16 = add(x = var_2621_cast_fp16, y = var_2622_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
            tensor<int32, [4]> var_2626 = const()[name = tensor<string, []>("op_2626"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_45_cast_fp16 = reshape(shape = var_2626, x = query_45_cast_fp16)[name = tensor<string, []>("mh_q_45_cast_fp16")];
            tensor<fp16, []> var_2628_to_fp16 = const()[name = tensor<string, []>("op_2628_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2629_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2628_to_fp16)[name = tensor<string, []>("op_2629_cast_fp16")];
            tensor<int32, [4]> var_2632 = const()[name = tensor<string, []>("op_2632"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2633_cast_fp16 = reshape(shape = var_2632, x = key_45_cast_fp16)[name = tensor<string, []>("op_2633_cast_fp16")];
            tensor<bool, []> mh_w_67_transpose_x_0 = const()[name = tensor<string, []>("mh_w_67_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_67_transpose_y_0 = const()[name = tensor<string, []>("mh_w_67_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2629_cast_fp16, y = var_2633_cast_fp16)[name = tensor<string, []>("mh_w_67_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_69_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_2641_cast_fp16 = softmax(axis = var_2553, x = mh_w_69_cast_fp16)[name = tensor<string, []>("op_2641_cast_fp16")];
            tensor<int32, [4]> var_2642 = const()[name = tensor<string, []>("op_2642"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2643_cast_fp16 = reshape(shape = var_2642, x = value_45_cast_fp16)[name = tensor<string, []>("op_2643_cast_fp16")];
            tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2643_cast_fp16, y = var_2641_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<int32, [4]> var_2646 = const()[name = tensor<string, []>("op_2646"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_111_cast_fp16 = reshape(shape = var_2646, x = attn_45_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<string, []> obj_161_pad_type_0 = const()[name = tensor<string, []>("obj_161_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_161_strides_0 = const()[name = tensor<string, []>("obj_161_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_161_pad_0 = const()[name = tensor<string, []>("obj_161_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_161_dilations_0 = const()[name = tensor<string, []>("obj_161_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_161_groups_0 = const()[name = tensor<string, []>("obj_161_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(482938112)))];
            tensor<fp16, [1024]> layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485035328)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("obj_161_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
            tensor<int32, [1]> out_69_axes_0 = const()[name = tensor<string, []>("out_69_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2668_to_fp16 = const()[name = tensor<string, []>("op_2668_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2668_to_fp16, x = inputs_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
            tensor<fp16, [1024]> obj_163_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_163_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485037440)))];
            tensor<fp16, [1024]> obj_163_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_163_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485039552)))];
            tensor<fp16, []> obj_163_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_163_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("obj_163_cast_fp16")];
            tensor<string, []> query_47_pad_type_0 = const()[name = tensor<string, []>("query_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_47_strides_0 = const()[name = tensor<string, []>("query_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_47_pad_0 = const()[name = tensor<string, []>("query_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_47_dilations_0 = const()[name = tensor<string, []>("query_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_47_groups_0 = const()[name = tensor<string, []>("query_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485041664)))];
            tensor<fp16, [1024]> layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487138880)))];
            tensor<fp16, [1, 1024, 1, 1]> query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
            tensor<string, []> key_47_pad_type_0 = const()[name = tensor<string, []>("key_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_47_strides_0 = const()[name = tensor<string, []>("key_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_47_pad_0 = const()[name = tensor<string, []>("key_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_47_dilations_0 = const()[name = tensor<string, []>("key_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_47_groups_0 = const()[name = tensor<string, []>("key_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487140992)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_47_cast_fp16")];
            tensor<string, []> value_47_pad_type_0 = const()[name = tensor<string, []>("value_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_47_strides_0 = const()[name = tensor<string, []>("value_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_47_pad_0 = const()[name = tensor<string, []>("value_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_47_dilations_0 = const()[name = tensor<string, []>("value_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_47_groups_0 = const()[name = tensor<string, []>("value_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(489238208)))];
            tensor<fp16, [1024]> layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(491335424)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_47_cast_fp16")];
            tensor<int32, [4]> var_2704 = const()[name = tensor<string, []>("op_2704"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_47_cast_fp16 = reshape(shape = var_2704, x = query_47_cast_fp16)[name = tensor<string, []>("mh_q_47_cast_fp16")];
            tensor<fp16, []> var_2706_to_fp16 = const()[name = tensor<string, []>("op_2706_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2707_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2706_to_fp16)[name = tensor<string, []>("op_2707_cast_fp16")];
            tensor<int32, [4]> var_2710 = const()[name = tensor<string, []>("op_2710"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2711_cast_fp16 = reshape(shape = var_2710, x = key_47_cast_fp16)[name = tensor<string, []>("op_2711_cast_fp16")];
            tensor<bool, []> mh_w_71_transpose_x_0 = const()[name = tensor<string, []>("mh_w_71_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_71_transpose_y_0 = const()[name = tensor<string, []>("mh_w_71_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2707_cast_fp16, y = var_2711_cast_fp16)[name = tensor<string, []>("mh_w_71_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_167_cast_fp16 = softmax(axis = var_2553, x = mh_w_71_cast_fp16)[name = tensor<string, []>("obj_167_cast_fp16")];
            tensor<int32, [4]> var_2715 = const()[name = tensor<string, []>("op_2715"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2716_cast_fp16 = reshape(shape = var_2715, x = value_47_cast_fp16)[name = tensor<string, []>("op_2716_cast_fp16")];
            tensor<bool, []> attn_47_transpose_x_0 = const()[name = tensor<string, []>("attn_47_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_47_transpose_y_0 = const()[name = tensor<string, []>("attn_47_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2716_cast_fp16, y = obj_167_cast_fp16)[name = tensor<string, []>("attn_47_cast_fp16")];
            tensor<int32, [4]> var_2719 = const()[name = tensor<string, []>("op_2719"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_113_cast_fp16 = reshape(shape = var_2719, x = attn_47_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<string, []> obj_165_pad_type_0 = const()[name = tensor<string, []>("obj_165_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_165_strides_0 = const()[name = tensor<string, []>("obj_165_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_165_pad_0 = const()[name = tensor<string, []>("obj_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_165_dilations_0 = const()[name = tensor<string, []>("obj_165_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_165_groups_0 = const()[name = tensor<string, []>("obj_165_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(491337536)))];
            tensor<fp16, [1024]> layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493434752)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("obj_165_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
            tensor<int32, [1]> out_71_axes_0 = const()[name = tensor<string, []>("out_71_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2740_to_fp16 = const()[name = tensor<string, []>("op_2740_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2740_to_fp16, x = inputs_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
            tensor<fp16, [1024]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493436864)))];
            tensor<fp16, [1024]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493438976)))];
            tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_117_strides_0 = const()[name = tensor<string, []>("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_117_dilations_0 = const()[name = tensor<string, []>("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_117_groups_0 = const()[name = tensor<string, []>("input_117_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493441088)))];
            tensor<fp16, [4096]> layers_11_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501829760)))];
            tensor<fp16, [1, 4096, 1, 1]> input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<string, []> input_119_mode_0 = const()[name = tensor<string, []>("input_119_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<string, []> hidden_states_25_pad_type_0 = const()[name = tensor<string, []>("hidden_states_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_25_strides_0 = const()[name = tensor<string, []>("hidden_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = tensor<string, []>("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_25_dilations_0 = const()[name = tensor<string, []>("hidden_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_25_groups_0 = const()[name = tensor<string, []>("hidden_states_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_11_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501838016)))];
            tensor<fp16, [1024]> layers_11_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510226688)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
            tensor<int32, []> var_2776 = const()[name = tensor<string, []>("op_2776"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_73_axes_0 = const()[name = tensor<string, []>("out_73_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2801_to_fp16 = const()[name = tensor<string, []>("op_2801_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2801_to_fp16, x = inputs_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
            tensor<fp16, [1024]> obj_169_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_169_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510228800)))];
            tensor<fp16, [1024]> obj_169_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_169_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510230912)))];
            tensor<fp16, []> obj_169_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_169_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_169_cast_fp16")];
            tensor<string, []> query_49_pad_type_0 = const()[name = tensor<string, []>("query_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_49_strides_0 = const()[name = tensor<string, []>("query_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_49_pad_0 = const()[name = tensor<string, []>("query_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_49_dilations_0 = const()[name = tensor<string, []>("query_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_49_groups_0 = const()[name = tensor<string, []>("query_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510233024)))];
            tensor<fp16, [1024]> layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512330240)))];
            tensor<fp16, [1, 1024, 1, 1]> query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
            tensor<string, []> current_key_25_pad_type_0 = const()[name = tensor<string, []>("current_key_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_25_strides_0 = const()[name = tensor<string, []>("current_key_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_25_pad_0 = const()[name = tensor<string, []>("current_key_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_25_dilations_0 = const()[name = tensor<string, []>("current_key_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_25_groups_0 = const()[name = tensor<string, []>("current_key_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512332352)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_25_cast_fp16 = conv(dilations = current_key_25_dilations_0, groups = current_key_25_groups_0, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = current_key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor<string, []>("current_key_25_cast_fp16")];
            tensor<string, []> current_value_25_pad_type_0 = const()[name = tensor<string, []>("current_value_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_25_strides_0 = const()[name = tensor<string, []>("current_value_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_25_pad_0 = const()[name = tensor<string, []>("current_value_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_25_dilations_0 = const()[name = tensor<string, []>("current_value_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_25_groups_0 = const()[name = tensor<string, []>("current_value_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514429568)))];
            tensor<fp16, [1024]> layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(516526784)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = current_value_25_dilations_0, groups = current_value_25_groups_0, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = current_value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor<string, []>("current_value_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2840_cast_fp16 = mul(x = var_87_cast_fp16_12, y = var_207_cast_fp16)[name = tensor<string, []>("op_2840_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2841_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2841_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_49_cast_fp16 = add(x = var_2840_cast_fp16, y = var_2841_cast_fp16)[name = tensor<string, []>("key_49_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2844_cast_fp16 = mul(x = var_114_cast_fp16_12, y = var_207_cast_fp16)[name = tensor<string, []>("op_2844_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_2845_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_2845_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_49_cast_fp16 = add(x = var_2844_cast_fp16, y = var_2845_cast_fp16)[name = tensor<string, []>("value_49_cast_fp16")];
            tensor<int32, [4]> var_2849 = const()[name = tensor<string, []>("op_2849"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_49_cast_fp16 = reshape(shape = var_2849, x = query_49_cast_fp16)[name = tensor<string, []>("mh_q_49_cast_fp16")];
            tensor<fp16, []> var_2851_to_fp16 = const()[name = tensor<string, []>("op_2851_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2852_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_2851_to_fp16)[name = tensor<string, []>("op_2852_cast_fp16")];
            tensor<int32, [4]> var_2855 = const()[name = tensor<string, []>("op_2855"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2856_cast_fp16 = reshape(shape = var_2855, x = key_49_cast_fp16)[name = tensor<string, []>("op_2856_cast_fp16")];
            tensor<bool, []> mh_w_73_transpose_x_0 = const()[name = tensor<string, []>("mh_w_73_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_73_transpose_y_0 = const()[name = tensor<string, []>("mh_w_73_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = tensor<string, []>("mh_w_73_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_75_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_2864_cast_fp16 = softmax(axis = var_2776, x = mh_w_75_cast_fp16)[name = tensor<string, []>("op_2864_cast_fp16")];
            tensor<int32, [4]> var_2865 = const()[name = tensor<string, []>("op_2865"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_2866_cast_fp16 = reshape(shape = var_2865, x = value_49_cast_fp16)[name = tensor<string, []>("op_2866_cast_fp16")];
            tensor<bool, []> attn_49_transpose_x_0 = const()[name = tensor<string, []>("attn_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_49_transpose_y_0 = const()[name = tensor<string, []>("attn_49_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2866_cast_fp16, y = var_2864_cast_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = tensor<string, []>("op_2869"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_121_cast_fp16 = reshape(shape = var_2869, x = attn_49_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<string, []> obj_175_pad_type_0 = const()[name = tensor<string, []>("obj_175_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_175_strides_0 = const()[name = tensor<string, []>("obj_175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_175_pad_0 = const()[name = tensor<string, []>("obj_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_175_dilations_0 = const()[name = tensor<string, []>("obj_175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_175_groups_0 = const()[name = tensor<string, []>("obj_175_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(516528896)))];
            tensor<fp16, [1024]> layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518626112)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_175_dilations_0, groups = obj_175_groups_0, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = obj_175_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("obj_175_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor<string, []>("inputs_75_cast_fp16")];
            tensor<int32, [1]> out_75_axes_0 = const()[name = tensor<string, []>("out_75_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2891_to_fp16 = const()[name = tensor<string, []>("op_2891_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2891_to_fp16, x = inputs_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
            tensor<fp16, [1024]> obj_177_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_177_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518628224)))];
            tensor<fp16, [1024]> obj_177_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_177_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518630336)))];
            tensor<fp16, []> obj_177_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_177_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("obj_177_cast_fp16")];
            tensor<string, []> query_51_pad_type_0 = const()[name = tensor<string, []>("query_51_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_51_strides_0 = const()[name = tensor<string, []>("query_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_51_pad_0 = const()[name = tensor<string, []>("query_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_51_dilations_0 = const()[name = tensor<string, []>("query_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_51_groups_0 = const()[name = tensor<string, []>("query_51_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518632448)))];
            tensor<fp16, [1024]> layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520729664)))];
            tensor<fp16, [1, 1024, 1, 1]> query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
            tensor<string, []> key_51_pad_type_0 = const()[name = tensor<string, []>("key_51_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_51_strides_0 = const()[name = tensor<string, []>("key_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_51_pad_0 = const()[name = tensor<string, []>("key_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_51_dilations_0 = const()[name = tensor<string, []>("key_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_51_groups_0 = const()[name = tensor<string, []>("key_51_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520731776)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_51_cast_fp16")];
            tensor<string, []> value_51_pad_type_0 = const()[name = tensor<string, []>("value_51_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_51_strides_0 = const()[name = tensor<string, []>("value_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_51_pad_0 = const()[name = tensor<string, []>("value_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_51_dilations_0 = const()[name = tensor<string, []>("value_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_51_groups_0 = const()[name = tensor<string, []>("value_51_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522828992)))];
            tensor<fp16, [1024]> layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(524926208)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_51_cast_fp16")];
            tensor<int32, [4]> var_2927 = const()[name = tensor<string, []>("op_2927"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_51_cast_fp16 = reshape(shape = var_2927, x = query_51_cast_fp16)[name = tensor<string, []>("mh_q_51_cast_fp16")];
            tensor<fp16, []> var_2929_to_fp16 = const()[name = tensor<string, []>("op_2929_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_2930_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_2929_to_fp16)[name = tensor<string, []>("op_2930_cast_fp16")];
            tensor<int32, [4]> var_2933 = const()[name = tensor<string, []>("op_2933"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2934_cast_fp16 = reshape(shape = var_2933, x = key_51_cast_fp16)[name = tensor<string, []>("op_2934_cast_fp16")];
            tensor<bool, []> mh_w_77_transpose_x_0 = const()[name = tensor<string, []>("mh_w_77_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_77_transpose_y_0 = const()[name = tensor<string, []>("mh_w_77_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2930_cast_fp16, y = var_2934_cast_fp16)[name = tensor<string, []>("mh_w_77_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_181_cast_fp16 = softmax(axis = var_2776, x = mh_w_77_cast_fp16)[name = tensor<string, []>("obj_181_cast_fp16")];
            tensor<int32, [4]> var_2938 = const()[name = tensor<string, []>("op_2938"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_2939_cast_fp16 = reshape(shape = var_2938, x = value_51_cast_fp16)[name = tensor<string, []>("op_2939_cast_fp16")];
            tensor<bool, []> attn_51_transpose_x_0 = const()[name = tensor<string, []>("attn_51_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_51_transpose_y_0 = const()[name = tensor<string, []>("attn_51_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2939_cast_fp16, y = obj_181_cast_fp16)[name = tensor<string, []>("attn_51_cast_fp16")];
            tensor<int32, [4]> var_2942 = const()[name = tensor<string, []>("op_2942"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_123_cast_fp16 = reshape(shape = var_2942, x = attn_51_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<string, []> obj_179_pad_type_0 = const()[name = tensor<string, []>("obj_179_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_179_strides_0 = const()[name = tensor<string, []>("obj_179_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_179_pad_0 = const()[name = tensor<string, []>("obj_179_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_179_dilations_0 = const()[name = tensor<string, []>("obj_179_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_179_groups_0 = const()[name = tensor<string, []>("obj_179_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(524928320)))];
            tensor<fp16, [1024]> layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527025536)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = obj_179_dilations_0, groups = obj_179_groups_0, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = obj_179_strides_0, weight = layers_12_encoder_attn_o_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("obj_179_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor<string, []>("inputs_77_cast_fp16")];
            tensor<int32, [1]> out_77_axes_0 = const()[name = tensor<string, []>("out_77_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_2960_to_fp16 = const()[name = tensor<string, []>("op_2960_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2960_to_fp16, x = inputs_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
            tensor<fp16, [1024]> input_125_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_125_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527027648)))];
            tensor<fp16, [1024]> input_125_beta_0_to_fp16 = const()[name = tensor<string, []>("input_125_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527029760)))];
            tensor<fp16, []> input_125_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_125_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<string, []> input_127_pad_type_0 = const()[name = tensor<string, []>("input_127_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_127_strides_0 = const()[name = tensor<string, []>("input_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_127_pad_0 = const()[name = tensor<string, []>("input_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_127_dilations_0 = const()[name = tensor<string, []>("input_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_127_groups_0 = const()[name = tensor<string, []>("input_127_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527031872)))];
            tensor<fp16, [4096]> layers_12_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535420544)))];
            tensor<fp16, [1, 4096, 1, 1]> input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<string, []> input_129_mode_0 = const()[name = tensor<string, []>("input_129_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<string, []> hidden_states_27_pad_type_0 = const()[name = tensor<string, []>("hidden_states_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = tensor<string, []>("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = tensor<string, []>("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = tensor<string, []>("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_27_groups_0 = const()[name = tensor<string, []>("hidden_states_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_12_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535428800)))];
            tensor<fp16, [1024]> layers_12_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(543817472)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
            tensor<int32, []> var_2995 = const()[name = tensor<string, []>("op_2995"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_79_axes_0 = const()[name = tensor<string, []>("out_79_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3020_to_fp16 = const()[name = tensor<string, []>("op_3020_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_3020_to_fp16, x = inputs_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
            tensor<fp16, [1024]> obj_183_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_183_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(543819584)))];
            tensor<fp16, [1024]> obj_183_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_183_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(543821696)))];
            tensor<fp16, []> obj_183_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_183_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("obj_183_cast_fp16")];
            tensor<string, []> query_53_pad_type_0 = const()[name = tensor<string, []>("query_53_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_53_strides_0 = const()[name = tensor<string, []>("query_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_53_pad_0 = const()[name = tensor<string, []>("query_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_53_dilations_0 = const()[name = tensor<string, []>("query_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_53_groups_0 = const()[name = tensor<string, []>("query_53_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(543823808)))];
            tensor<fp16, [1024]> layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545921024)))];
            tensor<fp16, [1, 1024, 1, 1]> query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
            tensor<string, []> current_key_27_pad_type_0 = const()[name = tensor<string, []>("current_key_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_27_strides_0 = const()[name = tensor<string, []>("current_key_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_27_pad_0 = const()[name = tensor<string, []>("current_key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_27_dilations_0 = const()[name = tensor<string, []>("current_key_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_27_groups_0 = const()[name = tensor<string, []>("current_key_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(545923136)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_27_cast_fp16 = conv(dilations = current_key_27_dilations_0, groups = current_key_27_groups_0, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = current_key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor<string, []>("current_key_27_cast_fp16")];
            tensor<string, []> current_value_27_pad_type_0 = const()[name = tensor<string, []>("current_value_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_27_strides_0 = const()[name = tensor<string, []>("current_value_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_27_pad_0 = const()[name = tensor<string, []>("current_value_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_27_dilations_0 = const()[name = tensor<string, []>("current_value_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_27_groups_0 = const()[name = tensor<string, []>("current_value_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548020352)))];
            tensor<fp16, [1024]> layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550117568)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = current_value_27_dilations_0, groups = current_value_27_groups_0, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = current_value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor<string, []>("current_value_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3059_cast_fp16 = mul(x = var_87_cast_fp16_13, y = var_207_cast_fp16)[name = tensor<string, []>("op_3059_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3060_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3060_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_53_cast_fp16 = add(x = var_3059_cast_fp16, y = var_3060_cast_fp16)[name = tensor<string, []>("key_53_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3063_cast_fp16 = mul(x = var_114_cast_fp16_13, y = var_207_cast_fp16)[name = tensor<string, []>("op_3063_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3064_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3064_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_53_cast_fp16 = add(x = var_3063_cast_fp16, y = var_3064_cast_fp16)[name = tensor<string, []>("value_53_cast_fp16")];
            tensor<int32, [4]> var_3068 = const()[name = tensor<string, []>("op_3068"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_53_cast_fp16 = reshape(shape = var_3068, x = query_53_cast_fp16)[name = tensor<string, []>("mh_q_53_cast_fp16")];
            tensor<fp16, []> var_3070_to_fp16 = const()[name = tensor<string, []>("op_3070_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3071_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3070_to_fp16)[name = tensor<string, []>("op_3071_cast_fp16")];
            tensor<int32, [4]> var_3074 = const()[name = tensor<string, []>("op_3074"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3075_cast_fp16 = reshape(shape = var_3074, x = key_53_cast_fp16)[name = tensor<string, []>("op_3075_cast_fp16")];
            tensor<bool, []> mh_w_79_transpose_x_0 = const()[name = tensor<string, []>("mh_w_79_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_79_transpose_y_0 = const()[name = tensor<string, []>("mh_w_79_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3071_cast_fp16, y = var_3075_cast_fp16)[name = tensor<string, []>("mh_w_79_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_81_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_3083_cast_fp16 = softmax(axis = var_2995, x = mh_w_81_cast_fp16)[name = tensor<string, []>("op_3083_cast_fp16")];
            tensor<int32, [4]> var_3084 = const()[name = tensor<string, []>("op_3084"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3085_cast_fp16 = reshape(shape = var_3084, x = value_53_cast_fp16)[name = tensor<string, []>("op_3085_cast_fp16")];
            tensor<bool, []> attn_53_transpose_x_0 = const()[name = tensor<string, []>("attn_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_53_transpose_y_0 = const()[name = tensor<string, []>("attn_53_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3085_cast_fp16, y = var_3083_cast_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
            tensor<int32, [4]> var_3088 = const()[name = tensor<string, []>("op_3088"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_131_cast_fp16 = reshape(shape = var_3088, x = attn_53_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<string, []> obj_189_pad_type_0 = const()[name = tensor<string, []>("obj_189_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_189_strides_0 = const()[name = tensor<string, []>("obj_189_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_189_pad_0 = const()[name = tensor<string, []>("obj_189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_189_dilations_0 = const()[name = tensor<string, []>("obj_189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_189_groups_0 = const()[name = tensor<string, []>("obj_189_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550119680)))];
            tensor<fp16, [1024]> layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552216896)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_189_dilations_0, groups = obj_189_groups_0, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = obj_189_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("obj_189_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
            tensor<int32, [1]> out_81_axes_0 = const()[name = tensor<string, []>("out_81_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3110_to_fp16 = const()[name = tensor<string, []>("op_3110_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_3110_to_fp16, x = inputs_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
            tensor<fp16, [1024]> obj_191_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_191_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552219008)))];
            tensor<fp16, [1024]> obj_191_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_191_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552221120)))];
            tensor<fp16, []> obj_191_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_191_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("obj_191_cast_fp16")];
            tensor<string, []> query_55_pad_type_0 = const()[name = tensor<string, []>("query_55_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_55_strides_0 = const()[name = tensor<string, []>("query_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_55_pad_0 = const()[name = tensor<string, []>("query_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_55_dilations_0 = const()[name = tensor<string, []>("query_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_55_groups_0 = const()[name = tensor<string, []>("query_55_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552223232)))];
            tensor<fp16, [1024]> layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554320448)))];
            tensor<fp16, [1, 1024, 1, 1]> query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
            tensor<string, []> key_55_pad_type_0 = const()[name = tensor<string, []>("key_55_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_55_strides_0 = const()[name = tensor<string, []>("key_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_55_pad_0 = const()[name = tensor<string, []>("key_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_55_dilations_0 = const()[name = tensor<string, []>("key_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_55_groups_0 = const()[name = tensor<string, []>("key_55_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554322560)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_55_cast_fp16")];
            tensor<string, []> value_55_pad_type_0 = const()[name = tensor<string, []>("value_55_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_55_strides_0 = const()[name = tensor<string, []>("value_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_55_pad_0 = const()[name = tensor<string, []>("value_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_55_dilations_0 = const()[name = tensor<string, []>("value_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_55_groups_0 = const()[name = tensor<string, []>("value_55_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(556419776)))];
            tensor<fp16, [1024]> layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558516992)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_55_cast_fp16")];
            tensor<int32, [4]> var_3146 = const()[name = tensor<string, []>("op_3146"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_55_cast_fp16 = reshape(shape = var_3146, x = query_55_cast_fp16)[name = tensor<string, []>("mh_q_55_cast_fp16")];
            tensor<fp16, []> var_3148_to_fp16 = const()[name = tensor<string, []>("op_3148_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3149_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3148_to_fp16)[name = tensor<string, []>("op_3149_cast_fp16")];
            tensor<int32, [4]> var_3152 = const()[name = tensor<string, []>("op_3152"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3153_cast_fp16 = reshape(shape = var_3152, x = key_55_cast_fp16)[name = tensor<string, []>("op_3153_cast_fp16")];
            tensor<bool, []> mh_w_83_transpose_x_0 = const()[name = tensor<string, []>("mh_w_83_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_83_transpose_y_0 = const()[name = tensor<string, []>("mh_w_83_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3149_cast_fp16, y = var_3153_cast_fp16)[name = tensor<string, []>("mh_w_83_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_195_cast_fp16 = softmax(axis = var_2995, x = mh_w_83_cast_fp16)[name = tensor<string, []>("obj_195_cast_fp16")];
            tensor<int32, [4]> var_3157 = const()[name = tensor<string, []>("op_3157"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3158_cast_fp16 = reshape(shape = var_3157, x = value_55_cast_fp16)[name = tensor<string, []>("op_3158_cast_fp16")];
            tensor<bool, []> attn_55_transpose_x_0 = const()[name = tensor<string, []>("attn_55_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_55_transpose_y_0 = const()[name = tensor<string, []>("attn_55_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3158_cast_fp16, y = obj_195_cast_fp16)[name = tensor<string, []>("attn_55_cast_fp16")];
            tensor<int32, [4]> var_3161 = const()[name = tensor<string, []>("op_3161"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_133_cast_fp16 = reshape(shape = var_3161, x = attn_55_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<string, []> obj_193_pad_type_0 = const()[name = tensor<string, []>("obj_193_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_193_strides_0 = const()[name = tensor<string, []>("obj_193_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_193_pad_0 = const()[name = tensor<string, []>("obj_193_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_193_dilations_0 = const()[name = tensor<string, []>("obj_193_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_193_groups_0 = const()[name = tensor<string, []>("obj_193_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558519104)))];
            tensor<fp16, [1024]> layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560616320)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = obj_193_dilations_0, groups = obj_193_groups_0, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = obj_193_strides_0, weight = layers_13_encoder_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("obj_193_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
            tensor<int32, [1]> out_83_axes_0 = const()[name = tensor<string, []>("out_83_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3179_to_fp16 = const()[name = tensor<string, []>("op_3179_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_3179_to_fp16, x = inputs_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
            tensor<fp16, [1024]> input_135_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_135_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560618432)))];
            tensor<fp16, [1024]> input_135_beta_0_to_fp16 = const()[name = tensor<string, []>("input_135_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560620544)))];
            tensor<fp16, []> input_135_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_135_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<string, []> input_137_pad_type_0 = const()[name = tensor<string, []>("input_137_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_137_strides_0 = const()[name = tensor<string, []>("input_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_137_pad_0 = const()[name = tensor<string, []>("input_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_137_dilations_0 = const()[name = tensor<string, []>("input_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_137_groups_0 = const()[name = tensor<string, []>("input_137_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560622656)))];
            tensor<fp16, [4096]> layers_13_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569011328)))];
            tensor<fp16, [1, 4096, 1, 1]> input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<string, []> input_139_mode_0 = const()[name = tensor<string, []>("input_139_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<string, []> hidden_states_29_pad_type_0 = const()[name = tensor<string, []>("hidden_states_29_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_29_strides_0 = const()[name = tensor<string, []>("hidden_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = tensor<string, []>("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_29_dilations_0 = const()[name = tensor<string, []>("hidden_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_29_groups_0 = const()[name = tensor<string, []>("hidden_states_29_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_13_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569019584)))];
            tensor<fp16, [1024]> layers_13_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577408256)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("hidden_states_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor<string, []>("inputs_85_cast_fp16")];
            tensor<int32, []> var_3214 = const()[name = tensor<string, []>("op_3214"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_85_axes_0 = const()[name = tensor<string, []>("out_85_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3239_to_fp16 = const()[name = tensor<string, []>("op_3239_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_3239_to_fp16, x = inputs_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
            tensor<fp16, [1024]> obj_197_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_197_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577410368)))];
            tensor<fp16, [1024]> obj_197_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_197_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577412480)))];
            tensor<fp16, []> obj_197_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_197_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("obj_197_cast_fp16")];
            tensor<string, []> query_57_pad_type_0 = const()[name = tensor<string, []>("query_57_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_57_strides_0 = const()[name = tensor<string, []>("query_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_57_pad_0 = const()[name = tensor<string, []>("query_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_57_dilations_0 = const()[name = tensor<string, []>("query_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_57_groups_0 = const()[name = tensor<string, []>("query_57_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577414592)))];
            tensor<fp16, [1024]> layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579511808)))];
            tensor<fp16, [1, 1024, 1, 1]> query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
            tensor<string, []> current_key_29_pad_type_0 = const()[name = tensor<string, []>("current_key_29_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_29_strides_0 = const()[name = tensor<string, []>("current_key_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_29_pad_0 = const()[name = tensor<string, []>("current_key_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_29_dilations_0 = const()[name = tensor<string, []>("current_key_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_29_groups_0 = const()[name = tensor<string, []>("current_key_29_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579513920)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_29_cast_fp16 = conv(dilations = current_key_29_dilations_0, groups = current_key_29_groups_0, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = current_key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor<string, []>("current_key_29_cast_fp16")];
            tensor<string, []> current_value_29_pad_type_0 = const()[name = tensor<string, []>("current_value_29_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_29_strides_0 = const()[name = tensor<string, []>("current_value_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_29_pad_0 = const()[name = tensor<string, []>("current_value_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_29_dilations_0 = const()[name = tensor<string, []>("current_value_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_29_groups_0 = const()[name = tensor<string, []>("current_value_29_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581611136)))];
            tensor<fp16, [1024]> layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(583708352)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = current_value_29_dilations_0, groups = current_value_29_groups_0, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = current_value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor<string, []>("current_value_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3278_cast_fp16 = mul(x = var_87_cast_fp16_14, y = var_207_cast_fp16)[name = tensor<string, []>("op_3278_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3279_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3279_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_57_cast_fp16 = add(x = var_3278_cast_fp16, y = var_3279_cast_fp16)[name = tensor<string, []>("key_57_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3282_cast_fp16 = mul(x = var_114_cast_fp16_14, y = var_207_cast_fp16)[name = tensor<string, []>("op_3282_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3283_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3283_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_57_cast_fp16 = add(x = var_3282_cast_fp16, y = var_3283_cast_fp16)[name = tensor<string, []>("value_57_cast_fp16")];
            tensor<int32, [4]> var_3287 = const()[name = tensor<string, []>("op_3287"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_57_cast_fp16 = reshape(shape = var_3287, x = query_57_cast_fp16)[name = tensor<string, []>("mh_q_57_cast_fp16")];
            tensor<fp16, []> var_3289_to_fp16 = const()[name = tensor<string, []>("op_3289_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3290_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3289_to_fp16)[name = tensor<string, []>("op_3290_cast_fp16")];
            tensor<int32, [4]> var_3293 = const()[name = tensor<string, []>("op_3293"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3294_cast_fp16 = reshape(shape = var_3293, x = key_57_cast_fp16)[name = tensor<string, []>("op_3294_cast_fp16")];
            tensor<bool, []> mh_w_85_transpose_x_0 = const()[name = tensor<string, []>("mh_w_85_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_85_transpose_y_0 = const()[name = tensor<string, []>("mh_w_85_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3290_cast_fp16, y = var_3294_cast_fp16)[name = tensor<string, []>("mh_w_85_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_87_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_3302_cast_fp16 = softmax(axis = var_3214, x = mh_w_87_cast_fp16)[name = tensor<string, []>("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3303 = const()[name = tensor<string, []>("op_3303"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3304_cast_fp16 = reshape(shape = var_3303, x = value_57_cast_fp16)[name = tensor<string, []>("op_3304_cast_fp16")];
            tensor<bool, []> attn_57_transpose_x_0 = const()[name = tensor<string, []>("attn_57_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_57_transpose_y_0 = const()[name = tensor<string, []>("attn_57_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3304_cast_fp16, y = var_3302_cast_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
            tensor<int32, [4]> var_3307 = const()[name = tensor<string, []>("op_3307"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_141_cast_fp16 = reshape(shape = var_3307, x = attn_57_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<string, []> obj_203_pad_type_0 = const()[name = tensor<string, []>("obj_203_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_203_strides_0 = const()[name = tensor<string, []>("obj_203_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_203_pad_0 = const()[name = tensor<string, []>("obj_203_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_203_dilations_0 = const()[name = tensor<string, []>("obj_203_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_203_groups_0 = const()[name = tensor<string, []>("obj_203_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(583710464)))];
            tensor<fp16, [1024]> layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(585807680)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_203_dilations_0, groups = obj_203_groups_0, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = obj_203_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("obj_203_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor<string, []>("inputs_87_cast_fp16")];
            tensor<int32, [1]> out_87_axes_0 = const()[name = tensor<string, []>("out_87_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3329_to_fp16 = const()[name = tensor<string, []>("op_3329_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_3329_to_fp16, x = inputs_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
            tensor<fp16, [1024]> obj_205_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_205_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(585809792)))];
            tensor<fp16, [1024]> obj_205_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_205_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(585811904)))];
            tensor<fp16, []> obj_205_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_205_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("obj_205_cast_fp16")];
            tensor<string, []> query_59_pad_type_0 = const()[name = tensor<string, []>("query_59_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_59_strides_0 = const()[name = tensor<string, []>("query_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_59_pad_0 = const()[name = tensor<string, []>("query_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_59_dilations_0 = const()[name = tensor<string, []>("query_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_59_groups_0 = const()[name = tensor<string, []>("query_59_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(585814016)))];
            tensor<fp16, [1024]> layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587911232)))];
            tensor<fp16, [1, 1024, 1, 1]> query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
            tensor<string, []> key_59_pad_type_0 = const()[name = tensor<string, []>("key_59_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_59_strides_0 = const()[name = tensor<string, []>("key_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_59_pad_0 = const()[name = tensor<string, []>("key_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_59_dilations_0 = const()[name = tensor<string, []>("key_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_59_groups_0 = const()[name = tensor<string, []>("key_59_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587913344)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_59_cast_fp16")];
            tensor<string, []> value_59_pad_type_0 = const()[name = tensor<string, []>("value_59_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_59_strides_0 = const()[name = tensor<string, []>("value_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_59_pad_0 = const()[name = tensor<string, []>("value_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_59_dilations_0 = const()[name = tensor<string, []>("value_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_59_groups_0 = const()[name = tensor<string, []>("value_59_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590010560)))];
            tensor<fp16, [1024]> layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(592107776)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_59_cast_fp16")];
            tensor<int32, [4]> var_3365 = const()[name = tensor<string, []>("op_3365"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_59_cast_fp16 = reshape(shape = var_3365, x = query_59_cast_fp16)[name = tensor<string, []>("mh_q_59_cast_fp16")];
            tensor<fp16, []> var_3367_to_fp16 = const()[name = tensor<string, []>("op_3367_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3368_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3367_to_fp16)[name = tensor<string, []>("op_3368_cast_fp16")];
            tensor<int32, [4]> var_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3372_cast_fp16 = reshape(shape = var_3371, x = key_59_cast_fp16)[name = tensor<string, []>("op_3372_cast_fp16")];
            tensor<bool, []> mh_w_89_transpose_x_0 = const()[name = tensor<string, []>("mh_w_89_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_89_transpose_y_0 = const()[name = tensor<string, []>("mh_w_89_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3368_cast_fp16, y = var_3372_cast_fp16)[name = tensor<string, []>("mh_w_89_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_209_cast_fp16 = softmax(axis = var_3214, x = mh_w_89_cast_fp16)[name = tensor<string, []>("obj_209_cast_fp16")];
            tensor<int32, [4]> var_3376 = const()[name = tensor<string, []>("op_3376"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3377_cast_fp16 = reshape(shape = var_3376, x = value_59_cast_fp16)[name = tensor<string, []>("op_3377_cast_fp16")];
            tensor<bool, []> attn_59_transpose_x_0 = const()[name = tensor<string, []>("attn_59_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_59_transpose_y_0 = const()[name = tensor<string, []>("attn_59_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3377_cast_fp16, y = obj_209_cast_fp16)[name = tensor<string, []>("attn_59_cast_fp16")];
            tensor<int32, [4]> var_3380 = const()[name = tensor<string, []>("op_3380"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_143_cast_fp16 = reshape(shape = var_3380, x = attn_59_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<string, []> obj_207_pad_type_0 = const()[name = tensor<string, []>("obj_207_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_207_strides_0 = const()[name = tensor<string, []>("obj_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_207_pad_0 = const()[name = tensor<string, []>("obj_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_207_dilations_0 = const()[name = tensor<string, []>("obj_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_207_groups_0 = const()[name = tensor<string, []>("obj_207_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(592109888)))];
            tensor<fp16, [1024]> layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594207104)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = obj_207_dilations_0, groups = obj_207_groups_0, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = obj_207_strides_0, weight = layers_14_encoder_attn_o_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("obj_207_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
            tensor<int32, [1]> out_89_axes_0 = const()[name = tensor<string, []>("out_89_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3401_to_fp16 = const()[name = tensor<string, []>("op_3401_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_3401_to_fp16, x = inputs_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
            tensor<fp16, [1024]> input_145_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_145_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594209216)))];
            tensor<fp16, [1024]> input_145_beta_0_to_fp16 = const()[name = tensor<string, []>("input_145_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594211328)))];
            tensor<fp16, []> input_145_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_145_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<string, []> input_147_pad_type_0 = const()[name = tensor<string, []>("input_147_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_147_strides_0 = const()[name = tensor<string, []>("input_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_147_pad_0 = const()[name = tensor<string, []>("input_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_147_dilations_0 = const()[name = tensor<string, []>("input_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_147_groups_0 = const()[name = tensor<string, []>("input_147_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594213440)))];
            tensor<fp16, [4096]> layers_14_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602602112)))];
            tensor<fp16, [1, 4096, 1, 1]> input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<string, []> input_149_mode_0 = const()[name = tensor<string, []>("input_149_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<string, []> hidden_states_31_pad_type_0 = const()[name = tensor<string, []>("hidden_states_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_31_strides_0 = const()[name = tensor<string, []>("hidden_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_31_pad_0 = const()[name = tensor<string, []>("hidden_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_31_dilations_0 = const()[name = tensor<string, []>("hidden_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_31_groups_0 = const()[name = tensor<string, []>("hidden_states_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_14_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602610368)))];
            tensor<fp16, [1024]> layers_14_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(610999040)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
            tensor<int32, []> var_3437 = const()[name = tensor<string, []>("op_3437"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_91_axes_0 = const()[name = tensor<string, []>("out_91_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3462_to_fp16 = const()[name = tensor<string, []>("op_3462_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_3462_to_fp16, x = inputs_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
            tensor<fp16, [1024]> obj_211_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_211_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(611001152)))];
            tensor<fp16, [1024]> obj_211_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_211_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(611003264)))];
            tensor<fp16, []> obj_211_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_211_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("obj_211_cast_fp16")];
            tensor<string, []> query_61_pad_type_0 = const()[name = tensor<string, []>("query_61_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_61_strides_0 = const()[name = tensor<string, []>("query_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_61_pad_0 = const()[name = tensor<string, []>("query_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_61_dilations_0 = const()[name = tensor<string, []>("query_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_61_groups_0 = const()[name = tensor<string, []>("query_61_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(611005376)))];
            tensor<fp16, [1024]> layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613102592)))];
            tensor<fp16, [1, 1024, 1, 1]> query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
            tensor<string, []> current_key_31_pad_type_0 = const()[name = tensor<string, []>("current_key_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_31_strides_0 = const()[name = tensor<string, []>("current_key_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_31_pad_0 = const()[name = tensor<string, []>("current_key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_31_dilations_0 = const()[name = tensor<string, []>("current_key_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_31_groups_0 = const()[name = tensor<string, []>("current_key_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613104704)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_31_cast_fp16 = conv(dilations = current_key_31_dilations_0, groups = current_key_31_groups_0, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = current_key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor<string, []>("current_key_31_cast_fp16")];
            tensor<string, []> current_value_31_pad_type_0 = const()[name = tensor<string, []>("current_value_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_31_strides_0 = const()[name = tensor<string, []>("current_value_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_31_pad_0 = const()[name = tensor<string, []>("current_value_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_31_dilations_0 = const()[name = tensor<string, []>("current_value_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_31_groups_0 = const()[name = tensor<string, []>("current_value_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(615201920)))];
            tensor<fp16, [1024]> layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617299136)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = current_value_31_dilations_0, groups = current_value_31_groups_0, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = current_value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor<string, []>("current_value_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3501_cast_fp16 = mul(x = var_87_cast_fp16_15, y = var_207_cast_fp16)[name = tensor<string, []>("op_3501_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3502_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3502_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_61_cast_fp16 = add(x = var_3501_cast_fp16, y = var_3502_cast_fp16)[name = tensor<string, []>("key_61_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3505_cast_fp16 = mul(x = var_114_cast_fp16_15, y = var_207_cast_fp16)[name = tensor<string, []>("op_3505_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3506_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3506_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_61_cast_fp16 = add(x = var_3505_cast_fp16, y = var_3506_cast_fp16)[name = tensor<string, []>("value_61_cast_fp16")];
            tensor<int32, [4]> var_3510 = const()[name = tensor<string, []>("op_3510"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_61_cast_fp16 = reshape(shape = var_3510, x = query_61_cast_fp16)[name = tensor<string, []>("mh_q_61_cast_fp16")];
            tensor<fp16, []> var_3512_to_fp16 = const()[name = tensor<string, []>("op_3512_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3513_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3512_to_fp16)[name = tensor<string, []>("op_3513_cast_fp16")];
            tensor<int32, [4]> var_3516 = const()[name = tensor<string, []>("op_3516"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3517_cast_fp16 = reshape(shape = var_3516, x = key_61_cast_fp16)[name = tensor<string, []>("op_3517_cast_fp16")];
            tensor<bool, []> mh_w_91_transpose_x_0 = const()[name = tensor<string, []>("mh_w_91_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_91_transpose_y_0 = const()[name = tensor<string, []>("mh_w_91_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3513_cast_fp16, y = var_3517_cast_fp16)[name = tensor<string, []>("mh_w_91_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_93_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_3525_cast_fp16 = softmax(axis = var_3437, x = mh_w_93_cast_fp16)[name = tensor<string, []>("op_3525_cast_fp16")];
            tensor<int32, [4]> var_3526 = const()[name = tensor<string, []>("op_3526"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3527_cast_fp16 = reshape(shape = var_3526, x = value_61_cast_fp16)[name = tensor<string, []>("op_3527_cast_fp16")];
            tensor<bool, []> attn_61_transpose_x_0 = const()[name = tensor<string, []>("attn_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_61_transpose_y_0 = const()[name = tensor<string, []>("attn_61_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3527_cast_fp16, y = var_3525_cast_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
            tensor<int32, [4]> var_3530 = const()[name = tensor<string, []>("op_3530"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_151_cast_fp16 = reshape(shape = var_3530, x = attn_61_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<string, []> obj_217_pad_type_0 = const()[name = tensor<string, []>("obj_217_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_217_strides_0 = const()[name = tensor<string, []>("obj_217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_217_pad_0 = const()[name = tensor<string, []>("obj_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_217_dilations_0 = const()[name = tensor<string, []>("obj_217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_217_groups_0 = const()[name = tensor<string, []>("obj_217_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(617301248)))];
            tensor<fp16, [1024]> layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(619398464)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_217_dilations_0, groups = obj_217_groups_0, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = obj_217_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("obj_217_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
            tensor<int32, [1]> out_93_axes_0 = const()[name = tensor<string, []>("out_93_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3552_to_fp16 = const()[name = tensor<string, []>("op_3552_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_3552_to_fp16, x = inputs_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
            tensor<fp16, [1024]> obj_219_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_219_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(619400576)))];
            tensor<fp16, [1024]> obj_219_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_219_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(619402688)))];
            tensor<fp16, []> obj_219_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_219_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_219_cast_fp16")];
            tensor<string, []> query_63_pad_type_0 = const()[name = tensor<string, []>("query_63_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_63_strides_0 = const()[name = tensor<string, []>("query_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_63_pad_0 = const()[name = tensor<string, []>("query_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_63_dilations_0 = const()[name = tensor<string, []>("query_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_63_groups_0 = const()[name = tensor<string, []>("query_63_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(619404800)))];
            tensor<fp16, [1024]> layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(621502016)))];
            tensor<fp16, [1, 1024, 1, 1]> query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = query_63_dilations_0, groups = query_63_groups_0, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = query_63_strides_0, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor<string, []>("query_63_cast_fp16")];
            tensor<string, []> key_63_pad_type_0 = const()[name = tensor<string, []>("key_63_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_63_strides_0 = const()[name = tensor<string, []>("key_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_63_pad_0 = const()[name = tensor<string, []>("key_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_63_dilations_0 = const()[name = tensor<string, []>("key_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_63_groups_0 = const()[name = tensor<string, []>("key_63_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(621504128)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_63_cast_fp16 = conv(dilations = key_63_dilations_0, groups = key_63_groups_0, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = key_63_strides_0, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_63_cast_fp16")];
            tensor<string, []> value_63_pad_type_0 = const()[name = tensor<string, []>("value_63_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_63_strides_0 = const()[name = tensor<string, []>("value_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_63_pad_0 = const()[name = tensor<string, []>("value_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_63_dilations_0 = const()[name = tensor<string, []>("value_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_63_groups_0 = const()[name = tensor<string, []>("value_63_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(623601344)))];
            tensor<fp16, [1024]> layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(625698560)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = value_63_dilations_0, groups = value_63_groups_0, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = value_63_strides_0, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_63_cast_fp16")];
            tensor<int32, [4]> var_3588 = const()[name = tensor<string, []>("op_3588"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_63_cast_fp16 = reshape(shape = var_3588, x = query_63_cast_fp16)[name = tensor<string, []>("mh_q_63_cast_fp16")];
            tensor<fp16, []> var_3590_to_fp16 = const()[name = tensor<string, []>("op_3590_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3591_cast_fp16 = mul(x = mh_q_63_cast_fp16, y = var_3590_to_fp16)[name = tensor<string, []>("op_3591_cast_fp16")];
            tensor<int32, [4]> var_3594 = const()[name = tensor<string, []>("op_3594"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3595_cast_fp16 = reshape(shape = var_3594, x = key_63_cast_fp16)[name = tensor<string, []>("op_3595_cast_fp16")];
            tensor<bool, []> mh_w_95_transpose_x_0 = const()[name = tensor<string, []>("mh_w_95_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_95_transpose_y_0 = const()[name = tensor<string, []>("mh_w_95_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3591_cast_fp16, y = var_3595_cast_fp16)[name = tensor<string, []>("mh_w_95_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_223_cast_fp16 = softmax(axis = var_3437, x = mh_w_95_cast_fp16)[name = tensor<string, []>("obj_223_cast_fp16")];
            tensor<int32, [4]> var_3599 = const()[name = tensor<string, []>("op_3599"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3600_cast_fp16 = reshape(shape = var_3599, x = value_63_cast_fp16)[name = tensor<string, []>("op_3600_cast_fp16")];
            tensor<bool, []> attn_63_transpose_x_0 = const()[name = tensor<string, []>("attn_63_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_63_transpose_y_0 = const()[name = tensor<string, []>("attn_63_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3600_cast_fp16, y = obj_223_cast_fp16)[name = tensor<string, []>("attn_63_cast_fp16")];
            tensor<int32, [4]> var_3603 = const()[name = tensor<string, []>("op_3603"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_153_cast_fp16 = reshape(shape = var_3603, x = attn_63_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<string, []> obj_221_pad_type_0 = const()[name = tensor<string, []>("obj_221_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_221_strides_0 = const()[name = tensor<string, []>("obj_221_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_221_pad_0 = const()[name = tensor<string, []>("obj_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_221_dilations_0 = const()[name = tensor<string, []>("obj_221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_221_groups_0 = const()[name = tensor<string, []>("obj_221_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(625700672)))];
            tensor<fp16, [1024]> layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627797888)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = obj_221_dilations_0, groups = obj_221_groups_0, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = obj_221_strides_0, weight = layers_15_encoder_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("obj_221_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor<string, []>("inputs_95_cast_fp16")];
            tensor<int32, [1]> out_95_axes_0 = const()[name = tensor<string, []>("out_95_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3624_to_fp16 = const()[name = tensor<string, []>("op_3624_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_3624_to_fp16, x = inputs_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
            tensor<fp16, [1024]> input_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627800000)))];
            tensor<fp16, [1024]> input_155_beta_0_to_fp16 = const()[name = tensor<string, []>("input_155_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627802112)))];
            tensor<fp16, []> input_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_157_strides_0 = const()[name = tensor<string, []>("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_157_dilations_0 = const()[name = tensor<string, []>("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_157_groups_0 = const()[name = tensor<string, []>("input_157_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627804224)))];
            tensor<fp16, [4096]> layers_15_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636192896)))];
            tensor<fp16, [1, 4096, 1, 1]> input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<string, []> input_159_mode_0 = const()[name = tensor<string, []>("input_159_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<string, []> hidden_states_33_pad_type_0 = const()[name = tensor<string, []>("hidden_states_33_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = tensor<string, []>("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = tensor<string, []>("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = tensor<string, []>("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_33_groups_0 = const()[name = tensor<string, []>("hidden_states_33_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_15_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(636201152)))];
            tensor<fp16, [1024]> layers_15_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644589824)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor<string, []>("inputs_97_cast_fp16")];
            tensor<int32, []> var_3660 = const()[name = tensor<string, []>("op_3660"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_97_axes_0 = const()[name = tensor<string, []>("out_97_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3685_to_fp16 = const()[name = tensor<string, []>("op_3685_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3685_to_fp16, x = inputs_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
            tensor<fp16, [1024]> obj_225_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_225_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644591936)))];
            tensor<fp16, [1024]> obj_225_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_225_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644594048)))];
            tensor<fp16, []> obj_225_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_225_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("obj_225_cast_fp16")];
            tensor<string, []> query_65_pad_type_0 = const()[name = tensor<string, []>("query_65_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_65_strides_0 = const()[name = tensor<string, []>("query_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_65_pad_0 = const()[name = tensor<string, []>("query_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_65_dilations_0 = const()[name = tensor<string, []>("query_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_65_groups_0 = const()[name = tensor<string, []>("query_65_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(644596160)))];
            tensor<fp16, [1024]> layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(646693376)))];
            tensor<fp16, [1, 1024, 1, 1]> query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor<string, []>("query_65_cast_fp16")];
            tensor<string, []> current_key_33_pad_type_0 = const()[name = tensor<string, []>("current_key_33_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_33_strides_0 = const()[name = tensor<string, []>("current_key_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_33_pad_0 = const()[name = tensor<string, []>("current_key_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_33_dilations_0 = const()[name = tensor<string, []>("current_key_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_33_groups_0 = const()[name = tensor<string, []>("current_key_33_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(646695488)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_33_cast_fp16 = conv(dilations = current_key_33_dilations_0, groups = current_key_33_groups_0, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = current_key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor<string, []>("current_key_33_cast_fp16")];
            tensor<string, []> current_value_33_pad_type_0 = const()[name = tensor<string, []>("current_value_33_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_33_strides_0 = const()[name = tensor<string, []>("current_value_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_33_pad_0 = const()[name = tensor<string, []>("current_value_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_33_dilations_0 = const()[name = tensor<string, []>("current_value_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_33_groups_0 = const()[name = tensor<string, []>("current_value_33_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648792704)))];
            tensor<fp16, [1024]> layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(650889920)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = current_value_33_dilations_0, groups = current_value_33_groups_0, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = current_value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor<string, []>("current_value_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3724_cast_fp16 = mul(x = var_87_cast_fp16_16, y = var_207_cast_fp16)[name = tensor<string, []>("op_3724_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3725_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3725_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_65_cast_fp16 = add(x = var_3724_cast_fp16, y = var_3725_cast_fp16)[name = tensor<string, []>("key_65_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3728_cast_fp16 = mul(x = var_114_cast_fp16_16, y = var_207_cast_fp16)[name = tensor<string, []>("op_3728_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3729_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3729_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_65_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_cast_fp16)[name = tensor<string, []>("value_65_cast_fp16")];
            tensor<int32, [4]> var_3733 = const()[name = tensor<string, []>("op_3733"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_65_cast_fp16 = reshape(shape = var_3733, x = query_65_cast_fp16)[name = tensor<string, []>("mh_q_65_cast_fp16")];
            tensor<fp16, []> var_3735_to_fp16 = const()[name = tensor<string, []>("op_3735_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3736_cast_fp16 = mul(x = mh_q_65_cast_fp16, y = var_3735_to_fp16)[name = tensor<string, []>("op_3736_cast_fp16")];
            tensor<int32, [4]> var_3739 = const()[name = tensor<string, []>("op_3739"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3740_cast_fp16 = reshape(shape = var_3739, x = key_65_cast_fp16)[name = tensor<string, []>("op_3740_cast_fp16")];
            tensor<bool, []> mh_w_97_transpose_x_0 = const()[name = tensor<string, []>("mh_w_97_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_97_transpose_y_0 = const()[name = tensor<string, []>("mh_w_97_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3736_cast_fp16, y = var_3740_cast_fp16)[name = tensor<string, []>("mh_w_97_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_99_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_3748_cast_fp16 = softmax(axis = var_3660, x = mh_w_99_cast_fp16)[name = tensor<string, []>("op_3748_cast_fp16")];
            tensor<int32, [4]> var_3749 = const()[name = tensor<string, []>("op_3749"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3750_cast_fp16 = reshape(shape = var_3749, x = value_65_cast_fp16)[name = tensor<string, []>("op_3750_cast_fp16")];
            tensor<bool, []> attn_65_transpose_x_0 = const()[name = tensor<string, []>("attn_65_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_65_transpose_y_0 = const()[name = tensor<string, []>("attn_65_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3750_cast_fp16, y = var_3748_cast_fp16)[name = tensor<string, []>("attn_65_cast_fp16")];
            tensor<int32, [4]> var_3753 = const()[name = tensor<string, []>("op_3753"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_161_cast_fp16 = reshape(shape = var_3753, x = attn_65_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<string, []> obj_231_pad_type_0 = const()[name = tensor<string, []>("obj_231_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_231_strides_0 = const()[name = tensor<string, []>("obj_231_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_231_pad_0 = const()[name = tensor<string, []>("obj_231_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_231_dilations_0 = const()[name = tensor<string, []>("obj_231_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_231_groups_0 = const()[name = tensor<string, []>("obj_231_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(650892032)))];
            tensor<fp16, [1024]> layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652989248)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_231_dilations_0, groups = obj_231_groups_0, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = obj_231_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("obj_231_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
            tensor<int32, [1]> out_99_axes_0 = const()[name = tensor<string, []>("out_99_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3775_to_fp16 = const()[name = tensor<string, []>("op_3775_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3775_to_fp16, x = inputs_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
            tensor<fp16, [1024]> obj_233_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_233_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652991360)))];
            tensor<fp16, [1024]> obj_233_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_233_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652993472)))];
            tensor<fp16, []> obj_233_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_233_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("obj_233_cast_fp16")];
            tensor<string, []> query_67_pad_type_0 = const()[name = tensor<string, []>("query_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_67_strides_0 = const()[name = tensor<string, []>("query_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_67_pad_0 = const()[name = tensor<string, []>("query_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_67_dilations_0 = const()[name = tensor<string, []>("query_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_67_groups_0 = const()[name = tensor<string, []>("query_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652995584)))];
            tensor<fp16, [1024]> layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655092800)))];
            tensor<fp16, [1, 1024, 1, 1]> query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = query_67_dilations_0, groups = query_67_groups_0, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = query_67_strides_0, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor<string, []>("query_67_cast_fp16")];
            tensor<string, []> key_67_pad_type_0 = const()[name = tensor<string, []>("key_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_67_strides_0 = const()[name = tensor<string, []>("key_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_67_pad_0 = const()[name = tensor<string, []>("key_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_67_dilations_0 = const()[name = tensor<string, []>("key_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_67_groups_0 = const()[name = tensor<string, []>("key_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655094912)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_67_cast_fp16 = conv(dilations = key_67_dilations_0, groups = key_67_groups_0, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = key_67_strides_0, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_67_cast_fp16")];
            tensor<string, []> value_67_pad_type_0 = const()[name = tensor<string, []>("value_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_67_strides_0 = const()[name = tensor<string, []>("value_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_67_pad_0 = const()[name = tensor<string, []>("value_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_67_dilations_0 = const()[name = tensor<string, []>("value_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_67_groups_0 = const()[name = tensor<string, []>("value_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(657192128)))];
            tensor<fp16, [1024]> layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659289344)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = value_67_dilations_0, groups = value_67_groups_0, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = value_67_strides_0, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_67_cast_fp16")];
            tensor<int32, [4]> var_3811 = const()[name = tensor<string, []>("op_3811"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_67_cast_fp16 = reshape(shape = var_3811, x = query_67_cast_fp16)[name = tensor<string, []>("mh_q_67_cast_fp16")];
            tensor<fp16, []> var_3813_to_fp16 = const()[name = tensor<string, []>("op_3813_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3814_cast_fp16 = mul(x = mh_q_67_cast_fp16, y = var_3813_to_fp16)[name = tensor<string, []>("op_3814_cast_fp16")];
            tensor<int32, [4]> var_3817 = const()[name = tensor<string, []>("op_3817"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3818_cast_fp16 = reshape(shape = var_3817, x = key_67_cast_fp16)[name = tensor<string, []>("op_3818_cast_fp16")];
            tensor<bool, []> mh_w_101_transpose_x_0 = const()[name = tensor<string, []>("mh_w_101_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_101_transpose_y_0 = const()[name = tensor<string, []>("mh_w_101_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3814_cast_fp16, y = var_3818_cast_fp16)[name = tensor<string, []>("mh_w_101_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_237_cast_fp16 = softmax(axis = var_3660, x = mh_w_101_cast_fp16)[name = tensor<string, []>("obj_237_cast_fp16")];
            tensor<int32, [4]> var_3822 = const()[name = tensor<string, []>("op_3822"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_3823_cast_fp16 = reshape(shape = var_3822, x = value_67_cast_fp16)[name = tensor<string, []>("op_3823_cast_fp16")];
            tensor<bool, []> attn_67_transpose_x_0 = const()[name = tensor<string, []>("attn_67_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_67_transpose_y_0 = const()[name = tensor<string, []>("attn_67_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3823_cast_fp16, y = obj_237_cast_fp16)[name = tensor<string, []>("attn_67_cast_fp16")];
            tensor<int32, [4]> var_3826 = const()[name = tensor<string, []>("op_3826"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_163_cast_fp16 = reshape(shape = var_3826, x = attn_67_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<string, []> obj_235_pad_type_0 = const()[name = tensor<string, []>("obj_235_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_235_strides_0 = const()[name = tensor<string, []>("obj_235_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_235_pad_0 = const()[name = tensor<string, []>("obj_235_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_235_dilations_0 = const()[name = tensor<string, []>("obj_235_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_235_groups_0 = const()[name = tensor<string, []>("obj_235_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(659291456)))];
            tensor<fp16, [1024]> layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(661388672)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = obj_235_dilations_0, groups = obj_235_groups_0, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = obj_235_strides_0, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("obj_235_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
            tensor<int32, [1]> out_101_axes_0 = const()[name = tensor<string, []>("out_101_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3847_to_fp16 = const()[name = tensor<string, []>("op_3847_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3847_to_fp16, x = inputs_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
            tensor<fp16, [1024]> input_165_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_165_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(661390784)))];
            tensor<fp16, [1024]> input_165_beta_0_to_fp16 = const()[name = tensor<string, []>("input_165_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(661392896)))];
            tensor<fp16, []> input_165_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_165_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<string, []> input_167_pad_type_0 = const()[name = tensor<string, []>("input_167_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_167_strides_0 = const()[name = tensor<string, []>("input_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_167_pad_0 = const()[name = tensor<string, []>("input_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_167_dilations_0 = const()[name = tensor<string, []>("input_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_167_groups_0 = const()[name = tensor<string, []>("input_167_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(661395008)))];
            tensor<fp16, [4096]> layers_16_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669783680)))];
            tensor<fp16, [1, 4096, 1, 1]> input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<string, []> input_169_mode_0 = const()[name = tensor<string, []>("input_169_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<string, []> hidden_states_35_pad_type_0 = const()[name = tensor<string, []>("hidden_states_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_35_strides_0 = const()[name = tensor<string, []>("hidden_states_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = tensor<string, []>("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_35_dilations_0 = const()[name = tensor<string, []>("hidden_states_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_35_groups_0 = const()[name = tensor<string, []>("hidden_states_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_16_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669791936)))];
            tensor<fp16, [1024]> layers_16_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678180608)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("hidden_states_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
            tensor<int32, []> var_3883 = const()[name = tensor<string, []>("op_3883"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_103_axes_0 = const()[name = tensor<string, []>("out_103_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3908_to_fp16 = const()[name = tensor<string, []>("op_3908_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3908_to_fp16, x = inputs_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
            tensor<fp16, [1024]> obj_239_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_239_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678182720)))];
            tensor<fp16, [1024]> obj_239_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_239_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678184832)))];
            tensor<fp16, []> obj_239_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_239_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("obj_239_cast_fp16")];
            tensor<string, []> query_69_pad_type_0 = const()[name = tensor<string, []>("query_69_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_69_strides_0 = const()[name = tensor<string, []>("query_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_69_pad_0 = const()[name = tensor<string, []>("query_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_69_dilations_0 = const()[name = tensor<string, []>("query_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_69_groups_0 = const()[name = tensor<string, []>("query_69_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678186944)))];
            tensor<fp16, [1024]> layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680284160)))];
            tensor<fp16, [1, 1024, 1, 1]> query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor<string, []>("query_69_cast_fp16")];
            tensor<string, []> current_key_35_pad_type_0 = const()[name = tensor<string, []>("current_key_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_35_strides_0 = const()[name = tensor<string, []>("current_key_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_35_pad_0 = const()[name = tensor<string, []>("current_key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_35_dilations_0 = const()[name = tensor<string, []>("current_key_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_35_groups_0 = const()[name = tensor<string, []>("current_key_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(680286272)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_35_cast_fp16 = conv(dilations = current_key_35_dilations_0, groups = current_key_35_groups_0, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = current_key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor<string, []>("current_key_35_cast_fp16")];
            tensor<string, []> current_value_35_pad_type_0 = const()[name = tensor<string, []>("current_value_35_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_35_strides_0 = const()[name = tensor<string, []>("current_value_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_35_pad_0 = const()[name = tensor<string, []>("current_value_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_35_dilations_0 = const()[name = tensor<string, []>("current_value_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_35_groups_0 = const()[name = tensor<string, []>("current_value_35_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(682383488)))];
            tensor<fp16, [1024]> layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684480704)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = current_value_35_dilations_0, groups = current_value_35_groups_0, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = current_value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor<string, []>("current_value_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3947_cast_fp16 = mul(x = var_87_cast_fp16_17, y = var_207_cast_fp16)[name = tensor<string, []>("op_3947_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3948_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3948_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_69_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3948_cast_fp16)[name = tensor<string, []>("key_69_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3951_cast_fp16 = mul(x = var_114_cast_fp16_17, y = var_207_cast_fp16)[name = tensor<string, []>("op_3951_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_3952_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_3952_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_69_cast_fp16 = add(x = var_3951_cast_fp16, y = var_3952_cast_fp16)[name = tensor<string, []>("value_69_cast_fp16")];
            tensor<int32, [4]> var_3956 = const()[name = tensor<string, []>("op_3956"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_69_cast_fp16 = reshape(shape = var_3956, x = query_69_cast_fp16)[name = tensor<string, []>("mh_q_69_cast_fp16")];
            tensor<fp16, []> var_3958_to_fp16 = const()[name = tensor<string, []>("op_3958_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_3959_cast_fp16 = mul(x = mh_q_69_cast_fp16, y = var_3958_to_fp16)[name = tensor<string, []>("op_3959_cast_fp16")];
            tensor<int32, [4]> var_3962 = const()[name = tensor<string, []>("op_3962"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3963_cast_fp16 = reshape(shape = var_3962, x = key_69_cast_fp16)[name = tensor<string, []>("op_3963_cast_fp16")];
            tensor<bool, []> mh_w_103_transpose_x_0 = const()[name = tensor<string, []>("mh_w_103_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_103_transpose_y_0 = const()[name = tensor<string, []>("mh_w_103_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3959_cast_fp16, y = var_3963_cast_fp16)[name = tensor<string, []>("mh_w_103_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_105_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_3971_cast_fp16 = softmax(axis = var_3883, x = mh_w_105_cast_fp16)[name = tensor<string, []>("op_3971_cast_fp16")];
            tensor<int32, [4]> var_3972 = const()[name = tensor<string, []>("op_3972"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_3973_cast_fp16 = reshape(shape = var_3972, x = value_69_cast_fp16)[name = tensor<string, []>("op_3973_cast_fp16")];
            tensor<bool, []> attn_69_transpose_x_0 = const()[name = tensor<string, []>("attn_69_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_69_transpose_y_0 = const()[name = tensor<string, []>("attn_69_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_3973_cast_fp16, y = var_3971_cast_fp16)[name = tensor<string, []>("attn_69_cast_fp16")];
            tensor<int32, [4]> var_3976 = const()[name = tensor<string, []>("op_3976"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_171_cast_fp16 = reshape(shape = var_3976, x = attn_69_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<string, []> obj_245_pad_type_0 = const()[name = tensor<string, []>("obj_245_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_245_strides_0 = const()[name = tensor<string, []>("obj_245_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_245_pad_0 = const()[name = tensor<string, []>("obj_245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_245_dilations_0 = const()[name = tensor<string, []>("obj_245_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_245_groups_0 = const()[name = tensor<string, []>("obj_245_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(684482816)))];
            tensor<fp16, [1024]> layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686580032)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_245_dilations_0, groups = obj_245_groups_0, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = obj_245_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("obj_245_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor<string, []>("inputs_105_cast_fp16")];
            tensor<int32, [1]> out_105_axes_0 = const()[name = tensor<string, []>("out_105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_3998_to_fp16 = const()[name = tensor<string, []>("op_3998_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3998_to_fp16, x = inputs_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
            tensor<fp16, [1024]> obj_247_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_247_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686582144)))];
            tensor<fp16, [1024]> obj_247_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_247_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686584256)))];
            tensor<fp16, []> obj_247_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_247_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("obj_247_cast_fp16")];
            tensor<string, []> query_71_pad_type_0 = const()[name = tensor<string, []>("query_71_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_71_strides_0 = const()[name = tensor<string, []>("query_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_71_pad_0 = const()[name = tensor<string, []>("query_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_71_dilations_0 = const()[name = tensor<string, []>("query_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_71_groups_0 = const()[name = tensor<string, []>("query_71_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(686586368)))];
            tensor<fp16, [1024]> layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688683584)))];
            tensor<fp16, [1, 1024, 1, 1]> query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = query_71_dilations_0, groups = query_71_groups_0, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = query_71_strides_0, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor<string, []>("query_71_cast_fp16")];
            tensor<string, []> key_71_pad_type_0 = const()[name = tensor<string, []>("key_71_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_71_strides_0 = const()[name = tensor<string, []>("key_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_71_pad_0 = const()[name = tensor<string, []>("key_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_71_dilations_0 = const()[name = tensor<string, []>("key_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_71_groups_0 = const()[name = tensor<string, []>("key_71_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688685696)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_71_cast_fp16 = conv(dilations = key_71_dilations_0, groups = key_71_groups_0, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = key_71_strides_0, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_71_cast_fp16")];
            tensor<string, []> value_71_pad_type_0 = const()[name = tensor<string, []>("value_71_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_71_strides_0 = const()[name = tensor<string, []>("value_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_71_pad_0 = const()[name = tensor<string, []>("value_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_71_dilations_0 = const()[name = tensor<string, []>("value_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_71_groups_0 = const()[name = tensor<string, []>("value_71_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(690782912)))];
            tensor<fp16, [1024]> layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(692880128)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = value_71_dilations_0, groups = value_71_groups_0, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = value_71_strides_0, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_71_cast_fp16")];
            tensor<int32, [4]> var_4034 = const()[name = tensor<string, []>("op_4034"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_71_cast_fp16 = reshape(shape = var_4034, x = query_71_cast_fp16)[name = tensor<string, []>("mh_q_71_cast_fp16")];
            tensor<fp16, []> var_4036_to_fp16 = const()[name = tensor<string, []>("op_4036_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4037_cast_fp16 = mul(x = mh_q_71_cast_fp16, y = var_4036_to_fp16)[name = tensor<string, []>("op_4037_cast_fp16")];
            tensor<int32, [4]> var_4040 = const()[name = tensor<string, []>("op_4040"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4041_cast_fp16 = reshape(shape = var_4040, x = key_71_cast_fp16)[name = tensor<string, []>("op_4041_cast_fp16")];
            tensor<bool, []> mh_w_107_transpose_x_0 = const()[name = tensor<string, []>("mh_w_107_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_107_transpose_y_0 = const()[name = tensor<string, []>("mh_w_107_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_4037_cast_fp16, y = var_4041_cast_fp16)[name = tensor<string, []>("mh_w_107_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_251_cast_fp16 = softmax(axis = var_3883, x = mh_w_107_cast_fp16)[name = tensor<string, []>("obj_251_cast_fp16")];
            tensor<int32, [4]> var_4045 = const()[name = tensor<string, []>("op_4045"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4046_cast_fp16 = reshape(shape = var_4045, x = value_71_cast_fp16)[name = tensor<string, []>("op_4046_cast_fp16")];
            tensor<bool, []> attn_71_transpose_x_0 = const()[name = tensor<string, []>("attn_71_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_71_transpose_y_0 = const()[name = tensor<string, []>("attn_71_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_4046_cast_fp16, y = obj_251_cast_fp16)[name = tensor<string, []>("attn_71_cast_fp16")];
            tensor<int32, [4]> var_4049 = const()[name = tensor<string, []>("op_4049"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_173_cast_fp16 = reshape(shape = var_4049, x = attn_71_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<string, []> obj_249_pad_type_0 = const()[name = tensor<string, []>("obj_249_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_249_strides_0 = const()[name = tensor<string, []>("obj_249_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_249_pad_0 = const()[name = tensor<string, []>("obj_249_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_249_dilations_0 = const()[name = tensor<string, []>("obj_249_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_249_groups_0 = const()[name = tensor<string, []>("obj_249_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(692882240)))];
            tensor<fp16, [1024]> layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694979456)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = obj_249_dilations_0, groups = obj_249_groups_0, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = obj_249_strides_0, weight = layers_17_encoder_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("obj_249_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor<string, []>("inputs_107_cast_fp16")];
            tensor<int32, [1]> out_107_axes_0 = const()[name = tensor<string, []>("out_107_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4070_to_fp16 = const()[name = tensor<string, []>("op_4070_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_4070_to_fp16, x = inputs_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
            tensor<fp16, [1024]> input_175_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_175_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694981568)))];
            tensor<fp16, [1024]> input_175_beta_0_to_fp16 = const()[name = tensor<string, []>("input_175_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694983680)))];
            tensor<fp16, []> input_175_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_175_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<string, []> input_177_pad_type_0 = const()[name = tensor<string, []>("input_177_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_177_strides_0 = const()[name = tensor<string, []>("input_177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_177_pad_0 = const()[name = tensor<string, []>("input_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_177_dilations_0 = const()[name = tensor<string, []>("input_177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_177_groups_0 = const()[name = tensor<string, []>("input_177_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694985792)))];
            tensor<fp16, [4096]> layers_17_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703374464)))];
            tensor<fp16, [1, 4096, 1, 1]> input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<string, []> input_179_mode_0 = const()[name = tensor<string, []>("input_179_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<string, []> hidden_states_37_pad_type_0 = const()[name = tensor<string, []>("hidden_states_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = tensor<string, []>("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = tensor<string, []>("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = tensor<string, []>("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_37_groups_0 = const()[name = tensor<string, []>("hidden_states_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_17_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703382720)))];
            tensor<fp16, [1024]> layers_17_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711771392)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
            tensor<int32, []> var_4106 = const()[name = tensor<string, []>("op_4106"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_109_axes_0 = const()[name = tensor<string, []>("out_109_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4131_to_fp16 = const()[name = tensor<string, []>("op_4131_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_4131_to_fp16, x = inputs_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
            tensor<fp16, [1024]> obj_253_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_253_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711773504)))];
            tensor<fp16, [1024]> obj_253_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_253_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711775616)))];
            tensor<fp16, []> obj_253_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_253_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("obj_253_cast_fp16")];
            tensor<string, []> query_73_pad_type_0 = const()[name = tensor<string, []>("query_73_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_73_strides_0 = const()[name = tensor<string, []>("query_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_73_pad_0 = const()[name = tensor<string, []>("query_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_73_dilations_0 = const()[name = tensor<string, []>("query_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_73_groups_0 = const()[name = tensor<string, []>("query_73_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(711777728)))];
            tensor<fp16, [1024]> layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713874944)))];
            tensor<fp16, [1, 1024, 1, 1]> query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor<string, []>("query_73_cast_fp16")];
            tensor<string, []> current_key_37_pad_type_0 = const()[name = tensor<string, []>("current_key_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_37_strides_0 = const()[name = tensor<string, []>("current_key_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_37_pad_0 = const()[name = tensor<string, []>("current_key_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_37_dilations_0 = const()[name = tensor<string, []>("current_key_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_37_groups_0 = const()[name = tensor<string, []>("current_key_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713877056)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_37_cast_fp16 = conv(dilations = current_key_37_dilations_0, groups = current_key_37_groups_0, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = current_key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor<string, []>("current_key_37_cast_fp16")];
            tensor<string, []> current_value_37_pad_type_0 = const()[name = tensor<string, []>("current_value_37_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_37_strides_0 = const()[name = tensor<string, []>("current_value_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_37_pad_0 = const()[name = tensor<string, []>("current_value_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_37_dilations_0 = const()[name = tensor<string, []>("current_value_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_37_groups_0 = const()[name = tensor<string, []>("current_value_37_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(715974272)))];
            tensor<fp16, [1024]> layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(718071488)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = current_value_37_dilations_0, groups = current_value_37_groups_0, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = current_value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor<string, []>("current_value_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4170_cast_fp16 = mul(x = var_87_cast_fp16_18, y = var_207_cast_fp16)[name = tensor<string, []>("op_4170_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4171_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4171_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_73_cast_fp16 = add(x = var_4170_cast_fp16, y = var_4171_cast_fp16)[name = tensor<string, []>("key_73_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4174_cast_fp16 = mul(x = var_114_cast_fp16_18, y = var_207_cast_fp16)[name = tensor<string, []>("op_4174_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4175_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4175_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_73_cast_fp16 = add(x = var_4174_cast_fp16, y = var_4175_cast_fp16)[name = tensor<string, []>("value_73_cast_fp16")];
            tensor<int32, [4]> var_4179 = const()[name = tensor<string, []>("op_4179"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_73_cast_fp16 = reshape(shape = var_4179, x = query_73_cast_fp16)[name = tensor<string, []>("mh_q_73_cast_fp16")];
            tensor<fp16, []> var_4181_to_fp16 = const()[name = tensor<string, []>("op_4181_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4182_cast_fp16 = mul(x = mh_q_73_cast_fp16, y = var_4181_to_fp16)[name = tensor<string, []>("op_4182_cast_fp16")];
            tensor<int32, [4]> var_4185 = const()[name = tensor<string, []>("op_4185"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4186_cast_fp16 = reshape(shape = var_4185, x = key_73_cast_fp16)[name = tensor<string, []>("op_4186_cast_fp16")];
            tensor<bool, []> mh_w_109_transpose_x_0 = const()[name = tensor<string, []>("mh_w_109_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_109_transpose_y_0 = const()[name = tensor<string, []>("mh_w_109_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_4182_cast_fp16, y = var_4186_cast_fp16)[name = tensor<string, []>("mh_w_109_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_111_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_4194_cast_fp16 = softmax(axis = var_4106, x = mh_w_111_cast_fp16)[name = tensor<string, []>("op_4194_cast_fp16")];
            tensor<int32, [4]> var_4195 = const()[name = tensor<string, []>("op_4195"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4196_cast_fp16 = reshape(shape = var_4195, x = value_73_cast_fp16)[name = tensor<string, []>("op_4196_cast_fp16")];
            tensor<bool, []> attn_73_transpose_x_0 = const()[name = tensor<string, []>("attn_73_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_73_transpose_y_0 = const()[name = tensor<string, []>("attn_73_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4196_cast_fp16, y = var_4194_cast_fp16)[name = tensor<string, []>("attn_73_cast_fp16")];
            tensor<int32, [4]> var_4199 = const()[name = tensor<string, []>("op_4199"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_181_cast_fp16 = reshape(shape = var_4199, x = attn_73_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<string, []> obj_259_pad_type_0 = const()[name = tensor<string, []>("obj_259_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_259_strides_0 = const()[name = tensor<string, []>("obj_259_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_259_pad_0 = const()[name = tensor<string, []>("obj_259_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_259_dilations_0 = const()[name = tensor<string, []>("obj_259_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_259_groups_0 = const()[name = tensor<string, []>("obj_259_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(718073600)))];
            tensor<fp16, [1024]> layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720170816)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_259_dilations_0, groups = obj_259_groups_0, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = obj_259_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("obj_259_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
            tensor<int32, [1]> out_111_axes_0 = const()[name = tensor<string, []>("out_111_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4221_to_fp16 = const()[name = tensor<string, []>("op_4221_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_4221_to_fp16, x = inputs_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
            tensor<fp16, [1024]> obj_261_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_261_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720172928)))];
            tensor<fp16, [1024]> obj_261_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_261_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720175040)))];
            tensor<fp16, []> obj_261_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_261_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("obj_261_cast_fp16")];
            tensor<string, []> query_75_pad_type_0 = const()[name = tensor<string, []>("query_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_75_strides_0 = const()[name = tensor<string, []>("query_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_75_pad_0 = const()[name = tensor<string, []>("query_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_75_dilations_0 = const()[name = tensor<string, []>("query_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_75_groups_0 = const()[name = tensor<string, []>("query_75_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720177152)))];
            tensor<fp16, [1024]> layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(722274368)))];
            tensor<fp16, [1, 1024, 1, 1]> query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = query_75_dilations_0, groups = query_75_groups_0, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = query_75_strides_0, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor<string, []>("query_75_cast_fp16")];
            tensor<string, []> key_75_pad_type_0 = const()[name = tensor<string, []>("key_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_75_strides_0 = const()[name = tensor<string, []>("key_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_75_pad_0 = const()[name = tensor<string, []>("key_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_75_dilations_0 = const()[name = tensor<string, []>("key_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_75_groups_0 = const()[name = tensor<string, []>("key_75_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(722276480)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_75_cast_fp16 = conv(dilations = key_75_dilations_0, groups = key_75_groups_0, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = key_75_strides_0, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_75_cast_fp16")];
            tensor<string, []> value_75_pad_type_0 = const()[name = tensor<string, []>("value_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_75_strides_0 = const()[name = tensor<string, []>("value_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_75_pad_0 = const()[name = tensor<string, []>("value_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_75_dilations_0 = const()[name = tensor<string, []>("value_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_75_groups_0 = const()[name = tensor<string, []>("value_75_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724373696)))];
            tensor<fp16, [1024]> layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726470912)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = value_75_dilations_0, groups = value_75_groups_0, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = value_75_strides_0, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_75_cast_fp16")];
            tensor<int32, [4]> var_4257 = const()[name = tensor<string, []>("op_4257"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_75_cast_fp16 = reshape(shape = var_4257, x = query_75_cast_fp16)[name = tensor<string, []>("mh_q_75_cast_fp16")];
            tensor<fp16, []> var_4259_to_fp16 = const()[name = tensor<string, []>("op_4259_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4260_cast_fp16 = mul(x = mh_q_75_cast_fp16, y = var_4259_to_fp16)[name = tensor<string, []>("op_4260_cast_fp16")];
            tensor<int32, [4]> var_4263 = const()[name = tensor<string, []>("op_4263"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4264_cast_fp16 = reshape(shape = var_4263, x = key_75_cast_fp16)[name = tensor<string, []>("op_4264_cast_fp16")];
            tensor<bool, []> mh_w_113_transpose_x_0 = const()[name = tensor<string, []>("mh_w_113_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_113_transpose_y_0 = const()[name = tensor<string, []>("mh_w_113_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4260_cast_fp16, y = var_4264_cast_fp16)[name = tensor<string, []>("mh_w_113_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_265_cast_fp16 = softmax(axis = var_4106, x = mh_w_113_cast_fp16)[name = tensor<string, []>("obj_265_cast_fp16")];
            tensor<int32, [4]> var_4268 = const()[name = tensor<string, []>("op_4268"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4269_cast_fp16 = reshape(shape = var_4268, x = value_75_cast_fp16)[name = tensor<string, []>("op_4269_cast_fp16")];
            tensor<bool, []> attn_75_transpose_x_0 = const()[name = tensor<string, []>("attn_75_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_75_transpose_y_0 = const()[name = tensor<string, []>("attn_75_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4269_cast_fp16, y = obj_265_cast_fp16)[name = tensor<string, []>("attn_75_cast_fp16")];
            tensor<int32, [4]> var_4272 = const()[name = tensor<string, []>("op_4272"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_183_cast_fp16 = reshape(shape = var_4272, x = attn_75_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<string, []> obj_263_pad_type_0 = const()[name = tensor<string, []>("obj_263_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_263_strides_0 = const()[name = tensor<string, []>("obj_263_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_263_pad_0 = const()[name = tensor<string, []>("obj_263_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_263_dilations_0 = const()[name = tensor<string, []>("obj_263_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_263_groups_0 = const()[name = tensor<string, []>("obj_263_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(726473024)))];
            tensor<fp16, [1024]> layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728570240)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = obj_263_dilations_0, groups = obj_263_groups_0, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = obj_263_strides_0, weight = layers_18_encoder_attn_o_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("obj_263_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
            tensor<int32, [1]> out_113_axes_0 = const()[name = tensor<string, []>("out_113_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4293_to_fp16 = const()[name = tensor<string, []>("op_4293_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_4293_to_fp16, x = inputs_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
            tensor<fp16, [1024]> input_185_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_185_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728572352)))];
            tensor<fp16, [1024]> input_185_beta_0_to_fp16 = const()[name = tensor<string, []>("input_185_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728574464)))];
            tensor<fp16, []> input_185_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_185_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<string, []> input_187_pad_type_0 = const()[name = tensor<string, []>("input_187_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_187_strides_0 = const()[name = tensor<string, []>("input_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_187_pad_0 = const()[name = tensor<string, []>("input_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_187_dilations_0 = const()[name = tensor<string, []>("input_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_187_groups_0 = const()[name = tensor<string, []>("input_187_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728576576)))];
            tensor<fp16, [4096]> layers_18_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736965248)))];
            tensor<fp16, [1, 4096, 1, 1]> input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<string, []> input_189_mode_0 = const()[name = tensor<string, []>("input_189_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<string, []> hidden_states_39_pad_type_0 = const()[name = tensor<string, []>("hidden_states_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_39_strides_0 = const()[name = tensor<string, []>("hidden_states_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = tensor<string, []>("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_39_dilations_0 = const()[name = tensor<string, []>("hidden_states_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_39_groups_0 = const()[name = tensor<string, []>("hidden_states_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_18_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(736973504)))];
            tensor<fp16, [1024]> layers_18_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_18_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(745362176)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("hidden_states_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor<string, []>("inputs_115_cast_fp16")];
            tensor<int32, []> var_4329 = const()[name = tensor<string, []>("op_4329"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_115_axes_0 = const()[name = tensor<string, []>("out_115_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4354_to_fp16 = const()[name = tensor<string, []>("op_4354_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_4354_to_fp16, x = inputs_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
            tensor<fp16, [1024]> obj_267_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_267_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(745364288)))];
            tensor<fp16, [1024]> obj_267_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_267_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(745366400)))];
            tensor<fp16, []> obj_267_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_267_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("obj_267_cast_fp16")];
            tensor<string, []> query_77_pad_type_0 = const()[name = tensor<string, []>("query_77_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_77_strides_0 = const()[name = tensor<string, []>("query_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_77_pad_0 = const()[name = tensor<string, []>("query_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_77_dilations_0 = const()[name = tensor<string, []>("query_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_77_groups_0 = const()[name = tensor<string, []>("query_77_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(745368512)))];
            tensor<fp16, [1024]> layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(747465728)))];
            tensor<fp16, [1, 1024, 1, 1]> query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor<string, []>("query_77_cast_fp16")];
            tensor<string, []> current_key_39_pad_type_0 = const()[name = tensor<string, []>("current_key_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_39_strides_0 = const()[name = tensor<string, []>("current_key_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_39_pad_0 = const()[name = tensor<string, []>("current_key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_39_dilations_0 = const()[name = tensor<string, []>("current_key_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_39_groups_0 = const()[name = tensor<string, []>("current_key_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(747467840)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_39_cast_fp16 = conv(dilations = current_key_39_dilations_0, groups = current_key_39_groups_0, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = current_key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor<string, []>("current_key_39_cast_fp16")];
            tensor<string, []> current_value_39_pad_type_0 = const()[name = tensor<string, []>("current_value_39_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_39_strides_0 = const()[name = tensor<string, []>("current_value_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_39_pad_0 = const()[name = tensor<string, []>("current_value_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_39_dilations_0 = const()[name = tensor<string, []>("current_value_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_39_groups_0 = const()[name = tensor<string, []>("current_value_39_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(749565056)))];
            tensor<fp16, [1024]> layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(751662272)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = current_value_39_dilations_0, groups = current_value_39_groups_0, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = current_value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor<string, []>("current_value_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4393_cast_fp16 = mul(x = var_87_cast_fp16_19, y = var_207_cast_fp16)[name = tensor<string, []>("op_4393_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4394_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4394_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_77_cast_fp16 = add(x = var_4393_cast_fp16, y = var_4394_cast_fp16)[name = tensor<string, []>("key_77_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4397_cast_fp16 = mul(x = var_114_cast_fp16_19, y = var_207_cast_fp16)[name = tensor<string, []>("op_4397_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4398_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4398_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_77_cast_fp16 = add(x = var_4397_cast_fp16, y = var_4398_cast_fp16)[name = tensor<string, []>("value_77_cast_fp16")];
            tensor<int32, [4]> var_4402 = const()[name = tensor<string, []>("op_4402"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_77_cast_fp16 = reshape(shape = var_4402, x = query_77_cast_fp16)[name = tensor<string, []>("mh_q_77_cast_fp16")];
            tensor<fp16, []> var_4404_to_fp16 = const()[name = tensor<string, []>("op_4404_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4405_cast_fp16 = mul(x = mh_q_77_cast_fp16, y = var_4404_to_fp16)[name = tensor<string, []>("op_4405_cast_fp16")];
            tensor<int32, [4]> var_4408 = const()[name = tensor<string, []>("op_4408"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4409_cast_fp16 = reshape(shape = var_4408, x = key_77_cast_fp16)[name = tensor<string, []>("op_4409_cast_fp16")];
            tensor<bool, []> mh_w_115_transpose_x_0 = const()[name = tensor<string, []>("mh_w_115_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_115_transpose_y_0 = const()[name = tensor<string, []>("mh_w_115_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4405_cast_fp16, y = var_4409_cast_fp16)[name = tensor<string, []>("mh_w_115_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_117_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_4417_cast_fp16 = softmax(axis = var_4329, x = mh_w_117_cast_fp16)[name = tensor<string, []>("op_4417_cast_fp16")];
            tensor<int32, [4]> var_4418 = const()[name = tensor<string, []>("op_4418"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4419_cast_fp16 = reshape(shape = var_4418, x = value_77_cast_fp16)[name = tensor<string, []>("op_4419_cast_fp16")];
            tensor<bool, []> attn_77_transpose_x_0 = const()[name = tensor<string, []>("attn_77_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_77_transpose_y_0 = const()[name = tensor<string, []>("attn_77_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4419_cast_fp16, y = var_4417_cast_fp16)[name = tensor<string, []>("attn_77_cast_fp16")];
            tensor<int32, [4]> var_4422 = const()[name = tensor<string, []>("op_4422"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_191_cast_fp16 = reshape(shape = var_4422, x = attn_77_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<string, []> obj_273_pad_type_0 = const()[name = tensor<string, []>("obj_273_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_273_strides_0 = const()[name = tensor<string, []>("obj_273_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_273_pad_0 = const()[name = tensor<string, []>("obj_273_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_273_dilations_0 = const()[name = tensor<string, []>("obj_273_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_273_groups_0 = const()[name = tensor<string, []>("obj_273_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(751664384)))];
            tensor<fp16, [1024]> layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753761600)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_273_dilations_0, groups = obj_273_groups_0, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = obj_273_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("obj_273_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor<string, []>("inputs_117_cast_fp16")];
            tensor<int32, [1]> out_117_axes_0 = const()[name = tensor<string, []>("out_117_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4444_to_fp16 = const()[name = tensor<string, []>("op_4444_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_4444_to_fp16, x = inputs_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
            tensor<fp16, [1024]> obj_275_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_275_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753763712)))];
            tensor<fp16, [1024]> obj_275_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_275_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753765824)))];
            tensor<fp16, []> obj_275_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_275_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("obj_275_cast_fp16")];
            tensor<string, []> query_79_pad_type_0 = const()[name = tensor<string, []>("query_79_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_79_strides_0 = const()[name = tensor<string, []>("query_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_79_pad_0 = const()[name = tensor<string, []>("query_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_79_dilations_0 = const()[name = tensor<string, []>("query_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_79_groups_0 = const()[name = tensor<string, []>("query_79_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(753767936)))];
            tensor<fp16, [1024]> layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755865152)))];
            tensor<fp16, [1, 1024, 1, 1]> query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = query_79_dilations_0, groups = query_79_groups_0, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = query_79_strides_0, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor<string, []>("query_79_cast_fp16")];
            tensor<string, []> key_79_pad_type_0 = const()[name = tensor<string, []>("key_79_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_79_strides_0 = const()[name = tensor<string, []>("key_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_79_pad_0 = const()[name = tensor<string, []>("key_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_79_dilations_0 = const()[name = tensor<string, []>("key_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_79_groups_0 = const()[name = tensor<string, []>("key_79_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(755867264)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_79_cast_fp16 = conv(dilations = key_79_dilations_0, groups = key_79_groups_0, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = key_79_strides_0, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_79_cast_fp16")];
            tensor<string, []> value_79_pad_type_0 = const()[name = tensor<string, []>("value_79_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_79_strides_0 = const()[name = tensor<string, []>("value_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_79_pad_0 = const()[name = tensor<string, []>("value_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_79_dilations_0 = const()[name = tensor<string, []>("value_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_79_groups_0 = const()[name = tensor<string, []>("value_79_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(757964480)))];
            tensor<fp16, [1024]> layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760061696)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = value_79_dilations_0, groups = value_79_groups_0, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = value_79_strides_0, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_79_cast_fp16")];
            tensor<int32, [4]> var_4480 = const()[name = tensor<string, []>("op_4480"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_79_cast_fp16 = reshape(shape = var_4480, x = query_79_cast_fp16)[name = tensor<string, []>("mh_q_79_cast_fp16")];
            tensor<fp16, []> var_4482_to_fp16 = const()[name = tensor<string, []>("op_4482_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4483_cast_fp16 = mul(x = mh_q_79_cast_fp16, y = var_4482_to_fp16)[name = tensor<string, []>("op_4483_cast_fp16")];
            tensor<int32, [4]> var_4486 = const()[name = tensor<string, []>("op_4486"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4487_cast_fp16 = reshape(shape = var_4486, x = key_79_cast_fp16)[name = tensor<string, []>("op_4487_cast_fp16")];
            tensor<bool, []> mh_w_119_transpose_x_0 = const()[name = tensor<string, []>("mh_w_119_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_119_transpose_y_0 = const()[name = tensor<string, []>("mh_w_119_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4483_cast_fp16, y = var_4487_cast_fp16)[name = tensor<string, []>("mh_w_119_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_279_cast_fp16 = softmax(axis = var_4329, x = mh_w_119_cast_fp16)[name = tensor<string, []>("obj_279_cast_fp16")];
            tensor<int32, [4]> var_4491 = const()[name = tensor<string, []>("op_4491"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4492_cast_fp16 = reshape(shape = var_4491, x = value_79_cast_fp16)[name = tensor<string, []>("op_4492_cast_fp16")];
            tensor<bool, []> attn_79_transpose_x_0 = const()[name = tensor<string, []>("attn_79_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_79_transpose_y_0 = const()[name = tensor<string, []>("attn_79_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4492_cast_fp16, y = obj_279_cast_fp16)[name = tensor<string, []>("attn_79_cast_fp16")];
            tensor<int32, [4]> var_4495 = const()[name = tensor<string, []>("op_4495"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_193_cast_fp16 = reshape(shape = var_4495, x = attn_79_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<string, []> obj_277_pad_type_0 = const()[name = tensor<string, []>("obj_277_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_277_strides_0 = const()[name = tensor<string, []>("obj_277_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_277_pad_0 = const()[name = tensor<string, []>("obj_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_277_dilations_0 = const()[name = tensor<string, []>("obj_277_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_277_groups_0 = const()[name = tensor<string, []>("obj_277_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(760063808)))];
            tensor<fp16, [1024]> layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762161024)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = obj_277_dilations_0, groups = obj_277_groups_0, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = obj_277_strides_0, weight = layers_19_encoder_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("obj_277_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
            tensor<int32, [1]> out_119_axes_0 = const()[name = tensor<string, []>("out_119_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4513_to_fp16 = const()[name = tensor<string, []>("op_4513_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_4513_to_fp16, x = inputs_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
            tensor<fp16, [1024]> input_195_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762163136)))];
            tensor<fp16, [1024]> input_195_beta_0_to_fp16 = const()[name = tensor<string, []>("input_195_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762165248)))];
            tensor<fp16, []> input_195_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_195_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<string, []> input_197_pad_type_0 = const()[name = tensor<string, []>("input_197_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_197_strides_0 = const()[name = tensor<string, []>("input_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_197_pad_0 = const()[name = tensor<string, []>("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_197_dilations_0 = const()[name = tensor<string, []>("input_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_197_groups_0 = const()[name = tensor<string, []>("input_197_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(762167360)))];
            tensor<fp16, [4096]> layers_19_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(770556032)))];
            tensor<fp16, [1, 4096, 1, 1]> input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<string, []> input_199_mode_0 = const()[name = tensor<string, []>("input_199_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<string, []> hidden_states_41_pad_type_0 = const()[name = tensor<string, []>("hidden_states_41_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_41_strides_0 = const()[name = tensor<string, []>("hidden_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_41_pad_0 = const()[name = tensor<string, []>("hidden_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_41_dilations_0 = const()[name = tensor<string, []>("hidden_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_41_groups_0 = const()[name = tensor<string, []>("hidden_states_41_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_19_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(770564288)))];
            tensor<fp16, [1024]> layers_19_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_19_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778952960)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
            tensor<int32, []> var_4548 = const()[name = tensor<string, []>("op_4548"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_121_axes_0 = const()[name = tensor<string, []>("out_121_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4573_to_fp16 = const()[name = tensor<string, []>("op_4573_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_4573_to_fp16, x = inputs_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
            tensor<fp16, [1024]> obj_281_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_281_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778955072)))];
            tensor<fp16, [1024]> obj_281_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_281_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778957184)))];
            tensor<fp16, []> obj_281_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_281_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("obj_281_cast_fp16")];
            tensor<string, []> query_81_pad_type_0 = const()[name = tensor<string, []>("query_81_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_81_strides_0 = const()[name = tensor<string, []>("query_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_81_pad_0 = const()[name = tensor<string, []>("query_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_81_dilations_0 = const()[name = tensor<string, []>("query_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_81_groups_0 = const()[name = tensor<string, []>("query_81_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778959296)))];
            tensor<fp16, [1024]> layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781056512)))];
            tensor<fp16, [1, 1024, 1, 1]> query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor<string, []>("query_81_cast_fp16")];
            tensor<string, []> current_key_41_pad_type_0 = const()[name = tensor<string, []>("current_key_41_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_41_strides_0 = const()[name = tensor<string, []>("current_key_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_41_pad_0 = const()[name = tensor<string, []>("current_key_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_41_dilations_0 = const()[name = tensor<string, []>("current_key_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_41_groups_0 = const()[name = tensor<string, []>("current_key_41_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781058624)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_41_cast_fp16 = conv(dilations = current_key_41_dilations_0, groups = current_key_41_groups_0, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = current_key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor<string, []>("current_key_41_cast_fp16")];
            tensor<string, []> current_value_41_pad_type_0 = const()[name = tensor<string, []>("current_value_41_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_41_strides_0 = const()[name = tensor<string, []>("current_value_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_41_pad_0 = const()[name = tensor<string, []>("current_value_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_41_dilations_0 = const()[name = tensor<string, []>("current_value_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_41_groups_0 = const()[name = tensor<string, []>("current_value_41_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783155840)))];
            tensor<fp16, [1024]> layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785253056)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = current_value_41_dilations_0, groups = current_value_41_groups_0, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = current_value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor<string, []>("current_value_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4612_cast_fp16 = mul(x = var_87_cast_fp16_20, y = var_207_cast_fp16)[name = tensor<string, []>("op_4612_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4613_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4613_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_81_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4613_cast_fp16)[name = tensor<string, []>("key_81_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4616_cast_fp16 = mul(x = var_114_cast_fp16_20, y = var_207_cast_fp16)[name = tensor<string, []>("op_4616_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4617_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4617_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_81_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4617_cast_fp16)[name = tensor<string, []>("value_81_cast_fp16")];
            tensor<int32, [4]> var_4621 = const()[name = tensor<string, []>("op_4621"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_81_cast_fp16 = reshape(shape = var_4621, x = query_81_cast_fp16)[name = tensor<string, []>("mh_q_81_cast_fp16")];
            tensor<fp16, []> var_4623_to_fp16 = const()[name = tensor<string, []>("op_4623_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4624_cast_fp16 = mul(x = mh_q_81_cast_fp16, y = var_4623_to_fp16)[name = tensor<string, []>("op_4624_cast_fp16")];
            tensor<int32, [4]> var_4627 = const()[name = tensor<string, []>("op_4627"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4628_cast_fp16 = reshape(shape = var_4627, x = key_81_cast_fp16)[name = tensor<string, []>("op_4628_cast_fp16")];
            tensor<bool, []> mh_w_121_transpose_x_0 = const()[name = tensor<string, []>("mh_w_121_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_121_transpose_y_0 = const()[name = tensor<string, []>("mh_w_121_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4624_cast_fp16, y = var_4628_cast_fp16)[name = tensor<string, []>("mh_w_121_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_123_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_4636_cast_fp16 = softmax(axis = var_4548, x = mh_w_123_cast_fp16)[name = tensor<string, []>("op_4636_cast_fp16")];
            tensor<int32, [4]> var_4637 = const()[name = tensor<string, []>("op_4637"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4638_cast_fp16 = reshape(shape = var_4637, x = value_81_cast_fp16)[name = tensor<string, []>("op_4638_cast_fp16")];
            tensor<bool, []> attn_81_transpose_x_0 = const()[name = tensor<string, []>("attn_81_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_81_transpose_y_0 = const()[name = tensor<string, []>("attn_81_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4638_cast_fp16, y = var_4636_cast_fp16)[name = tensor<string, []>("attn_81_cast_fp16")];
            tensor<int32, [4]> var_4641 = const()[name = tensor<string, []>("op_4641"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_201_cast_fp16 = reshape(shape = var_4641, x = attn_81_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
            tensor<string, []> obj_287_pad_type_0 = const()[name = tensor<string, []>("obj_287_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_287_strides_0 = const()[name = tensor<string, []>("obj_287_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_287_pad_0 = const()[name = tensor<string, []>("obj_287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_287_dilations_0 = const()[name = tensor<string, []>("obj_287_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_287_groups_0 = const()[name = tensor<string, []>("obj_287_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785255168)))];
            tensor<fp16, [1024]> layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787352384)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_287_dilations_0, groups = obj_287_groups_0, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = obj_287_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("obj_287_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
            tensor<int32, [1]> out_123_axes_0 = const()[name = tensor<string, []>("out_123_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4663_to_fp16 = const()[name = tensor<string, []>("op_4663_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_4663_to_fp16, x = inputs_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
            tensor<fp16, [1024]> obj_289_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_289_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787354496)))];
            tensor<fp16, [1024]> obj_289_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_289_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787356608)))];
            tensor<fp16, []> obj_289_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_289_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("obj_289_cast_fp16")];
            tensor<string, []> query_83_pad_type_0 = const()[name = tensor<string, []>("query_83_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_83_strides_0 = const()[name = tensor<string, []>("query_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_83_pad_0 = const()[name = tensor<string, []>("query_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_83_dilations_0 = const()[name = tensor<string, []>("query_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_83_groups_0 = const()[name = tensor<string, []>("query_83_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787358720)))];
            tensor<fp16, [1024]> layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(789455936)))];
            tensor<fp16, [1, 1024, 1, 1]> query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = query_83_dilations_0, groups = query_83_groups_0, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = query_83_strides_0, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor<string, []>("query_83_cast_fp16")];
            tensor<string, []> key_83_pad_type_0 = const()[name = tensor<string, []>("key_83_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_83_strides_0 = const()[name = tensor<string, []>("key_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_83_pad_0 = const()[name = tensor<string, []>("key_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_83_dilations_0 = const()[name = tensor<string, []>("key_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_83_groups_0 = const()[name = tensor<string, []>("key_83_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(789458048)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_83_cast_fp16 = conv(dilations = key_83_dilations_0, groups = key_83_groups_0, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = key_83_strides_0, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_83_cast_fp16")];
            tensor<string, []> value_83_pad_type_0 = const()[name = tensor<string, []>("value_83_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_83_strides_0 = const()[name = tensor<string, []>("value_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_83_pad_0 = const()[name = tensor<string, []>("value_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_83_dilations_0 = const()[name = tensor<string, []>("value_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_83_groups_0 = const()[name = tensor<string, []>("value_83_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(791555264)))];
            tensor<fp16, [1024]> layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(793652480)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = value_83_dilations_0, groups = value_83_groups_0, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = value_83_strides_0, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_83_cast_fp16")];
            tensor<int32, [4]> var_4699 = const()[name = tensor<string, []>("op_4699"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_83_cast_fp16 = reshape(shape = var_4699, x = query_83_cast_fp16)[name = tensor<string, []>("mh_q_83_cast_fp16")];
            tensor<fp16, []> var_4701_to_fp16 = const()[name = tensor<string, []>("op_4701_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4702_cast_fp16 = mul(x = mh_q_83_cast_fp16, y = var_4701_to_fp16)[name = tensor<string, []>("op_4702_cast_fp16")];
            tensor<int32, [4]> var_4705 = const()[name = tensor<string, []>("op_4705"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4706_cast_fp16 = reshape(shape = var_4705, x = key_83_cast_fp16)[name = tensor<string, []>("op_4706_cast_fp16")];
            tensor<bool, []> mh_w_125_transpose_x_0 = const()[name = tensor<string, []>("mh_w_125_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_125_transpose_y_0 = const()[name = tensor<string, []>("mh_w_125_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4702_cast_fp16, y = var_4706_cast_fp16)[name = tensor<string, []>("mh_w_125_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_293_cast_fp16 = softmax(axis = var_4548, x = mh_w_125_cast_fp16)[name = tensor<string, []>("obj_293_cast_fp16")];
            tensor<int32, [4]> var_4710 = const()[name = tensor<string, []>("op_4710"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4711_cast_fp16 = reshape(shape = var_4710, x = value_83_cast_fp16)[name = tensor<string, []>("op_4711_cast_fp16")];
            tensor<bool, []> attn_83_transpose_x_0 = const()[name = tensor<string, []>("attn_83_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_83_transpose_y_0 = const()[name = tensor<string, []>("attn_83_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4711_cast_fp16, y = obj_293_cast_fp16)[name = tensor<string, []>("attn_83_cast_fp16")];
            tensor<int32, [4]> var_4714 = const()[name = tensor<string, []>("op_4714"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_203_cast_fp16 = reshape(shape = var_4714, x = attn_83_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<string, []> obj_291_pad_type_0 = const()[name = tensor<string, []>("obj_291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_291_strides_0 = const()[name = tensor<string, []>("obj_291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_291_pad_0 = const()[name = tensor<string, []>("obj_291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_291_dilations_0 = const()[name = tensor<string, []>("obj_291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_291_groups_0 = const()[name = tensor<string, []>("obj_291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(793654592)))];
            tensor<fp16, [1024]> layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795751808)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = obj_291_dilations_0, groups = obj_291_groups_0, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = obj_291_strides_0, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("obj_291_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor<string, []>("inputs_125_cast_fp16")];
            tensor<int32, [1]> out_125_axes_0 = const()[name = tensor<string, []>("out_125_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4735_to_fp16 = const()[name = tensor<string, []>("op_4735_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_4735_to_fp16, x = inputs_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
            tensor<fp16, [1024]> input_205_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_205_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795753920)))];
            tensor<fp16, [1024]> input_205_beta_0_to_fp16 = const()[name = tensor<string, []>("input_205_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795756032)))];
            tensor<fp16, []> input_205_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_205_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795758144)))];
            tensor<fp16, [4096]> layers_20_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(804146816)))];
            tensor<fp16, [1, 4096, 1, 1]> input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<string, []> input_209_mode_0 = const()[name = tensor<string, []>("input_209_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<string, []> hidden_states_43_pad_type_0 = const()[name = tensor<string, []>("hidden_states_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = tensor<string, []>("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = tensor<string, []>("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = tensor<string, []>("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_43_groups_0 = const()[name = tensor<string, []>("hidden_states_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_20_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(804155072)))];
            tensor<fp16, [1024]> layers_20_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_20_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(812543744)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor<string, []>("inputs_127_cast_fp16")];
            tensor<int32, []> var_4771 = const()[name = tensor<string, []>("op_4771"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_127_axes_0 = const()[name = tensor<string, []>("out_127_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4796_to_fp16 = const()[name = tensor<string, []>("op_4796_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_4796_to_fp16, x = inputs_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
            tensor<fp16, [1024]> obj_295_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_295_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(812545856)))];
            tensor<fp16, [1024]> obj_295_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_295_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(812547968)))];
            tensor<fp16, []> obj_295_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_295_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("obj_295_cast_fp16")];
            tensor<string, []> query_85_pad_type_0 = const()[name = tensor<string, []>("query_85_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_85_strides_0 = const()[name = tensor<string, []>("query_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_85_pad_0 = const()[name = tensor<string, []>("query_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_85_dilations_0 = const()[name = tensor<string, []>("query_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_85_groups_0 = const()[name = tensor<string, []>("query_85_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(812550080)))];
            tensor<fp16, [1024]> layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814647296)))];
            tensor<fp16, [1, 1024, 1, 1]> query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor<string, []>("query_85_cast_fp16")];
            tensor<string, []> current_key_43_pad_type_0 = const()[name = tensor<string, []>("current_key_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_43_strides_0 = const()[name = tensor<string, []>("current_key_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_43_pad_0 = const()[name = tensor<string, []>("current_key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_43_dilations_0 = const()[name = tensor<string, []>("current_key_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_43_groups_0 = const()[name = tensor<string, []>("current_key_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(814649408)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_43_cast_fp16 = conv(dilations = current_key_43_dilations_0, groups = current_key_43_groups_0, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = current_key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor<string, []>("current_key_43_cast_fp16")];
            tensor<string, []> current_value_43_pad_type_0 = const()[name = tensor<string, []>("current_value_43_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_43_strides_0 = const()[name = tensor<string, []>("current_value_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_43_pad_0 = const()[name = tensor<string, []>("current_value_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_43_dilations_0 = const()[name = tensor<string, []>("current_value_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_43_groups_0 = const()[name = tensor<string, []>("current_value_43_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(816746624)))];
            tensor<fp16, [1024]> layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(818843840)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = current_value_43_dilations_0, groups = current_value_43_groups_0, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = current_value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor<string, []>("current_value_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4835_cast_fp16 = mul(x = var_87_cast_fp16_21, y = var_207_cast_fp16)[name = tensor<string, []>("op_4835_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4836_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4836_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_85_cast_fp16 = add(x = var_4835_cast_fp16, y = var_4836_cast_fp16)[name = tensor<string, []>("key_85_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4839_cast_fp16 = mul(x = var_114_cast_fp16_21, y = var_207_cast_fp16)[name = tensor<string, []>("op_4839_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_4840_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_4840_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_85_cast_fp16 = add(x = var_4839_cast_fp16, y = var_4840_cast_fp16)[name = tensor<string, []>("value_85_cast_fp16")];
            tensor<int32, [4]> var_4844 = const()[name = tensor<string, []>("op_4844"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_85_cast_fp16 = reshape(shape = var_4844, x = query_85_cast_fp16)[name = tensor<string, []>("mh_q_85_cast_fp16")];
            tensor<fp16, []> var_4846_to_fp16 = const()[name = tensor<string, []>("op_4846_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4847_cast_fp16 = mul(x = mh_q_85_cast_fp16, y = var_4846_to_fp16)[name = tensor<string, []>("op_4847_cast_fp16")];
            tensor<int32, [4]> var_4850 = const()[name = tensor<string, []>("op_4850"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4851_cast_fp16 = reshape(shape = var_4850, x = key_85_cast_fp16)[name = tensor<string, []>("op_4851_cast_fp16")];
            tensor<bool, []> mh_w_127_transpose_x_0 = const()[name = tensor<string, []>("mh_w_127_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_127_transpose_y_0 = const()[name = tensor<string, []>("mh_w_127_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4847_cast_fp16, y = var_4851_cast_fp16)[name = tensor<string, []>("mh_w_127_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_129_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_4859_cast_fp16 = softmax(axis = var_4771, x = mh_w_129_cast_fp16)[name = tensor<string, []>("op_4859_cast_fp16")];
            tensor<int32, [4]> var_4860 = const()[name = tensor<string, []>("op_4860"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_4861_cast_fp16 = reshape(shape = var_4860, x = value_85_cast_fp16)[name = tensor<string, []>("op_4861_cast_fp16")];
            tensor<bool, []> attn_85_transpose_x_0 = const()[name = tensor<string, []>("attn_85_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_85_transpose_y_0 = const()[name = tensor<string, []>("attn_85_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4861_cast_fp16, y = var_4859_cast_fp16)[name = tensor<string, []>("attn_85_cast_fp16")];
            tensor<int32, [4]> var_4864 = const()[name = tensor<string, []>("op_4864"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_211_cast_fp16 = reshape(shape = var_4864, x = attn_85_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<string, []> obj_301_pad_type_0 = const()[name = tensor<string, []>("obj_301_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_301_strides_0 = const()[name = tensor<string, []>("obj_301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_301_pad_0 = const()[name = tensor<string, []>("obj_301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_301_dilations_0 = const()[name = tensor<string, []>("obj_301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_301_groups_0 = const()[name = tensor<string, []>("obj_301_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(818845952)))];
            tensor<fp16, [1024]> layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(820943168)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_301_dilations_0, groups = obj_301_groups_0, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = obj_301_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("obj_301_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor<string, []>("inputs_129_cast_fp16")];
            tensor<int32, [1]> out_129_axes_0 = const()[name = tensor<string, []>("out_129_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4886_to_fp16 = const()[name = tensor<string, []>("op_4886_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_4886_to_fp16, x = inputs_129_cast_fp16)[name = tensor<string, []>("out_129_cast_fp16")];
            tensor<fp16, [1024]> obj_303_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_303_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(820945280)))];
            tensor<fp16, [1024]> obj_303_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_303_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(820947392)))];
            tensor<fp16, []> obj_303_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_303_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor<string, []>("obj_303_cast_fp16")];
            tensor<string, []> query_87_pad_type_0 = const()[name = tensor<string, []>("query_87_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_87_strides_0 = const()[name = tensor<string, []>("query_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_87_pad_0 = const()[name = tensor<string, []>("query_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_87_dilations_0 = const()[name = tensor<string, []>("query_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_87_groups_0 = const()[name = tensor<string, []>("query_87_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(820949504)))];
            tensor<fp16, [1024]> layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(823046720)))];
            tensor<fp16, [1, 1024, 1, 1]> query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = query_87_dilations_0, groups = query_87_groups_0, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = query_87_strides_0, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor<string, []>("query_87_cast_fp16")];
            tensor<string, []> key_87_pad_type_0 = const()[name = tensor<string, []>("key_87_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_87_strides_0 = const()[name = tensor<string, []>("key_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_87_pad_0 = const()[name = tensor<string, []>("key_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_87_dilations_0 = const()[name = tensor<string, []>("key_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_87_groups_0 = const()[name = tensor<string, []>("key_87_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(823048832)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_87_cast_fp16 = conv(dilations = key_87_dilations_0, groups = key_87_groups_0, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = key_87_strides_0, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_87_cast_fp16")];
            tensor<string, []> value_87_pad_type_0 = const()[name = tensor<string, []>("value_87_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_87_strides_0 = const()[name = tensor<string, []>("value_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_87_pad_0 = const()[name = tensor<string, []>("value_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_87_dilations_0 = const()[name = tensor<string, []>("value_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_87_groups_0 = const()[name = tensor<string, []>("value_87_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(825146048)))];
            tensor<fp16, [1024]> layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(827243264)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = value_87_dilations_0, groups = value_87_groups_0, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = value_87_strides_0, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_87_cast_fp16")];
            tensor<int32, [4]> var_4922 = const()[name = tensor<string, []>("op_4922"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_87_cast_fp16 = reshape(shape = var_4922, x = query_87_cast_fp16)[name = tensor<string, []>("mh_q_87_cast_fp16")];
            tensor<fp16, []> var_4924_to_fp16 = const()[name = tensor<string, []>("op_4924_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_4925_cast_fp16 = mul(x = mh_q_87_cast_fp16, y = var_4924_to_fp16)[name = tensor<string, []>("op_4925_cast_fp16")];
            tensor<int32, [4]> var_4928 = const()[name = tensor<string, []>("op_4928"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4929_cast_fp16 = reshape(shape = var_4928, x = key_87_cast_fp16)[name = tensor<string, []>("op_4929_cast_fp16")];
            tensor<bool, []> mh_w_131_transpose_x_0 = const()[name = tensor<string, []>("mh_w_131_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_131_transpose_y_0 = const()[name = tensor<string, []>("mh_w_131_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4925_cast_fp16, y = var_4929_cast_fp16)[name = tensor<string, []>("mh_w_131_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_307_cast_fp16 = softmax(axis = var_4771, x = mh_w_131_cast_fp16)[name = tensor<string, []>("obj_307_cast_fp16")];
            tensor<int32, [4]> var_4933 = const()[name = tensor<string, []>("op_4933"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_4934_cast_fp16 = reshape(shape = var_4933, x = value_87_cast_fp16)[name = tensor<string, []>("op_4934_cast_fp16")];
            tensor<bool, []> attn_87_transpose_x_0 = const()[name = tensor<string, []>("attn_87_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_87_transpose_y_0 = const()[name = tensor<string, []>("attn_87_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4934_cast_fp16, y = obj_307_cast_fp16)[name = tensor<string, []>("attn_87_cast_fp16")];
            tensor<int32, [4]> var_4937 = const()[name = tensor<string, []>("op_4937"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_213_cast_fp16 = reshape(shape = var_4937, x = attn_87_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<string, []> obj_305_pad_type_0 = const()[name = tensor<string, []>("obj_305_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_305_strides_0 = const()[name = tensor<string, []>("obj_305_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_305_pad_0 = const()[name = tensor<string, []>("obj_305_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_305_dilations_0 = const()[name = tensor<string, []>("obj_305_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_305_groups_0 = const()[name = tensor<string, []>("obj_305_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(827245376)))];
            tensor<fp16, [1024]> layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(829342592)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = obj_305_dilations_0, groups = obj_305_groups_0, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = obj_305_strides_0, weight = layers_21_encoder_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("obj_305_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor<string, []>("inputs_131_cast_fp16")];
            tensor<int32, [1]> out_131_axes_0 = const()[name = tensor<string, []>("out_131_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_4958_to_fp16 = const()[name = tensor<string, []>("op_4958_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_4958_to_fp16, x = inputs_131_cast_fp16)[name = tensor<string, []>("out_131_cast_fp16")];
            tensor<fp16, [1024]> input_215_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_215_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(829344704)))];
            tensor<fp16, [1024]> input_215_beta_0_to_fp16 = const()[name = tensor<string, []>("input_215_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(829346816)))];
            tensor<fp16, []> input_215_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_215_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<string, []> input_217_pad_type_0 = const()[name = tensor<string, []>("input_217_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_217_strides_0 = const()[name = tensor<string, []>("input_217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_217_pad_0 = const()[name = tensor<string, []>("input_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_217_dilations_0 = const()[name = tensor<string, []>("input_217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_217_groups_0 = const()[name = tensor<string, []>("input_217_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(829348928)))];
            tensor<fp16, [4096]> layers_21_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(837737600)))];
            tensor<fp16, [1, 4096, 1, 1]> input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<string, []> input_219_mode_0 = const()[name = tensor<string, []>("input_219_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<string, []> hidden_states_45_pad_type_0 = const()[name = tensor<string, []>("hidden_states_45_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_45_strides_0 = const()[name = tensor<string, []>("hidden_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = tensor<string, []>("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_45_dilations_0 = const()[name = tensor<string, []>("hidden_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_45_groups_0 = const()[name = tensor<string, []>("hidden_states_45_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_21_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(837745856)))];
            tensor<fp16, [1024]> layers_21_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_21_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846134528)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("hidden_states_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor<string, []>("inputs_133_cast_fp16")];
            tensor<int32, []> var_4994 = const()[name = tensor<string, []>("op_4994"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_133_axes_0 = const()[name = tensor<string, []>("out_133_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5019_to_fp16 = const()[name = tensor<string, []>("op_5019_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_5019_to_fp16, x = inputs_133_cast_fp16)[name = tensor<string, []>("out_133_cast_fp16")];
            tensor<fp16, [1024]> obj_309_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_309_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846136640)))];
            tensor<fp16, [1024]> obj_309_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_309_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846138752)))];
            tensor<fp16, []> obj_309_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_309_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor<string, []>("obj_309_cast_fp16")];
            tensor<string, []> query_89_pad_type_0 = const()[name = tensor<string, []>("query_89_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_89_strides_0 = const()[name = tensor<string, []>("query_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_89_pad_0 = const()[name = tensor<string, []>("query_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_89_dilations_0 = const()[name = tensor<string, []>("query_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_89_groups_0 = const()[name = tensor<string, []>("query_89_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846140864)))];
            tensor<fp16, [1024]> layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(848238080)))];
            tensor<fp16, [1, 1024, 1, 1]> query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor<string, []>("query_89_cast_fp16")];
            tensor<string, []> current_key_45_pad_type_0 = const()[name = tensor<string, []>("current_key_45_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_45_strides_0 = const()[name = tensor<string, []>("current_key_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_45_pad_0 = const()[name = tensor<string, []>("current_key_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_45_dilations_0 = const()[name = tensor<string, []>("current_key_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_45_groups_0 = const()[name = tensor<string, []>("current_key_45_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(848240192)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_45_cast_fp16 = conv(dilations = current_key_45_dilations_0, groups = current_key_45_groups_0, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = current_key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor<string, []>("current_key_45_cast_fp16")];
            tensor<string, []> current_value_45_pad_type_0 = const()[name = tensor<string, []>("current_value_45_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_45_strides_0 = const()[name = tensor<string, []>("current_value_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_45_pad_0 = const()[name = tensor<string, []>("current_value_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_45_dilations_0 = const()[name = tensor<string, []>("current_value_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_45_groups_0 = const()[name = tensor<string, []>("current_value_45_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(850337408)))];
            tensor<fp16, [1024]> layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852434624)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = current_value_45_dilations_0, groups = current_value_45_groups_0, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = current_value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor<string, []>("current_value_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5058_cast_fp16 = mul(x = var_87_cast_fp16_22, y = var_207_cast_fp16)[name = tensor<string, []>("op_5058_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5059_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_5059_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_89_cast_fp16 = add(x = var_5058_cast_fp16, y = var_5059_cast_fp16)[name = tensor<string, []>("key_89_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5062_cast_fp16 = mul(x = var_114_cast_fp16_22, y = var_207_cast_fp16)[name = tensor<string, []>("op_5062_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5063_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_5063_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_89_cast_fp16 = add(x = var_5062_cast_fp16, y = var_5063_cast_fp16)[name = tensor<string, []>("value_89_cast_fp16")];
            tensor<int32, [4]> var_5067 = const()[name = tensor<string, []>("op_5067"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_89_cast_fp16 = reshape(shape = var_5067, x = query_89_cast_fp16)[name = tensor<string, []>("mh_q_89_cast_fp16")];
            tensor<fp16, []> var_5069_to_fp16 = const()[name = tensor<string, []>("op_5069_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_5070_cast_fp16 = mul(x = mh_q_89_cast_fp16, y = var_5069_to_fp16)[name = tensor<string, []>("op_5070_cast_fp16")];
            tensor<int32, [4]> var_5073 = const()[name = tensor<string, []>("op_5073"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_5074_cast_fp16 = reshape(shape = var_5073, x = key_89_cast_fp16)[name = tensor<string, []>("op_5074_cast_fp16")];
            tensor<bool, []> mh_w_133_transpose_x_0 = const()[name = tensor<string, []>("mh_w_133_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_133_transpose_y_0 = const()[name = tensor<string, []>("mh_w_133_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_5070_cast_fp16, y = var_5074_cast_fp16)[name = tensor<string, []>("mh_w_133_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_135_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_5082_cast_fp16 = softmax(axis = var_4994, x = mh_w_135_cast_fp16)[name = tensor<string, []>("op_5082_cast_fp16")];
            tensor<int32, [4]> var_5083 = const()[name = tensor<string, []>("op_5083"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_5084_cast_fp16 = reshape(shape = var_5083, x = value_89_cast_fp16)[name = tensor<string, []>("op_5084_cast_fp16")];
            tensor<bool, []> attn_89_transpose_x_0 = const()[name = tensor<string, []>("attn_89_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_89_transpose_y_0 = const()[name = tensor<string, []>("attn_89_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_5084_cast_fp16, y = var_5082_cast_fp16)[name = tensor<string, []>("attn_89_cast_fp16")];
            tensor<int32, [4]> var_5087 = const()[name = tensor<string, []>("op_5087"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_221_cast_fp16 = reshape(shape = var_5087, x = attn_89_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<string, []> obj_315_pad_type_0 = const()[name = tensor<string, []>("obj_315_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_315_strides_0 = const()[name = tensor<string, []>("obj_315_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_315_pad_0 = const()[name = tensor<string, []>("obj_315_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_315_dilations_0 = const()[name = tensor<string, []>("obj_315_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_315_groups_0 = const()[name = tensor<string, []>("obj_315_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(852436736)))];
            tensor<fp16, [1024]> layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854533952)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_315_dilations_0, groups = obj_315_groups_0, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = obj_315_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("obj_315_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor<string, []>("inputs_135_cast_fp16")];
            tensor<int32, [1]> out_135_axes_0 = const()[name = tensor<string, []>("out_135_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5109_to_fp16 = const()[name = tensor<string, []>("op_5109_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_5109_to_fp16, x = inputs_135_cast_fp16)[name = tensor<string, []>("out_135_cast_fp16")];
            tensor<fp16, [1024]> obj_317_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_317_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854536064)))];
            tensor<fp16, [1024]> obj_317_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_317_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854538176)))];
            tensor<fp16, []> obj_317_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_317_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor<string, []>("obj_317_cast_fp16")];
            tensor<string, []> query_91_pad_type_0 = const()[name = tensor<string, []>("query_91_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_91_strides_0 = const()[name = tensor<string, []>("query_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_91_pad_0 = const()[name = tensor<string, []>("query_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_91_dilations_0 = const()[name = tensor<string, []>("query_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_91_groups_0 = const()[name = tensor<string, []>("query_91_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854540288)))];
            tensor<fp16, [1024]> layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(856637504)))];
            tensor<fp16, [1, 1024, 1, 1]> query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = query_91_dilations_0, groups = query_91_groups_0, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = query_91_strides_0, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor<string, []>("query_91_cast_fp16")];
            tensor<string, []> key_91_pad_type_0 = const()[name = tensor<string, []>("key_91_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_91_strides_0 = const()[name = tensor<string, []>("key_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_91_pad_0 = const()[name = tensor<string, []>("key_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_91_dilations_0 = const()[name = tensor<string, []>("key_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_91_groups_0 = const()[name = tensor<string, []>("key_91_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(856639616)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_91_cast_fp16 = conv(dilations = key_91_dilations_0, groups = key_91_groups_0, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = key_91_strides_0, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_91_cast_fp16")];
            tensor<string, []> value_91_pad_type_0 = const()[name = tensor<string, []>("value_91_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_91_strides_0 = const()[name = tensor<string, []>("value_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_91_pad_0 = const()[name = tensor<string, []>("value_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_91_dilations_0 = const()[name = tensor<string, []>("value_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_91_groups_0 = const()[name = tensor<string, []>("value_91_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(858736832)))];
            tensor<fp16, [1024]> layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(860834048)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = value_91_dilations_0, groups = value_91_groups_0, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = value_91_strides_0, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_91_cast_fp16")];
            tensor<int32, [4]> var_5145 = const()[name = tensor<string, []>("op_5145"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_91_cast_fp16 = reshape(shape = var_5145, x = query_91_cast_fp16)[name = tensor<string, []>("mh_q_91_cast_fp16")];
            tensor<fp16, []> var_5147_to_fp16 = const()[name = tensor<string, []>("op_5147_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_5148_cast_fp16 = mul(x = mh_q_91_cast_fp16, y = var_5147_to_fp16)[name = tensor<string, []>("op_5148_cast_fp16")];
            tensor<int32, [4]> var_5151 = const()[name = tensor<string, []>("op_5151"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_5152_cast_fp16 = reshape(shape = var_5151, x = key_91_cast_fp16)[name = tensor<string, []>("op_5152_cast_fp16")];
            tensor<bool, []> mh_w_137_transpose_x_0 = const()[name = tensor<string, []>("mh_w_137_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_137_transpose_y_0 = const()[name = tensor<string, []>("mh_w_137_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_5148_cast_fp16, y = var_5152_cast_fp16)[name = tensor<string, []>("mh_w_137_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_321_cast_fp16 = softmax(axis = var_4994, x = mh_w_137_cast_fp16)[name = tensor<string, []>("obj_321_cast_fp16")];
            tensor<int32, [4]> var_5156 = const()[name = tensor<string, []>("op_5156"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_5157_cast_fp16 = reshape(shape = var_5156, x = value_91_cast_fp16)[name = tensor<string, []>("op_5157_cast_fp16")];
            tensor<bool, []> attn_91_transpose_x_0 = const()[name = tensor<string, []>("attn_91_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_91_transpose_y_0 = const()[name = tensor<string, []>("attn_91_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_5157_cast_fp16, y = obj_321_cast_fp16)[name = tensor<string, []>("attn_91_cast_fp16")];
            tensor<int32, [4]> var_5160 = const()[name = tensor<string, []>("op_5160"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_223_cast_fp16 = reshape(shape = var_5160, x = attn_91_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
            tensor<string, []> obj_319_pad_type_0 = const()[name = tensor<string, []>("obj_319_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_319_strides_0 = const()[name = tensor<string, []>("obj_319_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_319_pad_0 = const()[name = tensor<string, []>("obj_319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_319_dilations_0 = const()[name = tensor<string, []>("obj_319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_319_groups_0 = const()[name = tensor<string, []>("obj_319_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(860836160)))];
            tensor<fp16, [1024]> layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862933376)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = obj_319_dilations_0, groups = obj_319_groups_0, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = obj_319_strides_0, weight = layers_22_encoder_attn_o_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("obj_319_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor<string, []>("inputs_137_cast_fp16")];
            tensor<int32, [1]> out_137_axes_0 = const()[name = tensor<string, []>("out_137_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5178_to_fp16 = const()[name = tensor<string, []>("op_5178_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_5178_to_fp16, x = inputs_137_cast_fp16)[name = tensor<string, []>("out_137_cast_fp16")];
            tensor<fp16, [1024]> input_225_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_225_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862935488)))];
            tensor<fp16, [1024]> input_225_beta_0_to_fp16 = const()[name = tensor<string, []>("input_225_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862937600)))];
            tensor<fp16, []> input_225_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_225_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<string, []> input_227_pad_type_0 = const()[name = tensor<string, []>("input_227_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_227_strides_0 = const()[name = tensor<string, []>("input_227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_227_pad_0 = const()[name = tensor<string, []>("input_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_227_dilations_0 = const()[name = tensor<string, []>("input_227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_227_groups_0 = const()[name = tensor<string, []>("input_227_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(862939712)))];
            tensor<fp16, [4096]> layers_22_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(871328384)))];
            tensor<fp16, [1, 4096, 1, 1]> input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<string, []> input_229_mode_0 = const()[name = tensor<string, []>("input_229_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
            tensor<string, []> hidden_states_47_pad_type_0 = const()[name = tensor<string, []>("hidden_states_47_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = tensor<string, []>("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = tensor<string, []>("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = tensor<string, []>("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_47_groups_0 = const()[name = tensor<string, []>("hidden_states_47_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_22_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(871336640)))];
            tensor<fp16, [1024]> layers_22_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_22_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(879725312)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor<string, []>("inputs_139_cast_fp16")];
            tensor<int32, []> var_5213 = const()[name = tensor<string, []>("op_5213"), val = tensor<int32, []>(3)];
            tensor<int32, [1]> out_139_axes_0 = const()[name = tensor<string, []>("out_139_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5238_to_fp16 = const()[name = tensor<string, []>("op_5238_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_5238_to_fp16, x = inputs_139_cast_fp16)[name = tensor<string, []>("out_139_cast_fp16")];
            tensor<fp16, [1024]> obj_323_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_323_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(879727424)))];
            tensor<fp16, [1024]> obj_323_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_323_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(879729536)))];
            tensor<fp16, []> obj_323_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_323_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor<string, []>("obj_323_cast_fp16")];
            tensor<string, []> query_93_pad_type_0 = const()[name = tensor<string, []>("query_93_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_93_strides_0 = const()[name = tensor<string, []>("query_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_93_pad_0 = const()[name = tensor<string, []>("query_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_93_dilations_0 = const()[name = tensor<string, []>("query_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_93_groups_0 = const()[name = tensor<string, []>("query_93_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(879731648)))];
            tensor<fp16, [1024]> layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881828864)))];
            tensor<fp16, [1, 1024, 1, 1]> query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor<string, []>("query_93_cast_fp16")];
            tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_key_strides_0 = const()[name = tensor<string, []>("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_key_dilations_0 = const()[name = tensor<string, []>("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_key_groups_0 = const()[name = tensor<string, []>("current_key_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881830976)))];
            tensor<fp16, [1, 1024, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
            tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> current_value_strides_0 = const()[name = tensor<string, []>("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> current_value_dilations_0 = const()[name = tensor<string, []>("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> current_value_groups_0 = const()[name = tensor<string, []>("current_value_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(883928192)))];
            tensor<fp16, [1024]> layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(886025408)))];
            tensor<fp16, [1, 1024, 1, 1]> current_value_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5277_cast_fp16 = mul(x = var_87_cast_fp16_23, y = var_207_cast_fp16)[name = tensor<string, []>("op_5277_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5278_cast_fp16 = mul(x = current_key_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_5278_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> key_93_cast_fp16 = add(x = var_5277_cast_fp16, y = var_5278_cast_fp16)[name = tensor<string, []>("key_93_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5281_cast_fp16 = mul(x = var_114_cast_fp16_23, y = var_207_cast_fp16)[name = tensor<string, []>("op_5281_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> var_5282_cast_fp16 = mul(x = current_value_cast_fp16, y = var_205_cast_fp16)[name = tensor<string, []>("op_5282_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 448]> value_93_cast_fp16 = add(x = var_5281_cast_fp16, y = var_5282_cast_fp16)[name = tensor<string, []>("value_93_cast_fp16")];
            tensor<int32, [4]> var_5286 = const()[name = tensor<string, []>("op_5286"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_93_cast_fp16 = reshape(shape = var_5286, x = query_93_cast_fp16)[name = tensor<string, []>("mh_q_93_cast_fp16")];
            tensor<fp16, []> var_5288_to_fp16 = const()[name = tensor<string, []>("op_5288_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_5289_cast_fp16 = mul(x = mh_q_93_cast_fp16, y = var_5288_to_fp16)[name = tensor<string, []>("op_5289_cast_fp16")];
            tensor<int32, [4]> var_5292 = const()[name = tensor<string, []>("op_5292"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_5293_cast_fp16 = reshape(shape = var_5292, x = key_93_cast_fp16)[name = tensor<string, []>("op_5293_cast_fp16")];
            tensor<bool, []> mh_w_139_transpose_x_0 = const()[name = tensor<string, []>("mh_w_139_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_139_transpose_y_0 = const()[name = tensor<string, []>("mh_w_139_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 448]> mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5289_cast_fp16, y = var_5293_cast_fp16)[name = tensor<string, []>("mh_w_139_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_229_cast_fp16)[name = tensor<string, []>("mh_w_141_cast_fp16")];
            tensor<fp16, [1, 16, 1, 448]> var_5301_cast_fp16 = softmax(axis = var_5213, x = mh_w_141_cast_fp16)[name = tensor<string, []>("op_5301_cast_fp16")];
            tensor<int32, [4]> var_5302 = const()[name = tensor<string, []>("op_5302"), val = tensor<int32, [4]>([1, 16, 64, 448])];
            tensor<fp16, [1, 16, 64, 448]> var_5303_cast_fp16 = reshape(shape = var_5302, x = value_93_cast_fp16)[name = tensor<string, []>("op_5303_cast_fp16")];
            tensor<bool, []> attn_93_transpose_x_0 = const()[name = tensor<string, []>("attn_93_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_93_transpose_y_0 = const()[name = tensor<string, []>("attn_93_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5303_cast_fp16, y = var_5301_cast_fp16)[name = tensor<string, []>("attn_93_cast_fp16")];
            tensor<int32, [4]> var_5306 = const()[name = tensor<string, []>("op_5306"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_231_cast_fp16 = reshape(shape = var_5306, x = attn_93_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<string, []> obj_329_pad_type_0 = const()[name = tensor<string, []>("obj_329_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_329_strides_0 = const()[name = tensor<string, []>("obj_329_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_329_pad_0 = const()[name = tensor<string, []>("obj_329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_329_dilations_0 = const()[name = tensor<string, []>("obj_329_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_329_groups_0 = const()[name = tensor<string, []>("obj_329_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(886027520)))];
            tensor<fp16, [1024]> layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888124736)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_329_dilations_0, groups = obj_329_groups_0, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = obj_329_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("obj_329_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor<string, []>("inputs_141_cast_fp16")];
            tensor<int32, [1]> out_141_axes_0 = const()[name = tensor<string, []>("out_141_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5328_to_fp16 = const()[name = tensor<string, []>("op_5328_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_5328_to_fp16, x = inputs_141_cast_fp16)[name = tensor<string, []>("out_141_cast_fp16")];
            tensor<fp16, [1024]> obj_331_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_331_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888126848)))];
            tensor<fp16, [1024]> obj_331_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_331_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888128960)))];
            tensor<fp16, []> obj_331_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_331_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor<string, []>("obj_331_cast_fp16")];
            tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888131072)))];
            tensor<fp16, [1024]> layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(890228288)))];
            tensor<fp16, [1, 1024, 1, 1]> query_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
            tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(890230400)))];
            tensor<fp16, [1, 1024, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
            tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(892327616)))];
            tensor<fp16, [1024]> layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894424832)))];
            tensor<fp16, [1, 1024, 1, 1500]> value_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
            tensor<int32, [4]> var_5364 = const()[name = tensor<string, []>("op_5364"), val = tensor<int32, [4]>([1, 16, 64, 1])];
            tensor<fp16, [1, 16, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_5364, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
            tensor<fp16, []> var_5366_to_fp16 = const()[name = tensor<string, []>("op_5366_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, 64, 1]> var_5367_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_5366_to_fp16)[name = tensor<string, []>("op_5367_cast_fp16")];
            tensor<int32, [4]> var_5370 = const()[name = tensor<string, []>("op_5370"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_5371_cast_fp16 = reshape(shape = var_5370, x = key_cast_fp16)[name = tensor<string, []>("op_5371_cast_fp16")];
            tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_5367_cast_fp16, y = var_5371_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
            tensor<fp16, [1, 16, 1, 1500]> obj_335_cast_fp16 = softmax(axis = var_5213, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_335_cast_fp16")];
            tensor<int32, [4]> var_5375 = const()[name = tensor<string, []>("op_5375"), val = tensor<int32, [4]>([1, 16, 64, 1500])];
            tensor<fp16, [1, 16, 64, 1500]> var_5376_cast_fp16 = reshape(shape = var_5375, x = value_cast_fp16)[name = tensor<string, []>("op_5376_cast_fp16")];
            tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_5376_cast_fp16, y = obj_335_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
            tensor<int32, [4]> var_5379 = const()[name = tensor<string, []>("op_5379"), val = tensor<int32, [4]>([1, 1024, 1, 1])];
            tensor<fp16, [1, 1024, 1, 1]> input_233_cast_fp16 = reshape(shape = var_5379, x = attn_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<string, []> obj_333_pad_type_0 = const()[name = tensor<string, []>("obj_333_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> obj_333_strides_0 = const()[name = tensor<string, []>("obj_333_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> obj_333_pad_0 = const()[name = tensor<string, []>("obj_333_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> obj_333_dilations_0 = const()[name = tensor<string, []>("obj_333_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> obj_333_groups_0 = const()[name = tensor<string, []>("obj_333_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(894426944)))];
            tensor<fp16, [1024]> layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(896524160)))];
            tensor<fp16, [1, 1024, 1, 1]> obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = obj_333_dilations_0, groups = obj_333_groups_0, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = obj_333_strides_0, weight = layers_23_encoder_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("obj_333_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor<string, []>("inputs_143_cast_fp16")];
            tensor<int32, [1]> out_143_axes_0 = const()[name = tensor<string, []>("out_143_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5397_to_fp16 = const()[name = tensor<string, []>("op_5397_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_5397_to_fp16, x = inputs_143_cast_fp16)[name = tensor<string, []>("out_143_cast_fp16")];
            tensor<fp16, [1024]> input_235_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(896526272)))];
            tensor<fp16, [1024]> input_235_beta_0_to_fp16 = const()[name = tensor<string, []>("input_235_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(896528384)))];
            tensor<fp16, []> input_235_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_235_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<string, []> input_237_pad_type_0 = const()[name = tensor<string, []>("input_237_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_237_strides_0 = const()[name = tensor<string, []>("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_237_pad_0 = const()[name = tensor<string, []>("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_237_dilations_0 = const()[name = tensor<string, []>("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_237_groups_0 = const()[name = tensor<string, []>("input_237_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(896530496)))];
            tensor<fp16, [4096]> layers_23_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904919168)))];
            tensor<fp16, [1, 4096, 1, 1]> input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
            tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 4096, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_237_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<string, []> hidden_states_49_pad_type_0 = const()[name = tensor<string, []>("hidden_states_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> hidden_states_49_strides_0 = const()[name = tensor<string, []>("hidden_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_49_pad_0 = const()[name = tensor<string, []>("hidden_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_49_dilations_0 = const()[name = tensor<string, []>("hidden_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> hidden_states_49_groups_0 = const()[name = tensor<string, []>("hidden_states_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_23_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904927424)))];
            tensor<fp16, [1024]> layers_23_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_23_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(913316096)))];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> inputs_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, []> var_5439_to_fp16 = const()[name = tensor<string, []>("op_5439_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5439_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<fp16, [1024]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(913318208)))];
            tensor<fp16, [1024]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(913320320)))];
            tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
            tensor<int32, [1]> var_5450_axes_0 = const()[name = tensor<string, []>("op_5450_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1]> var_5450_cast_fp16 = squeeze(axes = var_5450_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_5450_cast_fp16")];
            tensor<int32, [3]> var_5453_perm_0 = const()[name = tensor<string, []>("op_5453_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [51864]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(913322432)))];
            tensor<fp16, [1, 1, 1024]> var_5453_cast_fp16 = transpose(perm = var_5453_perm_0, x = var_5450_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 1, 51864]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_5453_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, []> var_5457 = const()[name = tensor<string, []>("op_5457"), val = tensor<int32, []>(1)];
            tensor<bool, []> obj_339_interleave_0 = const()[name = tensor<string, []>("obj_339_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 24576, 1, 1]> key_cache_updates = concat(axis = var_5457, interleave = obj_339_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_339_cast_fp16")];
            tensor<int32, []> var_5460 = const()[name = tensor<string, []>("op_5460"), val = tensor<int32, []>(1)];
            tensor<bool, []> obj_341_interleave_0 = const()[name = tensor<string, []>("obj_341_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 24576, 1, 1]> value_cache_updates = concat(axis = var_5460, interleave = obj_341_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_341_cast_fp16")];
            tensor<int32, [4]> var_5471_begin_0 = const()[name = tensor<string, []>("op_5471_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
            tensor<int32, [4]> var_5471_end_0 = const()[name = tensor<string, []>("op_5471_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
            tensor<bool, [4]> var_5471_end_mask_0 = const()[name = tensor<string, []>("op_5471_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5471_cast_fp16 = slice_by_index(begin = var_5471_begin_0, end = var_5471_end_0, end_mask = var_5471_end_mask_0, x = obj_167_cast_fp16)[name = tensor<string, []>("op_5471_cast_fp16")];
            tensor<int32, [4]> var_5474_begin_0 = const()[name = tensor<string, []>("op_5474_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5474_end_0 = const()[name = tensor<string, []>("op_5474_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5474_end_mask_0 = const()[name = tensor<string, []>("op_5474_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5474_squeeze_mask_0 = const()[name = tensor<string, []>("op_5474_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5474_cast_fp16 = slice_by_index(begin = var_5474_begin_0, end = var_5474_end_0, end_mask = var_5474_end_mask_0, squeeze_mask = var_5474_squeeze_mask_0, x = var_5471_cast_fp16)[name = tensor<string, []>("op_5474_cast_fp16")];
            tensor<int32, [4]> var_5489_begin_0 = const()[name = tensor<string, []>("op_5489_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5489_end_0 = const()[name = tensor<string, []>("op_5489_end_0"), val = tensor<int32, [4]>([1, 2, 1, 1500])];
            tensor<bool, [4]> var_5489_end_mask_0 = const()[name = tensor<string, []>("op_5489_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5489_cast_fp16 = slice_by_index(begin = var_5489_begin_0, end = var_5489_end_0, end_mask = var_5489_end_mask_0, x = obj_209_cast_fp16)[name = tensor<string, []>("op_5489_cast_fp16")];
            tensor<int32, [4]> var_5492_begin_0 = const()[name = tensor<string, []>("op_5492_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5492_end_0 = const()[name = tensor<string, []>("op_5492_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5492_end_mask_0 = const()[name = tensor<string, []>("op_5492_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5492_squeeze_mask_0 = const()[name = tensor<string, []>("op_5492_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5492_cast_fp16 = slice_by_index(begin = var_5492_begin_0, end = var_5492_end_0, end_mask = var_5492_end_mask_0, squeeze_mask = var_5492_squeeze_mask_0, x = var_5489_cast_fp16)[name = tensor<string, []>("op_5492_cast_fp16")];
            tensor<int32, [4]> var_5507_begin_0 = const()[name = tensor<string, []>("op_5507_begin_0"), val = tensor<int32, [4]>([0, 12, 0, 0])];
            tensor<int32, [4]> var_5507_end_0 = const()[name = tensor<string, []>("op_5507_end_0"), val = tensor<int32, [4]>([1, 13, 1, 1500])];
            tensor<bool, [4]> var_5507_end_mask_0 = const()[name = tensor<string, []>("op_5507_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5507_cast_fp16 = slice_by_index(begin = var_5507_begin_0, end = var_5507_end_0, end_mask = var_5507_end_mask_0, x = obj_209_cast_fp16)[name = tensor<string, []>("op_5507_cast_fp16")];
            tensor<int32, [4]> var_5510_begin_0 = const()[name = tensor<string, []>("op_5510_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5510_end_0 = const()[name = tensor<string, []>("op_5510_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5510_end_mask_0 = const()[name = tensor<string, []>("op_5510_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5510_squeeze_mask_0 = const()[name = tensor<string, []>("op_5510_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5510_cast_fp16 = slice_by_index(begin = var_5510_begin_0, end = var_5510_end_0, end_mask = var_5510_end_mask_0, squeeze_mask = var_5510_squeeze_mask_0, x = var_5507_cast_fp16)[name = tensor<string, []>("op_5510_cast_fp16")];
            tensor<int32, [4]> var_5525_begin_0 = const()[name = tensor<string, []>("op_5525_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
            tensor<int32, [4]> var_5525_end_0 = const()[name = tensor<string, []>("op_5525_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])];
            tensor<bool, [4]> var_5525_end_mask_0 = const()[name = tensor<string, []>("op_5525_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5525_cast_fp16 = slice_by_index(begin = var_5525_begin_0, end = var_5525_end_0, end_mask = var_5525_end_mask_0, x = obj_209_cast_fp16)[name = tensor<string, []>("op_5525_cast_fp16")];
            tensor<int32, [4]> var_5528_begin_0 = const()[name = tensor<string, []>("op_5528_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5528_end_0 = const()[name = tensor<string, []>("op_5528_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5528_end_mask_0 = const()[name = tensor<string, []>("op_5528_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5528_squeeze_mask_0 = const()[name = tensor<string, []>("op_5528_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5528_cast_fp16 = slice_by_index(begin = var_5528_begin_0, end = var_5528_end_0, end_mask = var_5528_end_mask_0, squeeze_mask = var_5528_squeeze_mask_0, x = var_5525_cast_fp16)[name = tensor<string, []>("op_5528_cast_fp16")];
            tensor<int32, [4]> var_5543_begin_0 = const()[name = tensor<string, []>("op_5543_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
            tensor<int32, [4]> var_5543_end_0 = const()[name = tensor<string, []>("op_5543_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
            tensor<bool, [4]> var_5543_end_mask_0 = const()[name = tensor<string, []>("op_5543_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5543_cast_fp16 = slice_by_index(begin = var_5543_begin_0, end = var_5543_end_0, end_mask = var_5543_end_mask_0, x = obj_223_cast_fp16)[name = tensor<string, []>("op_5543_cast_fp16")];
            tensor<int32, [4]> var_5546_begin_0 = const()[name = tensor<string, []>("op_5546_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5546_end_0 = const()[name = tensor<string, []>("op_5546_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5546_end_mask_0 = const()[name = tensor<string, []>("op_5546_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5546_squeeze_mask_0 = const()[name = tensor<string, []>("op_5546_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5546_cast_fp16 = slice_by_index(begin = var_5546_begin_0, end = var_5546_end_0, end_mask = var_5546_end_mask_0, squeeze_mask = var_5546_squeeze_mask_0, x = var_5543_cast_fp16)[name = tensor<string, []>("op_5546_cast_fp16")];
            tensor<int32, [4]> var_5561_begin_0 = const()[name = tensor<string, []>("op_5561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5561_end_0 = const()[name = tensor<string, []>("op_5561_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5561_end_mask_0 = const()[name = tensor<string, []>("op_5561_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, x = obj_237_cast_fp16)[name = tensor<string, []>("op_5561_cast_fp16")];
            tensor<int32, [4]> var_5564_begin_0 = const()[name = tensor<string, []>("op_5564_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5564_end_0 = const()[name = tensor<string, []>("op_5564_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5564_end_mask_0 = const()[name = tensor<string, []>("op_5564_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5564_squeeze_mask_0 = const()[name = tensor<string, []>("op_5564_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5564_cast_fp16 = slice_by_index(begin = var_5564_begin_0, end = var_5564_end_0, end_mask = var_5564_end_mask_0, squeeze_mask = var_5564_squeeze_mask_0, x = var_5561_cast_fp16)[name = tensor<string, []>("op_5564_cast_fp16")];
            tensor<int32, [4]> var_5579_begin_0 = const()[name = tensor<string, []>("op_5579_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
            tensor<int32, [4]> var_5579_end_0 = const()[name = tensor<string, []>("op_5579_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
            tensor<bool, [4]> var_5579_end_mask_0 = const()[name = tensor<string, []>("op_5579_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, x = obj_237_cast_fp16)[name = tensor<string, []>("op_5579_cast_fp16")];
            tensor<int32, [4]> var_5582_begin_0 = const()[name = tensor<string, []>("op_5582_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5582_end_0 = const()[name = tensor<string, []>("op_5582_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5582_end_mask_0 = const()[name = tensor<string, []>("op_5582_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5582_squeeze_mask_0 = const()[name = tensor<string, []>("op_5582_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5582_cast_fp16 = slice_by_index(begin = var_5582_begin_0, end = var_5582_end_0, end_mask = var_5582_end_mask_0, squeeze_mask = var_5582_squeeze_mask_0, x = var_5579_cast_fp16)[name = tensor<string, []>("op_5582_cast_fp16")];
            tensor<int32, [4]> var_5597_begin_0 = const()[name = tensor<string, []>("op_5597_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
            tensor<int32, [4]> var_5597_end_0 = const()[name = tensor<string, []>("op_5597_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1500])];
            tensor<bool, [4]> var_5597_end_mask_0 = const()[name = tensor<string, []>("op_5597_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5597_cast_fp16 = slice_by_index(begin = var_5597_begin_0, end = var_5597_end_0, end_mask = var_5597_end_mask_0, x = obj_237_cast_fp16)[name = tensor<string, []>("op_5597_cast_fp16")];
            tensor<int32, [4]> var_5600_begin_0 = const()[name = tensor<string, []>("op_5600_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5600_end_0 = const()[name = tensor<string, []>("op_5600_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5600_end_mask_0 = const()[name = tensor<string, []>("op_5600_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5600_squeeze_mask_0 = const()[name = tensor<string, []>("op_5600_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5600_cast_fp16 = slice_by_index(begin = var_5600_begin_0, end = var_5600_end_0, end_mask = var_5600_end_mask_0, squeeze_mask = var_5600_squeeze_mask_0, x = var_5597_cast_fp16)[name = tensor<string, []>("op_5600_cast_fp16")];
            tensor<int32, [4]> var_5615_begin_0 = const()[name = tensor<string, []>("op_5615_begin_0"), val = tensor<int32, [4]>([0, 12, 0, 0])];
            tensor<int32, [4]> var_5615_end_0 = const()[name = tensor<string, []>("op_5615_end_0"), val = tensor<int32, [4]>([1, 13, 1, 1500])];
            tensor<bool, [4]> var_5615_end_mask_0 = const()[name = tensor<string, []>("op_5615_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5615_cast_fp16 = slice_by_index(begin = var_5615_begin_0, end = var_5615_end_0, end_mask = var_5615_end_mask_0, x = obj_251_cast_fp16)[name = tensor<string, []>("op_5615_cast_fp16")];
            tensor<int32, [4]> var_5618_begin_0 = const()[name = tensor<string, []>("op_5618_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5618_end_0 = const()[name = tensor<string, []>("op_5618_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5618_end_mask_0 = const()[name = tensor<string, []>("op_5618_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5618_squeeze_mask_0 = const()[name = tensor<string, []>("op_5618_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5618_cast_fp16 = slice_by_index(begin = var_5618_begin_0, end = var_5618_end_0, end_mask = var_5618_end_mask_0, squeeze_mask = var_5618_squeeze_mask_0, x = var_5615_cast_fp16)[name = tensor<string, []>("op_5618_cast_fp16")];
            tensor<int32, [4]> var_5633_begin_0 = const()[name = tensor<string, []>("op_5633_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
            tensor<int32, [4]> var_5633_end_0 = const()[name = tensor<string, []>("op_5633_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])];
            tensor<bool, [4]> var_5633_end_mask_0 = const()[name = tensor<string, []>("op_5633_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, x = obj_251_cast_fp16)[name = tensor<string, []>("op_5633_cast_fp16")];
            tensor<int32, [4]> var_5636_begin_0 = const()[name = tensor<string, []>("op_5636_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5636_end_0 = const()[name = tensor<string, []>("op_5636_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5636_end_mask_0 = const()[name = tensor<string, []>("op_5636_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5636_squeeze_mask_0 = const()[name = tensor<string, []>("op_5636_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5636_cast_fp16 = slice_by_index(begin = var_5636_begin_0, end = var_5636_end_0, end_mask = var_5636_end_mask_0, squeeze_mask = var_5636_squeeze_mask_0, x = var_5633_cast_fp16)[name = tensor<string, []>("op_5636_cast_fp16")];
            tensor<int32, [4]> var_5651_begin_0 = const()[name = tensor<string, []>("op_5651_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])];
            tensor<int32, [4]> var_5651_end_0 = const()[name = tensor<string, []>("op_5651_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1500])];
            tensor<bool, [4]> var_5651_end_mask_0 = const()[name = tensor<string, []>("op_5651_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5651_cast_fp16 = slice_by_index(begin = var_5651_begin_0, end = var_5651_end_0, end_mask = var_5651_end_mask_0, x = obj_265_cast_fp16)[name = tensor<string, []>("op_5651_cast_fp16")];
            tensor<int32, [4]> var_5654_begin_0 = const()[name = tensor<string, []>("op_5654_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5654_end_0 = const()[name = tensor<string, []>("op_5654_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5654_end_mask_0 = const()[name = tensor<string, []>("op_5654_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5654_squeeze_mask_0 = const()[name = tensor<string, []>("op_5654_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5654_cast_fp16 = slice_by_index(begin = var_5654_begin_0, end = var_5654_end_0, end_mask = var_5654_end_mask_0, squeeze_mask = var_5654_squeeze_mask_0, x = var_5651_cast_fp16)[name = tensor<string, []>("op_5654_cast_fp16")];
            tensor<int32, [4]> var_5669_begin_0 = const()[name = tensor<string, []>("op_5669_begin_0"), val = tensor<int32, [4]>([0, 10, 0, 0])];
            tensor<int32, [4]> var_5669_end_0 = const()[name = tensor<string, []>("op_5669_end_0"), val = tensor<int32, [4]>([1, 11, 1, 1500])];
            tensor<bool, [4]> var_5669_end_mask_0 = const()[name = tensor<string, []>("op_5669_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5669_cast_fp16 = slice_by_index(begin = var_5669_begin_0, end = var_5669_end_0, end_mask = var_5669_end_mask_0, x = obj_265_cast_fp16)[name = tensor<string, []>("op_5669_cast_fp16")];
            tensor<int32, [4]> var_5672_begin_0 = const()[name = tensor<string, []>("op_5672_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5672_end_0 = const()[name = tensor<string, []>("op_5672_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5672_end_mask_0 = const()[name = tensor<string, []>("op_5672_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5672_squeeze_mask_0 = const()[name = tensor<string, []>("op_5672_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5672_cast_fp16 = slice_by_index(begin = var_5672_begin_0, end = var_5672_end_0, end_mask = var_5672_end_mask_0, squeeze_mask = var_5672_squeeze_mask_0, x = var_5669_cast_fp16)[name = tensor<string, []>("op_5672_cast_fp16")];
            tensor<int32, [4]> var_5687_begin_0 = const()[name = tensor<string, []>("op_5687_begin_0"), val = tensor<int32, [4]>([0, 15, 0, 0])];
            tensor<int32, [4]> var_5687_end_0 = const()[name = tensor<string, []>("op_5687_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5687_end_mask_0 = const()[name = tensor<string, []>("op_5687_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5687_cast_fp16 = slice_by_index(begin = var_5687_begin_0, end = var_5687_end_0, end_mask = var_5687_end_mask_0, x = obj_265_cast_fp16)[name = tensor<string, []>("op_5687_cast_fp16")];
            tensor<int32, [4]> var_5690_begin_0 = const()[name = tensor<string, []>("op_5690_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5690_end_0 = const()[name = tensor<string, []>("op_5690_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5690_end_mask_0 = const()[name = tensor<string, []>("op_5690_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5690_squeeze_mask_0 = const()[name = tensor<string, []>("op_5690_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5690_cast_fp16 = slice_by_index(begin = var_5690_begin_0, end = var_5690_end_0, end_mask = var_5690_end_mask_0, squeeze_mask = var_5690_squeeze_mask_0, x = var_5687_cast_fp16)[name = tensor<string, []>("op_5690_cast_fp16")];
            tensor<int32, [4]> var_5705_begin_0 = const()[name = tensor<string, []>("op_5705_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5705_end_0 = const()[name = tensor<string, []>("op_5705_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5705_end_mask_0 = const()[name = tensor<string, []>("op_5705_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5705_cast_fp16 = slice_by_index(begin = var_5705_begin_0, end = var_5705_end_0, end_mask = var_5705_end_mask_0, x = obj_293_cast_fp16)[name = tensor<string, []>("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5708_begin_0 = const()[name = tensor<string, []>("op_5708_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5708_end_0 = const()[name = tensor<string, []>("op_5708_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5708_end_mask_0 = const()[name = tensor<string, []>("op_5708_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5708_squeeze_mask_0 = const()[name = tensor<string, []>("op_5708_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5708_cast_fp16 = slice_by_index(begin = var_5708_begin_0, end = var_5708_end_0, end_mask = var_5708_end_mask_0, squeeze_mask = var_5708_squeeze_mask_0, x = var_5705_cast_fp16)[name = tensor<string, []>("op_5708_cast_fp16")];
            tensor<int32, [4]> var_5723_begin_0 = const()[name = tensor<string, []>("op_5723_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
            tensor<int32, [4]> var_5723_end_0 = const()[name = tensor<string, []>("op_5723_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])];
            tensor<bool, [4]> var_5723_end_mask_0 = const()[name = tensor<string, []>("op_5723_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5723_cast_fp16 = slice_by_index(begin = var_5723_begin_0, end = var_5723_end_0, end_mask = var_5723_end_mask_0, x = obj_293_cast_fp16)[name = tensor<string, []>("op_5723_cast_fp16")];
            tensor<int32, [4]> var_5726_begin_0 = const()[name = tensor<string, []>("op_5726_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5726_end_0 = const()[name = tensor<string, []>("op_5726_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5726_end_mask_0 = const()[name = tensor<string, []>("op_5726_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5726_squeeze_mask_0 = const()[name = tensor<string, []>("op_5726_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5726_cast_fp16 = slice_by_index(begin = var_5726_begin_0, end = var_5726_end_0, end_mask = var_5726_end_mask_0, squeeze_mask = var_5726_squeeze_mask_0, x = var_5723_cast_fp16)[name = tensor<string, []>("op_5726_cast_fp16")];
            tensor<int32, [4]> var_5741_begin_0 = const()[name = tensor<string, []>("op_5741_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
            tensor<int32, [4]> var_5741_end_0 = const()[name = tensor<string, []>("op_5741_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1500])];
            tensor<bool, [4]> var_5741_end_mask_0 = const()[name = tensor<string, []>("op_5741_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5741_cast_fp16 = slice_by_index(begin = var_5741_begin_0, end = var_5741_end_0, end_mask = var_5741_end_mask_0, x = obj_293_cast_fp16)[name = tensor<string, []>("op_5741_cast_fp16")];
            tensor<int32, [4]> var_5744_begin_0 = const()[name = tensor<string, []>("op_5744_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5744_end_0 = const()[name = tensor<string, []>("op_5744_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5744_end_mask_0 = const()[name = tensor<string, []>("op_5744_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5744_squeeze_mask_0 = const()[name = tensor<string, []>("op_5744_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5744_cast_fp16 = slice_by_index(begin = var_5744_begin_0, end = var_5744_end_0, end_mask = var_5744_end_mask_0, squeeze_mask = var_5744_squeeze_mask_0, x = var_5741_cast_fp16)[name = tensor<string, []>("op_5744_cast_fp16")];
            tensor<int32, [4]> var_5759_begin_0 = const()[name = tensor<string, []>("op_5759_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
            tensor<int32, [4]> var_5759_end_0 = const()[name = tensor<string, []>("op_5759_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])];
            tensor<bool, [4]> var_5759_end_mask_0 = const()[name = tensor<string, []>("op_5759_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5759_cast_fp16 = slice_by_index(begin = var_5759_begin_0, end = var_5759_end_0, end_mask = var_5759_end_mask_0, x = obj_293_cast_fp16)[name = tensor<string, []>("op_5759_cast_fp16")];
            tensor<int32, [4]> var_5762_begin_0 = const()[name = tensor<string, []>("op_5762_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5762_end_0 = const()[name = tensor<string, []>("op_5762_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5762_end_mask_0 = const()[name = tensor<string, []>("op_5762_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5762_squeeze_mask_0 = const()[name = tensor<string, []>("op_5762_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5762_cast_fp16 = slice_by_index(begin = var_5762_begin_0, end = var_5762_end_0, end_mask = var_5762_end_mask_0, squeeze_mask = var_5762_squeeze_mask_0, x = var_5759_cast_fp16)[name = tensor<string, []>("op_5762_cast_fp16")];
            tensor<int32, [4]> var_5777_begin_0 = const()[name = tensor<string, []>("op_5777_begin_0"), val = tensor<int32, [4]>([0, 12, 0, 0])];
            tensor<int32, [4]> var_5777_end_0 = const()[name = tensor<string, []>("op_5777_end_0"), val = tensor<int32, [4]>([1, 13, 1, 1500])];
            tensor<bool, [4]> var_5777_end_mask_0 = const()[name = tensor<string, []>("op_5777_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1, 1, 1500]> var_5777_cast_fp16 = slice_by_index(begin = var_5777_begin_0, end = var_5777_end_0, end_mask = var_5777_end_mask_0, x = obj_307_cast_fp16)[name = tensor<string, []>("op_5777_cast_fp16")];
            tensor<int32, [4]> var_5780_begin_0 = const()[name = tensor<string, []>("op_5780_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5780_end_0 = const()[name = tensor<string, []>("op_5780_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
            tensor<bool, [4]> var_5780_end_mask_0 = const()[name = tensor<string, []>("op_5780_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5780_squeeze_mask_0 = const()[name = tensor<string, []>("op_5780_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 1, 1500]> var_5780_cast_fp16 = slice_by_index(begin = var_5780_begin_0, end = var_5780_end_0, end_mask = var_5780_end_mask_0, squeeze_mask = var_5780_squeeze_mask_0, x = var_5777_cast_fp16)[name = tensor<string, []>("op_5780_cast_fp16")];
            tensor<int32, []> var_5787 = const()[name = tensor<string, []>("op_5787"), val = tensor<int32, []>(1)];
            tensor<bool, []> var_5788_interleave_0 = const()[name = tensor<string, []>("op_5788_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 18, 1500]> var_5788_cast_fp16 = concat(axis = var_5787, interleave = var_5788_interleave_0, values = (var_5474_cast_fp16, var_5492_cast_fp16, var_5510_cast_fp16, var_5528_cast_fp16, var_5546_cast_fp16, var_5564_cast_fp16, var_5582_cast_fp16, var_5600_cast_fp16, var_5618_cast_fp16, var_5636_cast_fp16, var_5654_cast_fp16, var_5672_cast_fp16, var_5690_cast_fp16, var_5708_cast_fp16, var_5726_cast_fp16, var_5744_cast_fp16, var_5762_cast_fp16, var_5780_cast_fp16))[name = tensor<string, []>("op_5788_cast_fp16")];
            tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> obj_keep_dims_0 = const()[name = tensor<string, []>("obj_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_5788_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
        } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
}