program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<int32, [16, 1, ?]> codes) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"codes", [16, 1, 57]}}), ("RangeDims", {{"codes", [[16, 16], [1, 1], [1, 125]]}})))] {
            tensor<int32, [3]> input_1_begin_0 = const()[name = tensor<string, []>("input_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_1_end_0 = const()[name = tensor<string, []>("input_1_end_0"), val = tensor<int32, [3]>([1, 1, 0])];
            tensor<bool, [3]> input_1_end_mask_0 = const()[name = tensor<string, []>("input_1_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_1_squeeze_mask_0 = const()[name = tensor<string, []>("input_1_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_1 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, squeeze_mask = input_1_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_1")];
            tensor<int32, []> var_118_batch_dims_0 = const()[name = tensor<string, []>("op_118_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_118_validate_indices_0 = const()[name = tensor<string, []>("op_118_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_0_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_0_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<string, []> input_1_to_int16_dtype_0 = const()[name = tensor<string, []>("input_1_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<string, []> cast_291_dtype_0 = const()[name = tensor<string, []>("cast_291_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, ?]> input_1_to_int16 = cast(dtype = input_1_to_int16_dtype_0, x = input_1)[name = tensor<string, []>("cast_419")];
            tensor<int32, [1, ?]> cast_291 = cast(dtype = cast_291_dtype_0, x = input_1_to_int16)[name = tensor<string, []>("cast_418")];
            tensor<bool, [1, ?]> greater_equal_0 = greater_equal(x = cast_291, 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, []>(1024)];
            tensor<int32, [1, ?]> add_12 = add(x = cast_291, y = slice_by_index_0)[name = tensor<string, []>("add_12")];
            tensor<int32, [1, ?]> select_0 = select(a = cast_291, b = add_12, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<int32, []> var_118_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_118_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
            tensor<int16, [1, ?]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_417")];
            tensor<fp16, [1, ?, 8]> var_118_cast_fp16_cast_uint16_cast_uint16 = gather(axis = var_118_cast_fp16_cast_uint16_axis_0, batch_dims = var_118_batch_dims_0, indices = select_0_to_int16, validate_indices = var_118_validate_indices_0, x = quantizers_0_codebook_weight_to_fp16)[name = tensor<string, []>("op_118_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> emb_1_pad_type_0 = const()[name = tensor<string, []>("emb_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> emb_1_strides_0 = const()[name = tensor<string, []>("emb_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> emb_1_pad_0 = const()[name = tensor<string, []>("emb_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> emb_1_dilations_0 = const()[name = tensor<string, []>("emb_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> emb_1_groups_0 = const()[name = tensor<string, []>("emb_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_0_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_0_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16512)))];
            tensor<fp16, [512]> quantizers_0_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_0_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24768)))];
            tensor<fp16, [1, 8, ?]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = var_118_cast_fp16_cast_uint16_cast_uint16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 512, ?]> emb_1_cast_fp16 = conv(bias = quantizers_0_out_proj_bias_to_fp16, dilations = emb_1_dilations_0, groups = emb_1_groups_0, pad = emb_1_pad_0, pad_type = emb_1_pad_type_0, strides = emb_1_strides_0, weight = quantizers_0_out_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("emb_1_cast_fp16")];
            tensor<int32, [3]> input_5_begin_0 = const()[name = tensor<string, []>("input_5_begin_0"), val = tensor<int32, [3]>([1, 0, 0])];
            tensor<int32, [3]> input_5_end_0 = const()[name = tensor<string, []>("input_5_end_0"), val = tensor<int32, [3]>([2, 1, 0])];
            tensor<bool, [3]> input_5_end_mask_0 = const()[name = tensor<string, []>("input_5_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_5_squeeze_mask_0 = const()[name = tensor<string, []>("input_5_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_5 = slice_by_index(begin = input_5_begin_0, end = input_5_end_0, end_mask = input_5_end_mask_0, squeeze_mask = input_5_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_5")];
            tensor<int32, []> var_139_axis_0 = const()[name = tensor<string, []>("op_139_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_139_batch_dims_0 = const()[name = tensor<string, []>("op_139_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_139_validate_indices_0 = const()[name = tensor<string, []>("op_139_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_1_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_1_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25856)))];
            tensor<string, []> input_5_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_5_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_5_to_uint16 = cast(dtype = input_5_to_uint16_dtype_0, x = input_5)[name = tensor<string, []>("cast_416")];
            tensor<fp16, [1, ?, 8]> var_139_cast_fp16_cast_uint16 = gather(axis = var_139_axis_0, batch_dims = var_139_batch_dims_0, indices = input_5_to_uint16, validate_indices = var_139_validate_indices_0, x = quantizers_1_codebook_weight_to_fp16)[name = tensor<string, []>("op_139_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_7_perm_0 = const()[name = tensor<string, []>("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_1_pad_type_0 = const()[name = tensor<string, []>("z_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_1_strides_0 = const()[name = tensor<string, []>("z_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_1_pad_0 = const()[name = tensor<string, []>("z_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_1_dilations_0 = const()[name = tensor<string, []>("z_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_1_groups_0 = const()[name = tensor<string, []>("z_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42304)))];
            tensor<fp16, [512]> quantizers_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50560)))];
            tensor<fp16, [1, 8, ?]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = var_139_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 512, ?]> z_1_cast_fp16 = conv(bias = quantizers_1_out_proj_bias_to_fp16, dilations = z_1_dilations_0, groups = z_1_groups_0, pad = z_1_pad_0, pad_type = z_1_pad_type_0, strides = z_1_strides_0, weight = quantizers_1_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("z_1_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_3_cast_fp16 = add(x = emb_1_cast_fp16, y = z_1_cast_fp16)[name = tensor<string, []>("emb_3_cast_fp16")];
            tensor<int32, [3]> input_9_begin_0 = const()[name = tensor<string, []>("input_9_begin_0"), val = tensor<int32, [3]>([2, 0, 0])];
            tensor<int32, [3]> input_9_end_0 = const()[name = tensor<string, []>("input_9_end_0"), val = tensor<int32, [3]>([3, 1, 0])];
            tensor<bool, [3]> input_9_end_mask_0 = const()[name = tensor<string, []>("input_9_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_9_squeeze_mask_0 = const()[name = tensor<string, []>("input_9_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_9 = slice_by_index(begin = input_9_begin_0, end = input_9_end_0, end_mask = input_9_end_mask_0, squeeze_mask = input_9_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_9")];
            tensor<int32, []> var_162_axis_0 = const()[name = tensor<string, []>("op_162_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_162_batch_dims_0 = const()[name = tensor<string, []>("op_162_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_162_validate_indices_0 = const()[name = tensor<string, []>("op_162_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_2_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_2_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51648)))];
            tensor<string, []> input_9_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_9_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_9_to_uint16 = cast(dtype = input_9_to_uint16_dtype_0, x = input_9)[name = tensor<string, []>("cast_415")];
            tensor<fp16, [1, ?, 8]> var_162_cast_fp16_cast_uint16 = gather(axis = var_162_axis_0, batch_dims = var_162_batch_dims_0, indices = input_9_to_uint16, validate_indices = var_162_validate_indices_0, x = quantizers_2_codebook_weight_to_fp16)[name = tensor<string, []>("op_162_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_11_perm_0 = const()[name = tensor<string, []>("input_11_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_3_pad_type_0 = const()[name = tensor<string, []>("z_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_3_strides_0 = const()[name = tensor<string, []>("z_3_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_3_pad_0 = const()[name = tensor<string, []>("z_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_3_dilations_0 = const()[name = tensor<string, []>("z_3_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_3_groups_0 = const()[name = tensor<string, []>("z_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68096)))];
            tensor<fp16, [512]> quantizers_2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76352)))];
            tensor<fp16, [1, 8, ?]> input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = var_162_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 512, ?]> z_3_cast_fp16 = conv(bias = quantizers_2_out_proj_bias_to_fp16, dilations = z_3_dilations_0, groups = z_3_groups_0, pad = z_3_pad_0, pad_type = z_3_pad_type_0, strides = z_3_strides_0, weight = quantizers_2_out_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("z_3_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_5_cast_fp16 = add(x = emb_3_cast_fp16, y = z_3_cast_fp16)[name = tensor<string, []>("emb_5_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = tensor<string, []>("input_13_begin_0"), val = tensor<int32, [3]>([3, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = tensor<string, []>("input_13_end_0"), val = tensor<int32, [3]>([4, 1, 0])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = tensor<string, []>("input_13_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_13_squeeze_mask_0 = const()[name = tensor<string, []>("input_13_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_13 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, squeeze_mask = input_13_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_13")];
            tensor<int32, []> var_185_axis_0 = const()[name = tensor<string, []>("op_185_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_185_batch_dims_0 = const()[name = tensor<string, []>("op_185_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_185_validate_indices_0 = const()[name = tensor<string, []>("op_185_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_3_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_3_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77440)))];
            tensor<string, []> input_13_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_13_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_13_to_uint16 = cast(dtype = input_13_to_uint16_dtype_0, x = input_13)[name = tensor<string, []>("cast_414")];
            tensor<fp16, [1, ?, 8]> var_185_cast_fp16_cast_uint16 = gather(axis = var_185_axis_0, batch_dims = var_185_batch_dims_0, indices = input_13_to_uint16, validate_indices = var_185_validate_indices_0, x = quantizers_3_codebook_weight_to_fp16)[name = tensor<string, []>("op_185_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_15_perm_0 = const()[name = tensor<string, []>("input_15_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_5_pad_type_0 = const()[name = tensor<string, []>("z_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_5_strides_0 = const()[name = tensor<string, []>("z_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_5_pad_0 = const()[name = tensor<string, []>("z_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_5_dilations_0 = const()[name = tensor<string, []>("z_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_5_groups_0 = const()[name = tensor<string, []>("z_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93888)))];
            tensor<fp16, [512]> quantizers_3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102144)))];
            tensor<fp16, [1, 8, ?]> input_15_cast_fp16 = transpose(perm = input_15_perm_0, x = var_185_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 512, ?]> z_5_cast_fp16 = conv(bias = quantizers_3_out_proj_bias_to_fp16, dilations = z_5_dilations_0, groups = z_5_groups_0, pad = z_5_pad_0, pad_type = z_5_pad_type_0, strides = z_5_strides_0, weight = quantizers_3_out_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("z_5_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_7_cast_fp16 = add(x = emb_5_cast_fp16, y = z_5_cast_fp16)[name = tensor<string, []>("emb_7_cast_fp16")];
            tensor<int32, [3]> input_17_begin_0 = const()[name = tensor<string, []>("input_17_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> input_17_end_0 = const()[name = tensor<string, []>("input_17_end_0"), val = tensor<int32, [3]>([5, 1, 0])];
            tensor<bool, [3]> input_17_end_mask_0 = const()[name = tensor<string, []>("input_17_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_17_squeeze_mask_0 = const()[name = tensor<string, []>("input_17_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_17 = slice_by_index(begin = input_17_begin_0, end = input_17_end_0, end_mask = input_17_end_mask_0, squeeze_mask = input_17_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_17")];
            tensor<int32, []> var_208_axis_0 = const()[name = tensor<string, []>("op_208_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_208_batch_dims_0 = const()[name = tensor<string, []>("op_208_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_208_validate_indices_0 = const()[name = tensor<string, []>("op_208_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_4_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_4_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103232)))];
            tensor<string, []> input_17_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_17_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_17_to_uint16 = cast(dtype = input_17_to_uint16_dtype_0, x = input_17)[name = tensor<string, []>("cast_413")];
            tensor<fp16, [1, ?, 8]> var_208_cast_fp16_cast_uint16 = gather(axis = var_208_axis_0, batch_dims = var_208_batch_dims_0, indices = input_17_to_uint16, validate_indices = var_208_validate_indices_0, x = quantizers_4_codebook_weight_to_fp16)[name = tensor<string, []>("op_208_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_19_perm_0 = const()[name = tensor<string, []>("input_19_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_7_pad_type_0 = const()[name = tensor<string, []>("z_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_7_strides_0 = const()[name = tensor<string, []>("z_7_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_7_pad_0 = const()[name = tensor<string, []>("z_7_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_7_dilations_0 = const()[name = tensor<string, []>("z_7_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_7_groups_0 = const()[name = tensor<string, []>("z_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_4_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_4_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119680)))];
            tensor<fp16, [512]> quantizers_4_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_4_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127936)))];
            tensor<fp16, [1, 8, ?]> input_19_cast_fp16 = transpose(perm = input_19_perm_0, x = var_208_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 512, ?]> z_7_cast_fp16 = conv(bias = quantizers_4_out_proj_bias_to_fp16, dilations = z_7_dilations_0, groups = z_7_groups_0, pad = z_7_pad_0, pad_type = z_7_pad_type_0, strides = z_7_strides_0, weight = quantizers_4_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("z_7_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_9_cast_fp16 = add(x = emb_7_cast_fp16, y = z_7_cast_fp16)[name = tensor<string, []>("emb_9_cast_fp16")];
            tensor<int32, [3]> input_21_begin_0 = const()[name = tensor<string, []>("input_21_begin_0"), val = tensor<int32, [3]>([5, 0, 0])];
            tensor<int32, [3]> input_21_end_0 = const()[name = tensor<string, []>("input_21_end_0"), val = tensor<int32, [3]>([6, 1, 0])];
            tensor<bool, [3]> input_21_end_mask_0 = const()[name = tensor<string, []>("input_21_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_21_squeeze_mask_0 = const()[name = tensor<string, []>("input_21_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_21 = slice_by_index(begin = input_21_begin_0, end = input_21_end_0, end_mask = input_21_end_mask_0, squeeze_mask = input_21_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_21")];
            tensor<int32, []> var_231_axis_0 = const()[name = tensor<string, []>("op_231_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_231_batch_dims_0 = const()[name = tensor<string, []>("op_231_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_231_validate_indices_0 = const()[name = tensor<string, []>("op_231_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_5_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_5_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129024)))];
            tensor<string, []> input_21_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_21_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_21_to_uint16 = cast(dtype = input_21_to_uint16_dtype_0, x = input_21)[name = tensor<string, []>("cast_412")];
            tensor<fp16, [1, ?, 8]> var_231_cast_fp16_cast_uint16 = gather(axis = var_231_axis_0, batch_dims = var_231_batch_dims_0, indices = input_21_to_uint16, validate_indices = var_231_validate_indices_0, x = quantizers_5_codebook_weight_to_fp16)[name = tensor<string, []>("op_231_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_23_perm_0 = const()[name = tensor<string, []>("input_23_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_9_pad_type_0 = const()[name = tensor<string, []>("z_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_9_strides_0 = const()[name = tensor<string, []>("z_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_9_pad_0 = const()[name = tensor<string, []>("z_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_9_dilations_0 = const()[name = tensor<string, []>("z_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_9_groups_0 = const()[name = tensor<string, []>("z_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_5_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_5_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145472)))];
            tensor<fp16, [512]> quantizers_5_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_5_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153728)))];
            tensor<fp16, [1, 8, ?]> input_23_cast_fp16 = transpose(perm = input_23_perm_0, x = var_231_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 512, ?]> z_9_cast_fp16 = conv(bias = quantizers_5_out_proj_bias_to_fp16, dilations = z_9_dilations_0, groups = z_9_groups_0, pad = z_9_pad_0, pad_type = z_9_pad_type_0, strides = z_9_strides_0, weight = quantizers_5_out_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("z_9_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_11_cast_fp16 = add(x = emb_9_cast_fp16, y = z_9_cast_fp16)[name = tensor<string, []>("emb_11_cast_fp16")];
            tensor<int32, [3]> input_25_begin_0 = const()[name = tensor<string, []>("input_25_begin_0"), val = tensor<int32, [3]>([6, 0, 0])];
            tensor<int32, [3]> input_25_end_0 = const()[name = tensor<string, []>("input_25_end_0"), val = tensor<int32, [3]>([7, 1, 0])];
            tensor<bool, [3]> input_25_end_mask_0 = const()[name = tensor<string, []>("input_25_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_25_squeeze_mask_0 = const()[name = tensor<string, []>("input_25_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, squeeze_mask = input_25_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_25")];
            tensor<int32, []> var_254_axis_0 = const()[name = tensor<string, []>("op_254_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_254_batch_dims_0 = const()[name = tensor<string, []>("op_254_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_254_validate_indices_0 = const()[name = tensor<string, []>("op_254_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_6_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_6_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154816)))];
            tensor<string, []> input_25_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_25_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_25_to_uint16 = cast(dtype = input_25_to_uint16_dtype_0, x = input_25)[name = tensor<string, []>("cast_411")];
            tensor<fp16, [1, ?, 8]> var_254_cast_fp16_cast_uint16 = gather(axis = var_254_axis_0, batch_dims = var_254_batch_dims_0, indices = input_25_to_uint16, validate_indices = var_254_validate_indices_0, x = quantizers_6_codebook_weight_to_fp16)[name = tensor<string, []>("op_254_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_27_perm_0 = const()[name = tensor<string, []>("input_27_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_11_pad_type_0 = const()[name = tensor<string, []>("z_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_11_strides_0 = const()[name = tensor<string, []>("z_11_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_11_pad_0 = const()[name = tensor<string, []>("z_11_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_11_dilations_0 = const()[name = tensor<string, []>("z_11_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_11_groups_0 = const()[name = tensor<string, []>("z_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_6_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_6_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171264)))];
            tensor<fp16, [512]> quantizers_6_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_6_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179520)))];
            tensor<fp16, [1, 8, ?]> input_27_cast_fp16 = transpose(perm = input_27_perm_0, x = var_254_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [1, 512, ?]> z_11_cast_fp16 = conv(bias = quantizers_6_out_proj_bias_to_fp16, dilations = z_11_dilations_0, groups = z_11_groups_0, pad = z_11_pad_0, pad_type = z_11_pad_type_0, strides = z_11_strides_0, weight = quantizers_6_out_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("z_11_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_13_cast_fp16 = add(x = emb_11_cast_fp16, y = z_11_cast_fp16)[name = tensor<string, []>("emb_13_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = tensor<string, []>("input_29_begin_0"), val = tensor<int32, [3]>([7, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = tensor<string, []>("input_29_end_0"), val = tensor<int32, [3]>([8, 1, 0])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = tensor<string, []>("input_29_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_29_squeeze_mask_0 = const()[name = tensor<string, []>("input_29_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, squeeze_mask = input_29_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_29")];
            tensor<int32, []> var_277_axis_0 = const()[name = tensor<string, []>("op_277_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_277_batch_dims_0 = const()[name = tensor<string, []>("op_277_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_277_validate_indices_0 = const()[name = tensor<string, []>("op_277_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_7_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_7_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180608)))];
            tensor<string, []> input_29_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_29_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_29_to_uint16 = cast(dtype = input_29_to_uint16_dtype_0, x = input_29)[name = tensor<string, []>("cast_410")];
            tensor<fp16, [1, ?, 8]> var_277_cast_fp16_cast_uint16 = gather(axis = var_277_axis_0, batch_dims = var_277_batch_dims_0, indices = input_29_to_uint16, validate_indices = var_277_validate_indices_0, x = quantizers_7_codebook_weight_to_fp16)[name = tensor<string, []>("op_277_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_31_perm_0 = const()[name = tensor<string, []>("input_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_13_pad_type_0 = const()[name = tensor<string, []>("z_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_13_strides_0 = const()[name = tensor<string, []>("z_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_13_pad_0 = const()[name = tensor<string, []>("z_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_13_dilations_0 = const()[name = tensor<string, []>("z_13_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_13_groups_0 = const()[name = tensor<string, []>("z_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_7_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_7_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197056)))];
            tensor<fp16, [512]> quantizers_7_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_7_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205312)))];
            tensor<fp16, [1, 8, ?]> input_31_cast_fp16 = transpose(perm = input_31_perm_0, x = var_277_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 512, ?]> z_13_cast_fp16 = conv(bias = quantizers_7_out_proj_bias_to_fp16, dilations = z_13_dilations_0, groups = z_13_groups_0, pad = z_13_pad_0, pad_type = z_13_pad_type_0, strides = z_13_strides_0, weight = quantizers_7_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("z_13_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_15_cast_fp16 = add(x = emb_13_cast_fp16, y = z_13_cast_fp16)[name = tensor<string, []>("emb_15_cast_fp16")];
            tensor<int32, [3]> input_33_begin_0 = const()[name = tensor<string, []>("input_33_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> input_33_end_0 = const()[name = tensor<string, []>("input_33_end_0"), val = tensor<int32, [3]>([9, 1, 0])];
            tensor<bool, [3]> input_33_end_mask_0 = const()[name = tensor<string, []>("input_33_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_33_squeeze_mask_0 = const()[name = tensor<string, []>("input_33_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_33 = slice_by_index(begin = input_33_begin_0, end = input_33_end_0, end_mask = input_33_end_mask_0, squeeze_mask = input_33_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_33")];
            tensor<int32, []> var_300_axis_0 = const()[name = tensor<string, []>("op_300_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_300_batch_dims_0 = const()[name = tensor<string, []>("op_300_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_300_validate_indices_0 = const()[name = tensor<string, []>("op_300_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_8_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_8_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206400)))];
            tensor<string, []> input_33_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_33_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_33_to_uint16 = cast(dtype = input_33_to_uint16_dtype_0, x = input_33)[name = tensor<string, []>("cast_409")];
            tensor<fp16, [1, ?, 8]> var_300_cast_fp16_cast_uint16 = gather(axis = var_300_axis_0, batch_dims = var_300_batch_dims_0, indices = input_33_to_uint16, validate_indices = var_300_validate_indices_0, x = quantizers_8_codebook_weight_to_fp16)[name = tensor<string, []>("op_300_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_35_perm_0 = const()[name = tensor<string, []>("input_35_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_15_pad_type_0 = const()[name = tensor<string, []>("z_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_15_strides_0 = const()[name = tensor<string, []>("z_15_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_15_pad_0 = const()[name = tensor<string, []>("z_15_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_15_dilations_0 = const()[name = tensor<string, []>("z_15_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_15_groups_0 = const()[name = tensor<string, []>("z_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_8_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_8_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222848)))];
            tensor<fp16, [512]> quantizers_8_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_8_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231104)))];
            tensor<fp16, [1, 8, ?]> input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = var_300_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [1, 512, ?]> z_15_cast_fp16 = conv(bias = quantizers_8_out_proj_bias_to_fp16, dilations = z_15_dilations_0, groups = z_15_groups_0, pad = z_15_pad_0, pad_type = z_15_pad_type_0, strides = z_15_strides_0, weight = quantizers_8_out_proj_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("z_15_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_17_cast_fp16 = add(x = emb_15_cast_fp16, y = z_15_cast_fp16)[name = tensor<string, []>("emb_17_cast_fp16")];
            tensor<int32, [3]> input_37_begin_0 = const()[name = tensor<string, []>("input_37_begin_0"), val = tensor<int32, [3]>([9, 0, 0])];
            tensor<int32, [3]> input_37_end_0 = const()[name = tensor<string, []>("input_37_end_0"), val = tensor<int32, [3]>([10, 1, 0])];
            tensor<bool, [3]> input_37_end_mask_0 = const()[name = tensor<string, []>("input_37_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_37_squeeze_mask_0 = const()[name = tensor<string, []>("input_37_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_37 = slice_by_index(begin = input_37_begin_0, end = input_37_end_0, end_mask = input_37_end_mask_0, squeeze_mask = input_37_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_37")];
            tensor<int32, []> var_323_axis_0 = const()[name = tensor<string, []>("op_323_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_323_batch_dims_0 = const()[name = tensor<string, []>("op_323_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_323_validate_indices_0 = const()[name = tensor<string, []>("op_323_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_9_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_9_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232192)))];
            tensor<string, []> input_37_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_37_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_37_to_uint16 = cast(dtype = input_37_to_uint16_dtype_0, x = input_37)[name = tensor<string, []>("cast_408")];
            tensor<fp16, [1, ?, 8]> var_323_cast_fp16_cast_uint16 = gather(axis = var_323_axis_0, batch_dims = var_323_batch_dims_0, indices = input_37_to_uint16, validate_indices = var_323_validate_indices_0, x = quantizers_9_codebook_weight_to_fp16)[name = tensor<string, []>("op_323_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_39_perm_0 = const()[name = tensor<string, []>("input_39_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_17_pad_type_0 = const()[name = tensor<string, []>("z_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_17_strides_0 = const()[name = tensor<string, []>("z_17_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_17_pad_0 = const()[name = tensor<string, []>("z_17_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_17_dilations_0 = const()[name = tensor<string, []>("z_17_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_17_groups_0 = const()[name = tensor<string, []>("z_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_9_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_9_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248640)))];
            tensor<fp16, [512]> quantizers_9_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_9_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256896)))];
            tensor<fp16, [1, 8, ?]> input_39_cast_fp16 = transpose(perm = input_39_perm_0, x = var_323_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [1, 512, ?]> z_17_cast_fp16 = conv(bias = quantizers_9_out_proj_bias_to_fp16, dilations = z_17_dilations_0, groups = z_17_groups_0, pad = z_17_pad_0, pad_type = z_17_pad_type_0, strides = z_17_strides_0, weight = quantizers_9_out_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("z_17_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_19_cast_fp16 = add(x = emb_17_cast_fp16, y = z_17_cast_fp16)[name = tensor<string, []>("emb_19_cast_fp16")];
            tensor<int32, [3]> input_41_begin_0 = const()[name = tensor<string, []>("input_41_begin_0"), val = tensor<int32, [3]>([10, 0, 0])];
            tensor<int32, [3]> input_41_end_0 = const()[name = tensor<string, []>("input_41_end_0"), val = tensor<int32, [3]>([11, 1, 0])];
            tensor<bool, [3]> input_41_end_mask_0 = const()[name = tensor<string, []>("input_41_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_41_squeeze_mask_0 = const()[name = tensor<string, []>("input_41_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_41 = slice_by_index(begin = input_41_begin_0, end = input_41_end_0, end_mask = input_41_end_mask_0, squeeze_mask = input_41_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_41")];
            tensor<int32, []> var_346_axis_0 = const()[name = tensor<string, []>("op_346_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_346_batch_dims_0 = const()[name = tensor<string, []>("op_346_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_346_validate_indices_0 = const()[name = tensor<string, []>("op_346_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_10_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_10_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257984)))];
            tensor<string, []> input_41_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_41_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_41_to_uint16 = cast(dtype = input_41_to_uint16_dtype_0, x = input_41)[name = tensor<string, []>("cast_407")];
            tensor<fp16, [1, ?, 8]> var_346_cast_fp16_cast_uint16 = gather(axis = var_346_axis_0, batch_dims = var_346_batch_dims_0, indices = input_41_to_uint16, validate_indices = var_346_validate_indices_0, x = quantizers_10_codebook_weight_to_fp16)[name = tensor<string, []>("op_346_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_43_perm_0 = const()[name = tensor<string, []>("input_43_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_19_pad_type_0 = const()[name = tensor<string, []>("z_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_19_strides_0 = const()[name = tensor<string, []>("z_19_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_19_pad_0 = const()[name = tensor<string, []>("z_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_19_dilations_0 = const()[name = tensor<string, []>("z_19_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_19_groups_0 = const()[name = tensor<string, []>("z_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_10_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_10_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274432)))];
            tensor<fp16, [512]> quantizers_10_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_10_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282688)))];
            tensor<fp16, [1, 8, ?]> input_43_cast_fp16 = transpose(perm = input_43_perm_0, x = var_346_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [1, 512, ?]> z_19_cast_fp16 = conv(bias = quantizers_10_out_proj_bias_to_fp16, dilations = z_19_dilations_0, groups = z_19_groups_0, pad = z_19_pad_0, pad_type = z_19_pad_type_0, strides = z_19_strides_0, weight = quantizers_10_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("z_19_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_21_cast_fp16 = add(x = emb_19_cast_fp16, y = z_19_cast_fp16)[name = tensor<string, []>("emb_21_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = tensor<string, []>("input_45_begin_0"), val = tensor<int32, [3]>([11, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = tensor<string, []>("input_45_end_0"), val = tensor<int32, [3]>([12, 1, 0])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = tensor<string, []>("input_45_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_45_squeeze_mask_0 = const()[name = tensor<string, []>("input_45_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_45 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, squeeze_mask = input_45_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_45")];
            tensor<int32, []> var_369_axis_0 = const()[name = tensor<string, []>("op_369_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_369_batch_dims_0 = const()[name = tensor<string, []>("op_369_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_369_validate_indices_0 = const()[name = tensor<string, []>("op_369_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_11_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_11_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283776)))];
            tensor<string, []> input_45_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_45_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_45_to_uint16 = cast(dtype = input_45_to_uint16_dtype_0, x = input_45)[name = tensor<string, []>("cast_406")];
            tensor<fp16, [1, ?, 8]> var_369_cast_fp16_cast_uint16 = gather(axis = var_369_axis_0, batch_dims = var_369_batch_dims_0, indices = input_45_to_uint16, validate_indices = var_369_validate_indices_0, x = quantizers_11_codebook_weight_to_fp16)[name = tensor<string, []>("op_369_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_47_perm_0 = const()[name = tensor<string, []>("input_47_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_21_pad_type_0 = const()[name = tensor<string, []>("z_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_21_strides_0 = const()[name = tensor<string, []>("z_21_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_21_pad_0 = const()[name = tensor<string, []>("z_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_21_dilations_0 = const()[name = tensor<string, []>("z_21_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_21_groups_0 = const()[name = tensor<string, []>("z_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_11_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_11_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300224)))];
            tensor<fp16, [512]> quantizers_11_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_11_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308480)))];
            tensor<fp16, [1, 8, ?]> input_47_cast_fp16 = transpose(perm = input_47_perm_0, x = var_369_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 512, ?]> z_21_cast_fp16 = conv(bias = quantizers_11_out_proj_bias_to_fp16, dilations = z_21_dilations_0, groups = z_21_groups_0, pad = z_21_pad_0, pad_type = z_21_pad_type_0, strides = z_21_strides_0, weight = quantizers_11_out_proj_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("z_21_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_23_cast_fp16 = add(x = emb_21_cast_fp16, y = z_21_cast_fp16)[name = tensor<string, []>("emb_23_cast_fp16")];
            tensor<int32, [3]> input_49_begin_0 = const()[name = tensor<string, []>("input_49_begin_0"), val = tensor<int32, [3]>([12, 0, 0])];
            tensor<int32, [3]> input_49_end_0 = const()[name = tensor<string, []>("input_49_end_0"), val = tensor<int32, [3]>([13, 1, 0])];
            tensor<bool, [3]> input_49_end_mask_0 = const()[name = tensor<string, []>("input_49_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_49_squeeze_mask_0 = const()[name = tensor<string, []>("input_49_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, squeeze_mask = input_49_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_49")];
            tensor<int32, []> var_392_axis_0 = const()[name = tensor<string, []>("op_392_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_392_batch_dims_0 = const()[name = tensor<string, []>("op_392_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_392_validate_indices_0 = const()[name = tensor<string, []>("op_392_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_12_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_12_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309568)))];
            tensor<string, []> input_49_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_49_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_49_to_uint16 = cast(dtype = input_49_to_uint16_dtype_0, x = input_49)[name = tensor<string, []>("cast_405")];
            tensor<fp16, [1, ?, 8]> var_392_cast_fp16_cast_uint16 = gather(axis = var_392_axis_0, batch_dims = var_392_batch_dims_0, indices = input_49_to_uint16, validate_indices = var_392_validate_indices_0, x = quantizers_12_codebook_weight_to_fp16)[name = tensor<string, []>("op_392_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_51_perm_0 = const()[name = tensor<string, []>("input_51_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_23_pad_type_0 = const()[name = tensor<string, []>("z_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_23_strides_0 = const()[name = tensor<string, []>("z_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_23_pad_0 = const()[name = tensor<string, []>("z_23_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_23_dilations_0 = const()[name = tensor<string, []>("z_23_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_23_groups_0 = const()[name = tensor<string, []>("z_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_12_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_12_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326016)))];
            tensor<fp16, [512]> quantizers_12_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_12_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334272)))];
            tensor<fp16, [1, 8, ?]> input_51_cast_fp16 = transpose(perm = input_51_perm_0, x = var_392_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [1, 512, ?]> z_23_cast_fp16 = conv(bias = quantizers_12_out_proj_bias_to_fp16, dilations = z_23_dilations_0, groups = z_23_groups_0, pad = z_23_pad_0, pad_type = z_23_pad_type_0, strides = z_23_strides_0, weight = quantizers_12_out_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("z_23_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_25_cast_fp16 = add(x = emb_23_cast_fp16, y = z_23_cast_fp16)[name = tensor<string, []>("emb_25_cast_fp16")];
            tensor<int32, [3]> input_53_begin_0 = const()[name = tensor<string, []>("input_53_begin_0"), val = tensor<int32, [3]>([13, 0, 0])];
            tensor<int32, [3]> input_53_end_0 = const()[name = tensor<string, []>("input_53_end_0"), val = tensor<int32, [3]>([14, 1, 0])];
            tensor<bool, [3]> input_53_end_mask_0 = const()[name = tensor<string, []>("input_53_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_53_squeeze_mask_0 = const()[name = tensor<string, []>("input_53_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_53 = slice_by_index(begin = input_53_begin_0, end = input_53_end_0, end_mask = input_53_end_mask_0, squeeze_mask = input_53_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_53")];
            tensor<int32, []> var_415_axis_0 = const()[name = tensor<string, []>("op_415_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_415_batch_dims_0 = const()[name = tensor<string, []>("op_415_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_415_validate_indices_0 = const()[name = tensor<string, []>("op_415_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_13_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_13_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335360)))];
            tensor<string, []> input_53_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_53_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_53_to_uint16 = cast(dtype = input_53_to_uint16_dtype_0, x = input_53)[name = tensor<string, []>("cast_404")];
            tensor<fp16, [1, ?, 8]> var_415_cast_fp16_cast_uint16 = gather(axis = var_415_axis_0, batch_dims = var_415_batch_dims_0, indices = input_53_to_uint16, validate_indices = var_415_validate_indices_0, x = quantizers_13_codebook_weight_to_fp16)[name = tensor<string, []>("op_415_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_25_pad_type_0 = const()[name = tensor<string, []>("z_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_25_strides_0 = const()[name = tensor<string, []>("z_25_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_25_pad_0 = const()[name = tensor<string, []>("z_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_25_dilations_0 = const()[name = tensor<string, []>("z_25_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_25_groups_0 = const()[name = tensor<string, []>("z_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_13_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_13_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351808)))];
            tensor<fp16, [512]> quantizers_13_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_13_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360064)))];
            tensor<fp16, [1, 8, ?]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = var_415_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [1, 512, ?]> z_25_cast_fp16 = conv(bias = quantizers_13_out_proj_bias_to_fp16, dilations = z_25_dilations_0, groups = z_25_groups_0, pad = z_25_pad_0, pad_type = z_25_pad_type_0, strides = z_25_strides_0, weight = quantizers_13_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("z_25_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_27_cast_fp16 = add(x = emb_25_cast_fp16, y = z_25_cast_fp16)[name = tensor<string, []>("emb_27_cast_fp16")];
            tensor<int32, [3]> input_57_begin_0 = const()[name = tensor<string, []>("input_57_begin_0"), val = tensor<int32, [3]>([14, 0, 0])];
            tensor<int32, [3]> input_57_end_0 = const()[name = tensor<string, []>("input_57_end_0"), val = tensor<int32, [3]>([15, 1, 0])];
            tensor<bool, [3]> input_57_end_mask_0 = const()[name = tensor<string, []>("input_57_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_57_squeeze_mask_0 = const()[name = tensor<string, []>("input_57_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_57 = slice_by_index(begin = input_57_begin_0, end = input_57_end_0, end_mask = input_57_end_mask_0, squeeze_mask = input_57_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_57")];
            tensor<int32, []> var_438_axis_0 = const()[name = tensor<string, []>("op_438_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_438_batch_dims_0 = const()[name = tensor<string, []>("op_438_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_438_validate_indices_0 = const()[name = tensor<string, []>("op_438_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_14_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_14_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361152)))];
            tensor<string, []> input_57_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_57_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_57_to_uint16 = cast(dtype = input_57_to_uint16_dtype_0, x = input_57)[name = tensor<string, []>("cast_403")];
            tensor<fp16, [1, ?, 8]> var_438_cast_fp16_cast_uint16 = gather(axis = var_438_axis_0, batch_dims = var_438_batch_dims_0, indices = input_57_to_uint16, validate_indices = var_438_validate_indices_0, x = quantizers_14_codebook_weight_to_fp16)[name = tensor<string, []>("op_438_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_59_perm_0 = const()[name = tensor<string, []>("input_59_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_27_pad_type_0 = const()[name = tensor<string, []>("z_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_27_strides_0 = const()[name = tensor<string, []>("z_27_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_27_pad_0 = const()[name = tensor<string, []>("z_27_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_27_dilations_0 = const()[name = tensor<string, []>("z_27_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_27_groups_0 = const()[name = tensor<string, []>("z_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_14_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_14_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377600)))];
            tensor<fp16, [512]> quantizers_14_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_14_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385856)))];
            tensor<fp16, [1, 8, ?]> input_59_cast_fp16 = transpose(perm = input_59_perm_0, x = var_438_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [1, 512, ?]> z_27_cast_fp16 = conv(bias = quantizers_14_out_proj_bias_to_fp16, dilations = z_27_dilations_0, groups = z_27_groups_0, pad = z_27_pad_0, pad_type = z_27_pad_type_0, strides = z_27_strides_0, weight = quantizers_14_out_proj_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("z_27_cast_fp16")];
            tensor<fp16, [1, 512, ?]> emb_cast_fp16 = add(x = emb_27_cast_fp16, y = z_27_cast_fp16)[name = tensor<string, []>("emb_cast_fp16")];
            tensor<int32, [3]> input_61_begin_0 = const()[name = tensor<string, []>("input_61_begin_0"), val = tensor<int32, [3]>([15, 0, 0])];
            tensor<int32, [3]> input_61_end_0 = const()[name = tensor<string, []>("input_61_end_0"), val = tensor<int32, [3]>([16, 1, 0])];
            tensor<bool, [3]> input_61_end_mask_0 = const()[name = tensor<string, []>("input_61_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> input_61_squeeze_mask_0 = const()[name = tensor<string, []>("input_61_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<int32, [1, ?]> input_61 = slice_by_index(begin = input_61_begin_0, end = input_61_end_0, end_mask = input_61_end_mask_0, squeeze_mask = input_61_squeeze_mask_0, x = codes)[name = tensor<string, []>("input_61")];
            tensor<int32, []> var_461_axis_0 = const()[name = tensor<string, []>("op_461_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_461_batch_dims_0 = const()[name = tensor<string, []>("op_461_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> var_461_validate_indices_0 = const()[name = tensor<string, []>("op_461_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1024, 8]> quantizers_15_codebook_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_15_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386944)))];
            tensor<string, []> input_61_to_uint16_dtype_0 = const()[name = tensor<string, []>("input_61_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, [1, ?]> input_61_to_uint16 = cast(dtype = input_61_to_uint16_dtype_0, x = input_61)[name = tensor<string, []>("cast_402")];
            tensor<fp16, [1, ?, 8]> var_461_cast_fp16_cast_uint16 = gather(axis = var_461_axis_0, batch_dims = var_461_batch_dims_0, indices = input_61_to_uint16, validate_indices = var_461_validate_indices_0, x = quantizers_15_codebook_weight_to_fp16)[name = tensor<string, []>("op_461_cast_fp16_cast_uint16")];
            tensor<int32, [3]> input_63_perm_0 = const()[name = tensor<string, []>("input_63_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> z_pad_type_0 = const()[name = tensor<string, []>("z_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> z_strides_0 = const()[name = tensor<string, []>("z_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> z_pad_0 = const()[name = tensor<string, []>("z_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> z_dilations_0 = const()[name = tensor<string, []>("z_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> z_groups_0 = const()[name = tensor<string, []>("z_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 8, 1]> quantizers_15_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("quantizers_15_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403392)))];
            tensor<fp16, [512]> quantizers_15_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("quantizers_15_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411648)))];
            tensor<fp16, [1, 8, ?]> input_63_cast_fp16 = transpose(perm = input_63_perm_0, x = var_461_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 512, ?]> z_cast_fp16 = conv(bias = quantizers_15_out_proj_bias_to_fp16, dilations = z_dilations_0, groups = z_groups_0, pad = z_pad_0, pad_type = z_pad_type_0, strides = z_strides_0, weight = quantizers_15_out_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("z_cast_fp16")];
            tensor<fp16, [1, 512, ?]> input_65_cast_fp16 = add(x = emb_cast_fp16, y = z_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<string, []> x_1_pad_type_0 = const()[name = tensor<string, []>("x_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> x_1_strides_0 = const()[name = tensor<string, []>("x_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_1_pad_0 = const()[name = tensor<string, []>("x_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_1_dilations_0 = const()[name = tensor<string, []>("x_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> x_1_groups_0 = const()[name = tensor<string, []>("x_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 512, 1]> output_proj_weight_to_fp16 = const()[name = tensor<string, []>("output_proj_weight_to_fp16"), val = tensor<fp16, [768, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412736)))];
            tensor<fp16, [768]> output_proj_bias_to_fp16 = const()[name = tensor<string, []>("output_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1199232)))];
            tensor<fp16, [1, 768, ?]> x_1_cast_fp16 = conv(bias = output_proj_bias_to_fp16, dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = output_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<int32, [3]> var_496_begin_0 = const()[name = tensor<string, []>("op_496_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_496_end_0 = const()[name = tensor<string, []>("op_496_end_0"), val = tensor<int32, [3]>([1, 1, 0])];
            tensor<bool, [3]> var_496_end_mask_0 = const()[name = tensor<string, []>("op_496_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_496_squeeze_mask_0 = const()[name = tensor<string, []>("op_496_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, ?]> var_496_cast_fp16 = slice_by_index(begin = var_496_begin_0, end = var_496_end_0, end_mask = var_496_end_mask_0, squeeze_mask = var_496_squeeze_mask_0, x = x_1_cast_fp16)[name = tensor<string, []>("op_496_cast_fp16")];
            tensor<fp16, []> fill_like_0_value_0_to_fp16 = const()[name = tensor<string, []>("fill_like_0_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?]> fill_like_0_cast_fp16 = fill_like(ref_tensor = var_496_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = tensor<string, []>("fill_like_0_cast_fp16")];
            tensor<int32, [1]> var_512_axes_0 = const()[name = tensor<string, []>("op_512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_512_keep_dims_0 = const()[name = tensor<string, []>("op_512_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1]> var_512_cast_fp16 = reduce_sum(axes = var_512_axes_0, keep_dims = var_512_keep_dims_0, x = fill_like_0_cast_fp16)[name = tensor<string, []>("op_512_cast_fp16")];
            tensor<string, []> input_lengths_1_dtype_0 = const()[name = tensor<string, []>("input_lengths_1_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [4]>([1, 192, 4, -1])];
            tensor<fp16, [1, 192, 4, ?]> var_533_cast_fp16 = reshape(shape = concat_0x, x = x_1_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
            tensor<int32, [4]> var_534 = const()[name = tensor<string, []>("op_534"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([1, 192, -1])];
            tensor<fp16, [1, 192, ?, 4]> var_535_cast_fp16 = transpose(perm = var_534, x = var_533_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [1, 192, ?]> x_3_cast_fp16 = reshape(shape = concat_1x, x = var_535_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, []> var_541 = const()[name = tensor<string, []>("op_541"), val = tensor<int32, []>(4)];
            tensor<int32, [1]> var_512_cast_fp16_to_int32 = cast(dtype = input_lengths_1_dtype_0, x = var_512_cast_fp16)[name = tensor<string, []>("cast_401")];
            tensor<int32, [1]> input_lengths_3 = mul(x = var_512_cast_fp16_to_int32, y = var_541)[name = tensor<string, []>("input_lengths_3")];
            tensor<int32, []> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_554 = const()[name = tensor<string, []>("op_554"), val = tensor<int32, []>(500)];
            tensor<int32, []> var_565 = const()[name = tensor<string, []>("op_565"), val = tensor<int32, []>(1)];
            tensor<int32, [3]> input_67_perm_0 = const()[name = tensor<string, []>("input_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [256, 192]> decoder_1_input_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_input_proj_weight_to_fp16"), val = tensor<fp16, [256, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1200832)))];
            tensor<fp16, [256]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1299200)))];
            tensor<fp16, [1, ?, 192]> input_67_cast_fp16 = transpose(perm = input_67_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, ?, 256]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1299776)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1300352)))];
            tensor<fp16, []> var_561_to_fp16 = const()[name = tensor<string, []>("op_561_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, ?, 256]> x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = decoder_1_transformer_layers_0_norm1_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_0_norm1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<int32, [3]> var_593_shape_cast_fp16 = shape(x = x_7_cast_fp16)[name = tensor<string, []>("op_593_shape_cast_fp16")];
            tensor<int32, []> gather_4_axis_0 = const()[name = tensor<string, []>("gather_4_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_4_batch_dims_0 = const()[name = tensor<string, []>("gather_4_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_4_validate_indices_0 = const()[name = tensor<string, []>("gather_4_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_593_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_593_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_4_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_4_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_593_shape_cast_fp16_to_uint16 = cast(dtype = var_593_shape_cast_fp16_to_uint16_dtype_0, x = var_593_shape_cast_fp16)[name = tensor<string, []>("cast_400")];
            tensor<uint16, []> gather_4_cast_uint16 = gather(axis = gather_4_axis_0, batch_dims = gather_4_batch_dims_0, indices = gather_4_indices_0_to_uint16, validate_indices = gather_4_validate_indices_0, x = var_593_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_4_cast_uint16")];
            tensor<string, []> gather_4_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_4_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_1_transformer_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1300928)))];
            tensor<fp16, [768]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1694208)))];
            tensor<fp16, [1, ?, 768]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_1_transformer_layers_0_self_attn_in_proj_weight_to_fp16, x = x_7_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<int32, [5]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_600_cast_fp16 = reshape(shape = concat_2x, x = linear_1_cast_fp16)[name = tensor<string, []>("op_600_cast_fp16")];
            tensor<int32, [5]> var_601 = const()[name = tensor<string, []>("op_601"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_3_cast_fp16 = transpose(perm = var_601, x = var_600_cast_fp16)[name = tensor<string, []>("transpose_35")];
            tensor<fp16, [1, 4, ?, 64]> q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = projected_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
            tensor<int32, [5]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = projected_3_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
            tensor<int32, [5]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = projected_3_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
            tensor<int32, [4]> var_613_shape_cast_fp16 = shape(x = q_1_cast_fp16)[name = tensor<string, []>("op_613_shape_cast_fp16")];
            tensor<int32, []> gather_8_axis_0 = const()[name = tensor<string, []>("gather_8_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_8_batch_dims_0 = const()[name = tensor<string, []>("gather_8_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_8_validate_indices_0 = const()[name = tensor<string, []>("gather_8_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_613_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_613_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_8_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_8_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_613_shape_cast_fp16_to_uint16 = cast(dtype = var_613_shape_cast_fp16_to_uint16_dtype_0, x = var_613_shape_cast_fp16)[name = tensor<string, []>("cast_398")];
            tensor<uint16, []> gather_8_cast_uint16 = gather(axis = gather_8_axis_0, batch_dims = gather_8_batch_dims_0, indices = gather_8_indices_0_to_uint16, validate_indices = gather_8_validate_indices_0, x = var_613_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_8_cast_uint16")];
            tensor<string, []> gather_8_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_8_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_8_cast_uint16_to_int32 = cast(dtype = gather_8_cast_uint16_to_int32_dtype_0, x = gather_8_cast_uint16)[name = tensor<string, []>("cast_397")];
            tensor<int32, [?]> var_617 = range_1d(end = gather_8_cast_uint16_to_int32, start = const_3, step = const_4)[name = tensor<string, []>("op_617")];
            tensor<fp16, [1, 1]> var_616_to_fp16 = const()[name = tensor<string, []>("op_616_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_617_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_617_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_617_to_fp16 = cast(dtype = var_617_promoted_to_fp16_dtype_0, x = var_617)[name = tensor<string, []>("cast_396")];
            tensor<fp16, [1, ?]> ts_1_cast_fp16 = add(x = var_616_to_fp16, y = var_617_to_fp16)[name = tensor<string, []>("ts_1_cast_fp16")];
            tensor<int32, [4]> var_620 = const()[name = tensor<string, []>("op_620"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_3_cast_fp16 = reshape(shape = var_620, x = ts_1_cast_fp16)[name = tensor<string, []>("ts_3_cast_fp16")];
            tensor<int32, [5]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_1_cast_fp16 = reshape(shape = concat_3x, x = q_1_cast_fp16)[name = tensor<string, []>("q2_1_cast_fp16")];
            tensor<int32, [5]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_1_cast_fp16 = reshape(shape = concat_4x, x = k_1_cast_fp16)[name = tensor<string, []>("k2_1_cast_fp16")];
            tensor<int32, [5]> qr_1_begin_0 = const()[name = tensor<string, []>("qr_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_1_end_0 = const()[name = tensor<string, []>("qr_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_1_end_mask_0 = const()[name = tensor<string, []>("qr_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_1_squeeze_mask_0 = const()[name = tensor<string, []>("qr_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_1_cast_fp16 = slice_by_index(begin = qr_1_begin_0, end = qr_1_end_0, end_mask = qr_1_end_mask_0, squeeze_mask = qr_1_squeeze_mask_0, x = q2_1_cast_fp16)[name = tensor<string, []>("qr_1_cast_fp16")];
            tensor<int32, [5]> qi_1_begin_0 = const()[name = tensor<string, []>("qi_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_1_end_0 = const()[name = tensor<string, []>("qi_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_1_end_mask_0 = const()[name = tensor<string, []>("qi_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_1_squeeze_mask_0 = const()[name = tensor<string, []>("qi_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_1_cast_fp16 = slice_by_index(begin = qi_1_begin_0, end = qi_1_end_0, end_mask = qi_1_end_mask_0, squeeze_mask = qi_1_squeeze_mask_0, x = q2_1_cast_fp16)[name = tensor<string, []>("qi_1_cast_fp16")];
            tensor<int32, [5]> kr_1_begin_0 = const()[name = tensor<string, []>("kr_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_1_end_0 = const()[name = tensor<string, []>("kr_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_1_end_mask_0 = const()[name = tensor<string, []>("kr_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_1_squeeze_mask_0 = const()[name = tensor<string, []>("kr_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_1_cast_fp16 = slice_by_index(begin = kr_1_begin_0, end = kr_1_end_0, end_mask = kr_1_end_mask_0, squeeze_mask = kr_1_squeeze_mask_0, x = k2_1_cast_fp16)[name = tensor<string, []>("kr_1_cast_fp16")];
            tensor<int32, [5]> ki_1_begin_0 = const()[name = tensor<string, []>("ki_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_1_end_0 = const()[name = tensor<string, []>("ki_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_1_end_mask_0 = const()[name = tensor<string, []>("ki_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_1_squeeze_mask_0 = const()[name = tensor<string, []>("ki_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_1_cast_fp16 = slice_by_index(begin = ki_1_begin_0, end = ki_1_end_0, end_mask = ki_1_end_mask_0, squeeze_mask = ki_1_squeeze_mask_0, x = k2_1_cast_fp16)[name = tensor<string, []>("ki_1_cast_fp16")];
            tensor<fp16, [32]> freqs_1_to_fp16 = const()[name = tensor<string, []>("freqs_1_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1695808)))];
            tensor<fp16, [1, 1, ?, 32]> var_633_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_3_cast_fp16)[name = tensor<string, []>("op_633_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_1_cast_fp16 = cos(x = var_633_cast_fp16)[name = tensor<string, []>("rotr_1_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_1_cast_fp16 = sin(x = var_633_cast_fp16)[name = tensor<string, []>("roti_1_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_637_cast_fp16 = mul(x = qr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor<string, []>("op_637_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_638_cast_fp16 = mul(x = qi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor<string, []>("op_638_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_639_cast_fp16 = sub(x = var_637_cast_fp16, y = var_638_cast_fp16)[name = tensor<string, []>("op_639_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_640_cast_fp16 = mul(x = qr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor<string, []>("op_640_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_641_cast_fp16 = mul(x = qi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor<string, []>("op_641_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_642_cast_fp16 = add(x = var_640_cast_fp16, y = var_641_cast_fp16)[name = tensor<string, []>("op_642_cast_fp16")];
            tensor<int32, []> qo_1_axis_0 = const()[name = tensor<string, []>("qo_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (var_639_cast_fp16, var_642_cast_fp16))[name = tensor<string, []>("qo_1_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_645_cast_fp16 = mul(x = kr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor<string, []>("op_645_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_646_cast_fp16 = mul(x = ki_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor<string, []>("op_646_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_647_cast_fp16 = sub(x = var_645_cast_fp16, y = var_646_cast_fp16)[name = tensor<string, []>("op_647_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_648_cast_fp16 = mul(x = kr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor<string, []>("op_648_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_649_cast_fp16 = mul(x = ki_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor<string, []>("op_649_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_650_cast_fp16 = add(x = var_648_cast_fp16, y = var_649_cast_fp16)[name = tensor<string, []>("op_650_cast_fp16")];
            tensor<int32, []> ko_1_axis_0 = const()[name = tensor<string, []>("ko_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (var_647_cast_fp16, var_650_cast_fp16))[name = tensor<string, []>("ko_1_cast_fp16")];
            tensor<int32, [4]> concat_5x = const()[name = tensor<string, []>("concat_5x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_3_cast_fp16 = reshape(shape = concat_5x, x = qo_1_cast_fp16)[name = tensor<string, []>("q_3_cast_fp16")];
            tensor<int32, [4]> concat_6x = const()[name = tensor<string, []>("concat_6x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_3_cast_fp16 = reshape(shape = concat_6x, x = ko_1_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
            tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_4_cast_uint16_to_int32 = cast(dtype = gather_4_cast_uint16_to_int32_dtype_0, x = gather_4_cast_uint16)[name = tensor<string, []>("cast_399")];
            tensor<int32, [?]> positions_1 = range_1d(end = gather_4_cast_uint16_to_int32, start = const_5, step = const_6)[name = tensor<string, []>("positions_1")];
            tensor<int32, [3]> concat_7x = const()[name = tensor<string, []>("concat_7x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_662 = reshape(shape = concat_7x, x = positions_1)[name = tensor<string, []>("op_662")];
            tensor<int32, [3]> var_663 = const()[name = tensor<string, []>("op_663"), val = tensor<int32, [3]>([-1, 1, 1])];
            tensor<int32, [1, 1, 1]> var_664 = reshape(shape = var_663, x = input_lengths_3)[name = tensor<string, []>("op_664")];
            tensor<bool, [1, 1, ?]> valid_k_1 = less(x = var_662, y = var_664)[name = tensor<string, []>("valid_k_1")];
            tensor<int32, [3]> concat_8x = const()[name = tensor<string, []>("concat_8x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_667 = reshape(shape = concat_8x, x = positions_1)[name = tensor<string, []>("op_667")];
            tensor<int32, [1, ?, ?]> delta_1 = sub(x = var_667, y = var_662)[name = tensor<string, []>("delta_1")];
            tensor<int32, []> concat_10_axis_0 = const()[name = tensor<string, []>("concat_10_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_10_interleave_0 = const()[name = tensor<string, []>("concat_10_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (var_565, gather_4_cast_uint16_to_int32, gather_4_cast_uint16_to_int32))[name = tensor<string, []>("concat_10")];
            tensor<fp16, []> fill_0_value_0_to_fp16 = const()[name = tensor<string, []>("fill_0_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_0_cast_fp16 = fill(shape = concat_10, value = fill_0_value_0_to_fp16)[name = tensor<string, []>("fill_0_cast_fp16")];
            tensor<string, []> attn_bias_1_dtype_0 = const()[name = tensor<string, []>("attn_bias_1_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_673 = greater_equal(x = delta_1, y = var_547)[name = tensor<string, []>("op_673")];
            tensor<bool, [1, ?, ?]> fill_0_cast_fp16_to_bool = cast(dtype = attn_bias_1_dtype_0, x = fill_0_cast_fp16)[name = tensor<string, []>("cast_395")];
            tensor<bool, [1, ?, ?]> attn_bias_3 = logical_and(x = fill_0_cast_fp16_to_bool, y = var_673)[name = tensor<string, []>("attn_bias_3")];
            tensor<bool, [1, ?, ?]> var_675 = less(x = delta_1, y = var_554)[name = tensor<string, []>("op_675")];
            tensor<bool, [1, ?, ?]> attn_bias_5 = logical_and(x = attn_bias_3, y = var_675)[name = tensor<string, []>("attn_bias_5")];
            tensor<bool, [1, ?, ?]> var_677 = logical_and(x = attn_bias_5, y = valid_k_1)[name = tensor<string, []>("op_677")];
            tensor<int32, [1]> var_679_axes_0 = const()[name = tensor<string, []>("op_679_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_679 = expand_dims(axes = var_679_axes_0, x = var_677)[name = tensor<string, []>("op_679")];
            tensor<fp16, []> sub_0_x_0_to_fp16 = const()[name = tensor<string, []>("sub_0_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_35_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_35_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_679_to_fp16 = cast(dtype = cast_35_to_fp16_dtype_0, x = var_679)[name = tensor<string, []>("cast_394")];
            tensor<fp16, [1, 1, ?, ?]> sub_0_cast_fp16 = sub(x = sub_0_x_0_to_fp16, y = var_679_to_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
            tensor<fp16, []> mul_0_x_0_to_fp16 = const()[name = tensor<string, []>("mul_0_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_0_cast_fp16 = mul(x = mul_0_x_0_to_fp16, y = sub_0_cast_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
            tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_1_cast_fp16 = mul(x = q_3_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")];
            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = k_3_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_cast_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
            tensor<string, []> add_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_0_cast_fp16_to_fp32 = cast(dtype = add_0_cast_fp16_to_fp32_dtype_0, x = add_0_cast_fp16)[name = tensor<string, []>("cast_393")];
            tensor<fp32, [1, 4, ?, ?]> softmax_0 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_0")];
            tensor<bool, []> out_1_transpose_x_0 = const()[name = tensor<string, []>("out_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_1_transpose_y_0 = const()[name = tensor<string, []>("out_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_0_to_fp16 = cast(dtype = softmax_0_to_fp16_dtype_0, x = softmax_0)[name = tensor<string, []>("cast_392")];
            tensor<fp16, [1, 4, ?, 64]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = softmax_0_to_fp16, y = v_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<int32, [2]> concat_11x = const()[name = tensor<string, []>("concat_11x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_685 = reshape(shape = concat_11x, x = positions_1)[name = tensor<string, []>("op_685")];
            tensor<int32, [2]> var_686 = const()[name = tensor<string, []>("op_686"), val = tensor<int32, [2]>([-1, 1])];
            tensor<int32, [1, 1]> var_687 = reshape(shape = var_686, x = input_lengths_3)[name = tensor<string, []>("op_687")];
            tensor<bool, [1, ?]> var_688 = less(x = var_685, y = var_687)[name = tensor<string, []>("op_688")];
            tensor<int32, [4]> concat_12x = const()[name = tensor<string, []>("concat_12x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_1 = reshape(shape = concat_12x, x = var_688)[name = tensor<string, []>("valid_q_1")];
            tensor<fp16, []> var_692_to_fp16 = const()[name = tensor<string, []>("op_692_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_3_cast_fp16 = select(a = out_1_cast_fp16, b = var_692_to_fp16, cond = valid_q_1)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<int32, [4]> var_694_perm_0 = const()[name = tensor<string, []>("op_694_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_13x = const()[name = tensor<string, []>("concat_13x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_694_cast_fp16 = transpose(perm = var_694_perm_0, x = out_3_cast_fp16)[name = tensor<string, []>("transpose_34")];
            tensor<fp16, [1, ?, 256]> input_69_cast_fp16 = reshape(shape = concat_13x, x = var_694_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_1_transformer_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1695936)))];
            tensor<fp16, [1, ?, 256]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_0_self_attn_out_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1827072)))];
            tensor<fp16, [1, ?, 256]> var_700_cast_fp16 = mul(x = decoder_1_transformer_layers_0_layer_scale_1_scale_to_fp16, y = linear_2_cast_fp16)[name = tensor<string, []>("op_700_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_71_cast_fp16 = add(x = linear_0_cast_fp16, y = var_700_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<int32, [1]> input_73_axes_0 = const()[name = tensor<string, []>("input_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1827648)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1828224)))];
            tensor<fp16, [1, ?, 256]> input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = decoder_1_transformer_layers_0_norm2_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_0_norm2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_1_transformer_layers_0_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1828800)))];
            tensor<fp16, [1024]> linear_3_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_3_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2353152)))];
            tensor<fp16, [1, ?, 1024]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_1_transformer_layers_0_ffn_0_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<string, []> input_77_mode_0 = const()[name = tensor<string, []>("input_77_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_77_cast_fp16 = gelu(mode = input_77_mode_0, x = linear_3_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_1_transformer_layers_0_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2355264)))];
            tensor<fp16, [1, ?, 256]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_0_ffn_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_0_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_0_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2879616)))];
            tensor<fp16, [1, ?, 256]> var_715_cast_fp16 = mul(x = decoder_1_transformer_layers_0_layer_scale_2_scale_to_fp16, y = linear_4_cast_fp16)[name = tensor<string, []>("op_715_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_79_cast_fp16 = add(x = input_71_cast_fp16, y = var_715_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2880192)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2880768)))];
            tensor<fp16, [1, ?, 256]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = decoder_1_transformer_layers_1_norm1_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_1_norm1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<int32, [3]> var_730_shape_cast_fp16 = shape(x = x_13_cast_fp16)[name = tensor<string, []>("op_730_shape_cast_fp16")];
            tensor<int32, []> gather_14_axis_0 = const()[name = tensor<string, []>("gather_14_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_14_batch_dims_0 = const()[name = tensor<string, []>("gather_14_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_14_validate_indices_0 = const()[name = tensor<string, []>("gather_14_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_730_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_730_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_14_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_14_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_730_shape_cast_fp16_to_uint16 = cast(dtype = var_730_shape_cast_fp16_to_uint16_dtype_0, x = var_730_shape_cast_fp16)[name = tensor<string, []>("cast_391")];
            tensor<uint16, []> gather_14_cast_uint16 = gather(axis = gather_14_axis_0, batch_dims = gather_14_batch_dims_0, indices = gather_14_indices_0_to_uint16, validate_indices = gather_14_validate_indices_0, x = var_730_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_14_cast_uint16")];
            tensor<string, []> gather_14_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_14_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_1_transformer_layers_1_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2881344)))];
            tensor<fp16, [1, ?, 768]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_1_transformer_layers_1_self_attn_in_proj_weight_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [5]> concat_14x = const()[name = tensor<string, []>("concat_14x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_737_cast_fp16 = reshape(shape = concat_14x, x = linear_5_cast_fp16)[name = tensor<string, []>("op_737_cast_fp16")];
            tensor<int32, [5]> var_738 = const()[name = tensor<string, []>("op_738"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_5_begin_0 = const()[name = tensor<string, []>("q_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_5_end_0 = const()[name = tensor<string, []>("q_5_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_5_end_mask_0 = const()[name = tensor<string, []>("q_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_5_squeeze_mask_0 = const()[name = tensor<string, []>("q_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_7_cast_fp16 = transpose(perm = var_738, x = var_737_cast_fp16)[name = tensor<string, []>("transpose_33")];
            tensor<fp16, [1, 4, ?, 64]> q_5_cast_fp16 = slice_by_index(begin = q_5_begin_0, end = q_5_end_0, end_mask = q_5_end_mask_0, squeeze_mask = q_5_squeeze_mask_0, x = projected_7_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")];
            tensor<int32, [5]> k_5_begin_0 = const()[name = tensor<string, []>("k_5_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_5_end_0 = const()[name = tensor<string, []>("k_5_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_5_end_mask_0 = const()[name = tensor<string, []>("k_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_5_squeeze_mask_0 = const()[name = tensor<string, []>("k_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = projected_7_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
            tensor<int32, [5]> v_3_begin_0 = const()[name = tensor<string, []>("v_3_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_3_end_0 = const()[name = tensor<string, []>("v_3_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_3_end_mask_0 = const()[name = tensor<string, []>("v_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_3_squeeze_mask_0 = const()[name = tensor<string, []>("v_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = projected_7_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
            tensor<int32, [4]> var_750_shape_cast_fp16 = shape(x = q_5_cast_fp16)[name = tensor<string, []>("op_750_shape_cast_fp16")];
            tensor<int32, []> gather_18_axis_0 = const()[name = tensor<string, []>("gather_18_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_18_batch_dims_0 = const()[name = tensor<string, []>("gather_18_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_18_validate_indices_0 = const()[name = tensor<string, []>("gather_18_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_750_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_750_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_18_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_18_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_750_shape_cast_fp16_to_uint16 = cast(dtype = var_750_shape_cast_fp16_to_uint16_dtype_0, x = var_750_shape_cast_fp16)[name = tensor<string, []>("cast_389")];
            tensor<uint16, []> gather_18_cast_uint16 = gather(axis = gather_18_axis_0, batch_dims = gather_18_batch_dims_0, indices = gather_18_indices_0_to_uint16, validate_indices = gather_18_validate_indices_0, x = var_750_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_18_cast_uint16")];
            tensor<string, []> gather_18_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_18_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_13 = const()[name = tensor<string, []>("const_13"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_18_cast_uint16_to_int32 = cast(dtype = gather_18_cast_uint16_to_int32_dtype_0, x = gather_18_cast_uint16)[name = tensor<string, []>("cast_388")];
            tensor<int32, [?]> var_754 = range_1d(end = gather_18_cast_uint16_to_int32, start = const_12, step = const_13)[name = tensor<string, []>("op_754")];
            tensor<fp16, [1, 1]> var_753_to_fp16 = const()[name = tensor<string, []>("op_753_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_754_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_754_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_754_to_fp16 = cast(dtype = var_754_promoted_to_fp16_dtype_0, x = var_754)[name = tensor<string, []>("cast_387")];
            tensor<fp16, [1, ?]> ts_5_cast_fp16 = add(x = var_753_to_fp16, y = var_754_to_fp16)[name = tensor<string, []>("ts_5_cast_fp16")];
            tensor<int32, [4]> var_757 = const()[name = tensor<string, []>("op_757"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_7_cast_fp16 = reshape(shape = var_757, x = ts_5_cast_fp16)[name = tensor<string, []>("ts_7_cast_fp16")];
            tensor<int32, [5]> concat_15x = const()[name = tensor<string, []>("concat_15x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_3_cast_fp16 = reshape(shape = concat_15x, x = q_5_cast_fp16)[name = tensor<string, []>("q2_3_cast_fp16")];
            tensor<int32, [5]> concat_16x = const()[name = tensor<string, []>("concat_16x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_3_cast_fp16 = reshape(shape = concat_16x, x = k_5_cast_fp16)[name = tensor<string, []>("k2_3_cast_fp16")];
            tensor<int32, [5]> qr_3_begin_0 = const()[name = tensor<string, []>("qr_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_3_end_0 = const()[name = tensor<string, []>("qr_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_3_end_mask_0 = const()[name = tensor<string, []>("qr_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_3_squeeze_mask_0 = const()[name = tensor<string, []>("qr_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_3_cast_fp16 = slice_by_index(begin = qr_3_begin_0, end = qr_3_end_0, end_mask = qr_3_end_mask_0, squeeze_mask = qr_3_squeeze_mask_0, x = q2_3_cast_fp16)[name = tensor<string, []>("qr_3_cast_fp16")];
            tensor<int32, [5]> qi_3_begin_0 = const()[name = tensor<string, []>("qi_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_3_end_0 = const()[name = tensor<string, []>("qi_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_3_end_mask_0 = const()[name = tensor<string, []>("qi_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_3_squeeze_mask_0 = const()[name = tensor<string, []>("qi_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_3_cast_fp16 = slice_by_index(begin = qi_3_begin_0, end = qi_3_end_0, end_mask = qi_3_end_mask_0, squeeze_mask = qi_3_squeeze_mask_0, x = q2_3_cast_fp16)[name = tensor<string, []>("qi_3_cast_fp16")];
            tensor<int32, [5]> kr_3_begin_0 = const()[name = tensor<string, []>("kr_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_3_end_0 = const()[name = tensor<string, []>("kr_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_3_end_mask_0 = const()[name = tensor<string, []>("kr_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_3_squeeze_mask_0 = const()[name = tensor<string, []>("kr_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_3_cast_fp16 = slice_by_index(begin = kr_3_begin_0, end = kr_3_end_0, end_mask = kr_3_end_mask_0, squeeze_mask = kr_3_squeeze_mask_0, x = k2_3_cast_fp16)[name = tensor<string, []>("kr_3_cast_fp16")];
            tensor<int32, [5]> ki_3_begin_0 = const()[name = tensor<string, []>("ki_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_3_end_0 = const()[name = tensor<string, []>("ki_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_3_end_mask_0 = const()[name = tensor<string, []>("ki_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_3_squeeze_mask_0 = const()[name = tensor<string, []>("ki_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_3_cast_fp16 = slice_by_index(begin = ki_3_begin_0, end = ki_3_end_0, end_mask = ki_3_end_mask_0, squeeze_mask = ki_3_squeeze_mask_0, x = k2_3_cast_fp16)[name = tensor<string, []>("ki_3_cast_fp16")];
            tensor<fp16, [32]> freqs_3_to_fp16 = const()[name = tensor<string, []>("freqs_3_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3274624)))];
            tensor<fp16, [1, 1, ?, 32]> var_770_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_7_cast_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_3_cast_fp16 = cos(x = var_770_cast_fp16)[name = tensor<string, []>("rotr_3_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_3_cast_fp16 = sin(x = var_770_cast_fp16)[name = tensor<string, []>("roti_3_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_774_cast_fp16 = mul(x = qr_3_cast_fp16, y = rotr_3_cast_fp16)[name = tensor<string, []>("op_774_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_775_cast_fp16 = mul(x = qi_3_cast_fp16, y = roti_3_cast_fp16)[name = tensor<string, []>("op_775_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_776_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor<string, []>("op_776_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_777_cast_fp16 = mul(x = qr_3_cast_fp16, y = roti_3_cast_fp16)[name = tensor<string, []>("op_777_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_778_cast_fp16 = mul(x = qi_3_cast_fp16, y = rotr_3_cast_fp16)[name = tensor<string, []>("op_778_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_779_cast_fp16 = add(x = var_777_cast_fp16, y = var_778_cast_fp16)[name = tensor<string, []>("op_779_cast_fp16")];
            tensor<int32, []> qo_3_axis_0 = const()[name = tensor<string, []>("qo_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (var_776_cast_fp16, var_779_cast_fp16))[name = tensor<string, []>("qo_3_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_782_cast_fp16 = mul(x = kr_3_cast_fp16, y = rotr_3_cast_fp16)[name = tensor<string, []>("op_782_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_783_cast_fp16 = mul(x = ki_3_cast_fp16, y = roti_3_cast_fp16)[name = tensor<string, []>("op_783_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_784_cast_fp16 = sub(x = var_782_cast_fp16, y = var_783_cast_fp16)[name = tensor<string, []>("op_784_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_785_cast_fp16 = mul(x = kr_3_cast_fp16, y = roti_3_cast_fp16)[name = tensor<string, []>("op_785_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_786_cast_fp16 = mul(x = ki_3_cast_fp16, y = rotr_3_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_787_cast_fp16 = add(x = var_785_cast_fp16, y = var_786_cast_fp16)[name = tensor<string, []>("op_787_cast_fp16")];
            tensor<int32, []> ko_3_axis_0 = const()[name = tensor<string, []>("ko_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (var_784_cast_fp16, var_787_cast_fp16))[name = tensor<string, []>("ko_3_cast_fp16")];
            tensor<int32, [4]> concat_17x = const()[name = tensor<string, []>("concat_17x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_7_cast_fp16 = reshape(shape = concat_17x, x = qo_3_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
            tensor<int32, [4]> concat_18x = const()[name = tensor<string, []>("concat_18x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_7_cast_fp16 = reshape(shape = concat_18x, x = ko_3_cast_fp16)[name = tensor<string, []>("k_7_cast_fp16")];
            tensor<int32, []> const_14 = const()[name = tensor<string, []>("const_14"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_14_cast_uint16_to_int32 = cast(dtype = gather_14_cast_uint16_to_int32_dtype_0, x = gather_14_cast_uint16)[name = tensor<string, []>("cast_390")];
            tensor<int32, [?]> positions_3 = range_1d(end = gather_14_cast_uint16_to_int32, start = const_14, step = const_15)[name = tensor<string, []>("positions_3")];
            tensor<int32, [3]> concat_19x = const()[name = tensor<string, []>("concat_19x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_799 = reshape(shape = concat_19x, x = positions_3)[name = tensor<string, []>("op_799")];
            tensor<bool, [1, 1, ?]> valid_k_3 = less(x = var_799, y = var_664)[name = tensor<string, []>("valid_k_3")];
            tensor<int32, [3]> concat_20x = const()[name = tensor<string, []>("concat_20x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_804 = reshape(shape = concat_20x, x = positions_3)[name = tensor<string, []>("op_804")];
            tensor<int32, [1, ?, ?]> delta_3 = sub(x = var_804, y = var_799)[name = tensor<string, []>("delta_3")];
            tensor<int32, []> concat_22_axis_0 = const()[name = tensor<string, []>("concat_22_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_22_interleave_0 = const()[name = tensor<string, []>("concat_22_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_22 = concat(axis = concat_22_axis_0, interleave = concat_22_interleave_0, values = (var_565, gather_14_cast_uint16_to_int32, gather_14_cast_uint16_to_int32))[name = tensor<string, []>("concat_22")];
            tensor<fp16, []> fill_1_value_0_to_fp16 = const()[name = tensor<string, []>("fill_1_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_1_cast_fp16 = fill(shape = concat_22, value = fill_1_value_0_to_fp16)[name = tensor<string, []>("fill_1_cast_fp16")];
            tensor<string, []> attn_bias_9_dtype_0 = const()[name = tensor<string, []>("attn_bias_9_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_810 = greater_equal(x = delta_3, y = var_547)[name = tensor<string, []>("op_810")];
            tensor<bool, [1, ?, ?]> fill_1_cast_fp16_to_bool = cast(dtype = attn_bias_9_dtype_0, x = fill_1_cast_fp16)[name = tensor<string, []>("cast_386")];
            tensor<bool, [1, ?, ?]> attn_bias_11 = logical_and(x = fill_1_cast_fp16_to_bool, y = var_810)[name = tensor<string, []>("attn_bias_11")];
            tensor<bool, [1, ?, ?]> var_812 = less(x = delta_3, y = var_554)[name = tensor<string, []>("op_812")];
            tensor<bool, [1, ?, ?]> attn_bias_13 = logical_and(x = attn_bias_11, y = var_812)[name = tensor<string, []>("attn_bias_13")];
            tensor<bool, [1, ?, ?]> var_814 = logical_and(x = attn_bias_13, y = valid_k_3)[name = tensor<string, []>("op_814")];
            tensor<int32, [1]> var_816_axes_0 = const()[name = tensor<string, []>("op_816_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_816 = expand_dims(axes = var_816_axes_0, x = var_814)[name = tensor<string, []>("op_816")];
            tensor<fp16, []> sub_1_x_0_to_fp16 = const()[name = tensor<string, []>("sub_1_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_57_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_57_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_816_to_fp16 = cast(dtype = cast_57_to_fp16_dtype_0, x = var_816)[name = tensor<string, []>("cast_385")];
            tensor<fp16, [1, 1, ?, ?]> sub_1_cast_fp16 = sub(x = sub_1_x_0_to_fp16, y = var_816_to_fp16)[name = tensor<string, []>("sub_1_cast_fp16")];
            tensor<fp16, []> mul_2_x_0_to_fp16 = const()[name = tensor<string, []>("mul_2_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_2_cast_fp16 = mul(x = mul_2_x_0_to_fp16, y = sub_1_cast_fp16)[name = tensor<string, []>("mul_2_cast_fp16")];
            tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_3_cast_fp16 = mul(x = q_7_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")];
            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = k_7_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_2_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
            tensor<string, []> add_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_1_cast_fp16_to_fp32 = cast(dtype = add_1_cast_fp16_to_fp32_dtype_0, x = add_1_cast_fp16)[name = tensor<string, []>("cast_384")];
            tensor<fp32, [1, 4, ?, ?]> softmax_1 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_1")];
            tensor<bool, []> out_5_transpose_x_0 = const()[name = tensor<string, []>("out_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_5_transpose_y_0 = const()[name = tensor<string, []>("out_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_1_to_fp16 = cast(dtype = softmax_1_to_fp16_dtype_0, x = softmax_1)[name = tensor<string, []>("cast_383")];
            tensor<fp16, [1, 4, ?, 64]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = softmax_1_to_fp16, y = v_3_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<int32, [2]> concat_23x = const()[name = tensor<string, []>("concat_23x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_822 = reshape(shape = concat_23x, x = positions_3)[name = tensor<string, []>("op_822")];
            tensor<bool, [1, ?]> var_825 = less(x = var_822, y = var_687)[name = tensor<string, []>("op_825")];
            tensor<int32, [4]> concat_24x = const()[name = tensor<string, []>("concat_24x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_3 = reshape(shape = concat_24x, x = var_825)[name = tensor<string, []>("valid_q_3")];
            tensor<fp16, []> var_829_to_fp16 = const()[name = tensor<string, []>("op_829_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_7_cast_fp16 = select(a = out_5_cast_fp16, b = var_829_to_fp16, cond = valid_q_3)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<int32, [4]> var_831_perm_0 = const()[name = tensor<string, []>("op_831_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_25x = const()[name = tensor<string, []>("concat_25x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_831_cast_fp16 = transpose(perm = var_831_perm_0, x = out_7_cast_fp16)[name = tensor<string, []>("transpose_32")];
            tensor<fp16, [1, ?, 256]> input_81_cast_fp16 = reshape(shape = concat_25x, x = var_831_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_1_transformer_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3274752)))];
            tensor<fp16, [1, ?, 256]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_1_self_attn_out_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3405888)))];
            tensor<fp16, [1, ?, 256]> var_837_cast_fp16 = mul(x = decoder_1_transformer_layers_1_layer_scale_1_scale_to_fp16, y = linear_6_cast_fp16)[name = tensor<string, []>("op_837_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_83_cast_fp16 = add(x = input_79_cast_fp16, y = var_837_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = tensor<string, []>("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3406464)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407040)))];
            tensor<fp16, [1, ?, 256]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, beta = decoder_1_transformer_layers_1_norm2_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_1_norm2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_1_transformer_layers_1_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407616)))];
            tensor<fp16, [1, ?, 1024]> linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_1_transformer_layers_1_ffn_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<string, []> input_89_mode_0 = const()[name = tensor<string, []>("input_89_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = linear_7_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_1_transformer_layers_1_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3931968)))];
            tensor<fp16, [1, ?, 256]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_1_ffn_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_1_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_1_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4456320)))];
            tensor<fp16, [1, ?, 256]> var_852_cast_fp16 = mul(x = decoder_1_transformer_layers_1_layer_scale_2_scale_to_fp16, y = linear_8_cast_fp16)[name = tensor<string, []>("op_852_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_91_cast_fp16 = add(x = input_83_cast_fp16, y = var_852_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4456896)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4457472)))];
            tensor<fp16, [1, ?, 256]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = decoder_1_transformer_layers_2_norm1_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_2_norm1_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<int32, [3]> var_867_shape_cast_fp16 = shape(x = x_19_cast_fp16)[name = tensor<string, []>("op_867_shape_cast_fp16")];
            tensor<int32, []> gather_24_axis_0 = const()[name = tensor<string, []>("gather_24_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_24_batch_dims_0 = const()[name = tensor<string, []>("gather_24_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_24_validate_indices_0 = const()[name = tensor<string, []>("gather_24_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_867_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_867_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_24_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_24_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_867_shape_cast_fp16_to_uint16 = cast(dtype = var_867_shape_cast_fp16_to_uint16_dtype_0, x = var_867_shape_cast_fp16)[name = tensor<string, []>("cast_382")];
            tensor<uint16, []> gather_24_cast_uint16 = gather(axis = gather_24_axis_0, batch_dims = gather_24_batch_dims_0, indices = gather_24_indices_0_to_uint16, validate_indices = gather_24_validate_indices_0, x = var_867_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_24_cast_uint16")];
            tensor<string, []> gather_24_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_24_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_1_transformer_layers_2_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4458048)))];
            tensor<fp16, [1, ?, 768]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_1_transformer_layers_2_self_attn_in_proj_weight_to_fp16, x = x_19_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<int32, [5]> concat_26x = const()[name = tensor<string, []>("concat_26x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_874_cast_fp16 = reshape(shape = concat_26x, x = linear_9_cast_fp16)[name = tensor<string, []>("op_874_cast_fp16")];
            tensor<int32, [5]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_9_begin_0 = const()[name = tensor<string, []>("q_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_9_end_0 = const()[name = tensor<string, []>("q_9_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_9_end_mask_0 = const()[name = tensor<string, []>("q_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_9_squeeze_mask_0 = const()[name = tensor<string, []>("q_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_11_cast_fp16 = transpose(perm = var_875, x = var_874_cast_fp16)[name = tensor<string, []>("transpose_31")];
            tensor<fp16, [1, 4, ?, 64]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, squeeze_mask = q_9_squeeze_mask_0, x = projected_11_cast_fp16)[name = tensor<string, []>("q_9_cast_fp16")];
            tensor<int32, [5]> k_9_begin_0 = const()[name = tensor<string, []>("k_9_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_9_end_0 = const()[name = tensor<string, []>("k_9_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_9_end_mask_0 = const()[name = tensor<string, []>("k_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_9_squeeze_mask_0 = const()[name = tensor<string, []>("k_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = projected_11_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
            tensor<int32, [5]> v_5_begin_0 = const()[name = tensor<string, []>("v_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_5_end_0 = const()[name = tensor<string, []>("v_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_5_end_mask_0 = const()[name = tensor<string, []>("v_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_5_squeeze_mask_0 = const()[name = tensor<string, []>("v_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = projected_11_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
            tensor<int32, [4]> var_887_shape_cast_fp16 = shape(x = q_9_cast_fp16)[name = tensor<string, []>("op_887_shape_cast_fp16")];
            tensor<int32, []> gather_28_axis_0 = const()[name = tensor<string, []>("gather_28_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_28_batch_dims_0 = const()[name = tensor<string, []>("gather_28_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_28_validate_indices_0 = const()[name = tensor<string, []>("gather_28_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_887_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_887_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_28_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_28_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_887_shape_cast_fp16_to_uint16 = cast(dtype = var_887_shape_cast_fp16_to_uint16_dtype_0, x = var_887_shape_cast_fp16)[name = tensor<string, []>("cast_380")];
            tensor<uint16, []> gather_28_cast_uint16 = gather(axis = gather_28_axis_0, batch_dims = gather_28_batch_dims_0, indices = gather_28_indices_0_to_uint16, validate_indices = gather_28_validate_indices_0, x = var_887_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_28_cast_uint16")];
            tensor<string, []> gather_28_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_28_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_21 = const()[name = tensor<string, []>("const_21"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_22 = const()[name = tensor<string, []>("const_22"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_28_cast_uint16_to_int32 = cast(dtype = gather_28_cast_uint16_to_int32_dtype_0, x = gather_28_cast_uint16)[name = tensor<string, []>("cast_379")];
            tensor<int32, [?]> var_891 = range_1d(end = gather_28_cast_uint16_to_int32, start = const_21, step = const_22)[name = tensor<string, []>("op_891")];
            tensor<fp16, [1, 1]> var_890_to_fp16 = const()[name = tensor<string, []>("op_890_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_891_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_891_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_891_to_fp16 = cast(dtype = var_891_promoted_to_fp16_dtype_0, x = var_891)[name = tensor<string, []>("cast_378")];
            tensor<fp16, [1, ?]> ts_9_cast_fp16 = add(x = var_890_to_fp16, y = var_891_to_fp16)[name = tensor<string, []>("ts_9_cast_fp16")];
            tensor<int32, [4]> var_894 = const()[name = tensor<string, []>("op_894"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_11_cast_fp16 = reshape(shape = var_894, x = ts_9_cast_fp16)[name = tensor<string, []>("ts_11_cast_fp16")];
            tensor<int32, [5]> concat_27x = const()[name = tensor<string, []>("concat_27x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_5_cast_fp16 = reshape(shape = concat_27x, x = q_9_cast_fp16)[name = tensor<string, []>("q2_5_cast_fp16")];
            tensor<int32, [5]> concat_28x = const()[name = tensor<string, []>("concat_28x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_5_cast_fp16 = reshape(shape = concat_28x, x = k_9_cast_fp16)[name = tensor<string, []>("k2_5_cast_fp16")];
            tensor<int32, [5]> qr_5_begin_0 = const()[name = tensor<string, []>("qr_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_5_end_0 = const()[name = tensor<string, []>("qr_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_5_end_mask_0 = const()[name = tensor<string, []>("qr_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_5_squeeze_mask_0 = const()[name = tensor<string, []>("qr_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_5_cast_fp16 = slice_by_index(begin = qr_5_begin_0, end = qr_5_end_0, end_mask = qr_5_end_mask_0, squeeze_mask = qr_5_squeeze_mask_0, x = q2_5_cast_fp16)[name = tensor<string, []>("qr_5_cast_fp16")];
            tensor<int32, [5]> qi_5_begin_0 = const()[name = tensor<string, []>("qi_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_5_end_0 = const()[name = tensor<string, []>("qi_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_5_end_mask_0 = const()[name = tensor<string, []>("qi_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_5_squeeze_mask_0 = const()[name = tensor<string, []>("qi_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_5_cast_fp16 = slice_by_index(begin = qi_5_begin_0, end = qi_5_end_0, end_mask = qi_5_end_mask_0, squeeze_mask = qi_5_squeeze_mask_0, x = q2_5_cast_fp16)[name = tensor<string, []>("qi_5_cast_fp16")];
            tensor<int32, [5]> kr_5_begin_0 = const()[name = tensor<string, []>("kr_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_5_end_0 = const()[name = tensor<string, []>("kr_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_5_end_mask_0 = const()[name = tensor<string, []>("kr_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_5_squeeze_mask_0 = const()[name = tensor<string, []>("kr_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_5_cast_fp16 = slice_by_index(begin = kr_5_begin_0, end = kr_5_end_0, end_mask = kr_5_end_mask_0, squeeze_mask = kr_5_squeeze_mask_0, x = k2_5_cast_fp16)[name = tensor<string, []>("kr_5_cast_fp16")];
            tensor<int32, [5]> ki_5_begin_0 = const()[name = tensor<string, []>("ki_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_5_end_0 = const()[name = tensor<string, []>("ki_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_5_end_mask_0 = const()[name = tensor<string, []>("ki_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_5_squeeze_mask_0 = const()[name = tensor<string, []>("ki_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_5_cast_fp16 = slice_by_index(begin = ki_5_begin_0, end = ki_5_end_0, end_mask = ki_5_end_mask_0, squeeze_mask = ki_5_squeeze_mask_0, x = k2_5_cast_fp16)[name = tensor<string, []>("ki_5_cast_fp16")];
            tensor<fp16, [32]> freqs_5_to_fp16 = const()[name = tensor<string, []>("freqs_5_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4851328)))];
            tensor<fp16, [1, 1, ?, 32]> var_907_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_11_cast_fp16)[name = tensor<string, []>("op_907_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_5_cast_fp16 = cos(x = var_907_cast_fp16)[name = tensor<string, []>("rotr_5_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_5_cast_fp16 = sin(x = var_907_cast_fp16)[name = tensor<string, []>("roti_5_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_911_cast_fp16 = mul(x = qr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor<string, []>("op_911_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_912_cast_fp16 = mul(x = qi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor<string, []>("op_912_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_913_cast_fp16 = sub(x = var_911_cast_fp16, y = var_912_cast_fp16)[name = tensor<string, []>("op_913_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_914_cast_fp16 = mul(x = qr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor<string, []>("op_914_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_915_cast_fp16 = mul(x = qi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor<string, []>("op_915_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_916_cast_fp16 = add(x = var_914_cast_fp16, y = var_915_cast_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
            tensor<int32, []> qo_5_axis_0 = const()[name = tensor<string, []>("qo_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (var_913_cast_fp16, var_916_cast_fp16))[name = tensor<string, []>("qo_5_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_919_cast_fp16 = mul(x = kr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor<string, []>("op_919_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_920_cast_fp16 = mul(x = ki_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor<string, []>("op_920_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_921_cast_fp16 = sub(x = var_919_cast_fp16, y = var_920_cast_fp16)[name = tensor<string, []>("op_921_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_922_cast_fp16 = mul(x = kr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor<string, []>("op_922_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_923_cast_fp16 = mul(x = ki_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor<string, []>("op_923_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_924_cast_fp16 = add(x = var_922_cast_fp16, y = var_923_cast_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
            tensor<int32, []> ko_5_axis_0 = const()[name = tensor<string, []>("ko_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (var_921_cast_fp16, var_924_cast_fp16))[name = tensor<string, []>("ko_5_cast_fp16")];
            tensor<int32, [4]> concat_29x = const()[name = tensor<string, []>("concat_29x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_11_cast_fp16 = reshape(shape = concat_29x, x = qo_5_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")];
            tensor<int32, [4]> concat_30x = const()[name = tensor<string, []>("concat_30x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_11_cast_fp16 = reshape(shape = concat_30x, x = ko_5_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")];
            tensor<int32, []> const_23 = const()[name = tensor<string, []>("const_23"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_24 = const()[name = tensor<string, []>("const_24"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_24_cast_uint16_to_int32 = cast(dtype = gather_24_cast_uint16_to_int32_dtype_0, x = gather_24_cast_uint16)[name = tensor<string, []>("cast_381")];
            tensor<int32, [?]> positions_5 = range_1d(end = gather_24_cast_uint16_to_int32, start = const_23, step = const_24)[name = tensor<string, []>("positions_5")];
            tensor<int32, [3]> concat_31x = const()[name = tensor<string, []>("concat_31x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_936 = reshape(shape = concat_31x, x = positions_5)[name = tensor<string, []>("op_936")];
            tensor<bool, [1, 1, ?]> valid_k_5 = less(x = var_936, y = var_664)[name = tensor<string, []>("valid_k_5")];
            tensor<int32, [3]> concat_32x = const()[name = tensor<string, []>("concat_32x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_941 = reshape(shape = concat_32x, x = positions_5)[name = tensor<string, []>("op_941")];
            tensor<int32, [1, ?, ?]> delta_5 = sub(x = var_941, y = var_936)[name = tensor<string, []>("delta_5")];
            tensor<int32, []> concat_34_axis_0 = const()[name = tensor<string, []>("concat_34_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_34_interleave_0 = const()[name = tensor<string, []>("concat_34_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_34 = concat(axis = concat_34_axis_0, interleave = concat_34_interleave_0, values = (var_565, gather_24_cast_uint16_to_int32, gather_24_cast_uint16_to_int32))[name = tensor<string, []>("concat_34")];
            tensor<fp16, []> fill_2_value_0_to_fp16 = const()[name = tensor<string, []>("fill_2_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_2_cast_fp16 = fill(shape = concat_34, value = fill_2_value_0_to_fp16)[name = tensor<string, []>("fill_2_cast_fp16")];
            tensor<string, []> attn_bias_17_dtype_0 = const()[name = tensor<string, []>("attn_bias_17_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_947 = greater_equal(x = delta_5, y = var_547)[name = tensor<string, []>("op_947")];
            tensor<bool, [1, ?, ?]> fill_2_cast_fp16_to_bool = cast(dtype = attn_bias_17_dtype_0, x = fill_2_cast_fp16)[name = tensor<string, []>("cast_377")];
            tensor<bool, [1, ?, ?]> attn_bias_19 = logical_and(x = fill_2_cast_fp16_to_bool, y = var_947)[name = tensor<string, []>("attn_bias_19")];
            tensor<bool, [1, ?, ?]> var_949 = less(x = delta_5, y = var_554)[name = tensor<string, []>("op_949")];
            tensor<bool, [1, ?, ?]> attn_bias_21 = logical_and(x = attn_bias_19, y = var_949)[name = tensor<string, []>("attn_bias_21")];
            tensor<bool, [1, ?, ?]> var_951 = logical_and(x = attn_bias_21, y = valid_k_5)[name = tensor<string, []>("op_951")];
            tensor<int32, [1]> var_953_axes_0 = const()[name = tensor<string, []>("op_953_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_953 = expand_dims(axes = var_953_axes_0, x = var_951)[name = tensor<string, []>("op_953")];
            tensor<fp16, []> sub_2_x_0_to_fp16 = const()[name = tensor<string, []>("sub_2_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_79_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_79_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_953_to_fp16 = cast(dtype = cast_79_to_fp16_dtype_0, x = var_953)[name = tensor<string, []>("cast_376")];
            tensor<fp16, [1, 1, ?, ?]> sub_2_cast_fp16 = sub(x = sub_2_x_0_to_fp16, y = var_953_to_fp16)[name = tensor<string, []>("sub_2_cast_fp16")];
            tensor<fp16, []> mul_4_x_0_to_fp16 = const()[name = tensor<string, []>("mul_4_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_4_cast_fp16 = mul(x = mul_4_x_0_to_fp16, y = sub_2_cast_fp16)[name = tensor<string, []>("mul_4_cast_fp16")];
            tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_5_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")];
            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = k_11_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_4_cast_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
            tensor<string, []> add_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_2_cast_fp16_to_fp32 = cast(dtype = add_2_cast_fp16_to_fp32_dtype_0, x = add_2_cast_fp16)[name = tensor<string, []>("cast_375")];
            tensor<fp32, [1, 4, ?, ?]> softmax_2 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_2")];
            tensor<bool, []> out_9_transpose_x_0 = const()[name = tensor<string, []>("out_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_9_transpose_y_0 = const()[name = tensor<string, []>("out_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_2_to_fp16 = cast(dtype = softmax_2_to_fp16_dtype_0, x = softmax_2)[name = tensor<string, []>("cast_374")];
            tensor<fp16, [1, 4, ?, 64]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = softmax_2_to_fp16, y = v_5_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<int32, [2]> concat_35x = const()[name = tensor<string, []>("concat_35x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_959 = reshape(shape = concat_35x, x = positions_5)[name = tensor<string, []>("op_959")];
            tensor<bool, [1, ?]> var_962 = less(x = var_959, y = var_687)[name = tensor<string, []>("op_962")];
            tensor<int32, [4]> concat_36x = const()[name = tensor<string, []>("concat_36x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_5 = reshape(shape = concat_36x, x = var_962)[name = tensor<string, []>("valid_q_5")];
            tensor<fp16, []> var_966_to_fp16 = const()[name = tensor<string, []>("op_966_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_11_cast_fp16 = select(a = out_9_cast_fp16, b = var_966_to_fp16, cond = valid_q_5)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<int32, [4]> var_968_perm_0 = const()[name = tensor<string, []>("op_968_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_37x = const()[name = tensor<string, []>("concat_37x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_968_cast_fp16 = transpose(perm = var_968_perm_0, x = out_11_cast_fp16)[name = tensor<string, []>("transpose_30")];
            tensor<fp16, [1, ?, 256]> input_93_cast_fp16 = reshape(shape = concat_37x, x = var_968_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_1_transformer_layers_2_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4851456)))];
            tensor<fp16, [1, ?, 256]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_2_self_attn_out_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_2_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4982592)))];
            tensor<fp16, [1, ?, 256]> var_974_cast_fp16 = mul(x = decoder_1_transformer_layers_2_layer_scale_1_scale_to_fp16, y = linear_10_cast_fp16)[name = tensor<string, []>("op_974_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_95_cast_fp16 = add(x = input_91_cast_fp16, y = var_974_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, [256]> decoder_1_transformer_layers_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4983168)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4983744)))];
            tensor<fp16, [1, ?, 256]> input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = decoder_1_transformer_layers_2_norm2_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_2_norm2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_1_transformer_layers_2_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4984320)))];
            tensor<fp16, [1, ?, 1024]> linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_1_transformer_layers_2_ffn_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<string, []> input_101_mode_0 = const()[name = tensor<string, []>("input_101_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_11_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_1_transformer_layers_2_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5508672)))];
            tensor<fp16, [1, ?, 256]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_2_ffn_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_2_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_2_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6033024)))];
            tensor<fp16, [1, ?, 256]> var_989_cast_fp16 = mul(x = decoder_1_transformer_layers_2_layer_scale_2_scale_to_fp16, y = linear_12_cast_fp16)[name = tensor<string, []>("op_989_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_103_cast_fp16 = add(x = input_95_cast_fp16, y = var_989_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6033600)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6034176)))];
            tensor<fp16, [1, ?, 256]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = decoder_1_transformer_layers_3_norm1_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_3_norm1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<int32, [3]> var_1004_shape_cast_fp16 = shape(x = x_25_cast_fp16)[name = tensor<string, []>("op_1004_shape_cast_fp16")];
            tensor<int32, []> gather_34_axis_0 = const()[name = tensor<string, []>("gather_34_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_34_batch_dims_0 = const()[name = tensor<string, []>("gather_34_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_34_validate_indices_0 = const()[name = tensor<string, []>("gather_34_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1004_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1004_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_34_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_34_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1004_shape_cast_fp16_to_uint16 = cast(dtype = var_1004_shape_cast_fp16_to_uint16_dtype_0, x = var_1004_shape_cast_fp16)[name = tensor<string, []>("cast_373")];
            tensor<uint16, []> gather_34_cast_uint16 = gather(axis = gather_34_axis_0, batch_dims = gather_34_batch_dims_0, indices = gather_34_indices_0_to_uint16, validate_indices = gather_34_validate_indices_0, x = var_1004_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_34_cast_uint16")];
            tensor<string, []> gather_34_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_34_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_1_transformer_layers_3_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6034752)))];
            tensor<fp16, [1, ?, 768]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_1_transformer_layers_3_self_attn_in_proj_weight_to_fp16, x = x_25_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<int32, [5]> concat_38x = const()[name = tensor<string, []>("concat_38x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1011_cast_fp16 = reshape(shape = concat_38x, x = linear_13_cast_fp16)[name = tensor<string, []>("op_1011_cast_fp16")];
            tensor<int32, [5]> var_1012 = const()[name = tensor<string, []>("op_1012"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_15_cast_fp16 = transpose(perm = var_1012, x = var_1011_cast_fp16)[name = tensor<string, []>("transpose_29")];
            tensor<fp16, [1, 4, ?, 64]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = projected_15_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
            tensor<int32, [5]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = projected_15_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
            tensor<int32, [5]> v_7_begin_0 = const()[name = tensor<string, []>("v_7_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_7_end_0 = const()[name = tensor<string, []>("v_7_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_7_end_mask_0 = const()[name = tensor<string, []>("v_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = projected_15_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
            tensor<int32, [4]> var_1024_shape_cast_fp16 = shape(x = q_13_cast_fp16)[name = tensor<string, []>("op_1024_shape_cast_fp16")];
            tensor<int32, []> gather_38_axis_0 = const()[name = tensor<string, []>("gather_38_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_38_batch_dims_0 = const()[name = tensor<string, []>("gather_38_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_38_validate_indices_0 = const()[name = tensor<string, []>("gather_38_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1024_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1024_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_38_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_38_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1024_shape_cast_fp16_to_uint16 = cast(dtype = var_1024_shape_cast_fp16_to_uint16_dtype_0, x = var_1024_shape_cast_fp16)[name = tensor<string, []>("cast_371")];
            tensor<uint16, []> gather_38_cast_uint16 = gather(axis = gather_38_axis_0, batch_dims = gather_38_batch_dims_0, indices = gather_38_indices_0_to_uint16, validate_indices = gather_38_validate_indices_0, x = var_1024_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_38_cast_uint16")];
            tensor<string, []> gather_38_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_38_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_30 = const()[name = tensor<string, []>("const_30"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_31 = const()[name = tensor<string, []>("const_31"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_38_cast_uint16_to_int32 = cast(dtype = gather_38_cast_uint16_to_int32_dtype_0, x = gather_38_cast_uint16)[name = tensor<string, []>("cast_370")];
            tensor<int32, [?]> var_1028 = range_1d(end = gather_38_cast_uint16_to_int32, start = const_30, step = const_31)[name = tensor<string, []>("op_1028")];
            tensor<fp16, [1, 1]> var_1027_to_fp16 = const()[name = tensor<string, []>("op_1027_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1028_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1028_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1028_to_fp16 = cast(dtype = var_1028_promoted_to_fp16_dtype_0, x = var_1028)[name = tensor<string, []>("cast_369")];
            tensor<fp16, [1, ?]> ts_13_cast_fp16 = add(x = var_1027_to_fp16, y = var_1028_to_fp16)[name = tensor<string, []>("ts_13_cast_fp16")];
            tensor<int32, [4]> var_1031 = const()[name = tensor<string, []>("op_1031"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_15_cast_fp16 = reshape(shape = var_1031, x = ts_13_cast_fp16)[name = tensor<string, []>("ts_15_cast_fp16")];
            tensor<int32, [5]> concat_39x = const()[name = tensor<string, []>("concat_39x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_7_cast_fp16 = reshape(shape = concat_39x, x = q_13_cast_fp16)[name = tensor<string, []>("q2_7_cast_fp16")];
            tensor<int32, [5]> concat_40x = const()[name = tensor<string, []>("concat_40x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_7_cast_fp16 = reshape(shape = concat_40x, x = k_13_cast_fp16)[name = tensor<string, []>("k2_7_cast_fp16")];
            tensor<int32, [5]> qr_7_begin_0 = const()[name = tensor<string, []>("qr_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_7_end_0 = const()[name = tensor<string, []>("qr_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_7_end_mask_0 = const()[name = tensor<string, []>("qr_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_7_squeeze_mask_0 = const()[name = tensor<string, []>("qr_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_7_cast_fp16 = slice_by_index(begin = qr_7_begin_0, end = qr_7_end_0, end_mask = qr_7_end_mask_0, squeeze_mask = qr_7_squeeze_mask_0, x = q2_7_cast_fp16)[name = tensor<string, []>("qr_7_cast_fp16")];
            tensor<int32, [5]> qi_7_begin_0 = const()[name = tensor<string, []>("qi_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_7_end_0 = const()[name = tensor<string, []>("qi_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_7_end_mask_0 = const()[name = tensor<string, []>("qi_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_7_squeeze_mask_0 = const()[name = tensor<string, []>("qi_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_7_cast_fp16 = slice_by_index(begin = qi_7_begin_0, end = qi_7_end_0, end_mask = qi_7_end_mask_0, squeeze_mask = qi_7_squeeze_mask_0, x = q2_7_cast_fp16)[name = tensor<string, []>("qi_7_cast_fp16")];
            tensor<int32, [5]> kr_7_begin_0 = const()[name = tensor<string, []>("kr_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_7_end_0 = const()[name = tensor<string, []>("kr_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_7_end_mask_0 = const()[name = tensor<string, []>("kr_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_7_squeeze_mask_0 = const()[name = tensor<string, []>("kr_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_7_cast_fp16 = slice_by_index(begin = kr_7_begin_0, end = kr_7_end_0, end_mask = kr_7_end_mask_0, squeeze_mask = kr_7_squeeze_mask_0, x = k2_7_cast_fp16)[name = tensor<string, []>("kr_7_cast_fp16")];
            tensor<int32, [5]> ki_7_begin_0 = const()[name = tensor<string, []>("ki_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_7_end_0 = const()[name = tensor<string, []>("ki_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_7_end_mask_0 = const()[name = tensor<string, []>("ki_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_7_squeeze_mask_0 = const()[name = tensor<string, []>("ki_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_7_cast_fp16 = slice_by_index(begin = ki_7_begin_0, end = ki_7_end_0, end_mask = ki_7_end_mask_0, squeeze_mask = ki_7_squeeze_mask_0, x = k2_7_cast_fp16)[name = tensor<string, []>("ki_7_cast_fp16")];
            tensor<fp16, [32]> freqs_7_to_fp16 = const()[name = tensor<string, []>("freqs_7_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6428032)))];
            tensor<fp16, [1, 1, ?, 32]> var_1044_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_15_cast_fp16)[name = tensor<string, []>("op_1044_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_7_cast_fp16 = cos(x = var_1044_cast_fp16)[name = tensor<string, []>("rotr_7_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_7_cast_fp16 = sin(x = var_1044_cast_fp16)[name = tensor<string, []>("roti_7_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1048_cast_fp16 = mul(x = qr_7_cast_fp16, y = rotr_7_cast_fp16)[name = tensor<string, []>("op_1048_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1049_cast_fp16 = mul(x = qi_7_cast_fp16, y = roti_7_cast_fp16)[name = tensor<string, []>("op_1049_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1050_cast_fp16 = sub(x = var_1048_cast_fp16, y = var_1049_cast_fp16)[name = tensor<string, []>("op_1050_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1051_cast_fp16 = mul(x = qr_7_cast_fp16, y = roti_7_cast_fp16)[name = tensor<string, []>("op_1051_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1052_cast_fp16 = mul(x = qi_7_cast_fp16, y = rotr_7_cast_fp16)[name = tensor<string, []>("op_1052_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1053_cast_fp16 = add(x = var_1051_cast_fp16, y = var_1052_cast_fp16)[name = tensor<string, []>("op_1053_cast_fp16")];
            tensor<int32, []> qo_7_axis_0 = const()[name = tensor<string, []>("qo_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (var_1050_cast_fp16, var_1053_cast_fp16))[name = tensor<string, []>("qo_7_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1056_cast_fp16 = mul(x = kr_7_cast_fp16, y = rotr_7_cast_fp16)[name = tensor<string, []>("op_1056_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1057_cast_fp16 = mul(x = ki_7_cast_fp16, y = roti_7_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1058_cast_fp16 = sub(x = var_1056_cast_fp16, y = var_1057_cast_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1059_cast_fp16 = mul(x = kr_7_cast_fp16, y = roti_7_cast_fp16)[name = tensor<string, []>("op_1059_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1060_cast_fp16 = mul(x = ki_7_cast_fp16, y = rotr_7_cast_fp16)[name = tensor<string, []>("op_1060_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1061_cast_fp16 = add(x = var_1059_cast_fp16, y = var_1060_cast_fp16)[name = tensor<string, []>("op_1061_cast_fp16")];
            tensor<int32, []> ko_7_axis_0 = const()[name = tensor<string, []>("ko_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (var_1058_cast_fp16, var_1061_cast_fp16))[name = tensor<string, []>("ko_7_cast_fp16")];
            tensor<int32, [4]> concat_41x = const()[name = tensor<string, []>("concat_41x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_15_cast_fp16 = reshape(shape = concat_41x, x = qo_7_cast_fp16)[name = tensor<string, []>("q_15_cast_fp16")];
            tensor<int32, [4]> concat_42x = const()[name = tensor<string, []>("concat_42x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_15_cast_fp16 = reshape(shape = concat_42x, x = ko_7_cast_fp16)[name = tensor<string, []>("k_15_cast_fp16")];
            tensor<int32, []> const_32 = const()[name = tensor<string, []>("const_32"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_33 = const()[name = tensor<string, []>("const_33"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_34_cast_uint16_to_int32 = cast(dtype = gather_34_cast_uint16_to_int32_dtype_0, x = gather_34_cast_uint16)[name = tensor<string, []>("cast_372")];
            tensor<int32, [?]> positions_7 = range_1d(end = gather_34_cast_uint16_to_int32, start = const_32, step = const_33)[name = tensor<string, []>("positions_7")];
            tensor<int32, [3]> concat_43x = const()[name = tensor<string, []>("concat_43x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1073 = reshape(shape = concat_43x, x = positions_7)[name = tensor<string, []>("op_1073")];
            tensor<bool, [1, 1, ?]> valid_k_7 = less(x = var_1073, y = var_664)[name = tensor<string, []>("valid_k_7")];
            tensor<int32, [3]> concat_44x = const()[name = tensor<string, []>("concat_44x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1078 = reshape(shape = concat_44x, x = positions_7)[name = tensor<string, []>("op_1078")];
            tensor<int32, [1, ?, ?]> delta_7 = sub(x = var_1078, y = var_1073)[name = tensor<string, []>("delta_7")];
            tensor<int32, []> concat_46_axis_0 = const()[name = tensor<string, []>("concat_46_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_46_interleave_0 = const()[name = tensor<string, []>("concat_46_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_46 = concat(axis = concat_46_axis_0, interleave = concat_46_interleave_0, values = (var_565, gather_34_cast_uint16_to_int32, gather_34_cast_uint16_to_int32))[name = tensor<string, []>("concat_46")];
            tensor<fp16, []> fill_3_value_0_to_fp16 = const()[name = tensor<string, []>("fill_3_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_3_cast_fp16 = fill(shape = concat_46, value = fill_3_value_0_to_fp16)[name = tensor<string, []>("fill_3_cast_fp16")];
            tensor<string, []> attn_bias_25_dtype_0 = const()[name = tensor<string, []>("attn_bias_25_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1084 = greater_equal(x = delta_7, y = var_547)[name = tensor<string, []>("op_1084")];
            tensor<bool, [1, ?, ?]> fill_3_cast_fp16_to_bool = cast(dtype = attn_bias_25_dtype_0, x = fill_3_cast_fp16)[name = tensor<string, []>("cast_368")];
            tensor<bool, [1, ?, ?]> attn_bias_27 = logical_and(x = fill_3_cast_fp16_to_bool, y = var_1084)[name = tensor<string, []>("attn_bias_27")];
            tensor<bool, [1, ?, ?]> var_1086 = less(x = delta_7, y = var_554)[name = tensor<string, []>("op_1086")];
            tensor<bool, [1, ?, ?]> attn_bias_29 = logical_and(x = attn_bias_27, y = var_1086)[name = tensor<string, []>("attn_bias_29")];
            tensor<bool, [1, ?, ?]> var_1088 = logical_and(x = attn_bias_29, y = valid_k_7)[name = tensor<string, []>("op_1088")];
            tensor<int32, [1]> var_1090_axes_0 = const()[name = tensor<string, []>("op_1090_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1090 = expand_dims(axes = var_1090_axes_0, x = var_1088)[name = tensor<string, []>("op_1090")];
            tensor<fp16, []> sub_3_x_0_to_fp16 = const()[name = tensor<string, []>("sub_3_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_101_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_101_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1090_to_fp16 = cast(dtype = cast_101_to_fp16_dtype_0, x = var_1090)[name = tensor<string, []>("cast_367")];
            tensor<fp16, [1, 1, ?, ?]> sub_3_cast_fp16 = sub(x = sub_3_x_0_to_fp16, y = var_1090_to_fp16)[name = tensor<string, []>("sub_3_cast_fp16")];
            tensor<fp16, []> mul_6_x_0_to_fp16 = const()[name = tensor<string, []>("mul_6_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_6_cast_fp16 = mul(x = mul_6_x_0_to_fp16, y = sub_3_cast_fp16)[name = tensor<string, []>("mul_6_cast_fp16")];
            tensor<fp16, []> mul_7_y_0_to_fp16 = const()[name = tensor<string, []>("mul_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_7_cast_fp16 = mul(x = q_15_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")];
            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = k_15_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_6_cast_fp16)[name = tensor<string, []>("add_3_cast_fp16")];
            tensor<string, []> add_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_3_cast_fp16_to_fp32 = cast(dtype = add_3_cast_fp16_to_fp32_dtype_0, x = add_3_cast_fp16)[name = tensor<string, []>("cast_366")];
            tensor<fp32, [1, 4, ?, ?]> softmax_3 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_3")];
            tensor<bool, []> out_13_transpose_x_0 = const()[name = tensor<string, []>("out_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_13_transpose_y_0 = const()[name = tensor<string, []>("out_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_3_to_fp16 = cast(dtype = softmax_3_to_fp16_dtype_0, x = softmax_3)[name = tensor<string, []>("cast_365")];
            tensor<fp16, [1, 4, ?, 64]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = softmax_3_to_fp16, y = v_7_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<int32, [2]> concat_47x = const()[name = tensor<string, []>("concat_47x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1096 = reshape(shape = concat_47x, x = positions_7)[name = tensor<string, []>("op_1096")];
            tensor<bool, [1, ?]> var_1099 = less(x = var_1096, y = var_687)[name = tensor<string, []>("op_1099")];
            tensor<int32, [4]> concat_48x = const()[name = tensor<string, []>("concat_48x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_7 = reshape(shape = concat_48x, x = var_1099)[name = tensor<string, []>("valid_q_7")];
            tensor<fp16, []> var_1103_to_fp16 = const()[name = tensor<string, []>("op_1103_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_15_cast_fp16 = select(a = out_13_cast_fp16, b = var_1103_to_fp16, cond = valid_q_7)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<int32, [4]> var_1105_perm_0 = const()[name = tensor<string, []>("op_1105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_49x = const()[name = tensor<string, []>("concat_49x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1105_cast_fp16 = transpose(perm = var_1105_perm_0, x = out_15_cast_fp16)[name = tensor<string, []>("transpose_28")];
            tensor<fp16, [1, ?, 256]> input_105_cast_fp16 = reshape(shape = concat_49x, x = var_1105_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_1_transformer_layers_3_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6428160)))];
            tensor<fp16, [1, ?, 256]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_3_self_attn_out_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6559296)))];
            tensor<fp16, [1, ?, 256]> var_1111_cast_fp16 = mul(x = decoder_1_transformer_layers_3_layer_scale_1_scale_to_fp16, y = linear_14_cast_fp16)[name = tensor<string, []>("op_1111_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_107_cast_fp16 = add(x = input_103_cast_fp16, y = var_1111_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<int32, [1]> input_109_axes_0 = const()[name = tensor<string, []>("input_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6559872)))];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6560448)))];
            tensor<fp16, [1, ?, 256]> input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = decoder_1_transformer_layers_3_norm2_bias_to_fp16, epsilon = var_561_to_fp16, gamma = decoder_1_transformer_layers_3_norm2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_1_transformer_layers_3_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6561024)))];
            tensor<fp16, [1, ?, 1024]> linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_1_transformer_layers_3_ffn_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<string, []> input_113_mode_0 = const()[name = tensor<string, []>("input_113_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_113_cast_fp16 = gelu(mode = input_113_mode_0, x = linear_15_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_1_transformer_layers_3_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7085376)))];
            tensor<fp16, [1, ?, 256]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_1_transformer_layers_3_ffn_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [256]> decoder_1_transformer_layers_3_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_1_transformer_layers_3_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7609728)))];
            tensor<fp16, [1, ?, 256]> var_1126_cast_fp16 = mul(x = decoder_1_transformer_layers_3_layer_scale_2_scale_to_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("op_1126_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_115_cast_fp16 = add(x = input_107_cast_fp16, y = var_1126_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<fp16, [768, 256]> decoder_1_output_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_1_output_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7610304)))];
            tensor<fp16, [1, ?, 768]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_1_output_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<int32, [3]> x_31_perm_0 = const()[name = tensor<string, []>("x_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [4]> concat_50x = const()[name = tensor<string, []>("concat_50x"), val = tensor<int32, [4]>([1, 384, 2, -1])];
            tensor<fp16, [1, 768, ?]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = linear_17_cast_fp16)[name = tensor<string, []>("transpose_27")];
            tensor<fp16, [1, 384, 2, ?]> var_1145_cast_fp16 = reshape(shape = concat_50x, x = x_31_cast_fp16)[name = tensor<string, []>("op_1145_cast_fp16")];
            tensor<int32, [4]> var_1146 = const()[name = tensor<string, []>("op_1146"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> concat_51x = const()[name = tensor<string, []>("concat_51x"), val = tensor<int32, [3]>([1, 384, -1])];
            tensor<fp16, [1, 384, ?, 2]> var_1147_cast_fp16 = transpose(perm = var_1146, x = var_1145_cast_fp16)[name = tensor<string, []>("transpose_26")];
            tensor<fp16, [1, 384, ?]> x_33_cast_fp16 = reshape(shape = concat_51x, x = var_1147_cast_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
            tensor<int32, []> var_1153 = const()[name = tensor<string, []>("op_1153"), val = tensor<int32, []>(2)];
            tensor<int32, [1]> input_lengths_5 = mul(x = input_lengths_3, y = var_1153)[name = tensor<string, []>("input_lengths_5")];
            tensor<int32, []> var_1159 = const()[name = tensor<string, []>("op_1159"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_1166 = const()[name = tensor<string, []>("op_1166"), val = tensor<int32, []>(800)];
            tensor<int32, []> var_1177 = const()[name = tensor<string, []>("op_1177"), val = tensor<int32, []>(1)];
            tensor<int32, [3]> input_117_perm_0 = const()[name = tensor<string, []>("input_117_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [256, 384]> decoder_3_input_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_input_proj_weight_to_fp16"), val = tensor<fp16, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8003584)))];
            tensor<fp16, [1, ?, 384]> input_117_cast_fp16 = transpose(perm = input_117_perm_0, x = x_33_cast_fp16)[name = tensor<string, []>("transpose_25")];
            tensor<fp16, [1, ?, 256]> linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_3_input_proj_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<int32, [1]> x_37_axes_0 = const()[name = tensor<string, []>("x_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8200256)))];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8200832)))];
            tensor<fp16, []> var_1173_to_fp16 = const()[name = tensor<string, []>("op_1173_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, ?, 256]> x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = decoder_3_transformer_layers_0_norm1_bias_to_fp16, epsilon = var_1173_to_fp16, gamma = decoder_3_transformer_layers_0_norm1_weight_to_fp16, x = linear_18_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<int32, [3]> var_1201_shape_cast_fp16 = shape(x = x_37_cast_fp16)[name = tensor<string, []>("op_1201_shape_cast_fp16")];
            tensor<int32, []> gather_47_axis_0 = const()[name = tensor<string, []>("gather_47_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_47_batch_dims_0 = const()[name = tensor<string, []>("gather_47_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_47_validate_indices_0 = const()[name = tensor<string, []>("gather_47_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1201_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1201_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_47_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_47_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1201_shape_cast_fp16_to_uint16 = cast(dtype = var_1201_shape_cast_fp16_to_uint16_dtype_0, x = var_1201_shape_cast_fp16)[name = tensor<string, []>("cast_364")];
            tensor<uint16, []> gather_47_cast_uint16 = gather(axis = gather_47_axis_0, batch_dims = gather_47_batch_dims_0, indices = gather_47_indices_0_to_uint16, validate_indices = gather_47_validate_indices_0, x = var_1201_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_47_cast_uint16")];
            tensor<string, []> gather_47_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_47_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_3_transformer_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8201408)))];
            tensor<fp16, [1, ?, 768]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_3_transformer_layers_0_self_attn_in_proj_weight_to_fp16, x = x_37_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<int32, [5]> concat_52x = const()[name = tensor<string, []>("concat_52x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1208_cast_fp16 = reshape(shape = concat_52x, x = linear_19_cast_fp16)[name = tensor<string, []>("op_1208_cast_fp16")];
            tensor<int32, [5]> var_1209 = const()[name = tensor<string, []>("op_1209"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_17_begin_0 = const()[name = tensor<string, []>("q_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_17_end_0 = const()[name = tensor<string, []>("q_17_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_17_end_mask_0 = const()[name = tensor<string, []>("q_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_17_squeeze_mask_0 = const()[name = tensor<string, []>("q_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_19_cast_fp16 = transpose(perm = var_1209, x = var_1208_cast_fp16)[name = tensor<string, []>("transpose_24")];
            tensor<fp16, [1, 4, ?, 64]> q_17_cast_fp16 = slice_by_index(begin = q_17_begin_0, end = q_17_end_0, end_mask = q_17_end_mask_0, squeeze_mask = q_17_squeeze_mask_0, x = projected_19_cast_fp16)[name = tensor<string, []>("q_17_cast_fp16")];
            tensor<int32, [5]> k_17_begin_0 = const()[name = tensor<string, []>("k_17_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_17_end_0 = const()[name = tensor<string, []>("k_17_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_17_end_mask_0 = const()[name = tensor<string, []>("k_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_17_squeeze_mask_0 = const()[name = tensor<string, []>("k_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = projected_19_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
            tensor<int32, [5]> v_9_begin_0 = const()[name = tensor<string, []>("v_9_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_9_end_0 = const()[name = tensor<string, []>("v_9_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_9_end_mask_0 = const()[name = tensor<string, []>("v_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_9_squeeze_mask_0 = const()[name = tensor<string, []>("v_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = projected_19_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
            tensor<int32, [4]> var_1221_shape_cast_fp16 = shape(x = q_17_cast_fp16)[name = tensor<string, []>("op_1221_shape_cast_fp16")];
            tensor<int32, []> gather_51_axis_0 = const()[name = tensor<string, []>("gather_51_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_51_batch_dims_0 = const()[name = tensor<string, []>("gather_51_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_51_validate_indices_0 = const()[name = tensor<string, []>("gather_51_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1221_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1221_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_51_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_51_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1221_shape_cast_fp16_to_uint16 = cast(dtype = var_1221_shape_cast_fp16_to_uint16_dtype_0, x = var_1221_shape_cast_fp16)[name = tensor<string, []>("cast_362")];
            tensor<uint16, []> gather_51_cast_uint16 = gather(axis = gather_51_axis_0, batch_dims = gather_51_batch_dims_0, indices = gather_51_indices_0_to_uint16, validate_indices = gather_51_validate_indices_0, x = var_1221_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_51_cast_uint16")];
            tensor<string, []> gather_51_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_51_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_39 = const()[name = tensor<string, []>("const_39"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_40 = const()[name = tensor<string, []>("const_40"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_51_cast_uint16_to_int32 = cast(dtype = gather_51_cast_uint16_to_int32_dtype_0, x = gather_51_cast_uint16)[name = tensor<string, []>("cast_361")];
            tensor<int32, [?]> var_1225 = range_1d(end = gather_51_cast_uint16_to_int32, start = const_39, step = const_40)[name = tensor<string, []>("op_1225")];
            tensor<fp16, [1, 1]> var_1224_to_fp16 = const()[name = tensor<string, []>("op_1224_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1225_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1225_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1225_to_fp16 = cast(dtype = var_1225_promoted_to_fp16_dtype_0, x = var_1225)[name = tensor<string, []>("cast_360")];
            tensor<fp16, [1, ?]> ts_17_cast_fp16 = add(x = var_1224_to_fp16, y = var_1225_to_fp16)[name = tensor<string, []>("ts_17_cast_fp16")];
            tensor<int32, [4]> var_1228 = const()[name = tensor<string, []>("op_1228"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_19_cast_fp16 = reshape(shape = var_1228, x = ts_17_cast_fp16)[name = tensor<string, []>("ts_19_cast_fp16")];
            tensor<int32, [5]> concat_53x = const()[name = tensor<string, []>("concat_53x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_9_cast_fp16 = reshape(shape = concat_53x, x = q_17_cast_fp16)[name = tensor<string, []>("q2_9_cast_fp16")];
            tensor<int32, [5]> concat_54x = const()[name = tensor<string, []>("concat_54x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_9_cast_fp16 = reshape(shape = concat_54x, x = k_17_cast_fp16)[name = tensor<string, []>("k2_9_cast_fp16")];
            tensor<int32, [5]> qr_9_begin_0 = const()[name = tensor<string, []>("qr_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_9_end_0 = const()[name = tensor<string, []>("qr_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_9_end_mask_0 = const()[name = tensor<string, []>("qr_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_9_squeeze_mask_0 = const()[name = tensor<string, []>("qr_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_9_cast_fp16 = slice_by_index(begin = qr_9_begin_0, end = qr_9_end_0, end_mask = qr_9_end_mask_0, squeeze_mask = qr_9_squeeze_mask_0, x = q2_9_cast_fp16)[name = tensor<string, []>("qr_9_cast_fp16")];
            tensor<int32, [5]> qi_9_begin_0 = const()[name = tensor<string, []>("qi_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_9_end_0 = const()[name = tensor<string, []>("qi_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_9_end_mask_0 = const()[name = tensor<string, []>("qi_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_9_squeeze_mask_0 = const()[name = tensor<string, []>("qi_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_9_cast_fp16 = slice_by_index(begin = qi_9_begin_0, end = qi_9_end_0, end_mask = qi_9_end_mask_0, squeeze_mask = qi_9_squeeze_mask_0, x = q2_9_cast_fp16)[name = tensor<string, []>("qi_9_cast_fp16")];
            tensor<int32, [5]> kr_9_begin_0 = const()[name = tensor<string, []>("kr_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_9_end_0 = const()[name = tensor<string, []>("kr_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_9_end_mask_0 = const()[name = tensor<string, []>("kr_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_9_squeeze_mask_0 = const()[name = tensor<string, []>("kr_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_9_cast_fp16 = slice_by_index(begin = kr_9_begin_0, end = kr_9_end_0, end_mask = kr_9_end_mask_0, squeeze_mask = kr_9_squeeze_mask_0, x = k2_9_cast_fp16)[name = tensor<string, []>("kr_9_cast_fp16")];
            tensor<int32, [5]> ki_9_begin_0 = const()[name = tensor<string, []>("ki_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_9_end_0 = const()[name = tensor<string, []>("ki_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_9_end_mask_0 = const()[name = tensor<string, []>("ki_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_9_squeeze_mask_0 = const()[name = tensor<string, []>("ki_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_9_cast_fp16 = slice_by_index(begin = ki_9_begin_0, end = ki_9_end_0, end_mask = ki_9_end_mask_0, squeeze_mask = ki_9_squeeze_mask_0, x = k2_9_cast_fp16)[name = tensor<string, []>("ki_9_cast_fp16")];
            tensor<fp16, [32]> freqs_9_to_fp16 = const()[name = tensor<string, []>("freqs_9_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8594688)))];
            tensor<fp16, [1, 1, ?, 32]> var_1241_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_19_cast_fp16)[name = tensor<string, []>("op_1241_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_9_cast_fp16 = cos(x = var_1241_cast_fp16)[name = tensor<string, []>("rotr_9_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_9_cast_fp16 = sin(x = var_1241_cast_fp16)[name = tensor<string, []>("roti_9_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1245_cast_fp16 = mul(x = qr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor<string, []>("op_1245_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1246_cast_fp16 = mul(x = qi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor<string, []>("op_1246_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1247_cast_fp16 = sub(x = var_1245_cast_fp16, y = var_1246_cast_fp16)[name = tensor<string, []>("op_1247_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1248_cast_fp16 = mul(x = qr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor<string, []>("op_1248_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1249_cast_fp16 = mul(x = qi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor<string, []>("op_1249_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1250_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1249_cast_fp16)[name = tensor<string, []>("op_1250_cast_fp16")];
            tensor<int32, []> qo_9_axis_0 = const()[name = tensor<string, []>("qo_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (var_1247_cast_fp16, var_1250_cast_fp16))[name = tensor<string, []>("qo_9_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1253_cast_fp16 = mul(x = kr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor<string, []>("op_1253_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1254_cast_fp16 = mul(x = ki_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor<string, []>("op_1254_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1255_cast_fp16 = sub(x = var_1253_cast_fp16, y = var_1254_cast_fp16)[name = tensor<string, []>("op_1255_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1256_cast_fp16 = mul(x = kr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor<string, []>("op_1256_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1257_cast_fp16 = mul(x = ki_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor<string, []>("op_1257_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1258_cast_fp16 = add(x = var_1256_cast_fp16, y = var_1257_cast_fp16)[name = tensor<string, []>("op_1258_cast_fp16")];
            tensor<int32, []> ko_9_axis_0 = const()[name = tensor<string, []>("ko_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (var_1255_cast_fp16, var_1258_cast_fp16))[name = tensor<string, []>("ko_9_cast_fp16")];
            tensor<int32, [4]> concat_55x = const()[name = tensor<string, []>("concat_55x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_19_cast_fp16 = reshape(shape = concat_55x, x = qo_9_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<int32, [4]> concat_56x = const()[name = tensor<string, []>("concat_56x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_19_cast_fp16 = reshape(shape = concat_56x, x = ko_9_cast_fp16)[name = tensor<string, []>("k_19_cast_fp16")];
            tensor<int32, []> const_41 = const()[name = tensor<string, []>("const_41"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_42 = const()[name = tensor<string, []>("const_42"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_47_cast_uint16_to_int32 = cast(dtype = gather_47_cast_uint16_to_int32_dtype_0, x = gather_47_cast_uint16)[name = tensor<string, []>("cast_363")];
            tensor<int32, [?]> positions_9 = range_1d(end = gather_47_cast_uint16_to_int32, start = const_41, step = const_42)[name = tensor<string, []>("positions_9")];
            tensor<int32, [3]> concat_57x = const()[name = tensor<string, []>("concat_57x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1270 = reshape(shape = concat_57x, x = positions_9)[name = tensor<string, []>("op_1270")];
            tensor<int32, [3]> var_1271 = const()[name = tensor<string, []>("op_1271"), val = tensor<int32, [3]>([-1, 1, 1])];
            tensor<int32, [1, 1, 1]> var_1272 = reshape(shape = var_1271, x = input_lengths_5)[name = tensor<string, []>("op_1272")];
            tensor<bool, [1, 1, ?]> valid_k_9 = less(x = var_1270, y = var_1272)[name = tensor<string, []>("valid_k_9")];
            tensor<int32, [3]> concat_58x = const()[name = tensor<string, []>("concat_58x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1275 = reshape(shape = concat_58x, x = positions_9)[name = tensor<string, []>("op_1275")];
            tensor<int32, [1, ?, ?]> delta_9 = sub(x = var_1275, y = var_1270)[name = tensor<string, []>("delta_9")];
            tensor<int32, []> concat_60_axis_0 = const()[name = tensor<string, []>("concat_60_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_60_interleave_0 = const()[name = tensor<string, []>("concat_60_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (var_1177, gather_47_cast_uint16_to_int32, gather_47_cast_uint16_to_int32))[name = tensor<string, []>("concat_60")];
            tensor<fp16, []> fill_4_value_0_to_fp16 = const()[name = tensor<string, []>("fill_4_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_4_cast_fp16 = fill(shape = concat_60, value = fill_4_value_0_to_fp16)[name = tensor<string, []>("fill_4_cast_fp16")];
            tensor<string, []> attn_bias_33_dtype_0 = const()[name = tensor<string, []>("attn_bias_33_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1281 = greater_equal(x = delta_9, y = var_1159)[name = tensor<string, []>("op_1281")];
            tensor<bool, [1, ?, ?]> fill_4_cast_fp16_to_bool = cast(dtype = attn_bias_33_dtype_0, x = fill_4_cast_fp16)[name = tensor<string, []>("cast_359")];
            tensor<bool, [1, ?, ?]> attn_bias_35 = logical_and(x = fill_4_cast_fp16_to_bool, y = var_1281)[name = tensor<string, []>("attn_bias_35")];
            tensor<bool, [1, ?, ?]> var_1283 = less(x = delta_9, y = var_1166)[name = tensor<string, []>("op_1283")];
            tensor<bool, [1, ?, ?]> attn_bias_37 = logical_and(x = attn_bias_35, y = var_1283)[name = tensor<string, []>("attn_bias_37")];
            tensor<bool, [1, ?, ?]> var_1285 = logical_and(x = attn_bias_37, y = valid_k_9)[name = tensor<string, []>("op_1285")];
            tensor<int32, [1]> var_1287_axes_0 = const()[name = tensor<string, []>("op_1287_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1287 = expand_dims(axes = var_1287_axes_0, x = var_1285)[name = tensor<string, []>("op_1287")];
            tensor<fp16, []> sub_4_x_0_to_fp16 = const()[name = tensor<string, []>("sub_4_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_125_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_125_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1287_to_fp16 = cast(dtype = cast_125_to_fp16_dtype_0, x = var_1287)[name = tensor<string, []>("cast_358")];
            tensor<fp16, [1, 1, ?, ?]> sub_4_cast_fp16 = sub(x = sub_4_x_0_to_fp16, y = var_1287_to_fp16)[name = tensor<string, []>("sub_4_cast_fp16")];
            tensor<fp16, []> mul_8_x_0_to_fp16 = const()[name = tensor<string, []>("mul_8_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_8_cast_fp16 = mul(x = mul_8_x_0_to_fp16, y = sub_4_cast_fp16)[name = tensor<string, []>("mul_8_cast_fp16")];
            tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_9_cast_fp16 = mul(x = q_19_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")];
            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = k_19_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_8_cast_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
            tensor<string, []> add_4_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_4_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_4_cast_fp16_to_fp32 = cast(dtype = add_4_cast_fp16_to_fp32_dtype_0, x = add_4_cast_fp16)[name = tensor<string, []>("cast_357")];
            tensor<fp32, [1, 4, ?, ?]> softmax_4 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_4")];
            tensor<bool, []> out_17_transpose_x_0 = const()[name = tensor<string, []>("out_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_17_transpose_y_0 = const()[name = tensor<string, []>("out_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_4_to_fp16 = cast(dtype = softmax_4_to_fp16_dtype_0, x = softmax_4)[name = tensor<string, []>("cast_356")];
            tensor<fp16, [1, 4, ?, 64]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = softmax_4_to_fp16, y = v_9_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<int32, [2]> concat_61x = const()[name = tensor<string, []>("concat_61x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1293 = reshape(shape = concat_61x, x = positions_9)[name = tensor<string, []>("op_1293")];
            tensor<int32, [2]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [2]>([-1, 1])];
            tensor<int32, [1, 1]> var_1295 = reshape(shape = var_1294, x = input_lengths_5)[name = tensor<string, []>("op_1295")];
            tensor<bool, [1, ?]> var_1296 = less(x = var_1293, y = var_1295)[name = tensor<string, []>("op_1296")];
            tensor<int32, [4]> concat_62x = const()[name = tensor<string, []>("concat_62x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_9 = reshape(shape = concat_62x, x = var_1296)[name = tensor<string, []>("valid_q_9")];
            tensor<fp16, []> var_1300_to_fp16 = const()[name = tensor<string, []>("op_1300_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_19_cast_fp16 = select(a = out_17_cast_fp16, b = var_1300_to_fp16, cond = valid_q_9)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<int32, [4]> var_1302_perm_0 = const()[name = tensor<string, []>("op_1302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_63x = const()[name = tensor<string, []>("concat_63x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1302_cast_fp16 = transpose(perm = var_1302_perm_0, x = out_19_cast_fp16)[name = tensor<string, []>("transpose_23")];
            tensor<fp16, [1, ?, 256]> input_119_cast_fp16 = reshape(shape = concat_63x, x = var_1302_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_3_transformer_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8594816)))];
            tensor<fp16, [1, ?, 256]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_3_transformer_layers_0_self_attn_out_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8725952)))];
            tensor<fp16, [1, ?, 256]> var_1308_cast_fp16 = mul(x = decoder_3_transformer_layers_0_layer_scale_1_scale_to_fp16, y = linear_20_cast_fp16)[name = tensor<string, []>("op_1308_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_121_cast_fp16 = add(x = linear_18_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<int32, [1]> input_123_axes_0 = const()[name = tensor<string, []>("input_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8726528)))];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8727104)))];
            tensor<fp16, [1, ?, 256]> input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = decoder_3_transformer_layers_0_norm2_bias_to_fp16, epsilon = var_1173_to_fp16, gamma = decoder_3_transformer_layers_0_norm2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_3_transformer_layers_0_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8727680)))];
            tensor<fp16, [1, ?, 1024]> linear_21_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_3_transformer_layers_0_ffn_0_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<string, []> input_127_mode_0 = const()[name = tensor<string, []>("input_127_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_21_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_3_transformer_layers_0_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9252032)))];
            tensor<fp16, [1, ?, 256]> linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_3_transformer_layers_0_ffn_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<fp16, [256]> decoder_3_transformer_layers_0_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_0_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9776384)))];
            tensor<fp16, [1, ?, 256]> var_1323_cast_fp16 = mul(x = decoder_3_transformer_layers_0_layer_scale_2_scale_to_fp16, y = linear_22_cast_fp16)[name = tensor<string, []>("op_1323_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_129_cast_fp16 = add(x = input_121_cast_fp16, y = var_1323_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9776960)))];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9777536)))];
            tensor<fp16, [1, ?, 256]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = decoder_3_transformer_layers_1_norm1_bias_to_fp16, epsilon = var_1173_to_fp16, gamma = decoder_3_transformer_layers_1_norm1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<int32, [3]> var_1338_shape_cast_fp16 = shape(x = x_43_cast_fp16)[name = tensor<string, []>("op_1338_shape_cast_fp16")];
            tensor<int32, []> gather_57_axis_0 = const()[name = tensor<string, []>("gather_57_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_57_batch_dims_0 = const()[name = tensor<string, []>("gather_57_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_57_validate_indices_0 = const()[name = tensor<string, []>("gather_57_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1338_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1338_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_57_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_57_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1338_shape_cast_fp16_to_uint16 = cast(dtype = var_1338_shape_cast_fp16_to_uint16_dtype_0, x = var_1338_shape_cast_fp16)[name = tensor<string, []>("cast_355")];
            tensor<uint16, []> gather_57_cast_uint16 = gather(axis = gather_57_axis_0, batch_dims = gather_57_batch_dims_0, indices = gather_57_indices_0_to_uint16, validate_indices = gather_57_validate_indices_0, x = var_1338_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_57_cast_uint16")];
            tensor<string, []> gather_57_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_57_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_3_transformer_layers_1_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9778112)))];
            tensor<fp16, [1, ?, 768]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_3_transformer_layers_1_self_attn_in_proj_weight_to_fp16, x = x_43_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [5]> concat_64x = const()[name = tensor<string, []>("concat_64x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1345_cast_fp16 = reshape(shape = concat_64x, x = linear_23_cast_fp16)[name = tensor<string, []>("op_1345_cast_fp16")];
            tensor<int32, [5]> var_1346 = const()[name = tensor<string, []>("op_1346"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_21_begin_0 = const()[name = tensor<string, []>("q_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_21_end_0 = const()[name = tensor<string, []>("q_21_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_21_end_mask_0 = const()[name = tensor<string, []>("q_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_21_squeeze_mask_0 = const()[name = tensor<string, []>("q_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_23_cast_fp16 = transpose(perm = var_1346, x = var_1345_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 4, ?, 64]> q_21_cast_fp16 = slice_by_index(begin = q_21_begin_0, end = q_21_end_0, end_mask = q_21_end_mask_0, squeeze_mask = q_21_squeeze_mask_0, x = projected_23_cast_fp16)[name = tensor<string, []>("q_21_cast_fp16")];
            tensor<int32, [5]> k_21_begin_0 = const()[name = tensor<string, []>("k_21_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_21_end_0 = const()[name = tensor<string, []>("k_21_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_21_end_mask_0 = const()[name = tensor<string, []>("k_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_21_squeeze_mask_0 = const()[name = tensor<string, []>("k_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = projected_23_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
            tensor<int32, [5]> v_11_begin_0 = const()[name = tensor<string, []>("v_11_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_11_end_0 = const()[name = tensor<string, []>("v_11_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_11_end_mask_0 = const()[name = tensor<string, []>("v_11_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_11_squeeze_mask_0 = const()[name = tensor<string, []>("v_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = projected_23_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
            tensor<int32, [4]> var_1358_shape_cast_fp16 = shape(x = q_21_cast_fp16)[name = tensor<string, []>("op_1358_shape_cast_fp16")];
            tensor<int32, []> gather_61_axis_0 = const()[name = tensor<string, []>("gather_61_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_61_batch_dims_0 = const()[name = tensor<string, []>("gather_61_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_61_validate_indices_0 = const()[name = tensor<string, []>("gather_61_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1358_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1358_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_61_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_61_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1358_shape_cast_fp16_to_uint16 = cast(dtype = var_1358_shape_cast_fp16_to_uint16_dtype_0, x = var_1358_shape_cast_fp16)[name = tensor<string, []>("cast_353")];
            tensor<uint16, []> gather_61_cast_uint16 = gather(axis = gather_61_axis_0, batch_dims = gather_61_batch_dims_0, indices = gather_61_indices_0_to_uint16, validate_indices = gather_61_validate_indices_0, x = var_1358_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_61_cast_uint16")];
            tensor<string, []> gather_61_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_61_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_48 = const()[name = tensor<string, []>("const_48"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_49 = const()[name = tensor<string, []>("const_49"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_61_cast_uint16_to_int32 = cast(dtype = gather_61_cast_uint16_to_int32_dtype_0, x = gather_61_cast_uint16)[name = tensor<string, []>("cast_352")];
            tensor<int32, [?]> var_1362 = range_1d(end = gather_61_cast_uint16_to_int32, start = const_48, step = const_49)[name = tensor<string, []>("op_1362")];
            tensor<fp16, [1, 1]> var_1361_to_fp16 = const()[name = tensor<string, []>("op_1361_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1362_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1362_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1362_to_fp16 = cast(dtype = var_1362_promoted_to_fp16_dtype_0, x = var_1362)[name = tensor<string, []>("cast_351")];
            tensor<fp16, [1, ?]> ts_21_cast_fp16 = add(x = var_1361_to_fp16, y = var_1362_to_fp16)[name = tensor<string, []>("ts_21_cast_fp16")];
            tensor<int32, [4]> var_1365 = const()[name = tensor<string, []>("op_1365"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_23_cast_fp16 = reshape(shape = var_1365, x = ts_21_cast_fp16)[name = tensor<string, []>("ts_23_cast_fp16")];
            tensor<int32, [5]> concat_65x = const()[name = tensor<string, []>("concat_65x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_11_cast_fp16 = reshape(shape = concat_65x, x = q_21_cast_fp16)[name = tensor<string, []>("q2_11_cast_fp16")];
            tensor<int32, [5]> concat_66x = const()[name = tensor<string, []>("concat_66x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_11_cast_fp16 = reshape(shape = concat_66x, x = k_21_cast_fp16)[name = tensor<string, []>("k2_11_cast_fp16")];
            tensor<int32, [5]> qr_11_begin_0 = const()[name = tensor<string, []>("qr_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_11_end_0 = const()[name = tensor<string, []>("qr_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_11_end_mask_0 = const()[name = tensor<string, []>("qr_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_11_squeeze_mask_0 = const()[name = tensor<string, []>("qr_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_11_cast_fp16 = slice_by_index(begin = qr_11_begin_0, end = qr_11_end_0, end_mask = qr_11_end_mask_0, squeeze_mask = qr_11_squeeze_mask_0, x = q2_11_cast_fp16)[name = tensor<string, []>("qr_11_cast_fp16")];
            tensor<int32, [5]> qi_11_begin_0 = const()[name = tensor<string, []>("qi_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_11_end_0 = const()[name = tensor<string, []>("qi_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_11_end_mask_0 = const()[name = tensor<string, []>("qi_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_11_squeeze_mask_0 = const()[name = tensor<string, []>("qi_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_11_cast_fp16 = slice_by_index(begin = qi_11_begin_0, end = qi_11_end_0, end_mask = qi_11_end_mask_0, squeeze_mask = qi_11_squeeze_mask_0, x = q2_11_cast_fp16)[name = tensor<string, []>("qi_11_cast_fp16")];
            tensor<int32, [5]> kr_11_begin_0 = const()[name = tensor<string, []>("kr_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_11_end_0 = const()[name = tensor<string, []>("kr_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_11_end_mask_0 = const()[name = tensor<string, []>("kr_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_11_squeeze_mask_0 = const()[name = tensor<string, []>("kr_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_11_cast_fp16 = slice_by_index(begin = kr_11_begin_0, end = kr_11_end_0, end_mask = kr_11_end_mask_0, squeeze_mask = kr_11_squeeze_mask_0, x = k2_11_cast_fp16)[name = tensor<string, []>("kr_11_cast_fp16")];
            tensor<int32, [5]> ki_11_begin_0 = const()[name = tensor<string, []>("ki_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_11_end_0 = const()[name = tensor<string, []>("ki_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_11_end_mask_0 = const()[name = tensor<string, []>("ki_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_11_squeeze_mask_0 = const()[name = tensor<string, []>("ki_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_11_cast_fp16 = slice_by_index(begin = ki_11_begin_0, end = ki_11_end_0, end_mask = ki_11_end_mask_0, squeeze_mask = ki_11_squeeze_mask_0, x = k2_11_cast_fp16)[name = tensor<string, []>("ki_11_cast_fp16")];
            tensor<fp16, [32]> freqs_11_to_fp16 = const()[name = tensor<string, []>("freqs_11_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10171392)))];
            tensor<fp16, [1, 1, ?, 32]> var_1378_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_23_cast_fp16)[name = tensor<string, []>("op_1378_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_11_cast_fp16 = cos(x = var_1378_cast_fp16)[name = tensor<string, []>("rotr_11_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_11_cast_fp16 = sin(x = var_1378_cast_fp16)[name = tensor<string, []>("roti_11_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1382_cast_fp16 = mul(x = qr_11_cast_fp16, y = rotr_11_cast_fp16)[name = tensor<string, []>("op_1382_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1383_cast_fp16 = mul(x = qi_11_cast_fp16, y = roti_11_cast_fp16)[name = tensor<string, []>("op_1383_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1384_cast_fp16 = sub(x = var_1382_cast_fp16, y = var_1383_cast_fp16)[name = tensor<string, []>("op_1384_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1385_cast_fp16 = mul(x = qr_11_cast_fp16, y = roti_11_cast_fp16)[name = tensor<string, []>("op_1385_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1386_cast_fp16 = mul(x = qi_11_cast_fp16, y = rotr_11_cast_fp16)[name = tensor<string, []>("op_1386_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1387_cast_fp16 = add(x = var_1385_cast_fp16, y = var_1386_cast_fp16)[name = tensor<string, []>("op_1387_cast_fp16")];
            tensor<int32, []> qo_11_axis_0 = const()[name = tensor<string, []>("qo_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (var_1384_cast_fp16, var_1387_cast_fp16))[name = tensor<string, []>("qo_11_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1390_cast_fp16 = mul(x = kr_11_cast_fp16, y = rotr_11_cast_fp16)[name = tensor<string, []>("op_1390_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1391_cast_fp16 = mul(x = ki_11_cast_fp16, y = roti_11_cast_fp16)[name = tensor<string, []>("op_1391_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1392_cast_fp16 = sub(x = var_1390_cast_fp16, y = var_1391_cast_fp16)[name = tensor<string, []>("op_1392_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1393_cast_fp16 = mul(x = kr_11_cast_fp16, y = roti_11_cast_fp16)[name = tensor<string, []>("op_1393_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1394_cast_fp16 = mul(x = ki_11_cast_fp16, y = rotr_11_cast_fp16)[name = tensor<string, []>("op_1394_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1395_cast_fp16 = add(x = var_1393_cast_fp16, y = var_1394_cast_fp16)[name = tensor<string, []>("op_1395_cast_fp16")];
            tensor<int32, []> ko_11_axis_0 = const()[name = tensor<string, []>("ko_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (var_1392_cast_fp16, var_1395_cast_fp16))[name = tensor<string, []>("ko_11_cast_fp16")];
            tensor<int32, [4]> concat_67x = const()[name = tensor<string, []>("concat_67x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_23_cast_fp16 = reshape(shape = concat_67x, x = qo_11_cast_fp16)[name = tensor<string, []>("q_23_cast_fp16")];
            tensor<int32, [4]> concat_68x = const()[name = tensor<string, []>("concat_68x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_23_cast_fp16 = reshape(shape = concat_68x, x = ko_11_cast_fp16)[name = tensor<string, []>("k_23_cast_fp16")];
            tensor<int32, []> const_50 = const()[name = tensor<string, []>("const_50"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_51 = const()[name = tensor<string, []>("const_51"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_57_cast_uint16_to_int32 = cast(dtype = gather_57_cast_uint16_to_int32_dtype_0, x = gather_57_cast_uint16)[name = tensor<string, []>("cast_354")];
            tensor<int32, [?]> positions_11 = range_1d(end = gather_57_cast_uint16_to_int32, start = const_50, step = const_51)[name = tensor<string, []>("positions_11")];
            tensor<int32, [3]> concat_69x = const()[name = tensor<string, []>("concat_69x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1407 = reshape(shape = concat_69x, x = positions_11)[name = tensor<string, []>("op_1407")];
            tensor<bool, [1, 1, ?]> valid_k_11 = less(x = var_1407, y = var_1272)[name = tensor<string, []>("valid_k_11")];
            tensor<int32, [3]> concat_70x = const()[name = tensor<string, []>("concat_70x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1412 = reshape(shape = concat_70x, x = positions_11)[name = tensor<string, []>("op_1412")];
            tensor<int32, [1, ?, ?]> delta_11 = sub(x = var_1412, y = var_1407)[name = tensor<string, []>("delta_11")];
            tensor<int32, []> concat_72_axis_0 = const()[name = tensor<string, []>("concat_72_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_72_interleave_0 = const()[name = tensor<string, []>("concat_72_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_72 = concat(axis = concat_72_axis_0, interleave = concat_72_interleave_0, values = (var_1177, gather_57_cast_uint16_to_int32, gather_57_cast_uint16_to_int32))[name = tensor<string, []>("concat_72")];
            tensor<fp16, []> fill_5_value_0_to_fp16 = const()[name = tensor<string, []>("fill_5_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_5_cast_fp16 = fill(shape = concat_72, value = fill_5_value_0_to_fp16)[name = tensor<string, []>("fill_5_cast_fp16")];
            tensor<string, []> attn_bias_41_dtype_0 = const()[name = tensor<string, []>("attn_bias_41_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1418 = greater_equal(x = delta_11, y = var_1159)[name = tensor<string, []>("op_1418")];
            tensor<bool, [1, ?, ?]> fill_5_cast_fp16_to_bool = cast(dtype = attn_bias_41_dtype_0, x = fill_5_cast_fp16)[name = tensor<string, []>("cast_350")];
            tensor<bool, [1, ?, ?]> attn_bias_43 = logical_and(x = fill_5_cast_fp16_to_bool, y = var_1418)[name = tensor<string, []>("attn_bias_43")];
            tensor<bool, [1, ?, ?]> var_1420 = less(x = delta_11, y = var_1166)[name = tensor<string, []>("op_1420")];
            tensor<bool, [1, ?, ?]> attn_bias_45 = logical_and(x = attn_bias_43, y = var_1420)[name = tensor<string, []>("attn_bias_45")];
            tensor<bool, [1, ?, ?]> var_1422 = logical_and(x = attn_bias_45, y = valid_k_11)[name = tensor<string, []>("op_1422")];
            tensor<int32, [1]> var_1424_axes_0 = const()[name = tensor<string, []>("op_1424_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1424 = expand_dims(axes = var_1424_axes_0, x = var_1422)[name = tensor<string, []>("op_1424")];
            tensor<fp16, []> sub_5_x_0_to_fp16 = const()[name = tensor<string, []>("sub_5_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_147_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_147_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1424_to_fp16 = cast(dtype = cast_147_to_fp16_dtype_0, x = var_1424)[name = tensor<string, []>("cast_349")];
            tensor<fp16, [1, 1, ?, ?]> sub_5_cast_fp16 = sub(x = sub_5_x_0_to_fp16, y = var_1424_to_fp16)[name = tensor<string, []>("sub_5_cast_fp16")];
            tensor<fp16, []> mul_10_x_0_to_fp16 = const()[name = tensor<string, []>("mul_10_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_10_cast_fp16 = mul(x = mul_10_x_0_to_fp16, y = sub_5_cast_fp16)[name = tensor<string, []>("mul_10_cast_fp16")];
            tensor<fp16, []> mul_11_y_0_to_fp16 = const()[name = tensor<string, []>("mul_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_11_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor<string, []>("mul_11_cast_fp16")];
            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = k_23_cast_fp16)[name = tensor<string, []>("matmul_5_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_10_cast_fp16)[name = tensor<string, []>("add_5_cast_fp16")];
            tensor<string, []> add_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_5_cast_fp16_to_fp32 = cast(dtype = add_5_cast_fp16_to_fp32_dtype_0, x = add_5_cast_fp16)[name = tensor<string, []>("cast_348")];
            tensor<fp32, [1, 4, ?, ?]> softmax_5 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_5")];
            tensor<bool, []> out_21_transpose_x_0 = const()[name = tensor<string, []>("out_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_21_transpose_y_0 = const()[name = tensor<string, []>("out_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_5_to_fp16 = cast(dtype = softmax_5_to_fp16_dtype_0, x = softmax_5)[name = tensor<string, []>("cast_347")];
            tensor<fp16, [1, 4, ?, 64]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = softmax_5_to_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<int32, [2]> concat_73x = const()[name = tensor<string, []>("concat_73x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1430 = reshape(shape = concat_73x, x = positions_11)[name = tensor<string, []>("op_1430")];
            tensor<bool, [1, ?]> var_1433 = less(x = var_1430, y = var_1295)[name = tensor<string, []>("op_1433")];
            tensor<int32, [4]> concat_74x = const()[name = tensor<string, []>("concat_74x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_11 = reshape(shape = concat_74x, x = var_1433)[name = tensor<string, []>("valid_q_11")];
            tensor<fp16, []> var_1437_to_fp16 = const()[name = tensor<string, []>("op_1437_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_23_cast_fp16 = select(a = out_21_cast_fp16, b = var_1437_to_fp16, cond = valid_q_11)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<int32, [4]> var_1439_perm_0 = const()[name = tensor<string, []>("op_1439_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_75x = const()[name = tensor<string, []>("concat_75x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1439_cast_fp16 = transpose(perm = var_1439_perm_0, x = out_23_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, ?, 256]> input_131_cast_fp16 = reshape(shape = concat_75x, x = var_1439_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_3_transformer_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10171520)))];
            tensor<fp16, [1, ?, 256]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_3_transformer_layers_1_self_attn_out_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10302656)))];
            tensor<fp16, [1, ?, 256]> var_1445_cast_fp16 = mul(x = decoder_3_transformer_layers_1_layer_scale_1_scale_to_fp16, y = linear_24_cast_fp16)[name = tensor<string, []>("op_1445_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_133_cast_fp16 = add(x = input_129_cast_fp16, y = var_1445_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<int32, [1]> input_135_axes_0 = const()[name = tensor<string, []>("input_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10303232)))];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10303808)))];
            tensor<fp16, [1, ?, 256]> input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = decoder_3_transformer_layers_1_norm2_bias_to_fp16, epsilon = var_1173_to_fp16, gamma = decoder_3_transformer_layers_1_norm2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_3_transformer_layers_1_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10304384)))];
            tensor<fp16, [1, ?, 1024]> linear_25_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_3_transformer_layers_1_ffn_0_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<string, []> input_139_mode_0 = const()[name = tensor<string, []>("input_139_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_25_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_3_transformer_layers_1_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10828736)))];
            tensor<fp16, [1, ?, 256]> linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_3_transformer_layers_1_ffn_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [256]> decoder_3_transformer_layers_1_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_3_transformer_layers_1_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11353088)))];
            tensor<fp16, [1, ?, 256]> var_1460_cast_fp16 = mul(x = decoder_3_transformer_layers_1_layer_scale_2_scale_to_fp16, y = linear_26_cast_fp16)[name = tensor<string, []>("op_1460_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_141_cast_fp16 = add(x = input_133_cast_fp16, y = var_1460_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<fp16, [768, 256]> decoder_3_output_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_3_output_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11353664)))];
            tensor<fp16, [1, ?, 768]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_3_output_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<int32, [3]> x_49_perm_0 = const()[name = tensor<string, []>("x_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [4]> concat_76x = const()[name = tensor<string, []>("concat_76x"), val = tensor<int32, [4]>([1, 384, 2, -1])];
            tensor<fp16, [1, 768, ?]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = linear_27_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 384, 2, ?]> var_1479_cast_fp16 = reshape(shape = concat_76x, x = x_49_cast_fp16)[name = tensor<string, []>("op_1479_cast_fp16")];
            tensor<int32, [4]> var_1480 = const()[name = tensor<string, []>("op_1480"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> concat_77x = const()[name = tensor<string, []>("concat_77x"), val = tensor<int32, [3]>([1, 384, -1])];
            tensor<fp16, [1, 384, ?, 2]> var_1481_cast_fp16 = transpose(perm = var_1480, x = var_1479_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, 384, ?]> x_51_cast_fp16 = reshape(shape = concat_77x, x = var_1481_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<int32, []> var_1487 = const()[name = tensor<string, []>("op_1487"), val = tensor<int32, []>(2)];
            tensor<int32, [1]> input_lengths_7 = mul(x = input_lengths_5, y = var_1487)[name = tensor<string, []>("input_lengths_7")];
            tensor<int32, []> var_1493 = const()[name = tensor<string, []>("op_1493"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_1500 = const()[name = tensor<string, []>("op_1500"), val = tensor<int32, []>(1200)];
            tensor<int32, []> var_1511 = const()[name = tensor<string, []>("op_1511"), val = tensor<int32, []>(1)];
            tensor<int32, [3]> input_143_perm_0 = const()[name = tensor<string, []>("input_143_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [256, 384]> decoder_5_input_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_input_proj_weight_to_fp16"), val = tensor<fp16, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11746944)))];
            tensor<fp16, [1, ?, 384]> input_143_cast_fp16 = transpose(perm = input_143_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, ?, 256]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_5_input_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<int32, [1]> x_55_axes_0 = const()[name = tensor<string, []>("x_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11943616)))];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11944192)))];
            tensor<fp16, []> var_1507_to_fp16 = const()[name = tensor<string, []>("op_1507_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, ?, 256]> x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, beta = decoder_5_transformer_layers_0_norm1_bias_to_fp16, epsilon = var_1507_to_fp16, gamma = decoder_5_transformer_layers_0_norm1_weight_to_fp16, x = linear_28_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<int32, [3]> var_1535_shape_cast_fp16 = shape(x = x_55_cast_fp16)[name = tensor<string, []>("op_1535_shape_cast_fp16")];
            tensor<int32, []> gather_70_axis_0 = const()[name = tensor<string, []>("gather_70_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_70_batch_dims_0 = const()[name = tensor<string, []>("gather_70_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_70_validate_indices_0 = const()[name = tensor<string, []>("gather_70_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1535_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1535_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_70_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_70_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1535_shape_cast_fp16_to_uint16 = cast(dtype = var_1535_shape_cast_fp16_to_uint16_dtype_0, x = var_1535_shape_cast_fp16)[name = tensor<string, []>("cast_346")];
            tensor<uint16, []> gather_70_cast_uint16 = gather(axis = gather_70_axis_0, batch_dims = gather_70_batch_dims_0, indices = gather_70_indices_0_to_uint16, validate_indices = gather_70_validate_indices_0, x = var_1535_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_70_cast_uint16")];
            tensor<string, []> gather_70_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_70_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_5_transformer_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11944768)))];
            tensor<fp16, [1, ?, 768]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_5_transformer_layers_0_self_attn_in_proj_weight_to_fp16, x = x_55_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<int32, [5]> concat_78x = const()[name = tensor<string, []>("concat_78x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1542_cast_fp16 = reshape(shape = concat_78x, x = linear_29_cast_fp16)[name = tensor<string, []>("op_1542_cast_fp16")];
            tensor<int32, [5]> var_1543 = const()[name = tensor<string, []>("op_1543"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_25_begin_0 = const()[name = tensor<string, []>("q_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_25_end_0 = const()[name = tensor<string, []>("q_25_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_25_end_mask_0 = const()[name = tensor<string, []>("q_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_25_squeeze_mask_0 = const()[name = tensor<string, []>("q_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_27_cast_fp16 = transpose(perm = var_1543, x = var_1542_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 4, ?, 64]> q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = projected_27_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
            tensor<int32, [5]> k_25_begin_0 = const()[name = tensor<string, []>("k_25_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_25_end_0 = const()[name = tensor<string, []>("k_25_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_25_end_mask_0 = const()[name = tensor<string, []>("k_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_25_squeeze_mask_0 = const()[name = tensor<string, []>("k_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = projected_27_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
            tensor<int32, [5]> v_13_begin_0 = const()[name = tensor<string, []>("v_13_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_13_end_0 = const()[name = tensor<string, []>("v_13_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_13_end_mask_0 = const()[name = tensor<string, []>("v_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = projected_27_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
            tensor<int32, [4]> var_1555_shape_cast_fp16 = shape(x = q_25_cast_fp16)[name = tensor<string, []>("op_1555_shape_cast_fp16")];
            tensor<int32, []> gather_74_axis_0 = const()[name = tensor<string, []>("gather_74_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_74_batch_dims_0 = const()[name = tensor<string, []>("gather_74_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_74_validate_indices_0 = const()[name = tensor<string, []>("gather_74_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1555_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1555_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_74_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_74_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1555_shape_cast_fp16_to_uint16 = cast(dtype = var_1555_shape_cast_fp16_to_uint16_dtype_0, x = var_1555_shape_cast_fp16)[name = tensor<string, []>("cast_344")];
            tensor<uint16, []> gather_74_cast_uint16 = gather(axis = gather_74_axis_0, batch_dims = gather_74_batch_dims_0, indices = gather_74_indices_0_to_uint16, validate_indices = gather_74_validate_indices_0, x = var_1555_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_74_cast_uint16")];
            tensor<string, []> gather_74_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_74_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_57 = const()[name = tensor<string, []>("const_57"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_58 = const()[name = tensor<string, []>("const_58"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_74_cast_uint16_to_int32 = cast(dtype = gather_74_cast_uint16_to_int32_dtype_0, x = gather_74_cast_uint16)[name = tensor<string, []>("cast_343")];
            tensor<int32, [?]> var_1559 = range_1d(end = gather_74_cast_uint16_to_int32, start = const_57, step = const_58)[name = tensor<string, []>("op_1559")];
            tensor<fp16, [1, 1]> var_1558_to_fp16 = const()[name = tensor<string, []>("op_1558_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1559_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1559_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1559_to_fp16 = cast(dtype = var_1559_promoted_to_fp16_dtype_0, x = var_1559)[name = tensor<string, []>("cast_342")];
            tensor<fp16, [1, ?]> ts_25_cast_fp16 = add(x = var_1558_to_fp16, y = var_1559_to_fp16)[name = tensor<string, []>("ts_25_cast_fp16")];
            tensor<int32, [4]> var_1562 = const()[name = tensor<string, []>("op_1562"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_27_cast_fp16 = reshape(shape = var_1562, x = ts_25_cast_fp16)[name = tensor<string, []>("ts_27_cast_fp16")];
            tensor<int32, [5]> concat_79x = const()[name = tensor<string, []>("concat_79x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_13_cast_fp16 = reshape(shape = concat_79x, x = q_25_cast_fp16)[name = tensor<string, []>("q2_13_cast_fp16")];
            tensor<int32, [5]> concat_80x = const()[name = tensor<string, []>("concat_80x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_13_cast_fp16 = reshape(shape = concat_80x, x = k_25_cast_fp16)[name = tensor<string, []>("k2_13_cast_fp16")];
            tensor<int32, [5]> qr_13_begin_0 = const()[name = tensor<string, []>("qr_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_13_end_0 = const()[name = tensor<string, []>("qr_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_13_end_mask_0 = const()[name = tensor<string, []>("qr_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_13_squeeze_mask_0 = const()[name = tensor<string, []>("qr_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_13_cast_fp16 = slice_by_index(begin = qr_13_begin_0, end = qr_13_end_0, end_mask = qr_13_end_mask_0, squeeze_mask = qr_13_squeeze_mask_0, x = q2_13_cast_fp16)[name = tensor<string, []>("qr_13_cast_fp16")];
            tensor<int32, [5]> qi_13_begin_0 = const()[name = tensor<string, []>("qi_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_13_end_0 = const()[name = tensor<string, []>("qi_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_13_end_mask_0 = const()[name = tensor<string, []>("qi_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_13_squeeze_mask_0 = const()[name = tensor<string, []>("qi_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_13_cast_fp16 = slice_by_index(begin = qi_13_begin_0, end = qi_13_end_0, end_mask = qi_13_end_mask_0, squeeze_mask = qi_13_squeeze_mask_0, x = q2_13_cast_fp16)[name = tensor<string, []>("qi_13_cast_fp16")];
            tensor<int32, [5]> kr_13_begin_0 = const()[name = tensor<string, []>("kr_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_13_end_0 = const()[name = tensor<string, []>("kr_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_13_end_mask_0 = const()[name = tensor<string, []>("kr_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_13_squeeze_mask_0 = const()[name = tensor<string, []>("kr_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_13_cast_fp16 = slice_by_index(begin = kr_13_begin_0, end = kr_13_end_0, end_mask = kr_13_end_mask_0, squeeze_mask = kr_13_squeeze_mask_0, x = k2_13_cast_fp16)[name = tensor<string, []>("kr_13_cast_fp16")];
            tensor<int32, [5]> ki_13_begin_0 = const()[name = tensor<string, []>("ki_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_13_end_0 = const()[name = tensor<string, []>("ki_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_13_end_mask_0 = const()[name = tensor<string, []>("ki_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_13_squeeze_mask_0 = const()[name = tensor<string, []>("ki_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_13_cast_fp16 = slice_by_index(begin = ki_13_begin_0, end = ki_13_end_0, end_mask = ki_13_end_mask_0, squeeze_mask = ki_13_squeeze_mask_0, x = k2_13_cast_fp16)[name = tensor<string, []>("ki_13_cast_fp16")];
            tensor<fp16, [32]> freqs_13_to_fp16 = const()[name = tensor<string, []>("freqs_13_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12338048)))];
            tensor<fp16, [1, 1, ?, 32]> var_1575_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_27_cast_fp16)[name = tensor<string, []>("op_1575_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_13_cast_fp16 = cos(x = var_1575_cast_fp16)[name = tensor<string, []>("rotr_13_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_13_cast_fp16 = sin(x = var_1575_cast_fp16)[name = tensor<string, []>("roti_13_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1579_cast_fp16 = mul(x = qr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor<string, []>("op_1579_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1580_cast_fp16 = mul(x = qi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor<string, []>("op_1580_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1581_cast_fp16 = sub(x = var_1579_cast_fp16, y = var_1580_cast_fp16)[name = tensor<string, []>("op_1581_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1582_cast_fp16 = mul(x = qr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor<string, []>("op_1582_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1583_cast_fp16 = mul(x = qi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor<string, []>("op_1583_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1584_cast_fp16 = add(x = var_1582_cast_fp16, y = var_1583_cast_fp16)[name = tensor<string, []>("op_1584_cast_fp16")];
            tensor<int32, []> qo_13_axis_0 = const()[name = tensor<string, []>("qo_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (var_1581_cast_fp16, var_1584_cast_fp16))[name = tensor<string, []>("qo_13_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1587_cast_fp16 = mul(x = kr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor<string, []>("op_1587_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1588_cast_fp16 = mul(x = ki_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor<string, []>("op_1588_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1589_cast_fp16 = sub(x = var_1587_cast_fp16, y = var_1588_cast_fp16)[name = tensor<string, []>("op_1589_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1590_cast_fp16 = mul(x = kr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor<string, []>("op_1590_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1591_cast_fp16 = mul(x = ki_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor<string, []>("op_1591_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1592_cast_fp16 = add(x = var_1590_cast_fp16, y = var_1591_cast_fp16)[name = tensor<string, []>("op_1592_cast_fp16")];
            tensor<int32, []> ko_13_axis_0 = const()[name = tensor<string, []>("ko_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (var_1589_cast_fp16, var_1592_cast_fp16))[name = tensor<string, []>("ko_13_cast_fp16")];
            tensor<int32, [4]> concat_81x = const()[name = tensor<string, []>("concat_81x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_27_cast_fp16 = reshape(shape = concat_81x, x = qo_13_cast_fp16)[name = tensor<string, []>("q_27_cast_fp16")];
            tensor<int32, [4]> concat_82x = const()[name = tensor<string, []>("concat_82x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_27_cast_fp16 = reshape(shape = concat_82x, x = ko_13_cast_fp16)[name = tensor<string, []>("k_27_cast_fp16")];
            tensor<int32, []> const_59 = const()[name = tensor<string, []>("const_59"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_60 = const()[name = tensor<string, []>("const_60"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_70_cast_uint16_to_int32 = cast(dtype = gather_70_cast_uint16_to_int32_dtype_0, x = gather_70_cast_uint16)[name = tensor<string, []>("cast_345")];
            tensor<int32, [?]> positions_13 = range_1d(end = gather_70_cast_uint16_to_int32, start = const_59, step = const_60)[name = tensor<string, []>("positions_13")];
            tensor<int32, [3]> concat_83x = const()[name = tensor<string, []>("concat_83x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1604 = reshape(shape = concat_83x, x = positions_13)[name = tensor<string, []>("op_1604")];
            tensor<int32, [3]> var_1605 = const()[name = tensor<string, []>("op_1605"), val = tensor<int32, [3]>([-1, 1, 1])];
            tensor<int32, [1, 1, 1]> var_1606 = reshape(shape = var_1605, x = input_lengths_7)[name = tensor<string, []>("op_1606")];
            tensor<bool, [1, 1, ?]> valid_k_13 = less(x = var_1604, y = var_1606)[name = tensor<string, []>("valid_k_13")];
            tensor<int32, [3]> concat_84x = const()[name = tensor<string, []>("concat_84x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1609 = reshape(shape = concat_84x, x = positions_13)[name = tensor<string, []>("op_1609")];
            tensor<int32, [1, ?, ?]> delta_13 = sub(x = var_1609, y = var_1604)[name = tensor<string, []>("delta_13")];
            tensor<int32, []> concat_86_axis_0 = const()[name = tensor<string, []>("concat_86_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_86_interleave_0 = const()[name = tensor<string, []>("concat_86_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_86 = concat(axis = concat_86_axis_0, interleave = concat_86_interleave_0, values = (var_1511, gather_70_cast_uint16_to_int32, gather_70_cast_uint16_to_int32))[name = tensor<string, []>("concat_86")];
            tensor<fp16, []> fill_6_value_0_to_fp16 = const()[name = tensor<string, []>("fill_6_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_6_cast_fp16 = fill(shape = concat_86, value = fill_6_value_0_to_fp16)[name = tensor<string, []>("fill_6_cast_fp16")];
            tensor<string, []> attn_bias_49_dtype_0 = const()[name = tensor<string, []>("attn_bias_49_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1615 = greater_equal(x = delta_13, y = var_1493)[name = tensor<string, []>("op_1615")];
            tensor<bool, [1, ?, ?]> fill_6_cast_fp16_to_bool = cast(dtype = attn_bias_49_dtype_0, x = fill_6_cast_fp16)[name = tensor<string, []>("cast_341")];
            tensor<bool, [1, ?, ?]> attn_bias_51 = logical_and(x = fill_6_cast_fp16_to_bool, y = var_1615)[name = tensor<string, []>("attn_bias_51")];
            tensor<bool, [1, ?, ?]> var_1617 = less(x = delta_13, y = var_1500)[name = tensor<string, []>("op_1617")];
            tensor<bool, [1, ?, ?]> attn_bias_53 = logical_and(x = attn_bias_51, y = var_1617)[name = tensor<string, []>("attn_bias_53")];
            tensor<bool, [1, ?, ?]> var_1619 = logical_and(x = attn_bias_53, y = valid_k_13)[name = tensor<string, []>("op_1619")];
            tensor<int32, [1]> var_1621_axes_0 = const()[name = tensor<string, []>("op_1621_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1621 = expand_dims(axes = var_1621_axes_0, x = var_1619)[name = tensor<string, []>("op_1621")];
            tensor<fp16, []> sub_6_x_0_to_fp16 = const()[name = tensor<string, []>("sub_6_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_171_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_171_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1621_to_fp16 = cast(dtype = cast_171_to_fp16_dtype_0, x = var_1621)[name = tensor<string, []>("cast_340")];
            tensor<fp16, [1, 1, ?, ?]> sub_6_cast_fp16 = sub(x = sub_6_x_0_to_fp16, y = var_1621_to_fp16)[name = tensor<string, []>("sub_6_cast_fp16")];
            tensor<fp16, []> mul_12_x_0_to_fp16 = const()[name = tensor<string, []>("mul_12_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_12_cast_fp16 = mul(x = mul_12_x_0_to_fp16, y = sub_6_cast_fp16)[name = tensor<string, []>("mul_12_cast_fp16")];
            tensor<fp16, []> mul_13_y_0_to_fp16 = const()[name = tensor<string, []>("mul_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_13_cast_fp16 = mul(x = q_27_cast_fp16, y = mul_13_y_0_to_fp16)[name = tensor<string, []>("mul_13_cast_fp16")];
            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = k_27_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_12_cast_fp16)[name = tensor<string, []>("add_6_cast_fp16")];
            tensor<string, []> add_6_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_6_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_6_cast_fp16_to_fp32 = cast(dtype = add_6_cast_fp16_to_fp32_dtype_0, x = add_6_cast_fp16)[name = tensor<string, []>("cast_339")];
            tensor<fp32, [1, 4, ?, ?]> softmax_6 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_6")];
            tensor<bool, []> out_25_transpose_x_0 = const()[name = tensor<string, []>("out_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_25_transpose_y_0 = const()[name = tensor<string, []>("out_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_6_to_fp16 = cast(dtype = softmax_6_to_fp16_dtype_0, x = softmax_6)[name = tensor<string, []>("cast_338")];
            tensor<fp16, [1, 4, ?, 64]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = softmax_6_to_fp16, y = v_13_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<int32, [2]> concat_87x = const()[name = tensor<string, []>("concat_87x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1627 = reshape(shape = concat_87x, x = positions_13)[name = tensor<string, []>("op_1627")];
            tensor<int32, [2]> var_1628 = const()[name = tensor<string, []>("op_1628"), val = tensor<int32, [2]>([-1, 1])];
            tensor<int32, [1, 1]> var_1629 = reshape(shape = var_1628, x = input_lengths_7)[name = tensor<string, []>("op_1629")];
            tensor<bool, [1, ?]> var_1630 = less(x = var_1627, y = var_1629)[name = tensor<string, []>("op_1630")];
            tensor<int32, [4]> concat_88x = const()[name = tensor<string, []>("concat_88x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_13 = reshape(shape = concat_88x, x = var_1630)[name = tensor<string, []>("valid_q_13")];
            tensor<fp16, []> var_1634_to_fp16 = const()[name = tensor<string, []>("op_1634_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_27_cast_fp16 = select(a = out_25_cast_fp16, b = var_1634_to_fp16, cond = valid_q_13)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<int32, [4]> var_1636_perm_0 = const()[name = tensor<string, []>("op_1636_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_89x = const()[name = tensor<string, []>("concat_89x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1636_cast_fp16 = transpose(perm = var_1636_perm_0, x = out_27_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, ?, 256]> input_145_cast_fp16 = reshape(shape = concat_89x, x = var_1636_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_5_transformer_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12338176)))];
            tensor<fp16, [1, ?, 256]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_5_transformer_layers_0_self_attn_out_proj_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12469312)))];
            tensor<fp16, [1, ?, 256]> var_1642_cast_fp16 = mul(x = decoder_5_transformer_layers_0_layer_scale_1_scale_to_fp16, y = linear_30_cast_fp16)[name = tensor<string, []>("op_1642_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_147_cast_fp16 = add(x = linear_28_cast_fp16, y = var_1642_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<int32, [1]> input_149_axes_0 = const()[name = tensor<string, []>("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12469888)))];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12470464)))];
            tensor<fp16, [1, ?, 256]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = decoder_5_transformer_layers_0_norm2_bias_to_fp16, epsilon = var_1507_to_fp16, gamma = decoder_5_transformer_layers_0_norm2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_5_transformer_layers_0_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12471040)))];
            tensor<fp16, [1, ?, 1024]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_5_transformer_layers_0_ffn_0_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<string, []> input_153_mode_0 = const()[name = tensor<string, []>("input_153_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_153_cast_fp16 = gelu(mode = input_153_mode_0, x = linear_31_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_5_transformer_layers_0_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12995392)))];
            tensor<fp16, [1, ?, 256]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_5_transformer_layers_0_ffn_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<fp16, [256]> decoder_5_transformer_layers_0_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_0_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13519744)))];
            tensor<fp16, [1, ?, 256]> var_1657_cast_fp16 = mul(x = decoder_5_transformer_layers_0_layer_scale_2_scale_to_fp16, y = linear_32_cast_fp16)[name = tensor<string, []>("op_1657_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_155_cast_fp16 = add(x = input_147_cast_fp16, y = var_1657_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<int32, [1]> x_61_axes_0 = const()[name = tensor<string, []>("x_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13520320)))];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13520896)))];
            tensor<fp16, [1, ?, 256]> x_61_cast_fp16 = layer_norm(axes = x_61_axes_0, beta = decoder_5_transformer_layers_1_norm1_bias_to_fp16, epsilon = var_1507_to_fp16, gamma = decoder_5_transformer_layers_1_norm1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<int32, [3]> var_1672_shape_cast_fp16 = shape(x = x_61_cast_fp16)[name = tensor<string, []>("op_1672_shape_cast_fp16")];
            tensor<int32, []> gather_80_axis_0 = const()[name = tensor<string, []>("gather_80_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_80_batch_dims_0 = const()[name = tensor<string, []>("gather_80_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_80_validate_indices_0 = const()[name = tensor<string, []>("gather_80_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1672_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1672_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_80_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_80_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1672_shape_cast_fp16_to_uint16 = cast(dtype = var_1672_shape_cast_fp16_to_uint16_dtype_0, x = var_1672_shape_cast_fp16)[name = tensor<string, []>("cast_337")];
            tensor<uint16, []> gather_80_cast_uint16 = gather(axis = gather_80_axis_0, batch_dims = gather_80_batch_dims_0, indices = gather_80_indices_0_to_uint16, validate_indices = gather_80_validate_indices_0, x = var_1672_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_80_cast_uint16")];
            tensor<string, []> gather_80_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_80_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_5_transformer_layers_1_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13521472)))];
            tensor<fp16, [1, ?, 768]> linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_5_transformer_layers_1_self_attn_in_proj_weight_to_fp16, x = x_61_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<int32, [5]> concat_90x = const()[name = tensor<string, []>("concat_90x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1679_cast_fp16 = reshape(shape = concat_90x, x = linear_33_cast_fp16)[name = tensor<string, []>("op_1679_cast_fp16")];
            tensor<int32, [5]> var_1680 = const()[name = tensor<string, []>("op_1680"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_29_begin_0 = const()[name = tensor<string, []>("q_29_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_29_end_0 = const()[name = tensor<string, []>("q_29_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_29_end_mask_0 = const()[name = tensor<string, []>("q_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_29_squeeze_mask_0 = const()[name = tensor<string, []>("q_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_31_cast_fp16 = transpose(perm = var_1680, x = var_1679_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, 4, ?, 64]> q_29_cast_fp16 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, squeeze_mask = q_29_squeeze_mask_0, x = projected_31_cast_fp16)[name = tensor<string, []>("q_29_cast_fp16")];
            tensor<int32, [5]> k_29_begin_0 = const()[name = tensor<string, []>("k_29_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_29_end_0 = const()[name = tensor<string, []>("k_29_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_29_end_mask_0 = const()[name = tensor<string, []>("k_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_29_squeeze_mask_0 = const()[name = tensor<string, []>("k_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = projected_31_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
            tensor<int32, [5]> v_15_begin_0 = const()[name = tensor<string, []>("v_15_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_15_end_0 = const()[name = tensor<string, []>("v_15_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_15_end_mask_0 = const()[name = tensor<string, []>("v_15_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_15_squeeze_mask_0 = const()[name = tensor<string, []>("v_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = projected_31_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
            tensor<int32, [4]> var_1692_shape_cast_fp16 = shape(x = q_29_cast_fp16)[name = tensor<string, []>("op_1692_shape_cast_fp16")];
            tensor<int32, []> gather_84_axis_0 = const()[name = tensor<string, []>("gather_84_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_84_batch_dims_0 = const()[name = tensor<string, []>("gather_84_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_84_validate_indices_0 = const()[name = tensor<string, []>("gather_84_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1692_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1692_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_84_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_84_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1692_shape_cast_fp16_to_uint16 = cast(dtype = var_1692_shape_cast_fp16_to_uint16_dtype_0, x = var_1692_shape_cast_fp16)[name = tensor<string, []>("cast_335")];
            tensor<uint16, []> gather_84_cast_uint16 = gather(axis = gather_84_axis_0, batch_dims = gather_84_batch_dims_0, indices = gather_84_indices_0_to_uint16, validate_indices = gather_84_validate_indices_0, x = var_1692_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_84_cast_uint16")];
            tensor<string, []> gather_84_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_84_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_66 = const()[name = tensor<string, []>("const_66"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_67 = const()[name = tensor<string, []>("const_67"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_84_cast_uint16_to_int32 = cast(dtype = gather_84_cast_uint16_to_int32_dtype_0, x = gather_84_cast_uint16)[name = tensor<string, []>("cast_334")];
            tensor<int32, [?]> var_1696 = range_1d(end = gather_84_cast_uint16_to_int32, start = const_66, step = const_67)[name = tensor<string, []>("op_1696")];
            tensor<fp16, [1, 1]> var_1695_to_fp16 = const()[name = tensor<string, []>("op_1695_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1696_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1696_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1696_to_fp16 = cast(dtype = var_1696_promoted_to_fp16_dtype_0, x = var_1696)[name = tensor<string, []>("cast_333")];
            tensor<fp16, [1, ?]> ts_29_cast_fp16 = add(x = var_1695_to_fp16, y = var_1696_to_fp16)[name = tensor<string, []>("ts_29_cast_fp16")];
            tensor<int32, [4]> var_1699 = const()[name = tensor<string, []>("op_1699"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_31_cast_fp16 = reshape(shape = var_1699, x = ts_29_cast_fp16)[name = tensor<string, []>("ts_31_cast_fp16")];
            tensor<int32, [5]> concat_91x = const()[name = tensor<string, []>("concat_91x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_15_cast_fp16 = reshape(shape = concat_91x, x = q_29_cast_fp16)[name = tensor<string, []>("q2_15_cast_fp16")];
            tensor<int32, [5]> concat_92x = const()[name = tensor<string, []>("concat_92x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_15_cast_fp16 = reshape(shape = concat_92x, x = k_29_cast_fp16)[name = tensor<string, []>("k2_15_cast_fp16")];
            tensor<int32, [5]> qr_15_begin_0 = const()[name = tensor<string, []>("qr_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_15_end_0 = const()[name = tensor<string, []>("qr_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_15_end_mask_0 = const()[name = tensor<string, []>("qr_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_15_squeeze_mask_0 = const()[name = tensor<string, []>("qr_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_15_cast_fp16 = slice_by_index(begin = qr_15_begin_0, end = qr_15_end_0, end_mask = qr_15_end_mask_0, squeeze_mask = qr_15_squeeze_mask_0, x = q2_15_cast_fp16)[name = tensor<string, []>("qr_15_cast_fp16")];
            tensor<int32, [5]> qi_15_begin_0 = const()[name = tensor<string, []>("qi_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_15_end_0 = const()[name = tensor<string, []>("qi_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_15_end_mask_0 = const()[name = tensor<string, []>("qi_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_15_squeeze_mask_0 = const()[name = tensor<string, []>("qi_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_15_cast_fp16 = slice_by_index(begin = qi_15_begin_0, end = qi_15_end_0, end_mask = qi_15_end_mask_0, squeeze_mask = qi_15_squeeze_mask_0, x = q2_15_cast_fp16)[name = tensor<string, []>("qi_15_cast_fp16")];
            tensor<int32, [5]> kr_15_begin_0 = const()[name = tensor<string, []>("kr_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_15_end_0 = const()[name = tensor<string, []>("kr_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_15_end_mask_0 = const()[name = tensor<string, []>("kr_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_15_squeeze_mask_0 = const()[name = tensor<string, []>("kr_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_15_cast_fp16 = slice_by_index(begin = kr_15_begin_0, end = kr_15_end_0, end_mask = kr_15_end_mask_0, squeeze_mask = kr_15_squeeze_mask_0, x = k2_15_cast_fp16)[name = tensor<string, []>("kr_15_cast_fp16")];
            tensor<int32, [5]> ki_15_begin_0 = const()[name = tensor<string, []>("ki_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_15_end_0 = const()[name = tensor<string, []>("ki_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_15_end_mask_0 = const()[name = tensor<string, []>("ki_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_15_squeeze_mask_0 = const()[name = tensor<string, []>("ki_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_15_cast_fp16 = slice_by_index(begin = ki_15_begin_0, end = ki_15_end_0, end_mask = ki_15_end_mask_0, squeeze_mask = ki_15_squeeze_mask_0, x = k2_15_cast_fp16)[name = tensor<string, []>("ki_15_cast_fp16")];
            tensor<fp16, [32]> freqs_15_to_fp16 = const()[name = tensor<string, []>("freqs_15_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13914752)))];
            tensor<fp16, [1, 1, ?, 32]> var_1712_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_31_cast_fp16)[name = tensor<string, []>("op_1712_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_15_cast_fp16 = cos(x = var_1712_cast_fp16)[name = tensor<string, []>("rotr_15_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_15_cast_fp16 = sin(x = var_1712_cast_fp16)[name = tensor<string, []>("roti_15_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1716_cast_fp16 = mul(x = qr_15_cast_fp16, y = rotr_15_cast_fp16)[name = tensor<string, []>("op_1716_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1717_cast_fp16 = mul(x = qi_15_cast_fp16, y = roti_15_cast_fp16)[name = tensor<string, []>("op_1717_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1718_cast_fp16 = sub(x = var_1716_cast_fp16, y = var_1717_cast_fp16)[name = tensor<string, []>("op_1718_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1719_cast_fp16 = mul(x = qr_15_cast_fp16, y = roti_15_cast_fp16)[name = tensor<string, []>("op_1719_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1720_cast_fp16 = mul(x = qi_15_cast_fp16, y = rotr_15_cast_fp16)[name = tensor<string, []>("op_1720_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1721_cast_fp16 = add(x = var_1719_cast_fp16, y = var_1720_cast_fp16)[name = tensor<string, []>("op_1721_cast_fp16")];
            tensor<int32, []> qo_15_axis_0 = const()[name = tensor<string, []>("qo_15_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (var_1718_cast_fp16, var_1721_cast_fp16))[name = tensor<string, []>("qo_15_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1724_cast_fp16 = mul(x = kr_15_cast_fp16, y = rotr_15_cast_fp16)[name = tensor<string, []>("op_1724_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1725_cast_fp16 = mul(x = ki_15_cast_fp16, y = roti_15_cast_fp16)[name = tensor<string, []>("op_1725_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1726_cast_fp16 = sub(x = var_1724_cast_fp16, y = var_1725_cast_fp16)[name = tensor<string, []>("op_1726_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1727_cast_fp16 = mul(x = kr_15_cast_fp16, y = roti_15_cast_fp16)[name = tensor<string, []>("op_1727_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1728_cast_fp16 = mul(x = ki_15_cast_fp16, y = rotr_15_cast_fp16)[name = tensor<string, []>("op_1728_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1729_cast_fp16 = add(x = var_1727_cast_fp16, y = var_1728_cast_fp16)[name = tensor<string, []>("op_1729_cast_fp16")];
            tensor<int32, []> ko_15_axis_0 = const()[name = tensor<string, []>("ko_15_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (var_1726_cast_fp16, var_1729_cast_fp16))[name = tensor<string, []>("ko_15_cast_fp16")];
            tensor<int32, [4]> concat_93x = const()[name = tensor<string, []>("concat_93x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_31_cast_fp16 = reshape(shape = concat_93x, x = qo_15_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
            tensor<int32, [4]> concat_94x = const()[name = tensor<string, []>("concat_94x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_31_cast_fp16 = reshape(shape = concat_94x, x = ko_15_cast_fp16)[name = tensor<string, []>("k_31_cast_fp16")];
            tensor<int32, []> const_68 = const()[name = tensor<string, []>("const_68"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_69 = const()[name = tensor<string, []>("const_69"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_80_cast_uint16_to_int32 = cast(dtype = gather_80_cast_uint16_to_int32_dtype_0, x = gather_80_cast_uint16)[name = tensor<string, []>("cast_336")];
            tensor<int32, [?]> positions_15 = range_1d(end = gather_80_cast_uint16_to_int32, start = const_68, step = const_69)[name = tensor<string, []>("positions_15")];
            tensor<int32, [3]> concat_95x = const()[name = tensor<string, []>("concat_95x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1741 = reshape(shape = concat_95x, x = positions_15)[name = tensor<string, []>("op_1741")];
            tensor<bool, [1, 1, ?]> valid_k_15 = less(x = var_1741, y = var_1606)[name = tensor<string, []>("valid_k_15")];
            tensor<int32, [3]> concat_96x = const()[name = tensor<string, []>("concat_96x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1746 = reshape(shape = concat_96x, x = positions_15)[name = tensor<string, []>("op_1746")];
            tensor<int32, [1, ?, ?]> delta_15 = sub(x = var_1746, y = var_1741)[name = tensor<string, []>("delta_15")];
            tensor<int32, []> concat_98_axis_0 = const()[name = tensor<string, []>("concat_98_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_98_interleave_0 = const()[name = tensor<string, []>("concat_98_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_98 = concat(axis = concat_98_axis_0, interleave = concat_98_interleave_0, values = (var_1511, gather_80_cast_uint16_to_int32, gather_80_cast_uint16_to_int32))[name = tensor<string, []>("concat_98")];
            tensor<fp16, []> fill_7_value_0_to_fp16 = const()[name = tensor<string, []>("fill_7_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_7_cast_fp16 = fill(shape = concat_98, value = fill_7_value_0_to_fp16)[name = tensor<string, []>("fill_7_cast_fp16")];
            tensor<string, []> attn_bias_57_dtype_0 = const()[name = tensor<string, []>("attn_bias_57_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1752 = greater_equal(x = delta_15, y = var_1493)[name = tensor<string, []>("op_1752")];
            tensor<bool, [1, ?, ?]> fill_7_cast_fp16_to_bool = cast(dtype = attn_bias_57_dtype_0, x = fill_7_cast_fp16)[name = tensor<string, []>("cast_332")];
            tensor<bool, [1, ?, ?]> attn_bias_59 = logical_and(x = fill_7_cast_fp16_to_bool, y = var_1752)[name = tensor<string, []>("attn_bias_59")];
            tensor<bool, [1, ?, ?]> var_1754 = less(x = delta_15, y = var_1500)[name = tensor<string, []>("op_1754")];
            tensor<bool, [1, ?, ?]> attn_bias_61 = logical_and(x = attn_bias_59, y = var_1754)[name = tensor<string, []>("attn_bias_61")];
            tensor<bool, [1, ?, ?]> var_1756 = logical_and(x = attn_bias_61, y = valid_k_15)[name = tensor<string, []>("op_1756")];
            tensor<int32, [1]> var_1758_axes_0 = const()[name = tensor<string, []>("op_1758_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1758 = expand_dims(axes = var_1758_axes_0, x = var_1756)[name = tensor<string, []>("op_1758")];
            tensor<fp16, []> sub_7_x_0_to_fp16 = const()[name = tensor<string, []>("sub_7_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_193_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_193_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1758_to_fp16 = cast(dtype = cast_193_to_fp16_dtype_0, x = var_1758)[name = tensor<string, []>("cast_331")];
            tensor<fp16, [1, 1, ?, ?]> sub_7_cast_fp16 = sub(x = sub_7_x_0_to_fp16, y = var_1758_to_fp16)[name = tensor<string, []>("sub_7_cast_fp16")];
            tensor<fp16, []> mul_14_x_0_to_fp16 = const()[name = tensor<string, []>("mul_14_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_14_cast_fp16 = mul(x = mul_14_x_0_to_fp16, y = sub_7_cast_fp16)[name = tensor<string, []>("mul_14_cast_fp16")];
            tensor<fp16, []> mul_15_y_0_to_fp16 = const()[name = tensor<string, []>("mul_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_15_cast_fp16 = mul(x = q_31_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")];
            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = k_31_cast_fp16)[name = tensor<string, []>("matmul_7_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_14_cast_fp16)[name = tensor<string, []>("add_7_cast_fp16")];
            tensor<string, []> add_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_7_cast_fp16_to_fp32 = cast(dtype = add_7_cast_fp16_to_fp32_dtype_0, x = add_7_cast_fp16)[name = tensor<string, []>("cast_330")];
            tensor<fp32, [1, 4, ?, ?]> softmax_7 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_7")];
            tensor<bool, []> out_29_transpose_x_0 = const()[name = tensor<string, []>("out_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_29_transpose_y_0 = const()[name = tensor<string, []>("out_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_7_to_fp16 = cast(dtype = softmax_7_to_fp16_dtype_0, x = softmax_7)[name = tensor<string, []>("cast_329")];
            tensor<fp16, [1, 4, ?, 64]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = softmax_7_to_fp16, y = v_15_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<int32, [2]> concat_99x = const()[name = tensor<string, []>("concat_99x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1764 = reshape(shape = concat_99x, x = positions_15)[name = tensor<string, []>("op_1764")];
            tensor<bool, [1, ?]> var_1767 = less(x = var_1764, y = var_1629)[name = tensor<string, []>("op_1767")];
            tensor<int32, [4]> concat_100x = const()[name = tensor<string, []>("concat_100x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_15 = reshape(shape = concat_100x, x = var_1767)[name = tensor<string, []>("valid_q_15")];
            tensor<fp16, []> var_1771_to_fp16 = const()[name = tensor<string, []>("op_1771_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_31_cast_fp16 = select(a = out_29_cast_fp16, b = var_1771_to_fp16, cond = valid_q_15)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<int32, [4]> var_1773_perm_0 = const()[name = tensor<string, []>("op_1773_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_101x = const()[name = tensor<string, []>("concat_101x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1773_cast_fp16 = transpose(perm = var_1773_perm_0, x = out_31_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, ?, 256]> input_157_cast_fp16 = reshape(shape = concat_101x, x = var_1773_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_5_transformer_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13914880)))];
            tensor<fp16, [1, ?, 256]> linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_5_transformer_layers_1_self_attn_out_proj_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14046016)))];
            tensor<fp16, [1, ?, 256]> var_1779_cast_fp16 = mul(x = decoder_5_transformer_layers_1_layer_scale_1_scale_to_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("op_1779_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_159_cast_fp16 = add(x = input_155_cast_fp16, y = var_1779_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<int32, [1]> input_161_axes_0 = const()[name = tensor<string, []>("input_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14046592)))];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14047168)))];
            tensor<fp16, [1, ?, 256]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = decoder_5_transformer_layers_1_norm2_bias_to_fp16, epsilon = var_1507_to_fp16, gamma = decoder_5_transformer_layers_1_norm2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_5_transformer_layers_1_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14047744)))];
            tensor<fp16, [1, ?, 1024]> linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_5_transformer_layers_1_ffn_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<string, []> input_165_mode_0 = const()[name = tensor<string, []>("input_165_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_165_cast_fp16 = gelu(mode = input_165_mode_0, x = linear_35_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_5_transformer_layers_1_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14572096)))];
            tensor<fp16, [1, ?, 256]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_5_transformer_layers_1_ffn_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [256]> decoder_5_transformer_layers_1_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_5_transformer_layers_1_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15096448)))];
            tensor<fp16, [1, ?, 256]> var_1794_cast_fp16 = mul(x = decoder_5_transformer_layers_1_layer_scale_2_scale_to_fp16, y = linear_36_cast_fp16)[name = tensor<string, []>("op_1794_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_167_cast_fp16 = add(x = input_159_cast_fp16, y = var_1794_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<fp16, [768, 256]> decoder_5_output_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_5_output_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15097024)))];
            tensor<fp16, [1, ?, 768]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_5_output_proj_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<int32, [3]> x_67_perm_0 = const()[name = tensor<string, []>("x_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [4]> concat_102x = const()[name = tensor<string, []>("concat_102x"), val = tensor<int32, [4]>([1, 384, 2, -1])];
            tensor<fp16, [1, 768, ?]> x_67_cast_fp16 = transpose(perm = x_67_perm_0, x = linear_37_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 384, 2, ?]> var_1813_cast_fp16 = reshape(shape = concat_102x, x = x_67_cast_fp16)[name = tensor<string, []>("op_1813_cast_fp16")];
            tensor<int32, [4]> var_1814 = const()[name = tensor<string, []>("op_1814"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> concat_103x = const()[name = tensor<string, []>("concat_103x"), val = tensor<int32, [3]>([1, 384, -1])];
            tensor<fp16, [1, 384, ?, 2]> var_1815_cast_fp16 = transpose(perm = var_1814, x = var_1813_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 384, ?]> x_69_cast_fp16 = reshape(shape = concat_103x, x = var_1815_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<int32, []> var_1821 = const()[name = tensor<string, []>("op_1821"), val = tensor<int32, []>(2)];
            tensor<int32, [1]> input_lengths = mul(x = input_lengths_7, y = var_1821)[name = tensor<string, []>("input_lengths")];
            tensor<int32, []> var_1827 = const()[name = tensor<string, []>("op_1827"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_1834 = const()[name = tensor<string, []>("op_1834"), val = tensor<int32, []>(1600)];
            tensor<int32, []> var_1845 = const()[name = tensor<string, []>("op_1845"), val = tensor<int32, []>(1)];
            tensor<int32, [3]> input_169_perm_0 = const()[name = tensor<string, []>("input_169_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [256, 384]> decoder_7_input_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_input_proj_weight_to_fp16"), val = tensor<fp16, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15490304)))];
            tensor<fp16, [1, ?, 384]> input_169_cast_fp16 = transpose(perm = input_169_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, ?, 256]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_input_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<int32, [1]> x_73_axes_0 = const()[name = tensor<string, []>("x_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15686976)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15687552)))];
            tensor<fp16, []> var_1841_to_fp16 = const()[name = tensor<string, []>("op_1841_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, ?, 256]> x_73_cast_fp16 = layer_norm(axes = x_73_axes_0, beta = decoder_7_transformer_layers_0_norm1_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_0_norm1_weight_to_fp16, x = linear_38_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<int32, [3]> var_1873_shape_cast_fp16 = shape(x = x_73_cast_fp16)[name = tensor<string, []>("op_1873_shape_cast_fp16")];
            tensor<int32, []> gather_93_axis_0 = const()[name = tensor<string, []>("gather_93_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_93_batch_dims_0 = const()[name = tensor<string, []>("gather_93_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_93_validate_indices_0 = const()[name = tensor<string, []>("gather_93_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1873_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1873_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_93_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_93_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_1873_shape_cast_fp16_to_uint16 = cast(dtype = var_1873_shape_cast_fp16_to_uint16_dtype_0, x = var_1873_shape_cast_fp16)[name = tensor<string, []>("cast_328")];
            tensor<uint16, []> gather_93_cast_uint16 = gather(axis = gather_93_axis_0, batch_dims = gather_93_batch_dims_0, indices = gather_93_indices_0_to_uint16, validate_indices = gather_93_validate_indices_0, x = var_1873_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_93_cast_uint16")];
            tensor<string, []> gather_93_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_93_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_7_transformer_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15688128)))];
            tensor<fp16, [1, ?, 768]> linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_7_transformer_layers_0_self_attn_in_proj_weight_to_fp16, x = x_73_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<int32, [5]> concat_104x = const()[name = tensor<string, []>("concat_104x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_1880_cast_fp16 = reshape(shape = concat_104x, x = linear_39_cast_fp16)[name = tensor<string, []>("op_1880_cast_fp16")];
            tensor<int32, [5]> var_1881 = const()[name = tensor<string, []>("op_1881"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_33_begin_0 = const()[name = tensor<string, []>("q_33_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_33_end_0 = const()[name = tensor<string, []>("q_33_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_33_end_mask_0 = const()[name = tensor<string, []>("q_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_33_squeeze_mask_0 = const()[name = tensor<string, []>("q_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_35_cast_fp16 = transpose(perm = var_1881, x = var_1880_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 4, ?, 64]> q_33_cast_fp16 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, squeeze_mask = q_33_squeeze_mask_0, x = projected_35_cast_fp16)[name = tensor<string, []>("q_33_cast_fp16")];
            tensor<int32, [5]> k_33_begin_0 = const()[name = tensor<string, []>("k_33_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_33_end_0 = const()[name = tensor<string, []>("k_33_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_33_end_mask_0 = const()[name = tensor<string, []>("k_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_33_squeeze_mask_0 = const()[name = tensor<string, []>("k_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = projected_35_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
            tensor<int32, [5]> v_17_begin_0 = const()[name = tensor<string, []>("v_17_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_17_end_0 = const()[name = tensor<string, []>("v_17_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_17_end_mask_0 = const()[name = tensor<string, []>("v_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_17_squeeze_mask_0 = const()[name = tensor<string, []>("v_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = projected_35_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
            tensor<int32, [4]> var_1893_shape_cast_fp16 = shape(x = q_33_cast_fp16)[name = tensor<string, []>("op_1893_shape_cast_fp16")];
            tensor<int32, []> gather_97_axis_0 = const()[name = tensor<string, []>("gather_97_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_97_batch_dims_0 = const()[name = tensor<string, []>("gather_97_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_97_validate_indices_0 = const()[name = tensor<string, []>("gather_97_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_1893_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_1893_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_97_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_97_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_1893_shape_cast_fp16_to_uint16 = cast(dtype = var_1893_shape_cast_fp16_to_uint16_dtype_0, x = var_1893_shape_cast_fp16)[name = tensor<string, []>("cast_326")];
            tensor<uint16, []> gather_97_cast_uint16 = gather(axis = gather_97_axis_0, batch_dims = gather_97_batch_dims_0, indices = gather_97_indices_0_to_uint16, validate_indices = gather_97_validate_indices_0, x = var_1893_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_97_cast_uint16")];
            tensor<string, []> gather_97_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_97_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_75 = const()[name = tensor<string, []>("const_75"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_76 = const()[name = tensor<string, []>("const_76"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_97_cast_uint16_to_int32 = cast(dtype = gather_97_cast_uint16_to_int32_dtype_0, x = gather_97_cast_uint16)[name = tensor<string, []>("cast_325")];
            tensor<int32, [?]> var_1897 = range_1d(end = gather_97_cast_uint16_to_int32, start = const_75, step = const_76)[name = tensor<string, []>("op_1897")];
            tensor<fp16, [1, 1]> var_1896_to_fp16 = const()[name = tensor<string, []>("op_1896_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_1897_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1897_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_1897_to_fp16 = cast(dtype = var_1897_promoted_to_fp16_dtype_0, x = var_1897)[name = tensor<string, []>("cast_324")];
            tensor<fp16, [1, ?]> ts_33_cast_fp16 = add(x = var_1896_to_fp16, y = var_1897_to_fp16)[name = tensor<string, []>("ts_33_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = tensor<string, []>("op_1900"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_35_cast_fp16 = reshape(shape = var_1900, x = ts_33_cast_fp16)[name = tensor<string, []>("ts_35_cast_fp16")];
            tensor<int32, [5]> concat_105x = const()[name = tensor<string, []>("concat_105x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_17_cast_fp16 = reshape(shape = concat_105x, x = q_33_cast_fp16)[name = tensor<string, []>("q2_17_cast_fp16")];
            tensor<int32, [5]> concat_106x = const()[name = tensor<string, []>("concat_106x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_17_cast_fp16 = reshape(shape = concat_106x, x = k_33_cast_fp16)[name = tensor<string, []>("k2_17_cast_fp16")];
            tensor<int32, [5]> qr_17_begin_0 = const()[name = tensor<string, []>("qr_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_17_end_0 = const()[name = tensor<string, []>("qr_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_17_end_mask_0 = const()[name = tensor<string, []>("qr_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_17_squeeze_mask_0 = const()[name = tensor<string, []>("qr_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_17_cast_fp16 = slice_by_index(begin = qr_17_begin_0, end = qr_17_end_0, end_mask = qr_17_end_mask_0, squeeze_mask = qr_17_squeeze_mask_0, x = q2_17_cast_fp16)[name = tensor<string, []>("qr_17_cast_fp16")];
            tensor<int32, [5]> qi_17_begin_0 = const()[name = tensor<string, []>("qi_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_17_end_0 = const()[name = tensor<string, []>("qi_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_17_end_mask_0 = const()[name = tensor<string, []>("qi_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_17_squeeze_mask_0 = const()[name = tensor<string, []>("qi_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_17_cast_fp16 = slice_by_index(begin = qi_17_begin_0, end = qi_17_end_0, end_mask = qi_17_end_mask_0, squeeze_mask = qi_17_squeeze_mask_0, x = q2_17_cast_fp16)[name = tensor<string, []>("qi_17_cast_fp16")];
            tensor<int32, [5]> kr_17_begin_0 = const()[name = tensor<string, []>("kr_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_17_end_0 = const()[name = tensor<string, []>("kr_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_17_end_mask_0 = const()[name = tensor<string, []>("kr_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_17_squeeze_mask_0 = const()[name = tensor<string, []>("kr_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_17_cast_fp16 = slice_by_index(begin = kr_17_begin_0, end = kr_17_end_0, end_mask = kr_17_end_mask_0, squeeze_mask = kr_17_squeeze_mask_0, x = k2_17_cast_fp16)[name = tensor<string, []>("kr_17_cast_fp16")];
            tensor<int32, [5]> ki_17_begin_0 = const()[name = tensor<string, []>("ki_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_17_end_0 = const()[name = tensor<string, []>("ki_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_17_end_mask_0 = const()[name = tensor<string, []>("ki_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_17_squeeze_mask_0 = const()[name = tensor<string, []>("ki_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_17_cast_fp16 = slice_by_index(begin = ki_17_begin_0, end = ki_17_end_0, end_mask = ki_17_end_mask_0, squeeze_mask = ki_17_squeeze_mask_0, x = k2_17_cast_fp16)[name = tensor<string, []>("ki_17_cast_fp16")];
            tensor<fp16, [32]> freqs_17_to_fp16 = const()[name = tensor<string, []>("freqs_17_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16081408)))];
            tensor<fp16, [1, 1, ?, 32]> var_1913_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_35_cast_fp16)[name = tensor<string, []>("op_1913_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_17_cast_fp16 = cos(x = var_1913_cast_fp16)[name = tensor<string, []>("rotr_17_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_17_cast_fp16 = sin(x = var_1913_cast_fp16)[name = tensor<string, []>("roti_17_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1917_cast_fp16 = mul(x = qr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor<string, []>("op_1917_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1918_cast_fp16 = mul(x = qi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor<string, []>("op_1918_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1919_cast_fp16 = sub(x = var_1917_cast_fp16, y = var_1918_cast_fp16)[name = tensor<string, []>("op_1919_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1920_cast_fp16 = mul(x = qr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor<string, []>("op_1920_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1921_cast_fp16 = mul(x = qi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor<string, []>("op_1921_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1922_cast_fp16 = add(x = var_1920_cast_fp16, y = var_1921_cast_fp16)[name = tensor<string, []>("op_1922_cast_fp16")];
            tensor<int32, []> qo_17_axis_0 = const()[name = tensor<string, []>("qo_17_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (var_1919_cast_fp16, var_1922_cast_fp16))[name = tensor<string, []>("qo_17_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1925_cast_fp16 = mul(x = kr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor<string, []>("op_1925_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1926_cast_fp16 = mul(x = ki_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor<string, []>("op_1926_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1927_cast_fp16 = sub(x = var_1925_cast_fp16, y = var_1926_cast_fp16)[name = tensor<string, []>("op_1927_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1928_cast_fp16 = mul(x = kr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor<string, []>("op_1928_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1929_cast_fp16 = mul(x = ki_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor<string, []>("op_1929_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_1930_cast_fp16 = add(x = var_1928_cast_fp16, y = var_1929_cast_fp16)[name = tensor<string, []>("op_1930_cast_fp16")];
            tensor<int32, []> ko_17_axis_0 = const()[name = tensor<string, []>("ko_17_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (var_1927_cast_fp16, var_1930_cast_fp16))[name = tensor<string, []>("ko_17_cast_fp16")];
            tensor<int32, [4]> concat_107x = const()[name = tensor<string, []>("concat_107x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_35_cast_fp16 = reshape(shape = concat_107x, x = qo_17_cast_fp16)[name = tensor<string, []>("q_35_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = tensor<string, []>("concat_108x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_35_cast_fp16 = reshape(shape = concat_108x, x = ko_17_cast_fp16)[name = tensor<string, []>("k_35_cast_fp16")];
            tensor<int32, []> const_77 = const()[name = tensor<string, []>("const_77"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_78 = const()[name = tensor<string, []>("const_78"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_93_cast_uint16_to_int32 = cast(dtype = gather_93_cast_uint16_to_int32_dtype_0, x = gather_93_cast_uint16)[name = tensor<string, []>("cast_327")];
            tensor<int32, [?]> positions_17 = range_1d(end = gather_93_cast_uint16_to_int32, start = const_77, step = const_78)[name = tensor<string, []>("positions_17")];
            tensor<int32, [3]> concat_109x = const()[name = tensor<string, []>("concat_109x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_1942 = reshape(shape = concat_109x, x = positions_17)[name = tensor<string, []>("op_1942")];
            tensor<int32, [3]> var_1943 = const()[name = tensor<string, []>("op_1943"), val = tensor<int32, [3]>([-1, 1, 1])];
            tensor<int32, [1, 1, 1]> var_1944 = reshape(shape = var_1943, x = input_lengths)[name = tensor<string, []>("op_1944")];
            tensor<bool, [1, 1, ?]> valid_k_17 = less(x = var_1942, y = var_1944)[name = tensor<string, []>("valid_k_17")];
            tensor<int32, [3]> concat_110x = const()[name = tensor<string, []>("concat_110x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_1947 = reshape(shape = concat_110x, x = positions_17)[name = tensor<string, []>("op_1947")];
            tensor<int32, [1, ?, ?]> delta_17 = sub(x = var_1947, y = var_1942)[name = tensor<string, []>("delta_17")];
            tensor<int32, []> concat_112_axis_0 = const()[name = tensor<string, []>("concat_112_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_112_interleave_0 = const()[name = tensor<string, []>("concat_112_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_112 = concat(axis = concat_112_axis_0, interleave = concat_112_interleave_0, values = (var_1845, gather_93_cast_uint16_to_int32, gather_93_cast_uint16_to_int32))[name = tensor<string, []>("concat_112")];
            tensor<fp16, []> fill_8_value_0_to_fp16 = const()[name = tensor<string, []>("fill_8_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_8_cast_fp16 = fill(shape = concat_112, value = fill_8_value_0_to_fp16)[name = tensor<string, []>("fill_8_cast_fp16")];
            tensor<string, []> attn_bias_65_dtype_0 = const()[name = tensor<string, []>("attn_bias_65_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_1953 = greater_equal(x = delta_17, y = var_1827)[name = tensor<string, []>("op_1953")];
            tensor<bool, [1, ?, ?]> fill_8_cast_fp16_to_bool = cast(dtype = attn_bias_65_dtype_0, x = fill_8_cast_fp16)[name = tensor<string, []>("cast_323")];
            tensor<bool, [1, ?, ?]> attn_bias_67 = logical_and(x = fill_8_cast_fp16_to_bool, y = var_1953)[name = tensor<string, []>("attn_bias_67")];
            tensor<bool, [1, ?, ?]> var_1955 = less(x = delta_17, y = var_1834)[name = tensor<string, []>("op_1955")];
            tensor<bool, [1, ?, ?]> attn_bias_69 = logical_and(x = attn_bias_67, y = var_1955)[name = tensor<string, []>("attn_bias_69")];
            tensor<bool, [1, ?, ?]> var_1957 = logical_and(x = attn_bias_69, y = valid_k_17)[name = tensor<string, []>("op_1957")];
            tensor<int32, [1]> var_1959_axes_0 = const()[name = tensor<string, []>("op_1959_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_1959 = expand_dims(axes = var_1959_axes_0, x = var_1957)[name = tensor<string, []>("op_1959")];
            tensor<fp16, []> sub_8_x_0_to_fp16 = const()[name = tensor<string, []>("sub_8_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_217_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_217_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_1959_to_fp16 = cast(dtype = cast_217_to_fp16_dtype_0, x = var_1959)[name = tensor<string, []>("cast_322")];
            tensor<fp16, [1, 1, ?, ?]> sub_8_cast_fp16 = sub(x = sub_8_x_0_to_fp16, y = var_1959_to_fp16)[name = tensor<string, []>("sub_8_cast_fp16")];
            tensor<fp16, []> mul_16_x_0_to_fp16 = const()[name = tensor<string, []>("mul_16_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_16_cast_fp16 = mul(x = mul_16_x_0_to_fp16, y = sub_8_cast_fp16)[name = tensor<string, []>("mul_16_cast_fp16")];
            tensor<fp16, []> mul_17_y_0_to_fp16 = const()[name = tensor<string, []>("mul_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_17_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor<string, []>("mul_17_cast_fp16")];
            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = k_35_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_16_cast_fp16)[name = tensor<string, []>("add_8_cast_fp16")];
            tensor<string, []> add_8_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_8_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_8_cast_fp16_to_fp32 = cast(dtype = add_8_cast_fp16_to_fp32_dtype_0, x = add_8_cast_fp16)[name = tensor<string, []>("cast_321")];
            tensor<fp32, [1, 4, ?, ?]> softmax_8 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_8")];
            tensor<bool, []> out_33_transpose_x_0 = const()[name = tensor<string, []>("out_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_33_transpose_y_0 = const()[name = tensor<string, []>("out_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_8_to_fp16 = cast(dtype = softmax_8_to_fp16_dtype_0, x = softmax_8)[name = tensor<string, []>("cast_320")];
            tensor<fp16, [1, 4, ?, 64]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = softmax_8_to_fp16, y = v_17_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<int32, [2]> concat_113x = const()[name = tensor<string, []>("concat_113x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_1965 = reshape(shape = concat_113x, x = positions_17)[name = tensor<string, []>("op_1965")];
            tensor<int32, [2]> var_1966 = const()[name = tensor<string, []>("op_1966"), val = tensor<int32, [2]>([-1, 1])];
            tensor<int32, [1, 1]> var_1967 = reshape(shape = var_1966, x = input_lengths)[name = tensor<string, []>("op_1967")];
            tensor<bool, [1, ?]> var_1968 = less(x = var_1965, y = var_1967)[name = tensor<string, []>("op_1968")];
            tensor<int32, [4]> concat_114x = const()[name = tensor<string, []>("concat_114x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_17 = reshape(shape = concat_114x, x = var_1968)[name = tensor<string, []>("valid_q_17")];
            tensor<fp16, []> var_1972_to_fp16 = const()[name = tensor<string, []>("op_1972_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_35_cast_fp16 = select(a = out_33_cast_fp16, b = var_1972_to_fp16, cond = valid_q_17)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<int32, [4]> var_1974_perm_0 = const()[name = tensor<string, []>("op_1974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_115x = const()[name = tensor<string, []>("concat_115x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_1974_cast_fp16 = transpose(perm = var_1974_perm_0, x = out_35_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, ?, 256]> input_171_cast_fp16 = reshape(shape = concat_115x, x = var_1974_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_7_transformer_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16081536)))];
            tensor<fp16, [1, ?, 256]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_0_self_attn_out_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16212672)))];
            tensor<fp16, [1, ?, 256]> var_1980_cast_fp16 = mul(x = decoder_7_transformer_layers_0_layer_scale_1_scale_to_fp16, y = linear_40_cast_fp16)[name = tensor<string, []>("op_1980_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_173_cast_fp16 = add(x = linear_38_cast_fp16, y = var_1980_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<int32, [1]> input_175_axes_0 = const()[name = tensor<string, []>("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16213248)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16213824)))];
            tensor<fp16, [1, ?, 256]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = decoder_7_transformer_layers_0_norm2_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_0_norm2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_7_transformer_layers_0_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16214400)))];
            tensor<fp16, [1, ?, 1024]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_7_transformer_layers_0_ffn_0_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<string, []> input_179_mode_0 = const()[name = tensor<string, []>("input_179_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_41_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_7_transformer_layers_0_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16738752)))];
            tensor<fp16, [1, ?, 256]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_0_ffn_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_0_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_0_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17263104)))];
            tensor<fp16, [1, ?, 256]> var_1995_cast_fp16 = mul(x = decoder_7_transformer_layers_0_layer_scale_2_scale_to_fp16, y = linear_42_cast_fp16)[name = tensor<string, []>("op_1995_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_181_cast_fp16 = add(x = input_173_cast_fp16, y = var_1995_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<int32, [1]> x_79_axes_0 = const()[name = tensor<string, []>("x_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17263680)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17264256)))];
            tensor<fp16, [1, ?, 256]> x_79_cast_fp16 = layer_norm(axes = x_79_axes_0, beta = decoder_7_transformer_layers_1_norm1_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_1_norm1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<int32, [3]> var_2010_shape_cast_fp16 = shape(x = x_79_cast_fp16)[name = tensor<string, []>("op_2010_shape_cast_fp16")];
            tensor<int32, []> gather_103_axis_0 = const()[name = tensor<string, []>("gather_103_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_103_batch_dims_0 = const()[name = tensor<string, []>("gather_103_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_103_validate_indices_0 = const()[name = tensor<string, []>("gather_103_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2010_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2010_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_103_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_103_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_2010_shape_cast_fp16_to_uint16 = cast(dtype = var_2010_shape_cast_fp16_to_uint16_dtype_0, x = var_2010_shape_cast_fp16)[name = tensor<string, []>("cast_319")];
            tensor<uint16, []> gather_103_cast_uint16 = gather(axis = gather_103_axis_0, batch_dims = gather_103_batch_dims_0, indices = gather_103_indices_0_to_uint16, validate_indices = gather_103_validate_indices_0, x = var_2010_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_103_cast_uint16")];
            tensor<string, []> gather_103_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_103_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_7_transformer_layers_1_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17264832)))];
            tensor<fp16, [1, ?, 768]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_7_transformer_layers_1_self_attn_in_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<int32, [5]> concat_116x = const()[name = tensor<string, []>("concat_116x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_2017_cast_fp16 = reshape(shape = concat_116x, x = linear_43_cast_fp16)[name = tensor<string, []>("op_2017_cast_fp16")];
            tensor<int32, [5]> var_2018 = const()[name = tensor<string, []>("op_2018"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_37_begin_0 = const()[name = tensor<string, []>("q_37_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_37_end_0 = const()[name = tensor<string, []>("q_37_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_37_end_mask_0 = const()[name = tensor<string, []>("q_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_37_squeeze_mask_0 = const()[name = tensor<string, []>("q_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_39_cast_fp16 = transpose(perm = var_2018, x = var_2017_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 4, ?, 64]> q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = projected_39_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
            tensor<int32, [5]> k_37_begin_0 = const()[name = tensor<string, []>("k_37_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_37_end_0 = const()[name = tensor<string, []>("k_37_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_37_end_mask_0 = const()[name = tensor<string, []>("k_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_37_squeeze_mask_0 = const()[name = tensor<string, []>("k_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = projected_39_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
            tensor<int32, [5]> v_19_begin_0 = const()[name = tensor<string, []>("v_19_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_19_end_0 = const()[name = tensor<string, []>("v_19_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_19_end_mask_0 = const()[name = tensor<string, []>("v_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = projected_39_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
            tensor<int32, [4]> var_2030_shape_cast_fp16 = shape(x = q_37_cast_fp16)[name = tensor<string, []>("op_2030_shape_cast_fp16")];
            tensor<int32, []> gather_107_axis_0 = const()[name = tensor<string, []>("gather_107_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_107_batch_dims_0 = const()[name = tensor<string, []>("gather_107_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_107_validate_indices_0 = const()[name = tensor<string, []>("gather_107_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2030_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2030_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_107_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_107_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_2030_shape_cast_fp16_to_uint16 = cast(dtype = var_2030_shape_cast_fp16_to_uint16_dtype_0, x = var_2030_shape_cast_fp16)[name = tensor<string, []>("cast_317")];
            tensor<uint16, []> gather_107_cast_uint16 = gather(axis = gather_107_axis_0, batch_dims = gather_107_batch_dims_0, indices = gather_107_indices_0_to_uint16, validate_indices = gather_107_validate_indices_0, x = var_2030_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_107_cast_uint16")];
            tensor<string, []> gather_107_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_107_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_84 = const()[name = tensor<string, []>("const_84"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_85 = const()[name = tensor<string, []>("const_85"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_107_cast_uint16_to_int32 = cast(dtype = gather_107_cast_uint16_to_int32_dtype_0, x = gather_107_cast_uint16)[name = tensor<string, []>("cast_316")];
            tensor<int32, [?]> var_2034 = range_1d(end = gather_107_cast_uint16_to_int32, start = const_84, step = const_85)[name = tensor<string, []>("op_2034")];
            tensor<fp16, [1, 1]> var_2033_to_fp16 = const()[name = tensor<string, []>("op_2033_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_2034_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2034_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_2034_to_fp16 = cast(dtype = var_2034_promoted_to_fp16_dtype_0, x = var_2034)[name = tensor<string, []>("cast_315")];
            tensor<fp16, [1, ?]> ts_37_cast_fp16 = add(x = var_2033_to_fp16, y = var_2034_to_fp16)[name = tensor<string, []>("ts_37_cast_fp16")];
            tensor<int32, [4]> var_2037 = const()[name = tensor<string, []>("op_2037"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_39_cast_fp16 = reshape(shape = var_2037, x = ts_37_cast_fp16)[name = tensor<string, []>("ts_39_cast_fp16")];
            tensor<int32, [5]> concat_117x = const()[name = tensor<string, []>("concat_117x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_19_cast_fp16 = reshape(shape = concat_117x, x = q_37_cast_fp16)[name = tensor<string, []>("q2_19_cast_fp16")];
            tensor<int32, [5]> concat_118x = const()[name = tensor<string, []>("concat_118x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_19_cast_fp16 = reshape(shape = concat_118x, x = k_37_cast_fp16)[name = tensor<string, []>("k2_19_cast_fp16")];
            tensor<int32, [5]> qr_19_begin_0 = const()[name = tensor<string, []>("qr_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_19_end_0 = const()[name = tensor<string, []>("qr_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_19_end_mask_0 = const()[name = tensor<string, []>("qr_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_19_squeeze_mask_0 = const()[name = tensor<string, []>("qr_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_19_cast_fp16 = slice_by_index(begin = qr_19_begin_0, end = qr_19_end_0, end_mask = qr_19_end_mask_0, squeeze_mask = qr_19_squeeze_mask_0, x = q2_19_cast_fp16)[name = tensor<string, []>("qr_19_cast_fp16")];
            tensor<int32, [5]> qi_19_begin_0 = const()[name = tensor<string, []>("qi_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_19_end_0 = const()[name = tensor<string, []>("qi_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_19_end_mask_0 = const()[name = tensor<string, []>("qi_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_19_squeeze_mask_0 = const()[name = tensor<string, []>("qi_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_19_cast_fp16 = slice_by_index(begin = qi_19_begin_0, end = qi_19_end_0, end_mask = qi_19_end_mask_0, squeeze_mask = qi_19_squeeze_mask_0, x = q2_19_cast_fp16)[name = tensor<string, []>("qi_19_cast_fp16")];
            tensor<int32, [5]> kr_19_begin_0 = const()[name = tensor<string, []>("kr_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_19_end_0 = const()[name = tensor<string, []>("kr_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_19_end_mask_0 = const()[name = tensor<string, []>("kr_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_19_squeeze_mask_0 = const()[name = tensor<string, []>("kr_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_19_cast_fp16 = slice_by_index(begin = kr_19_begin_0, end = kr_19_end_0, end_mask = kr_19_end_mask_0, squeeze_mask = kr_19_squeeze_mask_0, x = k2_19_cast_fp16)[name = tensor<string, []>("kr_19_cast_fp16")];
            tensor<int32, [5]> ki_19_begin_0 = const()[name = tensor<string, []>("ki_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_19_end_0 = const()[name = tensor<string, []>("ki_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_19_end_mask_0 = const()[name = tensor<string, []>("ki_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_19_squeeze_mask_0 = const()[name = tensor<string, []>("ki_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_19_cast_fp16 = slice_by_index(begin = ki_19_begin_0, end = ki_19_end_0, end_mask = ki_19_end_mask_0, squeeze_mask = ki_19_squeeze_mask_0, x = k2_19_cast_fp16)[name = tensor<string, []>("ki_19_cast_fp16")];
            tensor<fp16, [32]> freqs_19_to_fp16 = const()[name = tensor<string, []>("freqs_19_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17658112)))];
            tensor<fp16, [1, 1, ?, 32]> var_2050_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_39_cast_fp16)[name = tensor<string, []>("op_2050_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_19_cast_fp16 = cos(x = var_2050_cast_fp16)[name = tensor<string, []>("rotr_19_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_19_cast_fp16 = sin(x = var_2050_cast_fp16)[name = tensor<string, []>("roti_19_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2054_cast_fp16 = mul(x = qr_19_cast_fp16, y = rotr_19_cast_fp16)[name = tensor<string, []>("op_2054_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2055_cast_fp16 = mul(x = qi_19_cast_fp16, y = roti_19_cast_fp16)[name = tensor<string, []>("op_2055_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2056_cast_fp16 = sub(x = var_2054_cast_fp16, y = var_2055_cast_fp16)[name = tensor<string, []>("op_2056_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2057_cast_fp16 = mul(x = qr_19_cast_fp16, y = roti_19_cast_fp16)[name = tensor<string, []>("op_2057_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2058_cast_fp16 = mul(x = qi_19_cast_fp16, y = rotr_19_cast_fp16)[name = tensor<string, []>("op_2058_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2059_cast_fp16 = add(x = var_2057_cast_fp16, y = var_2058_cast_fp16)[name = tensor<string, []>("op_2059_cast_fp16")];
            tensor<int32, []> qo_19_axis_0 = const()[name = tensor<string, []>("qo_19_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (var_2056_cast_fp16, var_2059_cast_fp16))[name = tensor<string, []>("qo_19_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2062_cast_fp16 = mul(x = kr_19_cast_fp16, y = rotr_19_cast_fp16)[name = tensor<string, []>("op_2062_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2063_cast_fp16 = mul(x = ki_19_cast_fp16, y = roti_19_cast_fp16)[name = tensor<string, []>("op_2063_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2064_cast_fp16 = sub(x = var_2062_cast_fp16, y = var_2063_cast_fp16)[name = tensor<string, []>("op_2064_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2065_cast_fp16 = mul(x = kr_19_cast_fp16, y = roti_19_cast_fp16)[name = tensor<string, []>("op_2065_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2066_cast_fp16 = mul(x = ki_19_cast_fp16, y = rotr_19_cast_fp16)[name = tensor<string, []>("op_2066_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2067_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2066_cast_fp16)[name = tensor<string, []>("op_2067_cast_fp16")];
            tensor<int32, []> ko_19_axis_0 = const()[name = tensor<string, []>("ko_19_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (var_2064_cast_fp16, var_2067_cast_fp16))[name = tensor<string, []>("ko_19_cast_fp16")];
            tensor<int32, [4]> concat_119x = const()[name = tensor<string, []>("concat_119x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_39_cast_fp16 = reshape(shape = concat_119x, x = qo_19_cast_fp16)[name = tensor<string, []>("q_39_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = tensor<string, []>("concat_120x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_39_cast_fp16 = reshape(shape = concat_120x, x = ko_19_cast_fp16)[name = tensor<string, []>("k_39_cast_fp16")];
            tensor<int32, []> const_86 = const()[name = tensor<string, []>("const_86"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_87 = const()[name = tensor<string, []>("const_87"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_103_cast_uint16_to_int32 = cast(dtype = gather_103_cast_uint16_to_int32_dtype_0, x = gather_103_cast_uint16)[name = tensor<string, []>("cast_318")];
            tensor<int32, [?]> positions_19 = range_1d(end = gather_103_cast_uint16_to_int32, start = const_86, step = const_87)[name = tensor<string, []>("positions_19")];
            tensor<int32, [3]> concat_121x = const()[name = tensor<string, []>("concat_121x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_2079 = reshape(shape = concat_121x, x = positions_19)[name = tensor<string, []>("op_2079")];
            tensor<bool, [1, 1, ?]> valid_k_19 = less(x = var_2079, y = var_1944)[name = tensor<string, []>("valid_k_19")];
            tensor<int32, [3]> concat_122x = const()[name = tensor<string, []>("concat_122x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_2084 = reshape(shape = concat_122x, x = positions_19)[name = tensor<string, []>("op_2084")];
            tensor<int32, [1, ?, ?]> delta_19 = sub(x = var_2084, y = var_2079)[name = tensor<string, []>("delta_19")];
            tensor<int32, []> concat_124_axis_0 = const()[name = tensor<string, []>("concat_124_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_124_interleave_0 = const()[name = tensor<string, []>("concat_124_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_124 = concat(axis = concat_124_axis_0, interleave = concat_124_interleave_0, values = (var_1845, gather_103_cast_uint16_to_int32, gather_103_cast_uint16_to_int32))[name = tensor<string, []>("concat_124")];
            tensor<fp16, []> fill_9_value_0_to_fp16 = const()[name = tensor<string, []>("fill_9_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_9_cast_fp16 = fill(shape = concat_124, value = fill_9_value_0_to_fp16)[name = tensor<string, []>("fill_9_cast_fp16")];
            tensor<string, []> attn_bias_73_dtype_0 = const()[name = tensor<string, []>("attn_bias_73_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_2090 = greater_equal(x = delta_19, y = var_1827)[name = tensor<string, []>("op_2090")];
            tensor<bool, [1, ?, ?]> fill_9_cast_fp16_to_bool = cast(dtype = attn_bias_73_dtype_0, x = fill_9_cast_fp16)[name = tensor<string, []>("cast_314")];
            tensor<bool, [1, ?, ?]> attn_bias_75 = logical_and(x = fill_9_cast_fp16_to_bool, y = var_2090)[name = tensor<string, []>("attn_bias_75")];
            tensor<bool, [1, ?, ?]> var_2092 = less(x = delta_19, y = var_1834)[name = tensor<string, []>("op_2092")];
            tensor<bool, [1, ?, ?]> attn_bias_77 = logical_and(x = attn_bias_75, y = var_2092)[name = tensor<string, []>("attn_bias_77")];
            tensor<bool, [1, ?, ?]> var_2094 = logical_and(x = attn_bias_77, y = valid_k_19)[name = tensor<string, []>("op_2094")];
            tensor<int32, [1]> var_2096_axes_0 = const()[name = tensor<string, []>("op_2096_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_2096 = expand_dims(axes = var_2096_axes_0, x = var_2094)[name = tensor<string, []>("op_2096")];
            tensor<fp16, []> sub_9_x_0_to_fp16 = const()[name = tensor<string, []>("sub_9_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_239_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_239_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_2096_to_fp16 = cast(dtype = cast_239_to_fp16_dtype_0, x = var_2096)[name = tensor<string, []>("cast_313")];
            tensor<fp16, [1, 1, ?, ?]> sub_9_cast_fp16 = sub(x = sub_9_x_0_to_fp16, y = var_2096_to_fp16)[name = tensor<string, []>("sub_9_cast_fp16")];
            tensor<fp16, []> mul_18_x_0_to_fp16 = const()[name = tensor<string, []>("mul_18_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_18_cast_fp16 = mul(x = mul_18_x_0_to_fp16, y = sub_9_cast_fp16)[name = tensor<string, []>("mul_18_cast_fp16")];
            tensor<fp16, []> mul_19_y_0_to_fp16 = const()[name = tensor<string, []>("mul_19_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_19_cast_fp16 = mul(x = q_39_cast_fp16, y = mul_19_y_0_to_fp16)[name = tensor<string, []>("mul_19_cast_fp16")];
            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = k_39_cast_fp16)[name = tensor<string, []>("matmul_9_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_18_cast_fp16)[name = tensor<string, []>("add_9_cast_fp16")];
            tensor<string, []> add_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_9_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_9_cast_fp16_to_fp32 = cast(dtype = add_9_cast_fp16_to_fp32_dtype_0, x = add_9_cast_fp16)[name = tensor<string, []>("cast_312")];
            tensor<fp32, [1, 4, ?, ?]> softmax_9 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_9")];
            tensor<bool, []> out_37_transpose_x_0 = const()[name = tensor<string, []>("out_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_37_transpose_y_0 = const()[name = tensor<string, []>("out_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_9_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_9_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_9_to_fp16 = cast(dtype = softmax_9_to_fp16_dtype_0, x = softmax_9)[name = tensor<string, []>("cast_311")];
            tensor<fp16, [1, 4, ?, 64]> out_37_cast_fp16 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = softmax_9_to_fp16, y = v_19_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<int32, [2]> concat_125x = const()[name = tensor<string, []>("concat_125x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_2102 = reshape(shape = concat_125x, x = positions_19)[name = tensor<string, []>("op_2102")];
            tensor<bool, [1, ?]> var_2105 = less(x = var_2102, y = var_1967)[name = tensor<string, []>("op_2105")];
            tensor<int32, [4]> concat_126x = const()[name = tensor<string, []>("concat_126x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_19 = reshape(shape = concat_126x, x = var_2105)[name = tensor<string, []>("valid_q_19")];
            tensor<fp16, []> var_2109_to_fp16 = const()[name = tensor<string, []>("op_2109_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_39_cast_fp16 = select(a = out_37_cast_fp16, b = var_2109_to_fp16, cond = valid_q_19)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<int32, [4]> var_2111_perm_0 = const()[name = tensor<string, []>("op_2111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_127x = const()[name = tensor<string, []>("concat_127x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_2111_cast_fp16 = transpose(perm = var_2111_perm_0, x = out_39_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, ?, 256]> input_183_cast_fp16 = reshape(shape = concat_127x, x = var_2111_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_7_transformer_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17658240)))];
            tensor<fp16, [1, ?, 256]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_1_self_attn_out_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17789376)))];
            tensor<fp16, [1, ?, 256]> var_2117_cast_fp16 = mul(x = decoder_7_transformer_layers_1_layer_scale_1_scale_to_fp16, y = linear_44_cast_fp16)[name = tensor<string, []>("op_2117_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_185_cast_fp16 = add(x = input_181_cast_fp16, y = var_2117_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [1]> input_187_axes_0 = const()[name = tensor<string, []>("input_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17789952)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17790528)))];
            tensor<fp16, [1, ?, 256]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = decoder_7_transformer_layers_1_norm2_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_1_norm2_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_7_transformer_layers_1_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17791104)))];
            tensor<fp16, [1, ?, 1024]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_7_transformer_layers_1_ffn_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<string, []> input_191_mode_0 = const()[name = tensor<string, []>("input_191_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_45_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_7_transformer_layers_1_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18315456)))];
            tensor<fp16, [1, ?, 256]> linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_1_ffn_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_1_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_1_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18839808)))];
            tensor<fp16, [1, ?, 256]> var_2132_cast_fp16 = mul(x = decoder_7_transformer_layers_1_layer_scale_2_scale_to_fp16, y = linear_46_cast_fp16)[name = tensor<string, []>("op_2132_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_193_cast_fp16 = add(x = input_185_cast_fp16, y = var_2132_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<int32, [1]> x_85_axes_0 = const()[name = tensor<string, []>("x_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18840384)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18840960)))];
            tensor<fp16, [1, ?, 256]> x_85_cast_fp16 = layer_norm(axes = x_85_axes_0, beta = decoder_7_transformer_layers_2_norm1_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_2_norm1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<int32, [3]> var_2147_shape_cast_fp16 = shape(x = x_85_cast_fp16)[name = tensor<string, []>("op_2147_shape_cast_fp16")];
            tensor<int32, []> gather_113_axis_0 = const()[name = tensor<string, []>("gather_113_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_113_batch_dims_0 = const()[name = tensor<string, []>("gather_113_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_113_validate_indices_0 = const()[name = tensor<string, []>("gather_113_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2147_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2147_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_113_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_113_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_2147_shape_cast_fp16_to_uint16 = cast(dtype = var_2147_shape_cast_fp16_to_uint16_dtype_0, x = var_2147_shape_cast_fp16)[name = tensor<string, []>("cast_310")];
            tensor<uint16, []> gather_113_cast_uint16 = gather(axis = gather_113_axis_0, batch_dims = gather_113_batch_dims_0, indices = gather_113_indices_0_to_uint16, validate_indices = gather_113_validate_indices_0, x = var_2147_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_113_cast_uint16")];
            tensor<string, []> gather_113_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_113_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_7_transformer_layers_2_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18841536)))];
            tensor<fp16, [1, ?, 768]> linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_7_transformer_layers_2_self_attn_in_proj_weight_to_fp16, x = x_85_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<int32, [5]> concat_128x = const()[name = tensor<string, []>("concat_128x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_2154_cast_fp16 = reshape(shape = concat_128x, x = linear_47_cast_fp16)[name = tensor<string, []>("op_2154_cast_fp16")];
            tensor<int32, [5]> var_2155 = const()[name = tensor<string, []>("op_2155"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_41_begin_0 = const()[name = tensor<string, []>("q_41_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_41_end_0 = const()[name = tensor<string, []>("q_41_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_41_end_mask_0 = const()[name = tensor<string, []>("q_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_41_squeeze_mask_0 = const()[name = tensor<string, []>("q_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_43_cast_fp16 = transpose(perm = var_2155, x = var_2154_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 4, ?, 64]> q_41_cast_fp16 = slice_by_index(begin = q_41_begin_0, end = q_41_end_0, end_mask = q_41_end_mask_0, squeeze_mask = q_41_squeeze_mask_0, x = projected_43_cast_fp16)[name = tensor<string, []>("q_41_cast_fp16")];
            tensor<int32, [5]> k_41_begin_0 = const()[name = tensor<string, []>("k_41_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_41_end_0 = const()[name = tensor<string, []>("k_41_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_41_end_mask_0 = const()[name = tensor<string, []>("k_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_41_squeeze_mask_0 = const()[name = tensor<string, []>("k_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = projected_43_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
            tensor<int32, [5]> v_21_begin_0 = const()[name = tensor<string, []>("v_21_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_21_end_0 = const()[name = tensor<string, []>("v_21_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_21_end_mask_0 = const()[name = tensor<string, []>("v_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_21_squeeze_mask_0 = const()[name = tensor<string, []>("v_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = projected_43_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
            tensor<int32, [4]> var_2167_shape_cast_fp16 = shape(x = q_41_cast_fp16)[name = tensor<string, []>("op_2167_shape_cast_fp16")];
            tensor<int32, []> gather_117_axis_0 = const()[name = tensor<string, []>("gather_117_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_117_batch_dims_0 = const()[name = tensor<string, []>("gather_117_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_117_validate_indices_0 = const()[name = tensor<string, []>("gather_117_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2167_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2167_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_117_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_117_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_2167_shape_cast_fp16_to_uint16 = cast(dtype = var_2167_shape_cast_fp16_to_uint16_dtype_0, x = var_2167_shape_cast_fp16)[name = tensor<string, []>("cast_308")];
            tensor<uint16, []> gather_117_cast_uint16 = gather(axis = gather_117_axis_0, batch_dims = gather_117_batch_dims_0, indices = gather_117_indices_0_to_uint16, validate_indices = gather_117_validate_indices_0, x = var_2167_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_117_cast_uint16")];
            tensor<string, []> gather_117_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_117_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_93 = const()[name = tensor<string, []>("const_93"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_94 = const()[name = tensor<string, []>("const_94"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_117_cast_uint16_to_int32 = cast(dtype = gather_117_cast_uint16_to_int32_dtype_0, x = gather_117_cast_uint16)[name = tensor<string, []>("cast_307")];
            tensor<int32, [?]> var_2171 = range_1d(end = gather_117_cast_uint16_to_int32, start = const_93, step = const_94)[name = tensor<string, []>("op_2171")];
            tensor<fp16, [1, 1]> var_2170_to_fp16 = const()[name = tensor<string, []>("op_2170_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_2171_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2171_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_2171_to_fp16 = cast(dtype = var_2171_promoted_to_fp16_dtype_0, x = var_2171)[name = tensor<string, []>("cast_306")];
            tensor<fp16, [1, ?]> ts_41_cast_fp16 = add(x = var_2170_to_fp16, y = var_2171_to_fp16)[name = tensor<string, []>("ts_41_cast_fp16")];
            tensor<int32, [4]> var_2174 = const()[name = tensor<string, []>("op_2174"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_43_cast_fp16 = reshape(shape = var_2174, x = ts_41_cast_fp16)[name = tensor<string, []>("ts_43_cast_fp16")];
            tensor<int32, [5]> concat_129x = const()[name = tensor<string, []>("concat_129x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_21_cast_fp16 = reshape(shape = concat_129x, x = q_41_cast_fp16)[name = tensor<string, []>("q2_21_cast_fp16")];
            tensor<int32, [5]> concat_130x = const()[name = tensor<string, []>("concat_130x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_21_cast_fp16 = reshape(shape = concat_130x, x = k_41_cast_fp16)[name = tensor<string, []>("k2_21_cast_fp16")];
            tensor<int32, [5]> qr_21_begin_0 = const()[name = tensor<string, []>("qr_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_21_end_0 = const()[name = tensor<string, []>("qr_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_21_end_mask_0 = const()[name = tensor<string, []>("qr_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_21_squeeze_mask_0 = const()[name = tensor<string, []>("qr_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_21_cast_fp16 = slice_by_index(begin = qr_21_begin_0, end = qr_21_end_0, end_mask = qr_21_end_mask_0, squeeze_mask = qr_21_squeeze_mask_0, x = q2_21_cast_fp16)[name = tensor<string, []>("qr_21_cast_fp16")];
            tensor<int32, [5]> qi_21_begin_0 = const()[name = tensor<string, []>("qi_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_21_end_0 = const()[name = tensor<string, []>("qi_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_21_end_mask_0 = const()[name = tensor<string, []>("qi_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_21_squeeze_mask_0 = const()[name = tensor<string, []>("qi_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_21_cast_fp16 = slice_by_index(begin = qi_21_begin_0, end = qi_21_end_0, end_mask = qi_21_end_mask_0, squeeze_mask = qi_21_squeeze_mask_0, x = q2_21_cast_fp16)[name = tensor<string, []>("qi_21_cast_fp16")];
            tensor<int32, [5]> kr_21_begin_0 = const()[name = tensor<string, []>("kr_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_21_end_0 = const()[name = tensor<string, []>("kr_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_21_end_mask_0 = const()[name = tensor<string, []>("kr_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_21_squeeze_mask_0 = const()[name = tensor<string, []>("kr_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_21_cast_fp16 = slice_by_index(begin = kr_21_begin_0, end = kr_21_end_0, end_mask = kr_21_end_mask_0, squeeze_mask = kr_21_squeeze_mask_0, x = k2_21_cast_fp16)[name = tensor<string, []>("kr_21_cast_fp16")];
            tensor<int32, [5]> ki_21_begin_0 = const()[name = tensor<string, []>("ki_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_21_end_0 = const()[name = tensor<string, []>("ki_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_21_end_mask_0 = const()[name = tensor<string, []>("ki_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_21_squeeze_mask_0 = const()[name = tensor<string, []>("ki_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_21_cast_fp16 = slice_by_index(begin = ki_21_begin_0, end = ki_21_end_0, end_mask = ki_21_end_mask_0, squeeze_mask = ki_21_squeeze_mask_0, x = k2_21_cast_fp16)[name = tensor<string, []>("ki_21_cast_fp16")];
            tensor<fp16, [32]> freqs_21_to_fp16 = const()[name = tensor<string, []>("freqs_21_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19234816)))];
            tensor<fp16, [1, 1, ?, 32]> var_2187_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_43_cast_fp16)[name = tensor<string, []>("op_2187_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_21_cast_fp16 = cos(x = var_2187_cast_fp16)[name = tensor<string, []>("rotr_21_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_21_cast_fp16 = sin(x = var_2187_cast_fp16)[name = tensor<string, []>("roti_21_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2191_cast_fp16 = mul(x = qr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor<string, []>("op_2191_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2192_cast_fp16 = mul(x = qi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor<string, []>("op_2192_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2193_cast_fp16 = sub(x = var_2191_cast_fp16, y = var_2192_cast_fp16)[name = tensor<string, []>("op_2193_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2194_cast_fp16 = mul(x = qr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor<string, []>("op_2194_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2195_cast_fp16 = mul(x = qi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor<string, []>("op_2195_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2196_cast_fp16 = add(x = var_2194_cast_fp16, y = var_2195_cast_fp16)[name = tensor<string, []>("op_2196_cast_fp16")];
            tensor<int32, []> qo_21_axis_0 = const()[name = tensor<string, []>("qo_21_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (var_2193_cast_fp16, var_2196_cast_fp16))[name = tensor<string, []>("qo_21_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2199_cast_fp16 = mul(x = kr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor<string, []>("op_2199_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2200_cast_fp16 = mul(x = ki_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor<string, []>("op_2200_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2201_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor<string, []>("op_2201_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2202_cast_fp16 = mul(x = kr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor<string, []>("op_2202_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2203_cast_fp16 = mul(x = ki_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor<string, []>("op_2203_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2204_cast_fp16 = add(x = var_2202_cast_fp16, y = var_2203_cast_fp16)[name = tensor<string, []>("op_2204_cast_fp16")];
            tensor<int32, []> ko_21_axis_0 = const()[name = tensor<string, []>("ko_21_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (var_2201_cast_fp16, var_2204_cast_fp16))[name = tensor<string, []>("ko_21_cast_fp16")];
            tensor<int32, [4]> concat_131x = const()[name = tensor<string, []>("concat_131x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_43_cast_fp16 = reshape(shape = concat_131x, x = qo_21_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = tensor<string, []>("concat_132x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_43_cast_fp16 = reshape(shape = concat_132x, x = ko_21_cast_fp16)[name = tensor<string, []>("k_43_cast_fp16")];
            tensor<int32, []> const_95 = const()[name = tensor<string, []>("const_95"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_96 = const()[name = tensor<string, []>("const_96"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_113_cast_uint16_to_int32 = cast(dtype = gather_113_cast_uint16_to_int32_dtype_0, x = gather_113_cast_uint16)[name = tensor<string, []>("cast_309")];
            tensor<int32, [?]> positions_21 = range_1d(end = gather_113_cast_uint16_to_int32, start = const_95, step = const_96)[name = tensor<string, []>("positions_21")];
            tensor<int32, [3]> concat_133x = const()[name = tensor<string, []>("concat_133x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_2216 = reshape(shape = concat_133x, x = positions_21)[name = tensor<string, []>("op_2216")];
            tensor<bool, [1, 1, ?]> valid_k_21 = less(x = var_2216, y = var_1944)[name = tensor<string, []>("valid_k_21")];
            tensor<int32, [3]> concat_134x = const()[name = tensor<string, []>("concat_134x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_2221 = reshape(shape = concat_134x, x = positions_21)[name = tensor<string, []>("op_2221")];
            tensor<int32, [1, ?, ?]> delta_21 = sub(x = var_2221, y = var_2216)[name = tensor<string, []>("delta_21")];
            tensor<int32, []> concat_136_axis_0 = const()[name = tensor<string, []>("concat_136_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_136_interleave_0 = const()[name = tensor<string, []>("concat_136_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_136 = concat(axis = concat_136_axis_0, interleave = concat_136_interleave_0, values = (var_1845, gather_113_cast_uint16_to_int32, gather_113_cast_uint16_to_int32))[name = tensor<string, []>("concat_136")];
            tensor<fp16, []> fill_10_value_0_to_fp16 = const()[name = tensor<string, []>("fill_10_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_10_cast_fp16 = fill(shape = concat_136, value = fill_10_value_0_to_fp16)[name = tensor<string, []>("fill_10_cast_fp16")];
            tensor<string, []> attn_bias_81_dtype_0 = const()[name = tensor<string, []>("attn_bias_81_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_2227 = greater_equal(x = delta_21, y = var_1827)[name = tensor<string, []>("op_2227")];
            tensor<bool, [1, ?, ?]> fill_10_cast_fp16_to_bool = cast(dtype = attn_bias_81_dtype_0, x = fill_10_cast_fp16)[name = tensor<string, []>("cast_305")];
            tensor<bool, [1, ?, ?]> attn_bias_83 = logical_and(x = fill_10_cast_fp16_to_bool, y = var_2227)[name = tensor<string, []>("attn_bias_83")];
            tensor<bool, [1, ?, ?]> var_2229 = less(x = delta_21, y = var_1834)[name = tensor<string, []>("op_2229")];
            tensor<bool, [1, ?, ?]> attn_bias_85 = logical_and(x = attn_bias_83, y = var_2229)[name = tensor<string, []>("attn_bias_85")];
            tensor<bool, [1, ?, ?]> var_2231 = logical_and(x = attn_bias_85, y = valid_k_21)[name = tensor<string, []>("op_2231")];
            tensor<int32, [1]> var_2233_axes_0 = const()[name = tensor<string, []>("op_2233_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_2233 = expand_dims(axes = var_2233_axes_0, x = var_2231)[name = tensor<string, []>("op_2233")];
            tensor<fp16, []> sub_10_x_0_to_fp16 = const()[name = tensor<string, []>("sub_10_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_261_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_261_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_2233_to_fp16 = cast(dtype = cast_261_to_fp16_dtype_0, x = var_2233)[name = tensor<string, []>("cast_304")];
            tensor<fp16, [1, 1, ?, ?]> sub_10_cast_fp16 = sub(x = sub_10_x_0_to_fp16, y = var_2233_to_fp16)[name = tensor<string, []>("sub_10_cast_fp16")];
            tensor<fp16, []> mul_20_x_0_to_fp16 = const()[name = tensor<string, []>("mul_20_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_20_cast_fp16 = mul(x = mul_20_x_0_to_fp16, y = sub_10_cast_fp16)[name = tensor<string, []>("mul_20_cast_fp16")];
            tensor<fp16, []> mul_21_y_0_to_fp16 = const()[name = tensor<string, []>("mul_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_21_cast_fp16 = mul(x = q_43_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor<string, []>("mul_21_cast_fp16")];
            tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = k_43_cast_fp16)[name = tensor<string, []>("matmul_10_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_20_cast_fp16)[name = tensor<string, []>("add_10_cast_fp16")];
            tensor<string, []> add_10_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_10_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_10_cast_fp16_to_fp32 = cast(dtype = add_10_cast_fp16_to_fp32_dtype_0, x = add_10_cast_fp16)[name = tensor<string, []>("cast_303")];
            tensor<fp32, [1, 4, ?, ?]> softmax_10 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_10")];
            tensor<bool, []> out_41_transpose_x_0 = const()[name = tensor<string, []>("out_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_41_transpose_y_0 = const()[name = tensor<string, []>("out_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_10_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_10_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_10_to_fp16 = cast(dtype = softmax_10_to_fp16_dtype_0, x = softmax_10)[name = tensor<string, []>("cast_302")];
            tensor<fp16, [1, 4, ?, 64]> out_41_cast_fp16 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = softmax_10_to_fp16, y = v_21_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<int32, [2]> concat_137x = const()[name = tensor<string, []>("concat_137x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_2239 = reshape(shape = concat_137x, x = positions_21)[name = tensor<string, []>("op_2239")];
            tensor<bool, [1, ?]> var_2242 = less(x = var_2239, y = var_1967)[name = tensor<string, []>("op_2242")];
            tensor<int32, [4]> concat_138x = const()[name = tensor<string, []>("concat_138x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q_21 = reshape(shape = concat_138x, x = var_2242)[name = tensor<string, []>("valid_q_21")];
            tensor<fp16, []> var_2246_to_fp16 = const()[name = tensor<string, []>("op_2246_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_43_cast_fp16 = select(a = out_41_cast_fp16, b = var_2246_to_fp16, cond = valid_q_21)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<int32, [4]> var_2248_perm_0 = const()[name = tensor<string, []>("op_2248_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_139x = const()[name = tensor<string, []>("concat_139x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_2248_cast_fp16 = transpose(perm = var_2248_perm_0, x = out_43_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, ?, 256]> input_195_cast_fp16 = reshape(shape = concat_139x, x = var_2248_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_7_transformer_layers_2_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19234944)))];
            tensor<fp16, [1, ?, 256]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_2_self_attn_out_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19366080)))];
            tensor<fp16, [1, ?, 256]> var_2254_cast_fp16 = mul(x = decoder_7_transformer_layers_2_layer_scale_1_scale_to_fp16, y = linear_48_cast_fp16)[name = tensor<string, []>("op_2254_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_197_cast_fp16 = add(x = input_193_cast_fp16, y = var_2254_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<int32, [1]> input_199_axes_0 = const()[name = tensor<string, []>("input_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19366656)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19367232)))];
            tensor<fp16, [1, ?, 256]> input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = decoder_7_transformer_layers_2_norm2_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_2_norm2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_7_transformer_layers_2_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19367808)))];
            tensor<fp16, [1, ?, 1024]> linear_49_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_7_transformer_layers_2_ffn_0_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<string, []> input_203_mode_0 = const()[name = tensor<string, []>("input_203_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_49_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_7_transformer_layers_2_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19892160)))];
            tensor<fp16, [1, ?, 256]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_2_ffn_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_2_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_2_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20416512)))];
            tensor<fp16, [1, ?, 256]> var_2269_cast_fp16 = mul(x = decoder_7_transformer_layers_2_layer_scale_2_scale_to_fp16, y = linear_50_cast_fp16)[name = tensor<string, []>("op_2269_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_205_cast_fp16 = add(x = input_197_cast_fp16, y = var_2269_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<int32, [1]> x_91_axes_0 = const()[name = tensor<string, []>("x_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_norm1_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20417088)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_norm1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20417664)))];
            tensor<fp16, [1, ?, 256]> x_91_cast_fp16 = layer_norm(axes = x_91_axes_0, beta = decoder_7_transformer_layers_3_norm1_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_3_norm1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<int32, [3]> var_2284_shape_cast_fp16 = shape(x = x_91_cast_fp16)[name = tensor<string, []>("op_2284_shape_cast_fp16")];
            tensor<int32, []> gather_123_axis_0 = const()[name = tensor<string, []>("gather_123_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_123_batch_dims_0 = const()[name = tensor<string, []>("gather_123_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_123_validate_indices_0 = const()[name = tensor<string, []>("gather_123_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2284_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2284_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_123_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_123_indices_0_to_uint16"), val = tensor<uint16, []>(1)];
            tensor<uint16, [3]> var_2284_shape_cast_fp16_to_uint16 = cast(dtype = var_2284_shape_cast_fp16_to_uint16_dtype_0, x = var_2284_shape_cast_fp16)[name = tensor<string, []>("cast_301")];
            tensor<uint16, []> gather_123_cast_uint16 = gather(axis = gather_123_axis_0, batch_dims = gather_123_batch_dims_0, indices = gather_123_indices_0_to_uint16, validate_indices = gather_123_validate_indices_0, x = var_2284_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_123_cast_uint16")];
            tensor<string, []> gather_123_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_123_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<fp16, [768, 256]> decoder_7_transformer_layers_3_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20418240)))];
            tensor<fp16, [1, ?, 768]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = decoder_7_transformer_layers_3_self_attn_in_proj_weight_to_fp16, x = x_91_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<int32, [5]> concat_140x = const()[name = tensor<string, []>("concat_140x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])];
            tensor<fp16, [1, ?, 3, 4, 64]> var_2291_cast_fp16 = reshape(shape = concat_140x, x = linear_51_cast_fp16)[name = tensor<string, []>("op_2291_cast_fp16")];
            tensor<int32, [5]> var_2292 = const()[name = tensor<string, []>("op_2292"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [5]> q_45_begin_0 = const()[name = tensor<string, []>("q_45_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_45_end_0 = const()[name = tensor<string, []>("q_45_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])];
            tensor<bool, [5]> q_45_end_mask_0 = const()[name = tensor<string, []>("q_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> q_45_squeeze_mask_0 = const()[name = tensor<string, []>("q_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [3, 1, 4, ?, 64]> projected_cast_fp16 = transpose(perm = var_2292, x = var_2291_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 4, ?, 64]> q_45_cast_fp16 = slice_by_index(begin = q_45_begin_0, end = q_45_end_0, end_mask = q_45_end_mask_0, squeeze_mask = q_45_squeeze_mask_0, x = projected_cast_fp16)[name = tensor<string, []>("q_45_cast_fp16")];
            tensor<int32, [5]> k_45_begin_0 = const()[name = tensor<string, []>("k_45_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_45_end_0 = const()[name = tensor<string, []>("k_45_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])];
            tensor<bool, [5]> k_45_end_mask_0 = const()[name = tensor<string, []>("k_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_45_squeeze_mask_0 = const()[name = tensor<string, []>("k_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = projected_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
            tensor<int32, [5]> v_begin_0 = const()[name = tensor<string, []>("v_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_end_0 = const()[name = tensor<string, []>("v_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])];
            tensor<bool, [5]> v_end_mask_0 = const()[name = tensor<string, []>("v_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_squeeze_mask_0 = const()[name = tensor<string, []>("v_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, ?, 64]> v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = projected_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
            tensor<int32, [4]> var_2304_shape_cast_fp16 = shape(x = q_45_cast_fp16)[name = tensor<string, []>("op_2304_shape_cast_fp16")];
            tensor<int32, []> gather_127_axis_0 = const()[name = tensor<string, []>("gather_127_axis_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> gather_127_batch_dims_0 = const()[name = tensor<string, []>("gather_127_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> gather_127_validate_indices_0 = const()[name = tensor<string, []>("gather_127_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<string, []> var_2304_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_2304_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
            tensor<uint16, []> gather_127_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_127_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
            tensor<uint16, [4]> var_2304_shape_cast_fp16_to_uint16 = cast(dtype = var_2304_shape_cast_fp16_to_uint16_dtype_0, x = var_2304_shape_cast_fp16)[name = tensor<string, []>("cast_299")];
            tensor<uint16, []> gather_127_cast_uint16 = gather(axis = gather_127_axis_0, batch_dims = gather_127_batch_dims_0, indices = gather_127_indices_0_to_uint16, validate_indices = gather_127_validate_indices_0, x = var_2304_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_127_cast_uint16")];
            tensor<string, []> gather_127_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_127_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, []> const_102 = const()[name = tensor<string, []>("const_102"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_103 = const()[name = tensor<string, []>("const_103"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_127_cast_uint16_to_int32 = cast(dtype = gather_127_cast_uint16_to_int32_dtype_0, x = gather_127_cast_uint16)[name = tensor<string, []>("cast_298")];
            tensor<int32, [?]> var_2308 = range_1d(end = gather_127_cast_uint16_to_int32, start = const_102, step = const_103)[name = tensor<string, []>("op_2308")];
            tensor<fp16, [1, 1]> var_2307_to_fp16 = const()[name = tensor<string, []>("op_2307_to_fp16"), val = tensor<fp16, [1, 1]>([[0x0p+0]])];
            tensor<string, []> var_2308_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2308_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [?]> var_2308_to_fp16 = cast(dtype = var_2308_promoted_to_fp16_dtype_0, x = var_2308)[name = tensor<string, []>("cast_297")];
            tensor<fp16, [1, ?]> ts_45_cast_fp16 = add(x = var_2307_to_fp16, y = var_2308_to_fp16)[name = tensor<string, []>("ts_45_cast_fp16")];
            tensor<int32, [4]> var_2311 = const()[name = tensor<string, []>("op_2311"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, ?, 1]> ts_cast_fp16 = reshape(shape = var_2311, x = ts_45_cast_fp16)[name = tensor<string, []>("ts_cast_fp16")];
            tensor<int32, [5]> concat_141x = const()[name = tensor<string, []>("concat_141x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> q2_cast_fp16 = reshape(shape = concat_141x, x = q_45_cast_fp16)[name = tensor<string, []>("q2_cast_fp16")];
            tensor<int32, [5]> concat_142x = const()[name = tensor<string, []>("concat_142x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])];
            tensor<fp16, [1, 4, ?, 32, 2]> k2_cast_fp16 = reshape(shape = concat_142x, x = k_45_cast_fp16)[name = tensor<string, []>("k2_cast_fp16")];
            tensor<int32, [5]> qr_begin_0 = const()[name = tensor<string, []>("qr_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> qr_end_0 = const()[name = tensor<string, []>("qr_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> qr_end_mask_0 = const()[name = tensor<string, []>("qr_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qr_squeeze_mask_0 = const()[name = tensor<string, []>("qr_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qr_cast_fp16 = slice_by_index(begin = qr_begin_0, end = qr_end_0, end_mask = qr_end_mask_0, squeeze_mask = qr_squeeze_mask_0, x = q2_cast_fp16)[name = tensor<string, []>("qr_cast_fp16")];
            tensor<int32, [5]> qi_begin_0 = const()[name = tensor<string, []>("qi_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> qi_end_0 = const()[name = tensor<string, []>("qi_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> qi_end_mask_0 = const()[name = tensor<string, []>("qi_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> qi_squeeze_mask_0 = const()[name = tensor<string, []>("qi_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> qi_cast_fp16 = slice_by_index(begin = qi_begin_0, end = qi_end_0, end_mask = qi_end_mask_0, squeeze_mask = qi_squeeze_mask_0, x = q2_cast_fp16)[name = tensor<string, []>("qi_cast_fp16")];
            tensor<int32, [5]> kr_begin_0 = const()[name = tensor<string, []>("kr_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> kr_end_0 = const()[name = tensor<string, []>("kr_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])];
            tensor<bool, [5]> kr_end_mask_0 = const()[name = tensor<string, []>("kr_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> kr_squeeze_mask_0 = const()[name = tensor<string, []>("kr_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> kr_cast_fp16 = slice_by_index(begin = kr_begin_0, end = kr_end_0, end_mask = kr_end_mask_0, squeeze_mask = kr_squeeze_mask_0, x = k2_cast_fp16)[name = tensor<string, []>("kr_cast_fp16")];
            tensor<int32, [5]> ki_begin_0 = const()[name = tensor<string, []>("ki_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> ki_end_0 = const()[name = tensor<string, []>("ki_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])];
            tensor<bool, [5]> ki_end_mask_0 = const()[name = tensor<string, []>("ki_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> ki_squeeze_mask_0 = const()[name = tensor<string, []>("ki_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp16, [1, 4, ?, 32]> ki_cast_fp16 = slice_by_index(begin = ki_begin_0, end = ki_end_0, end_mask = ki_end_mask_0, squeeze_mask = ki_squeeze_mask_0, x = k2_cast_fp16)[name = tensor<string, []>("ki_cast_fp16")];
            tensor<fp16, [32]> freqs_to_fp16 = const()[name = tensor<string, []>("freqs_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20811520)))];
            tensor<fp16, [1, 1, ?, 32]> var_2324_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor<string, []>("op_2324_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> rotr_cast_fp16 = cos(x = var_2324_cast_fp16)[name = tensor<string, []>("rotr_cast_fp16")];
            tensor<fp16, [1, 1, ?, 32]> roti_cast_fp16 = sin(x = var_2324_cast_fp16)[name = tensor<string, []>("roti_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2328_cast_fp16 = mul(x = qr_cast_fp16, y = rotr_cast_fp16)[name = tensor<string, []>("op_2328_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2329_cast_fp16 = mul(x = qi_cast_fp16, y = roti_cast_fp16)[name = tensor<string, []>("op_2329_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2330_cast_fp16 = sub(x = var_2328_cast_fp16, y = var_2329_cast_fp16)[name = tensor<string, []>("op_2330_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2331_cast_fp16 = mul(x = qr_cast_fp16, y = roti_cast_fp16)[name = tensor<string, []>("op_2331_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2332_cast_fp16 = mul(x = qi_cast_fp16, y = rotr_cast_fp16)[name = tensor<string, []>("op_2332_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2333_cast_fp16 = add(x = var_2331_cast_fp16, y = var_2332_cast_fp16)[name = tensor<string, []>("op_2333_cast_fp16")];
            tensor<int32, []> qo_axis_0 = const()[name = tensor<string, []>("qo_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> qo_cast_fp16 = stack(axis = qo_axis_0, values = (var_2330_cast_fp16, var_2333_cast_fp16))[name = tensor<string, []>("qo_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2336_cast_fp16 = mul(x = kr_cast_fp16, y = rotr_cast_fp16)[name = tensor<string, []>("op_2336_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2337_cast_fp16 = mul(x = ki_cast_fp16, y = roti_cast_fp16)[name = tensor<string, []>("op_2337_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2338_cast_fp16 = sub(x = var_2336_cast_fp16, y = var_2337_cast_fp16)[name = tensor<string, []>("op_2338_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2339_cast_fp16 = mul(x = kr_cast_fp16, y = roti_cast_fp16)[name = tensor<string, []>("op_2339_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2340_cast_fp16 = mul(x = ki_cast_fp16, y = rotr_cast_fp16)[name = tensor<string, []>("op_2340_cast_fp16")];
            tensor<fp16, [1, 4, ?, 32]> var_2341_cast_fp16 = add(x = var_2339_cast_fp16, y = var_2340_cast_fp16)[name = tensor<string, []>("op_2341_cast_fp16")];
            tensor<int32, []> ko_axis_0 = const()[name = tensor<string, []>("ko_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 4, ?, 32, 2]> ko_cast_fp16 = stack(axis = ko_axis_0, values = (var_2338_cast_fp16, var_2341_cast_fp16))[name = tensor<string, []>("ko_cast_fp16")];
            tensor<int32, [4]> concat_143x = const()[name = tensor<string, []>("concat_143x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> q_cast_fp16 = reshape(shape = concat_143x, x = qo_cast_fp16)[name = tensor<string, []>("q_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = tensor<string, []>("concat_144x"), val = tensor<int32, [4]>([1, 4, -1, 64])];
            tensor<fp16, [1, 4, ?, 64]> k_cast_fp16 = reshape(shape = concat_144x, x = ko_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
            tensor<int32, []> const_104 = const()[name = tensor<string, []>("const_104"), val = tensor<int32, []>(0)];
            tensor<int32, []> const_105 = const()[name = tensor<string, []>("const_105"), val = tensor<int32, []>(1)];
            tensor<int32, []> gather_123_cast_uint16_to_int32 = cast(dtype = gather_123_cast_uint16_to_int32_dtype_0, x = gather_123_cast_uint16)[name = tensor<string, []>("cast_300")];
            tensor<int32, [?]> positions = range_1d(end = gather_123_cast_uint16_to_int32, start = const_104, step = const_105)[name = tensor<string, []>("positions")];
            tensor<int32, [3]> concat_145x = const()[name = tensor<string, []>("concat_145x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<int32, [1, 1, ?]> var_2353 = reshape(shape = concat_145x, x = positions)[name = tensor<string, []>("op_2353")];
            tensor<bool, [1, 1, ?]> valid_k = less(x = var_2353, y = var_1944)[name = tensor<string, []>("valid_k")];
            tensor<int32, [3]> concat_146x = const()[name = tensor<string, []>("concat_146x"), val = tensor<int32, [3]>([1, -1, 1])];
            tensor<int32, [1, ?, 1]> var_2358 = reshape(shape = concat_146x, x = positions)[name = tensor<string, []>("op_2358")];
            tensor<int32, [1, ?, ?]> delta = sub(x = var_2358, y = var_2353)[name = tensor<string, []>("delta")];
            tensor<int32, []> concat_148_axis_0 = const()[name = tensor<string, []>("concat_148_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> concat_148_interleave_0 = const()[name = tensor<string, []>("concat_148_interleave_0"), val = tensor<bool, []>(false)];
            tensor<int32, [3]> concat_148 = concat(axis = concat_148_axis_0, interleave = concat_148_interleave_0, values = (var_1845, gather_123_cast_uint16_to_int32, gather_123_cast_uint16_to_int32))[name = tensor<string, []>("concat_148")];
            tensor<fp16, []> fill_11_value_0_to_fp16 = const()[name = tensor<string, []>("fill_11_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, ?, ?]> fill_11_cast_fp16 = fill(shape = concat_148, value = fill_11_value_0_to_fp16)[name = tensor<string, []>("fill_11_cast_fp16")];
            tensor<string, []> attn_bias_89_dtype_0 = const()[name = tensor<string, []>("attn_bias_89_dtype_0"), val = tensor<string, []>("bool")];
            tensor<bool, [1, ?, ?]> var_2364 = greater_equal(x = delta, y = var_1827)[name = tensor<string, []>("op_2364")];
            tensor<bool, [1, ?, ?]> fill_11_cast_fp16_to_bool = cast(dtype = attn_bias_89_dtype_0, x = fill_11_cast_fp16)[name = tensor<string, []>("cast_296")];
            tensor<bool, [1, ?, ?]> attn_bias_91 = logical_and(x = fill_11_cast_fp16_to_bool, y = var_2364)[name = tensor<string, []>("attn_bias_91")];
            tensor<bool, [1, ?, ?]> var_2366 = less(x = delta, y = var_1834)[name = tensor<string, []>("op_2366")];
            tensor<bool, [1, ?, ?]> attn_bias_93 = logical_and(x = attn_bias_91, y = var_2366)[name = tensor<string, []>("attn_bias_93")];
            tensor<bool, [1, ?, ?]> var_2368 = logical_and(x = attn_bias_93, y = valid_k)[name = tensor<string, []>("op_2368")];
            tensor<int32, [1]> var_2370_axes_0 = const()[name = tensor<string, []>("op_2370_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, ?, ?]> var_2370 = expand_dims(axes = var_2370_axes_0, x = var_2368)[name = tensor<string, []>("op_2370")];
            tensor<fp16, []> sub_11_x_0_to_fp16 = const()[name = tensor<string, []>("sub_11_x_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> cast_283_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_283_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, ?, ?]> var_2370_to_fp16 = cast(dtype = cast_283_to_fp16_dtype_0, x = var_2370)[name = tensor<string, []>("cast_295")];
            tensor<fp16, [1, 1, ?, ?]> sub_11_cast_fp16 = sub(x = sub_11_x_0_to_fp16, y = var_2370_to_fp16)[name = tensor<string, []>("sub_11_cast_fp16")];
            tensor<fp16, []> mul_22_x_0_to_fp16 = const()[name = tensor<string, []>("mul_22_x_0_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
            tensor<fp16, [1, 1, ?, ?]> mul_22_cast_fp16 = mul(x = mul_22_x_0_to_fp16, y = sub_11_cast_fp16)[name = tensor<string, []>("mul_22_cast_fp16")];
            tensor<fp16, []> mul_23_y_0_to_fp16 = const()[name = tensor<string, []>("mul_23_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 4, ?, 64]> mul_23_cast_fp16 = mul(x = q_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor<string, []>("mul_23_cast_fp16")];
            tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, ?, ?]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = k_cast_fp16)[name = tensor<string, []>("matmul_11_cast_fp16")];
            tensor<fp16, [1, 4, ?, ?]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_22_cast_fp16)[name = tensor<string, []>("add_11_cast_fp16")];
            tensor<string, []> add_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("add_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 4, ?, ?]> add_11_cast_fp16_to_fp32 = cast(dtype = add_11_cast_fp16_to_fp32_dtype_0, x = add_11_cast_fp16)[name = tensor<string, []>("cast_294")];
            tensor<fp32, [1, 4, ?, ?]> softmax_11 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_11")];
            tensor<bool, []> out_45_transpose_x_0 = const()[name = tensor<string, []>("out_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_45_transpose_y_0 = const()[name = tensor<string, []>("out_45_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 4, ?, ?]> softmax_11_to_fp16 = cast(dtype = softmax_11_to_fp16_dtype_0, x = softmax_11)[name = tensor<string, []>("cast_293")];
            tensor<fp16, [1, 4, ?, 64]> out_45_cast_fp16 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = softmax_11_to_fp16, y = v_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<int32, [2]> concat_149x = const()[name = tensor<string, []>("concat_149x"), val = tensor<int32, [2]>([1, -1])];
            tensor<int32, [1, ?]> var_2376 = reshape(shape = concat_149x, x = positions)[name = tensor<string, []>("op_2376")];
            tensor<bool, [1, ?]> var_2379 = less(x = var_2376, y = var_1967)[name = tensor<string, []>("op_2379")];
            tensor<int32, [4]> concat_150x = const()[name = tensor<string, []>("concat_150x"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<bool, [1, 1, ?, 1]> valid_q = reshape(shape = concat_150x, x = var_2379)[name = tensor<string, []>("valid_q")];
            tensor<fp16, []> var_2383_to_fp16 = const()[name = tensor<string, []>("op_2383_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 4, ?, 64]> out_cast_fp16 = select(a = out_45_cast_fp16, b = var_2383_to_fp16, cond = valid_q)[name = tensor<string, []>("out_cast_fp16")];
            tensor<int32, [4]> var_2385_perm_0 = const()[name = tensor<string, []>("op_2385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_151x = const()[name = tensor<string, []>("concat_151x"), val = tensor<int32, [3]>([1, -1, 256])];
            tensor<fp16, [1, ?, 4, 64]> var_2385_cast_fp16 = transpose(perm = var_2385_perm_0, x = out_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, ?, 256]> input_207_cast_fp16 = reshape(shape = concat_151x, x = var_2385_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<fp16, [256, 256]> decoder_7_transformer_layers_3_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20811648)))];
            tensor<fp16, [1, ?, 256]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_3_self_attn_out_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_layer_scale_1_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_layer_scale_1_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20942784)))];
            tensor<fp16, [1, ?, 256]> var_2391_cast_fp16 = mul(x = decoder_7_transformer_layers_3_layer_scale_1_scale_to_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("op_2391_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_209_cast_fp16 = add(x = input_205_cast_fp16, y = var_2391_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<int32, [1]> input_211_axes_0 = const()[name = tensor<string, []>("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_norm2_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20943360)))];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_norm2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20943936)))];
            tensor<fp16, [1, ?, 256]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = decoder_7_transformer_layers_3_norm2_bias_to_fp16, epsilon = var_1841_to_fp16, gamma = decoder_7_transformer_layers_3_norm2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<fp16, [1024, 256]> decoder_7_transformer_layers_3_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_ffn_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20944512)))];
            tensor<fp16, [1, ?, 1024]> linear_53_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = decoder_7_transformer_layers_3_ffn_0_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<string, []> input_215_mode_0 = const()[name = tensor<string, []>("input_215_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, ?, 1024]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_53_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<fp16, [256, 1024]> decoder_7_transformer_layers_3_ffn_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_ffn_2_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21468864)))];
            tensor<fp16, [1, ?, 256]> linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = decoder_7_transformer_layers_3_ffn_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [256]> decoder_7_transformer_layers_3_layer_scale_2_scale_to_fp16 = const()[name = tensor<string, []>("decoder_7_transformer_layers_3_layer_scale_2_scale_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21993216)))];
            tensor<fp16, [1, ?, 256]> var_2406_cast_fp16 = mul(x = decoder_7_transformer_layers_3_layer_scale_2_scale_to_fp16, y = linear_54_cast_fp16)[name = tensor<string, []>("op_2406_cast_fp16")];
            tensor<fp16, [1, ?, 256]> input_cast_fp16 = add(x = input_209_cast_fp16, y = var_2406_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<fp16, [240, 256]> decoder_7_output_proj_weight_to_fp16 = const()[name = tensor<string, []>("decoder_7_output_proj_weight_to_fp16"), val = tensor<fp16, [240, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21993792)))];
            tensor<fp16, [240]> linear_55_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_55_bias_0_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22116736)))];
            tensor<fp16, [1, ?, 240]> linear_55_cast_fp16 = linear(bias = linear_55_bias_0_to_fp16, weight = decoder_7_output_proj_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<int32, [3]> x_97_perm_0 = const()[name = tensor<string, []>("x_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [4]> concat_152x = const()[name = tensor<string, []>("concat_152x"), val = tensor<int32, [4]>([1, 1, 240, -1])];
            tensor<fp16, [1, 240, ?]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = linear_55_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<fp16, [1, 1, 240, ?]> var_2426_cast_fp16 = reshape(shape = concat_152x, x = x_97_cast_fp16)[name = tensor<string, []>("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2427 = const()[name = tensor<string, []>("op_2427"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> concat_153x = const()[name = tensor<string, []>("concat_153x"), val = tensor<int32, [3]>([1, 1, -1])];
            tensor<fp16, [1, 1, ?, 240]> var_2428_cast_fp16 = transpose(perm = var_2427, x = var_2426_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp16, [1, 1, ?]> x_cast_fp16 = reshape(shape = concat_153x, x = var_2428_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = tensor<string, []>("op_2435_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, ?]> var_2435_cast_fp16 = squeeze(axes = var_2435_axes_0, x = x_cast_fp16)[name = tensor<string, []>("op_2435_cast_fp16")];
            tensor<int32, [3]> var_2442 = const()[name = tensor<string, []>("op_2442"), val = tensor<int32, [3]>([1, -1, 2])];
            tensor<fp16, [1, ?, 2]> var_2443_cast_fp16 = reshape(shape = var_2442, x = var_2435_cast_fp16)[name = tensor<string, []>("op_2443_cast_fp16")];
            tensor<int32, [3]> var_2446_perm_0 = const()[name = tensor<string, []>("op_2446_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_2446_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2446_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 2, ?]> var_2446_cast_fp16 = transpose(perm = var_2446_perm_0, x = var_2443_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp32, [1, 2, ?]> audio = cast(dtype = var_2446_cast_fp16_to_fp32_dtype_0, x = var_2446_cast_fp16)[name = tensor<string, []>("cast_292")];
        } -> (audio);
}