program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func main<ios17>(tensor<fp32, [5, 1, ?, 1024]> decoder_mems, tensor<fp32, [1, 125, 1024]> encoder_embeddings, tensor<fp32, [1, 125]> encoder_mask, tensor<int32, [1, 1]> input_ids, tensor<int32, [1]> start_pos) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"decoder_mems", [5, 1, 1, 1024]}}), ("RangeDims", {{"decoder_mems", [[5, 5], [1, 1], [1, 512], [1024, 1024]]}})))] {
            tensor<int32, [1, 1]> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<int32, [1, 1]>([[0]])];
            tensor<int32, [1, 1]> position_ids = add(x = var_40, y = start_pos)[name = tensor<string, []>("position_ids")];
            tensor<int32, []> var_46_batch_dims_0 = const()[name = tensor<string, []>("op_46_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_46_validate_indices_0 = const()[name = tensor<string, []>("op_46_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [5248, 1024]> embedding_token_embedding_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("embedding_token_embedding_weight_to_fp16_quantized"), quantized_data = tensor<int8, [5248, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), scale = tensor<fp16, [5248]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5379392))), zero_point = tensor<int8, [5248]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5374080)))];
            tensor<string, []> input_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("input_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<string, []> cast_70_dtype_0 = const()[name = tensor<string, []>("cast_70_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
            tensor<int16, [1, 1]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = tensor<string, []>("cast_9")];
            tensor<int32, [1, 1]> cast_70 = cast(dtype = cast_70_dtype_0, x = input_ids_to_int16)[name = tensor<string, []>("cast_8")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_70, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(5248)];
            tensor<int32, [1, 1]> add_0 = add(x = cast_70, y = slice_by_index_0)[name = tensor<string, []>("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = cast_70, b = add_0, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<string, []> cast_0_dtype_0 = const()[name = tensor<string, []>("cast_0_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> greater_equal_0_y_0_1 = const()[name = tensor<string, []>("greater_equal_0_y_0_1"), val = tensor<int32, []>(0)];
            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_7")];
            tensor<int32, [1, 1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = tensor<string, []>("cast_6")];
            tensor<bool, [1, 1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = tensor<string, []>("greater_equal_0_1")];
            tensor<int32, []> slice_by_index_0_1 = const()[name = tensor<string, []>("slice_by_index_0_1"), val = tensor<int32, []>(5248)];
            tensor<int32, [1, 1]> add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = tensor<string, []>("add_0_1")];
            tensor<int32, [1, 1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = tensor<string, []>("select_0_1")];
            tensor<int32, []> op_46_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_46_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [1, 1, 1024]> op_46_cast_fp16_cast_uint16_cast_uint16 = gather(axis = op_46_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = var_46_batch_dims_0, indices = select_0_1, validate_indices = var_46_validate_indices_0, x = embedding_token_embedding_weight_to_fp16_quantized)[name = tensor<string, []>("op_46_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, []> var_50_axis_0 = const()[name = tensor<string, []>("op_50_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_50_batch_dims_0 = const()[name = tensor<string, []>("op_50_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_50_validate_indices_0 = const()[name = tensor<string, []>("op_50_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 1024]> embedding_position_embedding_pos_enc_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("embedding_position_embedding_pos_enc_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5389952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6439680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<string, []> position_ids_to_uint16_dtype_0 = const()[name = tensor<string, []>("position_ids_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, 1]> position_ids_to_uint16 = cast(dtype = position_ids_to_uint16_dtype_0, x = position_ids)[name = tensor<string, []>("cast_5")];
            tensor<fp16, [1, 1, 1024]> var_50_cast_fp16_cast_uint16 = gather(axis = var_50_axis_0, batch_dims = var_50_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_50_validate_indices_0, x = embedding_position_embedding_pos_enc_to_fp16_quantized)[name = tensor<string, []>("op_50_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = add(x = op_46_cast_fp16_cast_uint16_cast_uint16, y = var_50_cast_fp16_cast_uint16)[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<int32, [1]> states_1_axes_0 = const()[name = tensor<string, []>("states_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> embedding_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("embedding_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6441792)))];
            tensor<fp16, [1024]> embedding_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("embedding_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6443904)))];
            tensor<fp16, []> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> states_1_cast_fp16 = layer_norm(axes = states_1_axes_0, beta = embedding_layer_norm_bias_to_fp16, epsilon = var_54_to_fp16, gamma = embedding_layer_norm_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("states_1_cast_fp16")];
            tensor<string, []> var_115_dtype_0 = const()[name = tensor<string, []>("op_115_dtype_0"), val = tensor<string, []>("bool")];
            tensor<int32, [1]> attn_mask_axes_0 = const()[name = tensor<string, []>("attn_mask_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 125]> var_115 = cast(dtype = var_115_dtype_0, x = encoder_mask)[name = tensor<string, []>("cast_4")];
            tensor<bool, [1, 1, 125]> attn_mask = expand_dims(axes = attn_mask_axes_0, x = var_115)[name = tensor<string, []>("attn_mask")];
            tensor<fp16, []> var_123_promoted_to_fp16 = const()[name = tensor<string, []>("op_123_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> var_122_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_122_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 125]> attn_mask_to_fp16 = cast(dtype = var_122_to_fp16_dtype_0, x = attn_mask)[name = tensor<string, []>("cast_3")];
            tensor<fp16, [1, 1, 125]> var_125_cast_fp16 = sub(x = var_123_promoted_to_fp16, y = attn_mask_to_fp16)[name = tensor<string, []>("op_125_cast_fp16")];
            tensor<fp16, []> var_126_to_fp16 = const()[name = tensor<string, []>("op_126_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
            tensor<fp16, [1, 1, 125]> attention_mask_3_cast_fp16 = mul(x = var_125_cast_fp16, y = var_126_to_fp16)[name = tensor<string, []>("attention_mask_3_cast_fp16")];
            tensor<int32, [1]> attention_mask_9_axes_0 = const()[name = tensor<string, []>("attention_mask_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 125]> attention_mask_9_cast_fp16 = expand_dims(axes = attention_mask_9_axes_0, x = attention_mask_3_cast_fp16)[name = tensor<string, []>("attention_mask_9_cast_fp16")];
            tensor<int32, [4]> prior_1_begin_0 = const()[name = tensor<string, []>("prior_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> prior_1_end_0 = const()[name = tensor<string, []>("prior_1_end_0"), val = tensor<int32, [4]>([1, 1, 0, 1024])];
            tensor<bool, [4]> prior_1_end_mask_0 = const()[name = tensor<string, []>("prior_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> prior_1_squeeze_mask_0 = const()[name = tensor<string, []>("prior_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<string, []> decoder_mems_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_mems_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [5, 1, ?, 1024]> decoder_mems_to_fp16 = cast(dtype = decoder_mems_to_fp16_dtype_0, x = decoder_mems)[name = tensor<string, []>("cast_2")];
            tensor<fp16, [1, ?, 1024]> prior_1_cast_fp16 = slice_by_index(begin = prior_1_begin_0, end = prior_1_end_0, end_mask = prior_1_end_mask_0, squeeze_mask = prior_1_squeeze_mask_0, x = decoder_mems_to_fp16)[name = tensor<string, []>("prior_1_cast_fp16")];
            tensor<bool, []> var_141_interleave_0 = const()[name = tensor<string, []>("op_141_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, []> const_69 = const()[name = tensor<string, []>("const_69"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, ?, 1024]> var_141_cast_fp16 = concat(axis = const_69, interleave = var_141_interleave_0, values = (prior_1_cast_fp16, states_1_cast_fp16))[name = tensor<string, []>("op_141_cast_fp16")];
            tensor<int32, []> var_151 = const()[name = tensor<string, []>("op_151"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6446016)))];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6448128)))];
            tensor<fp16, []> var_158_to_fp16 = const()[name = tensor<string, []>("op_158_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = decoder_layers_0_layer_norm_1_bias_to_fp16, epsilon = var_158_to_fp16, gamma = decoder_layers_0_layer_norm_1_weight_to_fp16, x = states_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, ?, 1024]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = decoder_layers_0_layer_norm_1_bias_to_fp16, epsilon = var_158_to_fp16, gamma = decoder_layers_0_layer_norm_1_weight_to_fp16, x = var_141_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_first_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_first_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6450240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7498880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7500992)))];
            tensor<fp16, [1, 1, 1024]> linear_0_cast_fp16 = linear(bias = decoder_layers_0_first_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_0_first_sub_layer_query_net_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_first_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_first_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7503104))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8551744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8553856)))];
            tensor<fp16, [1, ?, 1024]> linear_1_cast_fp16 = linear(bias = decoder_layers_0_first_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_0_first_sub_layer_key_net_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_first_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_first_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8555968))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9604608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9606720)))];
            tensor<fp16, [1, ?, 1024]> linear_2_cast_fp16 = linear(bias = decoder_layers_0_first_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_0_first_sub_layer_value_net_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [4]> var_189 = const()[name = tensor<string, []>("op_189"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_3_cast_fp16 = reshape(shape = var_189, x = linear_0_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<fp16, []> _inversed_query_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_1_cast_fp16 = mul(x = x_3_cast_fp16, y = _inversed_query_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_7_cast_fp16 = reshape(shape = concat_0x, x = linear_1_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<fp16, []> _inversed_key_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, ?, 8, 128]> _inversed_key_1_cast_fp16 = mul(x = x_7_cast_fp16, y = _inversed_key_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_11_cast_fp16 = reshape(shape = concat_1x, x = linear_2_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [4]> var_207 = const()[name = tensor<string, []>("op_207"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_1_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_1_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_1_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_1_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, ?, 128]> transpose_33 = transpose(perm = transpose_33_perm_0, x = _inversed_key_1_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [1, 8, 1, 128]> transpose_32 = transpose(perm = transpose_32_perm_0, x = _inversed_query_1_cast_fp16)[name = tensor<string, []>("transpose_79")];
            tensor<fp16, [1, 8, 1, ?]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_1, transpose_y = attention_scores_1_transpose_y_1, x = transpose_32, y = transpose_33)[name = tensor<string, []>("attention_scores_1_cast_fp16")];
            tensor<fp16, [1, 8, 1, ?]> input_9_cast_fp16 = softmax(axis = var_151, x = attention_scores_1_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<bool, []> context_1_transpose_x_0 = const()[name = tensor<string, []>("context_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_1_transpose_y_0 = const()[name = tensor<string, []>("context_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, ?, 128]> value_1_cast_fp16 = transpose(perm = var_207, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 8, 1, 128]> context_1_cast_fp16 = matmul(transpose_x = context_1_transpose_x_0, transpose_y = context_1_transpose_y_0, x = input_9_cast_fp16, y = value_1_cast_fp16)[name = tensor<string, []>("context_1_cast_fp16")];
            tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_222_cast_fp16 = transpose(perm = var_221, x = context_1_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [1, 1, 1024]> input_11_cast_fp16 = reshape(shape = var_226, x = var_222_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_first_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_first_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9608832))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10657472))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10659584)))];
            tensor<fp16, [1, 1, 1024]> linear_3_cast_fp16 = linear(bias = decoder_layers_0_first_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_0_first_sub_layer_out_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_15_cast_fp16 = add(x = linear_3_cast_fp16, y = states_1_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10661696)))];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10663808)))];
            tensor<fp16, [1, 1, 1024]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = decoder_layers_0_layer_norm_2_bias_to_fp16, epsilon = var_158_to_fp16, gamma = decoder_layers_0_layer_norm_2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_second_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_second_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10665920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11714560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11716672)))];
            tensor<fp16, [1, 1, 1024]> linear_4_cast_fp16 = linear(bias = decoder_layers_0_second_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_0_second_sub_layer_query_net_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<string, []> encoder_embeddings_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_embeddings_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_second_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_second_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11718784))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12767424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12769536)))];
            tensor<fp16, [1, 125, 1024]> encoder_embeddings_to_fp16 = cast(dtype = encoder_embeddings_to_fp16_dtype_0, x = encoder_embeddings)[name = tensor<string, []>("cast_1")];
            tensor<fp16, [1, 125, 1024]> linear_5_cast_fp16 = linear(bias = decoder_layers_0_second_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_0_second_sub_layer_key_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_second_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_second_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12771648))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13820288))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13822400)))];
            tensor<fp16, [1, 125, 1024]> linear_6_cast_fp16 = linear(bias = decoder_layers_0_second_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_0_second_sub_layer_value_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<int32, [4]> var_255 = const()[name = tensor<string, []>("op_255"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_15_cast_fp16 = reshape(shape = var_255, x = linear_4_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<fp16, []> _inversed_query_3_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_3_cast_fp16 = mul(x = x_15_cast_fp16, y = _inversed_query_3_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_3_cast_fp16")];
            tensor<int32, [4]> var_263 = const()[name = tensor<string, []>("op_263"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_19_cast_fp16 = reshape(shape = var_263, x = linear_5_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<fp16, []> _inversed_key_3_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 125, 8, 128]> _inversed_key_3_cast_fp16 = mul(x = x_19_cast_fp16, y = _inversed_key_3_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_3_cast_fp16")];
            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_23_cast_fp16 = reshape(shape = var_271, x = linear_6_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<int32, [4]> var_273 = const()[name = tensor<string, []>("op_273"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_5_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_5_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_5_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_5_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 125, 128]> transpose_35 = transpose(perm = transpose_35_perm_0, x = _inversed_key_3_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [1, 8, 1, 128]> transpose_34 = transpose(perm = transpose_34_perm_0, x = _inversed_query_3_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_5_cast_fp16 = matmul(transpose_x = attention_scores_5_transpose_x_1, transpose_y = attention_scores_5_transpose_y_1, x = transpose_34, y = transpose_35)[name = tensor<string, []>("attention_scores_5_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_7_cast_fp16 = add(x = attention_scores_5_cast_fp16, y = attention_mask_9_cast_fp16)[name = tensor<string, []>("attention_scores_7_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> input_19_cast_fp16 = softmax(axis = var_151, x = attention_scores_7_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<bool, []> context_5_transpose_x_0 = const()[name = tensor<string, []>("context_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_5_transpose_y_0 = const()[name = tensor<string, []>("context_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 125, 128]> value_3_cast_fp16 = transpose(perm = var_273, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [1, 8, 1, 128]> context_5_cast_fp16 = matmul(transpose_x = context_5_transpose_x_0, transpose_y = context_5_transpose_y_0, x = input_19_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("context_5_cast_fp16")];
            tensor<int32, [4]> var_287 = const()[name = tensor<string, []>("op_287"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_292 = const()[name = tensor<string, []>("op_292"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_288_cast_fp16 = transpose(perm = var_287, x = context_5_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 1, 1024]> input_21_cast_fp16 = reshape(shape = var_292, x = var_288_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_0_second_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_second_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13824512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14873152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14875264)))];
            tensor<fp16, [1, 1, 1024]> linear_7_cast_fp16 = linear(bias = decoder_layers_0_second_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_0_second_sub_layer_out_projection_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_25_cast_fp16 = add(x = linear_7_cast_fp16, y = input_15_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14877376)))];
            tensor<fp16, [1024]> decoder_layers_0_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14879488)))];
            tensor<fp16, [1, 1, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = decoder_layers_0_layer_norm_3_bias_to_fp16, epsilon = var_158_to_fp16, gamma = decoder_layers_0_layer_norm_3_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<fp16, [4096, 1024]> decoder_layers_0_third_sub_layer_dense_in_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_third_sub_layer_dense_in_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14881600))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19080128))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19075968)))];
            tensor<fp16, [4096]> decoder_layers_0_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19088384)))];
            tensor<fp16, [1, 1, 4096]> linear_8_cast_fp16 = linear(bias = decoder_layers_0_third_sub_layer_dense_in_bias_to_fp16, weight = decoder_layers_0_third_sub_layer_dense_in_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_29_cast_fp16 = relu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [1024, 4096]> decoder_layers_0_third_sub_layer_dense_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_0_third_sub_layer_dense_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19096640))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23291008))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_0_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_0_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23293120)))];
            tensor<fp16, [1, 1, 1024]> linear_9_cast_fp16 = linear(bias = decoder_layers_0_third_sub_layer_dense_out_bias_to_fp16, weight = decoder_layers_0_third_sub_layer_dense_out_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> states_3_cast_fp16 = add(x = linear_9_cast_fp16, y = input_25_cast_fp16)[name = tensor<string, []>("states_3_cast_fp16")];
            tensor<int32, [4]> prior_3_begin_0 = const()[name = tensor<string, []>("prior_3_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> prior_3_end_0 = const()[name = tensor<string, []>("prior_3_end_0"), val = tensor<int32, [4]>([2, 1, 0, 1024])];
            tensor<bool, [4]> prior_3_end_mask_0 = const()[name = tensor<string, []>("prior_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> prior_3_squeeze_mask_0 = const()[name = tensor<string, []>("prior_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, ?, 1024]> prior_3_cast_fp16 = slice_by_index(begin = prior_3_begin_0, end = prior_3_end_0, end_mask = prior_3_end_mask_0, squeeze_mask = prior_3_squeeze_mask_0, x = decoder_mems_to_fp16)[name = tensor<string, []>("prior_3_cast_fp16")];
            tensor<bool, []> var_332_interleave_0 = const()[name = tensor<string, []>("op_332_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, []> const_70 = const()[name = tensor<string, []>("const_70"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, ?, 1024]> var_332_cast_fp16 = concat(axis = const_70, interleave = var_332_interleave_0, values = (prior_3_cast_fp16, states_3_cast_fp16))[name = tensor<string, []>("op_332_cast_fp16")];
            tensor<int32, []> var_342 = const()[name = tensor<string, []>("op_342"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> input_35_axes_0 = const()[name = tensor<string, []>("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23295232)))];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23297344)))];
            tensor<fp16, []> var_349_to_fp16 = const()[name = tensor<string, []>("op_349_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = decoder_layers_1_layer_norm_1_bias_to_fp16, epsilon = var_349_to_fp16, gamma = decoder_layers_1_layer_norm_1_weight_to_fp16, x = states_3_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<int32, [1]> input_37_axes_0 = const()[name = tensor<string, []>("input_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, ?, 1024]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = decoder_layers_1_layer_norm_1_bias_to_fp16, epsilon = var_349_to_fp16, gamma = decoder_layers_1_layer_norm_1_weight_to_fp16, x = var_332_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_first_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_first_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23299456))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24348096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24350208)))];
            tensor<fp16, [1, 1, 1024]> linear_10_cast_fp16 = linear(bias = decoder_layers_1_first_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_1_first_sub_layer_query_net_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_first_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_first_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24352320))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25400960))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25403072)))];
            tensor<fp16, [1, ?, 1024]> linear_11_cast_fp16 = linear(bias = decoder_layers_1_first_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_1_first_sub_layer_key_net_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_first_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_first_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25405184))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26453824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26455936)))];
            tensor<fp16, [1, ?, 1024]> linear_12_cast_fp16 = linear(bias = decoder_layers_1_first_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_1_first_sub_layer_value_net_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [4]> var_380 = const()[name = tensor<string, []>("op_380"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_27_cast_fp16 = reshape(shape = var_380, x = linear_10_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<fp16, []> _inversed_query_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_5_cast_fp16 = mul(x = x_27_cast_fp16, y = _inversed_query_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_5_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_31_cast_fp16 = reshape(shape = concat_2x, x = linear_11_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<fp16, []> _inversed_key_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, ?, 8, 128]> _inversed_key_5_cast_fp16 = mul(x = x_31_cast_fp16, y = _inversed_key_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_5_cast_fp16")];
            tensor<int32, [4]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_35_cast_fp16 = reshape(shape = concat_3x, x = linear_12_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<int32, [4]> var_398 = const()[name = tensor<string, []>("op_398"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_9_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_9_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_9_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_9_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = tensor<string, []>("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_37_perm_0 = const()[name = tensor<string, []>("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, ?, 128]> transpose_37 = transpose(perm = transpose_37_perm_0, x = _inversed_key_5_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [1, 8, 1, 128]> transpose_36 = transpose(perm = transpose_36_perm_0, x = _inversed_query_5_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [1, 8, 1, ?]> attention_scores_9_cast_fp16 = matmul(transpose_x = attention_scores_9_transpose_x_1, transpose_y = attention_scores_9_transpose_y_1, x = transpose_36, y = transpose_37)[name = tensor<string, []>("attention_scores_9_cast_fp16")];
            tensor<fp16, [1, 8, 1, ?]> input_39_cast_fp16 = softmax(axis = var_342, x = attention_scores_9_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<bool, []> context_9_transpose_x_0 = const()[name = tensor<string, []>("context_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_9_transpose_y_0 = const()[name = tensor<string, []>("context_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, ?, 128]> value_5_cast_fp16 = transpose(perm = var_398, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [1, 8, 1, 128]> context_9_cast_fp16 = matmul(transpose_x = context_9_transpose_x_0, transpose_y = context_9_transpose_y_0, x = input_39_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("context_9_cast_fp16")];
            tensor<int32, [4]> var_412 = const()[name = tensor<string, []>("op_412"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_413_cast_fp16 = transpose(perm = var_412, x = context_9_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 1, 1024]> input_41_cast_fp16 = reshape(shape = var_417, x = var_413_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_first_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_first_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26458048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27506688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27508800)))];
            tensor<fp16, [1, 1, 1024]> linear_13_cast_fp16 = linear(bias = decoder_layers_1_first_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_1_first_sub_layer_out_projection_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_45_cast_fp16 = add(x = linear_13_cast_fp16, y = states_3_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = tensor<string, []>("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27510912)))];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27513024)))];
            tensor<fp16, [1, 1, 1024]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = decoder_layers_1_layer_norm_2_bias_to_fp16, epsilon = var_349_to_fp16, gamma = decoder_layers_1_layer_norm_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_second_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_second_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27515136))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28563776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28565888)))];
            tensor<fp16, [1, 1, 1024]> linear_14_cast_fp16 = linear(bias = decoder_layers_1_second_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_1_second_sub_layer_query_net_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_second_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_second_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28568000))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29616640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29618752)))];
            tensor<fp16, [1, 125, 1024]> linear_15_cast_fp16 = linear(bias = decoder_layers_1_second_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_1_second_sub_layer_key_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_second_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_second_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29620864))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30669504))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30671616)))];
            tensor<fp16, [1, 125, 1024]> linear_16_cast_fp16 = linear(bias = decoder_layers_1_second_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_1_second_sub_layer_value_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<int32, [4]> var_446 = const()[name = tensor<string, []>("op_446"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_39_cast_fp16 = reshape(shape = var_446, x = linear_14_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
            tensor<fp16, []> _inversed_query_7_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_7_cast_fp16 = mul(x = x_39_cast_fp16, y = _inversed_query_7_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_7_cast_fp16")];
            tensor<int32, [4]> var_454 = const()[name = tensor<string, []>("op_454"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_43_cast_fp16 = reshape(shape = var_454, x = linear_15_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<fp16, []> _inversed_key_7_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 125, 8, 128]> _inversed_key_7_cast_fp16 = mul(x = x_43_cast_fp16, y = _inversed_key_7_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_7_cast_fp16")];
            tensor<int32, [4]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_47_cast_fp16 = reshape(shape = var_462, x = linear_16_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<int32, [4]> var_464 = const()[name = tensor<string, []>("op_464"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_13_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_13_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_13_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_13_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = tensor<string, []>("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_39_perm_0 = const()[name = tensor<string, []>("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 125, 128]> transpose_39 = transpose(perm = transpose_39_perm_0, x = _inversed_key_7_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [1, 8, 1, 128]> transpose_38 = transpose(perm = transpose_38_perm_0, x = _inversed_query_7_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_1, transpose_y = attention_scores_13_transpose_y_1, x = transpose_38, y = transpose_39)[name = tensor<string, []>("attention_scores_13_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = attention_mask_9_cast_fp16)[name = tensor<string, []>("attention_scores_15_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> input_49_cast_fp16 = softmax(axis = var_342, x = attention_scores_15_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<bool, []> context_13_transpose_x_0 = const()[name = tensor<string, []>("context_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_13_transpose_y_0 = const()[name = tensor<string, []>("context_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 125, 128]> value_7_cast_fp16 = transpose(perm = var_464, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [1, 8, 1, 128]> context_13_cast_fp16 = matmul(transpose_x = context_13_transpose_x_0, transpose_y = context_13_transpose_y_0, x = input_49_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("context_13_cast_fp16")];
            tensor<int32, [4]> var_478 = const()[name = tensor<string, []>("op_478"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_483 = const()[name = tensor<string, []>("op_483"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_479_cast_fp16 = transpose(perm = var_478, x = context_13_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 1, 1024]> input_51_cast_fp16 = reshape(shape = var_483, x = var_479_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_1_second_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_second_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30673728))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31722368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31724480)))];
            tensor<fp16, [1, 1, 1024]> linear_17_cast_fp16 = linear(bias = decoder_layers_1_second_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_1_second_sub_layer_out_projection_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_55_cast_fp16 = add(x = linear_17_cast_fp16, y = input_45_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<int32, [1]> input_57_axes_0 = const()[name = tensor<string, []>("input_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31726592)))];
            tensor<fp16, [1024]> decoder_layers_1_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31728704)))];
            tensor<fp16, [1, 1, 1024]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = decoder_layers_1_layer_norm_3_bias_to_fp16, epsilon = var_349_to_fp16, gamma = decoder_layers_1_layer_norm_3_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<fp16, [4096, 1024]> decoder_layers_1_third_sub_layer_dense_in_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_third_sub_layer_dense_in_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31730816))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35925184))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19075968)))];
            tensor<fp16, [4096]> decoder_layers_1_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35933440)))];
            tensor<fp16, [1, 1, 4096]> linear_18_cast_fp16 = linear(bias = decoder_layers_1_third_sub_layer_dense_in_bias_to_fp16, weight = decoder_layers_1_third_sub_layer_dense_in_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_59_cast_fp16 = relu(x = linear_18_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<fp16, [1024, 4096]> decoder_layers_1_third_sub_layer_dense_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_1_third_sub_layer_dense_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35941696))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40136064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_1_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_1_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40138176)))];
            tensor<fp16, [1, 1, 1024]> linear_19_cast_fp16 = linear(bias = decoder_layers_1_third_sub_layer_dense_out_bias_to_fp16, weight = decoder_layers_1_third_sub_layer_dense_out_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> states_5_cast_fp16 = add(x = linear_19_cast_fp16, y = input_55_cast_fp16)[name = tensor<string, []>("states_5_cast_fp16")];
            tensor<int32, [4]> prior_5_begin_0 = const()[name = tensor<string, []>("prior_5_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> prior_5_end_0 = const()[name = tensor<string, []>("prior_5_end_0"), val = tensor<int32, [4]>([3, 1, 0, 1024])];
            tensor<bool, [4]> prior_5_end_mask_0 = const()[name = tensor<string, []>("prior_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> prior_5_squeeze_mask_0 = const()[name = tensor<string, []>("prior_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, ?, 1024]> prior_5_cast_fp16 = slice_by_index(begin = prior_5_begin_0, end = prior_5_end_0, end_mask = prior_5_end_mask_0, squeeze_mask = prior_5_squeeze_mask_0, x = decoder_mems_to_fp16)[name = tensor<string, []>("prior_5_cast_fp16")];
            tensor<bool, []> var_523_interleave_0 = const()[name = tensor<string, []>("op_523_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, []> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, ?, 1024]> var_523_cast_fp16 = concat(axis = const_71, interleave = var_523_interleave_0, values = (prior_5_cast_fp16, states_5_cast_fp16))[name = tensor<string, []>("op_523_cast_fp16")];
            tensor<int32, []> var_533 = const()[name = tensor<string, []>("op_533"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40140288)))];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40142400)))];
            tensor<fp16, []> var_540_to_fp16 = const()[name = tensor<string, []>("op_540_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = decoder_layers_2_layer_norm_1_bias_to_fp16, epsilon = var_540_to_fp16, gamma = decoder_layers_2_layer_norm_1_weight_to_fp16, x = states_5_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = tensor<string, []>("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, ?, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = decoder_layers_2_layer_norm_1_bias_to_fp16, epsilon = var_540_to_fp16, gamma = decoder_layers_2_layer_norm_1_weight_to_fp16, x = var_523_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_first_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_first_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40144512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41193152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41195264)))];
            tensor<fp16, [1, 1, 1024]> linear_20_cast_fp16 = linear(bias = decoder_layers_2_first_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_2_first_sub_layer_query_net_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_first_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_first_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41197376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42246016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42248128)))];
            tensor<fp16, [1, ?, 1024]> linear_21_cast_fp16 = linear(bias = decoder_layers_2_first_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_2_first_sub_layer_key_net_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_first_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_first_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42250240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43298880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43300992)))];
            tensor<fp16, [1, ?, 1024]> linear_22_cast_fp16 = linear(bias = decoder_layers_2_first_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_2_first_sub_layer_value_net_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_571 = const()[name = tensor<string, []>("op_571"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_51_cast_fp16 = reshape(shape = var_571, x = linear_20_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<fp16, []> _inversed_query_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_9_cast_fp16 = mul(x = x_51_cast_fp16, y = _inversed_query_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_9_cast_fp16")];
            tensor<int32, [4]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_55_cast_fp16 = reshape(shape = concat_4x, x = linear_21_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<fp16, []> _inversed_key_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, ?, 8, 128]> _inversed_key_9_cast_fp16 = mul(x = x_55_cast_fp16, y = _inversed_key_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_9_cast_fp16")];
            tensor<int32, [4]> concat_5x = const()[name = tensor<string, []>("concat_5x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_59_cast_fp16 = reshape(shape = concat_5x, x = linear_22_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<int32, [4]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_17_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_17_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_17_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_17_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = tensor<string, []>("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_41_perm_0 = const()[name = tensor<string, []>("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, ?, 128]> transpose_41 = transpose(perm = transpose_41_perm_0, x = _inversed_key_9_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [1, 8, 1, 128]> transpose_40 = transpose(perm = transpose_40_perm_0, x = _inversed_query_9_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [1, 8, 1, ?]> attention_scores_17_cast_fp16 = matmul(transpose_x = attention_scores_17_transpose_x_1, transpose_y = attention_scores_17_transpose_y_1, x = transpose_40, y = transpose_41)[name = tensor<string, []>("attention_scores_17_cast_fp16")];
            tensor<fp16, [1, 8, 1, ?]> input_69_cast_fp16 = softmax(axis = var_533, x = attention_scores_17_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<bool, []> context_17_transpose_x_0 = const()[name = tensor<string, []>("context_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_17_transpose_y_0 = const()[name = tensor<string, []>("context_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, ?, 128]> value_9_cast_fp16 = transpose(perm = var_589, x = x_59_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [1, 8, 1, 128]> context_17_cast_fp16 = matmul(transpose_x = context_17_transpose_x_0, transpose_y = context_17_transpose_y_0, x = input_69_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("context_17_cast_fp16")];
            tensor<int32, [4]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_608 = const()[name = tensor<string, []>("op_608"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_604_cast_fp16 = transpose(perm = var_603, x = context_17_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [1, 1, 1024]> input_71_cast_fp16 = reshape(shape = var_608, x = var_604_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_first_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_first_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43303104))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44351744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44353856)))];
            tensor<fp16, [1, 1, 1024]> linear_23_cast_fp16 = linear(bias = decoder_layers_2_first_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_2_first_sub_layer_out_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = states_5_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44355968)))];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44358080)))];
            tensor<fp16, [1, 1, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = decoder_layers_2_layer_norm_2_bias_to_fp16, epsilon = var_540_to_fp16, gamma = decoder_layers_2_layer_norm_2_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_second_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_second_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44360192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45408832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45410944)))];
            tensor<fp16, [1, 1, 1024]> linear_24_cast_fp16 = linear(bias = decoder_layers_2_second_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_2_second_sub_layer_query_net_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_second_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_second_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45413056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46461696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46463808)))];
            tensor<fp16, [1, 125, 1024]> linear_25_cast_fp16 = linear(bias = decoder_layers_2_second_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_2_second_sub_layer_key_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_second_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_second_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46465920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47514560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47516672)))];
            tensor<fp16, [1, 125, 1024]> linear_26_cast_fp16 = linear(bias = decoder_layers_2_second_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_2_second_sub_layer_value_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<int32, [4]> var_637 = const()[name = tensor<string, []>("op_637"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_63_cast_fp16 = reshape(shape = var_637, x = linear_24_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<fp16, []> _inversed_query_11_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_11_cast_fp16 = mul(x = x_63_cast_fp16, y = _inversed_query_11_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_11_cast_fp16")];
            tensor<int32, [4]> var_645 = const()[name = tensor<string, []>("op_645"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_67_cast_fp16 = reshape(shape = var_645, x = linear_25_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<fp16, []> _inversed_key_11_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 125, 8, 128]> _inversed_key_11_cast_fp16 = mul(x = x_67_cast_fp16, y = _inversed_key_11_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_11_cast_fp16")];
            tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_71_cast_fp16 = reshape(shape = var_653, x = linear_26_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<int32, [4]> var_655 = const()[name = tensor<string, []>("op_655"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_21_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_21_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_21_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_21_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = tensor<string, []>("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_43_perm_0 = const()[name = tensor<string, []>("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 125, 128]> transpose_43 = transpose(perm = transpose_43_perm_0, x = _inversed_key_11_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [1, 8, 1, 128]> transpose_42 = transpose(perm = transpose_42_perm_0, x = _inversed_query_11_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_21_cast_fp16 = matmul(transpose_x = attention_scores_21_transpose_x_1, transpose_y = attention_scores_21_transpose_y_1, x = transpose_42, y = transpose_43)[name = tensor<string, []>("attention_scores_21_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_23_cast_fp16 = add(x = attention_scores_21_cast_fp16, y = attention_mask_9_cast_fp16)[name = tensor<string, []>("attention_scores_23_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> input_79_cast_fp16 = softmax(axis = var_533, x = attention_scores_23_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<bool, []> context_21_transpose_x_0 = const()[name = tensor<string, []>("context_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_21_transpose_y_0 = const()[name = tensor<string, []>("context_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 125, 128]> value_11_cast_fp16 = transpose(perm = var_655, x = x_71_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [1, 8, 1, 128]> context_21_cast_fp16 = matmul(transpose_x = context_21_transpose_x_0, transpose_y = context_21_transpose_y_0, x = input_79_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("context_21_cast_fp16")];
            tensor<int32, [4]> var_669 = const()[name = tensor<string, []>("op_669"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_674 = const()[name = tensor<string, []>("op_674"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_670_cast_fp16 = transpose(perm = var_669, x = context_21_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [1, 1, 1024]> input_81_cast_fp16 = reshape(shape = var_674, x = var_670_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_2_second_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_second_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47518784))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48567424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48569536)))];
            tensor<fp16, [1, 1, 1024]> linear_27_cast_fp16 = linear(bias = decoder_layers_2_second_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_2_second_sub_layer_out_projection_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_85_cast_fp16 = add(x = linear_27_cast_fp16, y = input_75_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<int32, [1]> input_87_axes_0 = const()[name = tensor<string, []>("input_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48571648)))];
            tensor<fp16, [1024]> decoder_layers_2_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48573760)))];
            tensor<fp16, [1, 1, 1024]> input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = decoder_layers_2_layer_norm_3_bias_to_fp16, epsilon = var_540_to_fp16, gamma = decoder_layers_2_layer_norm_3_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<fp16, [4096, 1024]> decoder_layers_2_third_sub_layer_dense_in_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_third_sub_layer_dense_in_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48575872))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52770240))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19075968)))];
            tensor<fp16, [4096]> decoder_layers_2_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52778496)))];
            tensor<fp16, [1, 1, 4096]> linear_28_cast_fp16 = linear(bias = decoder_layers_2_third_sub_layer_dense_in_bias_to_fp16, weight = decoder_layers_2_third_sub_layer_dense_in_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_89_cast_fp16 = relu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<fp16, [1024, 4096]> decoder_layers_2_third_sub_layer_dense_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_2_third_sub_layer_dense_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52786752))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56981120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_2_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_2_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56983232)))];
            tensor<fp16, [1, 1, 1024]> linear_29_cast_fp16 = linear(bias = decoder_layers_2_third_sub_layer_dense_out_bias_to_fp16, weight = decoder_layers_2_third_sub_layer_dense_out_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> states_7_cast_fp16 = add(x = linear_29_cast_fp16, y = input_85_cast_fp16)[name = tensor<string, []>("states_7_cast_fp16")];
            tensor<int32, [4]> prior_7_begin_0 = const()[name = tensor<string, []>("prior_7_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> prior_7_end_0 = const()[name = tensor<string, []>("prior_7_end_0"), val = tensor<int32, [4]>([4, 1, 0, 1024])];
            tensor<bool, [4]> prior_7_end_mask_0 = const()[name = tensor<string, []>("prior_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> prior_7_squeeze_mask_0 = const()[name = tensor<string, []>("prior_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, ?, 1024]> prior_7_cast_fp16 = slice_by_index(begin = prior_7_begin_0, end = prior_7_end_0, end_mask = prior_7_end_mask_0, squeeze_mask = prior_7_squeeze_mask_0, x = decoder_mems_to_fp16)[name = tensor<string, []>("prior_7_cast_fp16")];
            tensor<bool, []> var_714_interleave_0 = const()[name = tensor<string, []>("op_714_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, []> const_72 = const()[name = tensor<string, []>("const_72"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, ?, 1024]> var_714_cast_fp16 = concat(axis = const_72, interleave = var_714_interleave_0, values = (prior_7_cast_fp16, states_7_cast_fp16))[name = tensor<string, []>("op_714_cast_fp16")];
            tensor<int32, []> var_724 = const()[name = tensor<string, []>("op_724"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56985344)))];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56987456)))];
            tensor<fp16, []> var_731_to_fp16 = const()[name = tensor<string, []>("op_731_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = decoder_layers_3_layer_norm_1_bias_to_fp16, epsilon = var_731_to_fp16, gamma = decoder_layers_3_layer_norm_1_weight_to_fp16, x = states_7_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<int32, [1]> input_97_axes_0 = const()[name = tensor<string, []>("input_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, ?, 1024]> input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = decoder_layers_3_layer_norm_1_bias_to_fp16, epsilon = var_731_to_fp16, gamma = decoder_layers_3_layer_norm_1_weight_to_fp16, x = var_714_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_first_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_first_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56989568))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58038208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_first_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_first_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58040320)))];
            tensor<fp16, [1, 1, 1024]> linear_30_cast_fp16 = linear(bias = decoder_layers_3_first_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_3_first_sub_layer_query_net_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_first_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_first_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58042432))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59091072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_first_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_first_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59093184)))];
            tensor<fp16, [1, ?, 1024]> linear_31_cast_fp16 = linear(bias = decoder_layers_3_first_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_3_first_sub_layer_key_net_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_first_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_first_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59095296))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60143936))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_first_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_first_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60146048)))];
            tensor<fp16, [1, ?, 1024]> linear_32_cast_fp16 = linear(bias = decoder_layers_3_first_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_3_first_sub_layer_value_net_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = tensor<string, []>("op_762"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_75_cast_fp16 = reshape(shape = var_762, x = linear_30_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<fp16, []> _inversed_query_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_13_cast_fp16 = mul(x = x_75_cast_fp16, y = _inversed_query_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_13_cast_fp16")];
            tensor<int32, [4]> concat_6x = const()[name = tensor<string, []>("concat_6x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_79_cast_fp16 = reshape(shape = concat_6x, x = linear_31_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<fp16, []> _inversed_key_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, ?, 8, 128]> _inversed_key_13_cast_fp16 = mul(x = x_79_cast_fp16, y = _inversed_key_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_13_cast_fp16")];
            tensor<int32, [4]> concat_7x = const()[name = tensor<string, []>("concat_7x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> x_83_cast_fp16 = reshape(shape = concat_7x, x = linear_32_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<int32, [4]> var_780 = const()[name = tensor<string, []>("op_780"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_25_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_25_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_25_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_25_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = tensor<string, []>("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_45_perm_0 = const()[name = tensor<string, []>("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, ?, 128]> transpose_45 = transpose(perm = transpose_45_perm_0, x = _inversed_key_13_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [1, 8, 1, 128]> transpose_44 = transpose(perm = transpose_44_perm_0, x = _inversed_query_13_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 8, 1, ?]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_1, transpose_y = attention_scores_25_transpose_y_1, x = transpose_44, y = transpose_45)[name = tensor<string, []>("attention_scores_25_cast_fp16")];
            tensor<fp16, [1, 8, 1, ?]> input_99_cast_fp16 = softmax(axis = var_724, x = attention_scores_25_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<bool, []> context_25_transpose_x_0 = const()[name = tensor<string, []>("context_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_25_transpose_y_0 = const()[name = tensor<string, []>("context_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, ?, 128]> value_13_cast_fp16 = transpose(perm = var_780, x = x_83_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 8, 1, 128]> context_25_cast_fp16 = matmul(transpose_x = context_25_transpose_x_0, transpose_y = context_25_transpose_y_0, x = input_99_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("context_25_cast_fp16")];
            tensor<int32, [4]> var_794 = const()[name = tensor<string, []>("op_794"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_799 = const()[name = tensor<string, []>("op_799"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_795_cast_fp16 = transpose(perm = var_794, x = context_25_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 1, 1024]> input_101_cast_fp16 = reshape(shape = var_799, x = var_795_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_first_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_first_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60148160))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61196800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_first_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_first_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61198912)))];
            tensor<fp16, [1, 1, 1024]> linear_33_cast_fp16 = linear(bias = decoder_layers_3_first_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_3_first_sub_layer_out_projection_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_105_cast_fp16 = add(x = linear_33_cast_fp16, y = states_7_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = tensor<string, []>("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61201024)))];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61203136)))];
            tensor<fp16, [1, 1, 1024]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = decoder_layers_3_layer_norm_2_bias_to_fp16, epsilon = var_731_to_fp16, gamma = decoder_layers_3_layer_norm_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_second_sub_layer_query_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_second_sub_layer_query_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61205248))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62253888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_second_sub_layer_query_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_second_sub_layer_query_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62256000)))];
            tensor<fp16, [1, 1, 1024]> linear_34_cast_fp16 = linear(bias = decoder_layers_3_second_sub_layer_query_net_bias_to_fp16, weight = decoder_layers_3_second_sub_layer_query_net_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_second_sub_layer_key_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_second_sub_layer_key_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62258112))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63306752))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_second_sub_layer_key_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63308864)))];
            tensor<fp16, [1, 125, 1024]> linear_35_cast_fp16 = linear(bias = decoder_layers_3_second_sub_layer_key_net_bias_to_fp16, weight = decoder_layers_3_second_sub_layer_key_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_second_sub_layer_value_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_second_sub_layer_value_net_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63310976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64359616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_second_sub_layer_value_net_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64361728)))];
            tensor<fp16, [1, 125, 1024]> linear_36_cast_fp16 = linear(bias = decoder_layers_3_second_sub_layer_value_net_bias_to_fp16, weight = decoder_layers_3_second_sub_layer_value_net_weight_to_fp16_quantized, x = encoder_embeddings_to_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<int32, [4]> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, [4]>([1, 1, 8, 128])];
            tensor<fp16, [1, 1, 8, 128]> x_87_cast_fp16 = reshape(shape = var_825, x = linear_34_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<fp16, []> _inversed_query_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_query_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 1, 8, 128]> _inversed_query_cast_fp16 = mul(x = x_87_cast_fp16, y = _inversed_query_y_0_to_fp16)[name = tensor<string, []>("_inversed_query_cast_fp16")];
            tensor<int32, [4]> var_833 = const()[name = tensor<string, []>("op_833"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_91_cast_fp16 = reshape(shape = var_833, x = linear_35_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<fp16, []> _inversed_key_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_key_y_0_to_fp16"), val = tensor<fp16, []>(0x1.308p-2)];
            tensor<fp16, [1, 125, 8, 128]> _inversed_key_cast_fp16 = mul(x = x_91_cast_fp16, y = _inversed_key_y_0_to_fp16)[name = tensor<string, []>("_inversed_key_cast_fp16")];
            tensor<int32, [4]> var_841 = const()[name = tensor<string, []>("op_841"), val = tensor<int32, [4]>([1, 125, 8, 128])];
            tensor<fp16, [1, 125, 8, 128]> x_cast_fp16 = reshape(shape = var_841, x = linear_36_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = tensor<string, []>("op_843"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<bool, []> attention_scores_29_transpose_x_1 = const()[name = tensor<string, []>("attention_scores_29_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> attention_scores_29_transpose_y_1 = const()[name = tensor<string, []>("attention_scores_29_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = tensor<string, []>("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_47_perm_0 = const()[name = tensor<string, []>("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 8, 125, 128]> transpose_47 = transpose(perm = transpose_47_perm_0, x = _inversed_key_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 8, 1, 128]> transpose_46 = transpose(perm = transpose_46_perm_0, x = _inversed_query_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_29_cast_fp16 = matmul(transpose_x = attention_scores_29_transpose_x_1, transpose_y = attention_scores_29_transpose_y_1, x = transpose_46, y = transpose_47)[name = tensor<string, []>("attention_scores_29_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> attention_scores_cast_fp16 = add(x = attention_scores_29_cast_fp16, y = attention_mask_9_cast_fp16)[name = tensor<string, []>("attention_scores_cast_fp16")];
            tensor<fp16, [1, 8, 1, 125]> input_109_cast_fp16 = softmax(axis = var_724, x = attention_scores_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<bool, []> context_29_transpose_x_0 = const()[name = tensor<string, []>("context_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> context_29_transpose_y_0 = const()[name = tensor<string, []>("context_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 8, 125, 128]> value_cast_fp16 = transpose(perm = var_843, x = x_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 8, 1, 128]> context_29_cast_fp16 = matmul(transpose_x = context_29_transpose_x_0, transpose_y = context_29_transpose_y_0, x = input_109_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("context_29_cast_fp16")];
            tensor<int32, [4]> var_857 = const()[name = tensor<string, []>("op_857"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_862 = const()[name = tensor<string, []>("op_862"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp16, [1, 1, 8, 128]> var_858_cast_fp16 = transpose(perm = var_857, x = context_29_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 1, 1024]> input_111_cast_fp16 = reshape(shape = var_862, x = var_858_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<fp16, [1024, 1024]> decoder_layers_3_second_sub_layer_out_projection_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_second_sub_layer_out_projection_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64363840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65412480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_second_sub_layer_out_projection_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_second_sub_layer_out_projection_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65414592)))];
            tensor<fp16, [1, 1, 1024]> linear_37_cast_fp16 = linear(bias = decoder_layers_3_second_sub_layer_out_projection_bias_to_fp16, weight = decoder_layers_3_second_sub_layer_out_projection_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> input_115_cast_fp16 = add(x = linear_37_cast_fp16, y = input_105_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_3_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_3_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65416704)))];
            tensor<fp16, [1024]> decoder_layers_3_layer_norm_3_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_layer_norm_3_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65418816)))];
            tensor<fp16, [1, 1, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = decoder_layers_3_layer_norm_3_bias_to_fp16, epsilon = var_731_to_fp16, gamma = decoder_layers_3_layer_norm_3_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<fp16, [4096, 1024]> decoder_layers_3_third_sub_layer_dense_in_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_third_sub_layer_dense_in_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65420928))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69615296))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19075968)))];
            tensor<fp16, [4096]> decoder_layers_3_third_sub_layer_dense_in_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_third_sub_layer_dense_in_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69623552)))];
            tensor<fp16, [1, 1, 4096]> linear_38_cast_fp16 = linear(bias = decoder_layers_3_third_sub_layer_dense_in_bias_to_fp16, weight = decoder_layers_3_third_sub_layer_dense_in_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, [1, 1, 4096]> input_119_cast_fp16 = relu(x = linear_38_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [1024, 4096]> decoder_layers_3_third_sub_layer_dense_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("decoder_layers_3_third_sub_layer_dense_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69631808))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73826176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6438592)))];
            tensor<fp16, [1024]> decoder_layers_3_third_sub_layer_dense_out_bias_to_fp16 = const()[name = tensor<string, []>("decoder_layers_3_third_sub_layer_dense_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73828288)))];
            tensor<fp16, [1, 1, 1024]> linear_39_cast_fp16 = linear(bias = decoder_layers_3_third_sub_layer_dense_out_bias_to_fp16, weight = decoder_layers_3_third_sub_layer_dense_out_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> states_9_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = tensor<string, []>("states_9_cast_fp16")];
            tensor<int32, [4]> prior_begin_0 = const()[name = tensor<string, []>("prior_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> prior_end_0 = const()[name = tensor<string, []>("prior_end_0"), val = tensor<int32, [4]>([5, 1, 0, 1024])];
            tensor<bool, [4]> prior_end_mask_0 = const()[name = tensor<string, []>("prior_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> prior_squeeze_mask_0 = const()[name = tensor<string, []>("prior_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, ?, 1024]> prior_cast_fp16 = slice_by_index(begin = prior_begin_0, end = prior_end_0, end_mask = prior_end_mask_0, squeeze_mask = prior_squeeze_mask_0, x = decoder_mems_to_fp16)[name = tensor<string, []>("prior_cast_fp16")];
            tensor<bool, []> var_895_interleave_0 = const()[name = tensor<string, []>("op_895_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, []> const_73 = const()[name = tensor<string, []>("const_73"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, ?, 1024]> var_895_cast_fp16 = concat(axis = const_73, interleave = var_895_interleave_0, values = (prior_cast_fp16, states_9_cast_fp16))[name = tensor<string, []>("op_895_cast_fp16")];
            tensor<int32, [1]> states_axes_0 = const()[name = tensor<string, []>("states_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> decoder_final_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("decoder_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73830400)))];
            tensor<fp16, [1024]> decoder_final_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("decoder_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73832512)))];
            tensor<fp16, []> var_900_to_fp16 = const()[name = tensor<string, []>("op_900_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 1, 1024]> states_cast_fp16 = layer_norm(axes = states_axes_0, beta = decoder_final_layer_norm_bias_to_fp16, epsilon = var_900_to_fp16, gamma = decoder_final_layer_norm_weight_to_fp16, x = states_9_cast_fp16)[name = tensor<string, []>("states_cast_fp16")];
            tensor<int32, []> var_916 = const()[name = tensor<string, []>("op_916"), val = tensor<int32, []>(-1)];
            tensor<fp16, [5248]> head_mlp_layer0_bias_to_fp16 = const()[name = tensor<string, []>("head_mlp_layer0_bias_to_fp16"), val = tensor<fp16, [5248]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73834624)))];
            tensor<fp16, [1, 1, 5248]> linear_40_cast_fp16 = linear(bias = head_mlp_layer0_bias_to_fp16, weight = embedding_token_embedding_weight_to_fp16_quantized, x = states_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [1, 1, 5248]> var_925_softmax_cast_fp16 = softmax(axis = var_916, x = linear_40_cast_fp16)[name = tensor<string, []>("op_925_softmax_cast_fp16")];
            tensor<fp32, []> var_925_epsilon_0 = const()[name = tensor<string, []>("op_925_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [1, 1, 5248]> logits = log(epsilon = var_925_epsilon_0, x = var_925_softmax_cast_fp16)[name = tensor<string, []>("op_925_cast_fp16")];
            tensor<int32, []> var_928_axis_0 = const()[name = tensor<string, []>("op_928_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [5, 1, ?, 1024]> decoder_hidden_states = stack(axis = var_928_axis_0, values = (var_141_cast_fp16, var_332_cast_fp16, var_523_cast_fp16, var_714_cast_fp16, var_895_cast_fp16))[name = tensor<string, []>("op_928_cast_fp16")];
        } -> (logits, decoder_hidden_states);
}