program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios18>(tensor<fp32, [1, 50, 256]> style_ttl, tensor<int32, [1, 128]> text_ids, tensor<fp32, [1, 1, 128]> text_mask) {
            int32 var_29 = const()[name = string("op_29"), val = int32(-1)];
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [8322, 256]> m_char_embedder_weight_to_fp16 = const()[name = string("m_char_embedder_weight_to_fp16"), val = tensor<fp16, [8322, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string text_ids_to_int16_dtype_0 = const()[name = string("text_ids_to_int16_dtype_0"), val = string("int16")];
            string cast_45_dtype_0 = const()[name = string("cast_45_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [1, 128]> text_ids_to_int16 = cast(dtype = text_ids_to_int16_dtype_0, x = text_ids)[name = string("cast_51")];
            tensor<int32, [1, 128]> cast_45 = cast(dtype = cast_45_dtype_0, x = text_ids_to_int16)[name = string("cast_50")];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = cast_45, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(8322)];
            tensor<int32, [1, 128]> add_0 = add(x = cast_45, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0 = select(a = cast_45, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 128]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_49")];
            tensor<fp16, [1, 128, 256]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = m_char_embedder_weight_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, [3]> var_55_perm_0 = const()[name = string("op_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string text_mask_to_fp16_dtype_0 = const()[name = string("text_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 128]> text_mask_to_fp16 = cast(dtype = text_mask_to_fp16_dtype_0, x = text_mask)[name = string("cast_48")];
            tensor<fp16, [1, 256, 128]> var_55_cast_fp16 = transpose(perm = var_55_perm_0, x = x_1_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_45")];
            tensor<fp16, [1, 256, 128]> x_3_cast_fp16 = mul(x = var_55_cast_fp16, y = text_mask_to_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_5_cast_fp16 = mul(x = x_3_cast_fp16, y = text_mask_to_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [3]> input_3_begin_0 = const()[name = string("input_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> input_3_end_0 = const()[name = string("input_3_end_0"), val = tensor<int32, [3]>([1, 1, 128])];
            tensor<bool, [3]> input_3_end_mask_0 = const()[name = string("input_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1, 127]> input_3_cast_fp16 = slice_by_index(begin = input_3_begin_0, end = input_3_end_0, end_mask = input_3_end_mask_0, x = text_mask_to_fp16)[name = string("input_3_cast_fp16")];
            tensor<int32, [6]> var_65_pad_0 = const()[name = string("op_65_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 1])];
            string var_65_mode_0 = const()[name = string("op_65_mode_0"), val = string("constant")];
            fp16 const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 128]> var_65_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = var_65_mode_0, pad = var_65_pad_0, x = input_3_cast_fp16)[name = string("op_65_cast_fp16")];
            fp16 var_21_to_fp16 = const()[name = string("op_21_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1, 128]> var_66_cast_fp16 = sub(x = var_21_to_fp16, y = var_65_cast_fp16)[name = string("op_66_cast_fp16")];
            tensor<fp16, [1, 1, 128]> last_1_cast_fp16 = mul(x = text_mask_to_fp16, y = var_66_cast_fp16)[name = string("last_1_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_68_cast_fp16 = mul(x = x_5_cast_fp16, y = last_1_cast_fp16)[name = string("op_68_cast_fp16")];
            tensor<int32, [1]> x_last_1_axes_0 = const()[name = string("x_last_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_1_keep_dims_0 = const()[name = string("x_last_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_1_cast_fp16 = reduce_sum(axes = x_last_1_axes_0, keep_dims = x_last_1_keep_dims_0, x = var_68_cast_fp16)[name = string("x_last_1_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_71_cast_fp16 = mul(x = x_5_cast_fp16, y = text_mask_to_fp16)[name = string("op_71_cast_fp16")];
            tensor<fp16, [1, 1, 128]> var_72_cast_fp16 = sub(x = var_21_to_fp16, y = text_mask_to_fp16)[name = string("op_72_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_73_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_1_cast_fp16)[name = string("op_73_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_5_cast_fp16 = add(x = var_71_cast_fp16, y = var_73_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<int32, [6]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 2])];
            string input_7_mode_0 = const()[name = string("input_7_mode_0"), val = string("replicate")];
            fp16 const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 132]> input_7_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_7_mode_0, pad = input_7_pad_0, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
            string h_1_pad_type_0 = const()[name = string("h_1_pad_type_0"), val = string("valid")];
            int32 h_1_groups_0 = const()[name = string("h_1_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_1_strides_0 = const()[name = string("h_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_1_pad_0 = const()[name = string("h_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_1_dilations_0 = const()[name = string("h_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [256, 1, 5]> m_convnext_0_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_0_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4260992)))];
            tensor<fp16, [256]> m_convnext_0_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_0_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4263616)))];
            tensor<fp16, [1, 256, 128]> h_1_cast_fp16 = conv(bias = m_convnext_0_dwconv__conv_bias_to_fp16, dilations = h_1_dilations_0, groups = h_1_groups_0, pad = h_1_pad_0, pad_type = h_1_pad_type_0, strides = h_1_strides_0, weight = m_convnext_0_dwconv__conv_weight_to_fp16, x = input_7_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_7_cast_fp16 = mul(x = h_1_cast_fp16, y = text_mask_to_fp16)[name = string("x_7_cast_fp16")];
            tensor<int32, [3]> input_9_perm_0 = const()[name = string("input_9_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_0_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_0_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4264192)))];
            tensor<fp16, [256]> m_convnext_0_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_0_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4264768)))];
            fp16 var_25_to_fp16 = const()[name = string("op_25_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 128, 256]> input_9_cast_fp16 = transpose(perm = input_9_perm_0, x = x_7_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 128, 256]> var_91_cast_fp16 = layer_norm(axes = var_91_axes_0, beta = m_convnext_0_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_0_norm_norm_weight_to_fp16, x = input_9_cast_fp16)[name = string("op_91_cast_fp16")];
            tensor<int32, [3]> input_11_perm_0 = const()[name = string("input_11_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_3_pad_type_0 = const()[name = string("h_3_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_3_strides_0 = const()[name = string("h_3_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_3_pad_0 = const()[name = string("h_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_3_dilations_0 = const()[name = string("h_3_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_3_groups_0 = const()[name = string("h_3_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_0_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_0_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4265344)))];
            tensor<fp16, [1024]> m_convnext_0_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_0_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4789696)))];
            tensor<fp16, [1, 256, 128]> input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = var_91_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 1024, 128]> h_3_cast_fp16 = conv(bias = m_convnext_0_pwconv1_bias_to_fp16, dilations = h_3_dilations_0, groups = h_3_groups_0, pad = h_3_pad_0, pad_type = h_3_pad_type_0, strides = h_3_strides_0, weight = m_convnext_0_pwconv1_weight_to_fp16, x = input_11_cast_fp16)[name = string("h_3_cast_fp16")];
            string input_13_mode_0 = const()[name = string("input_13_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = h_3_cast_fp16)[name = string("input_13_cast_fp16")];
            string h_5_pad_type_0 = const()[name = string("h_5_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_5_strides_0 = const()[name = string("h_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_5_pad_0 = const()[name = string("h_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_5_dilations_0 = const()[name = string("h_5_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_5_groups_0 = const()[name = string("h_5_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_108_weight_0_to_fp16 = const()[name = string("op_108_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4791808)))];
            tensor<fp16, [256]> var_108_bias_0_to_fp16 = const()[name = string("op_108_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5316160)))];
            tensor<fp16, [1, 256, 128]> var_108_cast_fp16 = conv(bias = var_108_bias_0_to_fp16, dilations = h_5_dilations_0, groups = h_5_groups_0, pad = h_5_pad_0, pad_type = h_5_pad_type_0, strides = h_5_strides_0, weight = var_108_weight_0_to_fp16, x = input_13_cast_fp16)[name = string("op_108_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_1_cast_fp16 = add(x = x_5_cast_fp16, y = var_108_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_9_cast_fp16 = mul(x = out_1_cast_fp16, y = text_mask_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = text_mask_to_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_122_cast_fp16 = mul(x = x_11_cast_fp16, y = last_1_cast_fp16)[name = string("op_122_cast_fp16")];
            tensor<int32, [1]> x_last_3_axes_0 = const()[name = string("x_last_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_3_keep_dims_0 = const()[name = string("x_last_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_3_cast_fp16 = reduce_sum(axes = x_last_3_axes_0, keep_dims = x_last_3_keep_dims_0, x = var_122_cast_fp16)[name = string("x_last_3_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_125_cast_fp16 = mul(x = x_11_cast_fp16, y = text_mask_to_fp16)[name = string("op_125_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_127_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_3_cast_fp16)[name = string("op_127_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_17_cast_fp16 = add(x = var_125_cast_fp16, y = var_127_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<int32, [6]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 2])];
            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("replicate")];
            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 132]> input_19_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_19_mode_0, pad = input_19_pad_0, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
            string h_7_pad_type_0 = const()[name = string("h_7_pad_type_0"), val = string("valid")];
            int32 h_7_groups_0 = const()[name = string("h_7_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_7_strides_0 = const()[name = string("h_7_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_7_pad_0 = const()[name = string("h_7_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_7_dilations_0 = const()[name = string("h_7_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [256, 1, 5]> m_convnext_1_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_1_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5316736)))];
            tensor<fp16, [256]> m_convnext_1_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_1_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5319360)))];
            tensor<fp16, [1, 256, 128]> h_7_cast_fp16 = conv(bias = m_convnext_1_dwconv__conv_bias_to_fp16, dilations = h_7_dilations_0, groups = h_7_groups_0, pad = h_7_pad_0, pad_type = h_7_pad_type_0, strides = h_7_strides_0, weight = m_convnext_1_dwconv__conv_weight_to_fp16, x = input_19_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_13_cast_fp16 = mul(x = h_7_cast_fp16, y = text_mask_to_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [3]> input_21_perm_0 = const()[name = string("input_21_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_145_axes_0 = const()[name = string("op_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_1_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_1_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5319936)))];
            tensor<fp16, [256]> m_convnext_1_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_1_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5320512)))];
            tensor<fp16, [1, 128, 256]> input_21_cast_fp16 = transpose(perm = input_21_perm_0, x = x_13_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 128, 256]> var_145_cast_fp16 = layer_norm(axes = var_145_axes_0, beta = m_convnext_1_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_1_norm_norm_weight_to_fp16, x = input_21_cast_fp16)[name = string("op_145_cast_fp16")];
            tensor<int32, [3]> input_23_perm_0 = const()[name = string("input_23_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_9_pad_type_0 = const()[name = string("h_9_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_9_strides_0 = const()[name = string("h_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_9_pad_0 = const()[name = string("h_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_9_dilations_0 = const()[name = string("h_9_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_9_groups_0 = const()[name = string("h_9_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_1_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_1_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5321088)))];
            tensor<fp16, [1024]> m_convnext_1_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_1_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5845440)))];
            tensor<fp16, [1, 256, 128]> input_23_cast_fp16 = transpose(perm = input_23_perm_0, x = var_145_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 1024, 128]> h_9_cast_fp16 = conv(bias = m_convnext_1_pwconv1_bias_to_fp16, dilations = h_9_dilations_0, groups = h_9_groups_0, pad = h_9_pad_0, pad_type = h_9_pad_type_0, strides = h_9_strides_0, weight = m_convnext_1_pwconv1_weight_to_fp16, x = input_23_cast_fp16)[name = string("h_9_cast_fp16")];
            string input_25_mode_0 = const()[name = string("input_25_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = h_9_cast_fp16)[name = string("input_25_cast_fp16")];
            string h_11_pad_type_0 = const()[name = string("h_11_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_11_strides_0 = const()[name = string("h_11_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_11_pad_0 = const()[name = string("h_11_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_11_dilations_0 = const()[name = string("h_11_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_11_groups_0 = const()[name = string("h_11_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_162_weight_0_to_fp16 = const()[name = string("op_162_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5847552)))];
            tensor<fp16, [256]> var_162_bias_0_to_fp16 = const()[name = string("op_162_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6371904)))];
            tensor<fp16, [1, 256, 128]> var_162_cast_fp16 = conv(bias = var_162_bias_0_to_fp16, dilations = h_11_dilations_0, groups = h_11_groups_0, pad = h_11_pad_0, pad_type = h_11_pad_type_0, strides = h_11_strides_0, weight = var_162_weight_0_to_fp16, x = input_25_cast_fp16)[name = string("op_162_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_3_cast_fp16 = add(x = x_11_cast_fp16, y = var_162_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_15_cast_fp16 = mul(x = out_3_cast_fp16, y = text_mask_to_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_17_cast_fp16 = mul(x = x_15_cast_fp16, y = text_mask_to_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_176_cast_fp16 = mul(x = x_17_cast_fp16, y = last_1_cast_fp16)[name = string("op_176_cast_fp16")];
            tensor<int32, [1]> x_last_5_axes_0 = const()[name = string("x_last_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_5_keep_dims_0 = const()[name = string("x_last_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_5_cast_fp16 = reduce_sum(axes = x_last_5_axes_0, keep_dims = x_last_5_keep_dims_0, x = var_176_cast_fp16)[name = string("x_last_5_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_179_cast_fp16 = mul(x = x_17_cast_fp16, y = text_mask_to_fp16)[name = string("op_179_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_181_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_5_cast_fp16)[name = string("op_181_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_29_cast_fp16 = add(x = var_179_cast_fp16, y = var_181_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [6]> input_31_pad_0 = const()[name = string("input_31_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 4, 4])];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("replicate")];
            fp16 const_5_to_fp16 = const()[name = string("const_5_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 136]> input_31_cast_fp16 = pad(constant_val = const_5_to_fp16, mode = input_31_mode_0, pad = input_31_pad_0, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
            string h_13_pad_type_0 = const()[name = string("h_13_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_13_dilations_0 = const()[name = string("h_13_dilations_0"), val = tensor<int32, [1]>([2])];
            int32 h_13_groups_0 = const()[name = string("h_13_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_13_strides_0 = const()[name = string("h_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_13_pad_0 = const()[name = string("h_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [256, 1, 5]> m_convnext_2_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_2_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6372480)))];
            tensor<fp16, [256]> m_convnext_2_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_2_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6375104)))];
            tensor<fp16, [1, 256, 128]> h_13_cast_fp16 = conv(bias = m_convnext_2_dwconv__conv_bias_to_fp16, dilations = h_13_dilations_0, groups = h_13_groups_0, pad = h_13_pad_0, pad_type = h_13_pad_type_0, strides = h_13_strides_0, weight = m_convnext_2_dwconv__conv_weight_to_fp16, x = input_31_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_19_cast_fp16 = mul(x = h_13_cast_fp16, y = text_mask_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [3]> input_33_perm_0 = const()[name = string("input_33_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_199_axes_0 = const()[name = string("op_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_2_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_2_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6375680)))];
            tensor<fp16, [256]> m_convnext_2_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_2_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6376256)))];
            tensor<fp16, [1, 128, 256]> input_33_cast_fp16 = transpose(perm = input_33_perm_0, x = x_19_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 128, 256]> var_199_cast_fp16 = layer_norm(axes = var_199_axes_0, beta = m_convnext_2_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_2_norm_norm_weight_to_fp16, x = input_33_cast_fp16)[name = string("op_199_cast_fp16")];
            tensor<int32, [3]> input_35_perm_0 = const()[name = string("input_35_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_15_pad_type_0 = const()[name = string("h_15_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_15_strides_0 = const()[name = string("h_15_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_15_pad_0 = const()[name = string("h_15_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_15_dilations_0 = const()[name = string("h_15_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_15_groups_0 = const()[name = string("h_15_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_2_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_2_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6376832)))];
            tensor<fp16, [1024]> m_convnext_2_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_2_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6901184)))];
            tensor<fp16, [1, 256, 128]> input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = var_199_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 1024, 128]> h_15_cast_fp16 = conv(bias = m_convnext_2_pwconv1_bias_to_fp16, dilations = h_15_dilations_0, groups = h_15_groups_0, pad = h_15_pad_0, pad_type = h_15_pad_type_0, strides = h_15_strides_0, weight = m_convnext_2_pwconv1_weight_to_fp16, x = input_35_cast_fp16)[name = string("h_15_cast_fp16")];
            string input_37_mode_0 = const()[name = string("input_37_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = h_15_cast_fp16)[name = string("input_37_cast_fp16")];
            string h_17_pad_type_0 = const()[name = string("h_17_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_17_strides_0 = const()[name = string("h_17_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_17_pad_0 = const()[name = string("h_17_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_17_dilations_0 = const()[name = string("h_17_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_17_groups_0 = const()[name = string("h_17_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_216_weight_0_to_fp16 = const()[name = string("op_216_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6903296)))];
            tensor<fp16, [256]> var_216_bias_0_to_fp16 = const()[name = string("op_216_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7427648)))];
            tensor<fp16, [1, 256, 128]> var_216_cast_fp16 = conv(bias = var_216_bias_0_to_fp16, dilations = h_17_dilations_0, groups = h_17_groups_0, pad = h_17_pad_0, pad_type = h_17_pad_type_0, strides = h_17_strides_0, weight = var_216_weight_0_to_fp16, x = input_37_cast_fp16)[name = string("op_216_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_5_cast_fp16 = add(x = x_17_cast_fp16, y = var_216_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_21_cast_fp16 = mul(x = out_5_cast_fp16, y = text_mask_to_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_23_cast_fp16 = mul(x = x_21_cast_fp16, y = text_mask_to_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_230_cast_fp16 = mul(x = x_23_cast_fp16, y = last_1_cast_fp16)[name = string("op_230_cast_fp16")];
            tensor<int32, [1]> x_last_7_axes_0 = const()[name = string("x_last_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_7_keep_dims_0 = const()[name = string("x_last_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_7_cast_fp16 = reduce_sum(axes = x_last_7_axes_0, keep_dims = x_last_7_keep_dims_0, x = var_230_cast_fp16)[name = string("x_last_7_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_233_cast_fp16 = mul(x = x_23_cast_fp16, y = text_mask_to_fp16)[name = string("op_233_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_235_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_7_cast_fp16)[name = string("op_235_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_41_cast_fp16 = add(x = var_233_cast_fp16, y = var_235_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<int32, [6]> input_43_pad_0 = const()[name = string("input_43_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 4, 4])];
            string input_43_mode_0 = const()[name = string("input_43_mode_0"), val = string("replicate")];
            fp16 const_7_to_fp16 = const()[name = string("const_7_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 136]> input_43_cast_fp16 = pad(constant_val = const_7_to_fp16, mode = input_43_mode_0, pad = input_43_pad_0, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
            string h_19_pad_type_0 = const()[name = string("h_19_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_19_dilations_0 = const()[name = string("h_19_dilations_0"), val = tensor<int32, [1]>([2])];
            int32 h_19_groups_0 = const()[name = string("h_19_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_19_strides_0 = const()[name = string("h_19_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_19_pad_0 = const()[name = string("h_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [256, 1, 5]> m_convnext_3_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_3_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7428224)))];
            tensor<fp16, [256]> m_convnext_3_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_3_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7430848)))];
            tensor<fp16, [1, 256, 128]> h_19_cast_fp16 = conv(bias = m_convnext_3_dwconv__conv_bias_to_fp16, dilations = h_19_dilations_0, groups = h_19_groups_0, pad = h_19_pad_0, pad_type = h_19_pad_type_0, strides = h_19_strides_0, weight = m_convnext_3_dwconv__conv_weight_to_fp16, x = input_43_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_25_cast_fp16 = mul(x = h_19_cast_fp16, y = text_mask_to_fp16)[name = string("x_25_cast_fp16")];
            tensor<int32, [3]> input_45_perm_0 = const()[name = string("input_45_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_3_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_3_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7431424)))];
            tensor<fp16, [256]> m_convnext_3_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_3_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7432000)))];
            tensor<fp16, [1, 128, 256]> input_45_cast_fp16 = transpose(perm = input_45_perm_0, x = x_25_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 128, 256]> var_253_cast_fp16 = layer_norm(axes = var_253_axes_0, beta = m_convnext_3_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_3_norm_norm_weight_to_fp16, x = input_45_cast_fp16)[name = string("op_253_cast_fp16")];
            tensor<int32, [3]> input_47_perm_0 = const()[name = string("input_47_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_21_pad_type_0 = const()[name = string("h_21_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_21_strides_0 = const()[name = string("h_21_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_21_pad_0 = const()[name = string("h_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_21_dilations_0 = const()[name = string("h_21_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_21_groups_0 = const()[name = string("h_21_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_3_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_3_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7432576)))];
            tensor<fp16, [1024]> m_convnext_3_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_3_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7956928)))];
            tensor<fp16, [1, 256, 128]> input_47_cast_fp16 = transpose(perm = input_47_perm_0, x = var_253_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1024, 128]> h_21_cast_fp16 = conv(bias = m_convnext_3_pwconv1_bias_to_fp16, dilations = h_21_dilations_0, groups = h_21_groups_0, pad = h_21_pad_0, pad_type = h_21_pad_type_0, strides = h_21_strides_0, weight = m_convnext_3_pwconv1_weight_to_fp16, x = input_47_cast_fp16)[name = string("h_21_cast_fp16")];
            string input_49_mode_0 = const()[name = string("input_49_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = h_21_cast_fp16)[name = string("input_49_cast_fp16")];
            string h_23_pad_type_0 = const()[name = string("h_23_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_23_strides_0 = const()[name = string("h_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_23_pad_0 = const()[name = string("h_23_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_23_dilations_0 = const()[name = string("h_23_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_23_groups_0 = const()[name = string("h_23_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_270_weight_0_to_fp16 = const()[name = string("op_270_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7959040)))];
            tensor<fp16, [256]> var_270_bias_0_to_fp16 = const()[name = string("op_270_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8483392)))];
            tensor<fp16, [1, 256, 128]> var_270_cast_fp16 = conv(bias = var_270_bias_0_to_fp16, dilations = h_23_dilations_0, groups = h_23_groups_0, pad = h_23_pad_0, pad_type = h_23_pad_type_0, strides = h_23_strides_0, weight = var_270_weight_0_to_fp16, x = input_49_cast_fp16)[name = string("op_270_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_7_cast_fp16 = add(x = x_23_cast_fp16, y = var_270_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_27_cast_fp16 = mul(x = out_7_cast_fp16, y = text_mask_to_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_29_cast_fp16 = mul(x = x_27_cast_fp16, y = text_mask_to_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_284_cast_fp16 = mul(x = x_29_cast_fp16, y = last_1_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> x_last_9_axes_0 = const()[name = string("x_last_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_9_keep_dims_0 = const()[name = string("x_last_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_9_cast_fp16 = reduce_sum(axes = x_last_9_axes_0, keep_dims = x_last_9_keep_dims_0, x = var_284_cast_fp16)[name = string("x_last_9_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_287_cast_fp16 = mul(x = x_29_cast_fp16, y = text_mask_to_fp16)[name = string("op_287_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_289_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_9_cast_fp16)[name = string("op_289_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_53_cast_fp16 = add(x = var_287_cast_fp16, y = var_289_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<int32, [6]> input_55_pad_0 = const()[name = string("input_55_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 8, 8])];
            string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("replicate")];
            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 144]> input_55_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_55_mode_0, pad = input_55_pad_0, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")];
            string h_25_pad_type_0 = const()[name = string("h_25_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_25_dilations_0 = const()[name = string("h_25_dilations_0"), val = tensor<int32, [1]>([4])];
            int32 h_25_groups_0 = const()[name = string("h_25_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_25_strides_0 = const()[name = string("h_25_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_25_pad_0 = const()[name = string("h_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [256, 1, 5]> m_convnext_4_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_4_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8483968)))];
            tensor<fp16, [256]> m_convnext_4_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_4_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8486592)))];
            tensor<fp16, [1, 256, 128]> h_25_cast_fp16 = conv(bias = m_convnext_4_dwconv__conv_bias_to_fp16, dilations = h_25_dilations_0, groups = h_25_groups_0, pad = h_25_pad_0, pad_type = h_25_pad_type_0, strides = h_25_strides_0, weight = m_convnext_4_dwconv__conv_weight_to_fp16, x = input_55_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_31_cast_fp16 = mul(x = h_25_cast_fp16, y = text_mask_to_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_307_axes_0 = const()[name = string("op_307_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_4_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_4_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8487168)))];
            tensor<fp16, [256]> m_convnext_4_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_4_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8487744)))];
            tensor<fp16, [1, 128, 256]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_31_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 128, 256]> var_307_cast_fp16 = layer_norm(axes = var_307_axes_0, beta = m_convnext_4_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_4_norm_norm_weight_to_fp16, x = input_57_cast_fp16)[name = string("op_307_cast_fp16")];
            tensor<int32, [3]> input_59_perm_0 = const()[name = string("input_59_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_27_pad_type_0 = const()[name = string("h_27_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_27_strides_0 = const()[name = string("h_27_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_27_pad_0 = const()[name = string("h_27_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_27_dilations_0 = const()[name = string("h_27_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_27_groups_0 = const()[name = string("h_27_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_4_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_4_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8488320)))];
            tensor<fp16, [1024]> m_convnext_4_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_4_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9012672)))];
            tensor<fp16, [1, 256, 128]> input_59_cast_fp16 = transpose(perm = input_59_perm_0, x = var_307_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 1024, 128]> h_27_cast_fp16 = conv(bias = m_convnext_4_pwconv1_bias_to_fp16, dilations = h_27_dilations_0, groups = h_27_groups_0, pad = h_27_pad_0, pad_type = h_27_pad_type_0, strides = h_27_strides_0, weight = m_convnext_4_pwconv1_weight_to_fp16, x = input_59_cast_fp16)[name = string("h_27_cast_fp16")];
            string input_61_mode_0 = const()[name = string("input_61_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = h_27_cast_fp16)[name = string("input_61_cast_fp16")];
            string h_29_pad_type_0 = const()[name = string("h_29_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_29_strides_0 = const()[name = string("h_29_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_29_pad_0 = const()[name = string("h_29_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_29_dilations_0 = const()[name = string("h_29_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_29_groups_0 = const()[name = string("h_29_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_324_weight_0_to_fp16 = const()[name = string("op_324_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9014784)))];
            tensor<fp16, [256]> var_324_bias_0_to_fp16 = const()[name = string("op_324_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9539136)))];
            tensor<fp16, [1, 256, 128]> var_324_cast_fp16 = conv(bias = var_324_bias_0_to_fp16, dilations = h_29_dilations_0, groups = h_29_groups_0, pad = h_29_pad_0, pad_type = h_29_pad_type_0, strides = h_29_strides_0, weight = var_324_weight_0_to_fp16, x = input_61_cast_fp16)[name = string("op_324_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_9_cast_fp16 = add(x = x_29_cast_fp16, y = var_324_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_33_cast_fp16 = mul(x = out_9_cast_fp16, y = text_mask_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_35_cast_fp16 = mul(x = x_33_cast_fp16, y = text_mask_to_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_338_cast_fp16 = mul(x = x_35_cast_fp16, y = last_1_cast_fp16)[name = string("op_338_cast_fp16")];
            tensor<int32, [1]> x_last_axes_0 = const()[name = string("x_last_axes_0"), val = tensor<int32, [1]>([-1])];
            bool x_last_keep_dims_0 = const()[name = string("x_last_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> x_last_cast_fp16 = reduce_sum(axes = x_last_axes_0, keep_dims = x_last_keep_dims_0, x = var_338_cast_fp16)[name = string("x_last_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_341_cast_fp16 = mul(x = x_35_cast_fp16, y = text_mask_to_fp16)[name = string("op_341_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_343_cast_fp16 = mul(x = var_72_cast_fp16, y = x_last_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<fp16, [1, 256, 128]> input_65_cast_fp16 = add(x = var_341_cast_fp16, y = var_343_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [6]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 8, 8])];
            string input_67_mode_0 = const()[name = string("input_67_mode_0"), val = string("replicate")];
            fp16 const_11_to_fp16 = const()[name = string("const_11_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 144]> input_67_cast_fp16 = pad(constant_val = const_11_to_fp16, mode = input_67_mode_0, pad = input_67_pad_0, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            string h_31_pad_type_0 = const()[name = string("h_31_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_31_dilations_0 = const()[name = string("h_31_dilations_0"), val = tensor<int32, [1]>([4])];
            int32 h_31_groups_0 = const()[name = string("h_31_groups_0"), val = int32(256)];
            tensor<int32, [1]> h_31_strides_0 = const()[name = string("h_31_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_31_pad_0 = const()[name = string("h_31_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<fp16, [256, 1, 5]> m_convnext_5_dwconv__conv_weight_to_fp16 = const()[name = string("m_convnext_5_dwconv__conv_weight_to_fp16"), val = tensor<fp16, [256, 1, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9539712)))];
            tensor<fp16, [256]> m_convnext_5_dwconv__conv_bias_to_fp16 = const()[name = string("m_convnext_5_dwconv__conv_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9542336)))];
            tensor<fp16, [1, 256, 128]> h_31_cast_fp16 = conv(bias = m_convnext_5_dwconv__conv_bias_to_fp16, dilations = h_31_dilations_0, groups = h_31_groups_0, pad = h_31_pad_0, pad_type = h_31_pad_type_0, strides = h_31_strides_0, weight = m_convnext_5_dwconv__conv_weight_to_fp16, x = input_67_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_37_cast_fp16 = mul(x = h_31_cast_fp16, y = text_mask_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<int32, [3]> input_69_perm_0 = const()[name = string("input_69_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_361_axes_0 = const()[name = string("op_361_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_convnext_5_norm_norm_weight_to_fp16 = const()[name = string("m_convnext_5_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9542912)))];
            tensor<fp16, [256]> m_convnext_5_norm_norm_bias_to_fp16 = const()[name = string("m_convnext_5_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9543488)))];
            tensor<fp16, [1, 128, 256]> input_69_cast_fp16 = transpose(perm = input_69_perm_0, x = x_37_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 128, 256]> var_361_cast_fp16 = layer_norm(axes = var_361_axes_0, beta = m_convnext_5_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_convnext_5_norm_norm_weight_to_fp16, x = input_69_cast_fp16)[name = string("op_361_cast_fp16")];
            tensor<int32, [3]> input_71_perm_0 = const()[name = string("input_71_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string h_33_pad_type_0 = const()[name = string("h_33_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_33_strides_0 = const()[name = string("h_33_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_33_pad_0 = const()[name = string("h_33_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_33_dilations_0 = const()[name = string("h_33_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_33_groups_0 = const()[name = string("h_33_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_convnext_5_pwconv1_weight_to_fp16 = const()[name = string("m_convnext_5_pwconv1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9544064)))];
            tensor<fp16, [1024]> m_convnext_5_pwconv1_bias_to_fp16 = const()[name = string("m_convnext_5_pwconv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10068416)))];
            tensor<fp16, [1, 256, 128]> input_71_cast_fp16 = transpose(perm = input_71_perm_0, x = var_361_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 1024, 128]> h_33_cast_fp16 = conv(bias = m_convnext_5_pwconv1_bias_to_fp16, dilations = h_33_dilations_0, groups = h_33_groups_0, pad = h_33_pad_0, pad_type = h_33_pad_type_0, strides = h_33_strides_0, weight = m_convnext_5_pwconv1_weight_to_fp16, x = input_71_cast_fp16)[name = string("h_33_cast_fp16")];
            string input_73_mode_0 = const()[name = string("input_73_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 128]> input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = h_33_cast_fp16)[name = string("input_73_cast_fp16")];
            string h_35_pad_type_0 = const()[name = string("h_35_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_35_strides_0 = const()[name = string("h_35_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_35_pad_0 = const()[name = string("h_35_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_35_dilations_0 = const()[name = string("h_35_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_35_groups_0 = const()[name = string("h_35_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> var_378_weight_0_to_fp16 = const()[name = string("op_378_weight_0_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10070528)))];
            tensor<fp16, [256]> var_378_bias_0_to_fp16 = const()[name = string("op_378_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10594880)))];
            tensor<fp16, [1, 256, 128]> var_378_cast_fp16 = conv(bias = var_378_bias_0_to_fp16, dilations = h_35_dilations_0, groups = h_35_groups_0, pad = h_35_pad_0, pad_type = h_35_pad_type_0, strides = h_35_strides_0, weight = var_378_weight_0_to_fp16, x = input_73_cast_fp16)[name = string("op_378_cast_fp16")];
            tensor<fp16, [1, 256, 128]> out_11_cast_fp16 = add(x = x_35_cast_fp16, y = var_378_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_39_cast_fp16 = mul(x = out_11_cast_fp16, y = text_mask_to_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_41_cast_fp16 = mul(x = x_39_cast_fp16, y = text_mask_to_fp16)[name = string("x_41_cast_fp16")];
            string var_401_pad_type_0 = const()[name = string("op_401_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_401_strides_0 = const()[name = string("op_401_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_401_pad_0 = const()[name = string("op_401_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_401_dilations_0 = const()[name = string("op_401_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_401_groups_0 = const()[name = string("op_401_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_0_attn_conv_q_weight_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_q_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10595456)))];
            tensor<fp16, [256]> m_attn_layers_0_attn_conv_q_bias_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_q_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10726592)))];
            tensor<fp16, [1, 256, 128]> var_401_cast_fp16 = conv(bias = m_attn_layers_0_attn_conv_q_bias_to_fp16, dilations = var_401_dilations_0, groups = var_401_groups_0, pad = var_401_pad_0, pad_type = var_401_pad_type_0, strides = var_401_strides_0, weight = m_attn_layers_0_attn_conv_q_weight_to_fp16, x = x_41_cast_fp16)[name = string("op_401_cast_fp16")];
            tensor<int32, [4]> var_402 = const()[name = string("op_402"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_403_cast_fp16 = reshape(shape = var_402, x = var_401_cast_fp16)[name = string("op_403_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_411_pad_type_0 = const()[name = string("op_411_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_411_strides_0 = const()[name = string("op_411_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_411_pad_0 = const()[name = string("op_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_411_dilations_0 = const()[name = string("op_411_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_411_groups_0 = const()[name = string("op_411_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_0_attn_conv_k_weight_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_k_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10727168)))];
            tensor<fp16, [256]> m_attn_layers_0_attn_conv_k_bias_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_k_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10858304)))];
            tensor<fp16, [1, 256, 128]> var_411_cast_fp16 = conv(bias = m_attn_layers_0_attn_conv_k_bias_to_fp16, dilations = var_411_dilations_0, groups = var_411_groups_0, pad = var_411_pad_0, pad_type = var_411_pad_type_0, strides = var_411_strides_0, weight = m_attn_layers_0_attn_conv_k_weight_to_fp16, x = x_41_cast_fp16)[name = string("op_411_cast_fp16")];
            tensor<int32, [4]> var_412 = const()[name = string("op_412"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_413_cast_fp16 = reshape(shape = var_412, x = var_411_cast_fp16)[name = string("op_413_cast_fp16")];
            string var_421_pad_type_0 = const()[name = string("op_421_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_421_strides_0 = const()[name = string("op_421_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_421_pad_0 = const()[name = string("op_421_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_421_dilations_0 = const()[name = string("op_421_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_421_groups_0 = const()[name = string("op_421_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_0_attn_conv_v_weight_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_v_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10858880)))];
            tensor<fp16, [256]> m_attn_layers_0_attn_conv_v_bias_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_v_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10990016)))];
            tensor<fp16, [1, 256, 128]> var_421_cast_fp16 = conv(bias = m_attn_layers_0_attn_conv_v_bias_to_fp16, dilations = var_421_dilations_0, groups = var_421_groups_0, pad = var_421_pad_0, pad_type = var_421_pad_type_0, strides = var_421_strides_0, weight = m_attn_layers_0_attn_conv_v_weight_to_fp16, x = x_41_cast_fp16)[name = string("op_421_cast_fp16")];
            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_423_cast_fp16 = reshape(shape = var_422, x = var_421_cast_fp16)[name = string("op_423_cast_fp16")];
            fp16 var_425_to_fp16 = const()[name = string("op_425_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 4, 128, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_403_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 128, 64]> var_426_cast_fp16 = mul(x = q_1_cast_fp16, y = var_425_to_fp16)[name = string("op_426_cast_fp16")];
            bool scores_1_transpose_x_0 = const()[name = string("scores_1_transpose_x_0"), val = bool(false)];
            bool scores_1_transpose_y_0 = const()[name = string("scores_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> scores_1_cast_fp16 = matmul(transpose_x = scores_1_transpose_x_0, transpose_y = scores_1_transpose_y_0, x = var_426_cast_fp16, y = var_413_cast_fp16)[name = string("scores_1_cast_fp16")];
            bool x_43_transpose_x_0 = const()[name = string("x_43_transpose_x_0"), val = bool(false)];
            bool x_43_transpose_y_0 = const()[name = string("x_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 255]> var_450_to_fp16 = const()[name = string("op_450_to_fp16"), val = tensor<fp16, [1, 1, 64, 255]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10990592)))];
            tensor<fp16, [1, 4, 128, 255]> x_43_cast_fp16 = matmul(transpose_x = x_43_transpose_x_0, transpose_y = x_43_transpose_y_0, x = var_426_cast_fp16, y = var_450_to_fp16)[name = string("x_43_cast_fp16")];
            tensor<int32, [8]> x_45_pad_0 = const()[name = string("x_45_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            string x_45_mode_0 = const()[name = string("x_45_mode_0"), val = string("constant")];
            fp16 const_18_to_fp16 = const()[name = string("const_18_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 256]> x_45_cast_fp16 = pad(constant_val = const_18_to_fp16, mode = x_45_mode_0, pad = x_45_pad_0, x = x_43_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [3]> var_462 = const()[name = string("op_462"), val = tensor<int32, [3]>([1, 4, 32768])];
            tensor<fp16, [1, 4, 32768]> input_77_cast_fp16 = reshape(shape = var_462, x = x_45_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<int32, [6]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 127])];
            string x_47_mode_0 = const()[name = string("x_47_mode_0"), val = string("constant")];
            fp16 const_19_to_fp16 = const()[name = string("const_19_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32895]> x_47_cast_fp16 = pad(constant_val = const_19_to_fp16, mode = x_47_mode_0, pad = x_47_pad_0, x = input_77_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<int32, [4]> var_477 = const()[name = string("op_477"), val = tensor<int32, [4]>([1, 4, 129, 255])];
            tensor<fp16, [1, 4, 129, 255]> x_49_cast_fp16 = reshape(shape = var_477, x = x_47_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [4]> var_484_begin_0 = const()[name = string("op_484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_484_end_0 = const()[name = string("op_484_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_484_end_mask_0 = const()[name = string("op_484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 128, 255]> var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = x_49_cast_fp16)[name = string("op_484_cast_fp16")];
            tensor<int32, [4]> var_485_begin_0 = const()[name = string("op_485_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 127])];
            tensor<int32, [4]> var_485_end_0 = const()[name = string("op_485_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_485_end_mask_0 = const()[name = string("op_485_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_485_cast_fp16 = slice_by_index(begin = var_485_begin_0, end = var_485_end_0, end_mask = var_485_end_mask_0, x = var_484_cast_fp16)[name = string("op_485_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_485_cast_fp16)[name = string("scores_3_cast_fp16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 128]> var_487_cast_fp16 = expand_dims(axes = var_487_axes_0, x = text_mask_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<fp16, [1, 1, 1, 128]> var_488_cast_fp16 = sub(x = var_21_to_fp16, y = var_487_cast_fp16)[name = string("op_488_cast_fp16")];
            fp16 var_489_to_fp16 = const()[name = string("op_489_to_fp16"), val = fp16(0x1.388p+13)];
            tensor<fp16, [1, 1, 1, 128]> var_490_cast_fp16 = mul(x = var_488_cast_fp16, y = var_489_to_fp16)[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> input_79_cast_fp16 = sub(x = scores_3_cast_fp16, y = var_490_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> p_attn_1_cast_fp16 = softmax(axis = var_29, x = input_79_cast_fp16)[name = string("p_attn_1_cast_fp16")];
            bool out_13_transpose_x_1 = const()[name = string("out_13_transpose_x_1"), val = bool(false)];
            bool out_13_transpose_y_1 = const()[name = string("out_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 64]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_1, transpose_y = out_13_transpose_y_1, x = p_attn_1_cast_fp16, y = var_423_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [8]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 127])];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("constant")];
            fp16 const_24_to_fp16 = const()[name = string("const_24_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 255]> x_51_cast_fp16 = pad(constant_val = const_24_to_fp16, mode = x_51_mode_0, pad = x_51_pad_0, x = p_attn_1_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [3]> var_527 = const()[name = string("op_527"), val = tensor<int32, [3]>([1, 4, 32640])];
            tensor<fp16, [1, 4, 32640]> input_83_cast_fp16 = reshape(shape = var_527, x = x_51_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [6]> x_53_pad_0 = const()[name = string("x_53_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 128, 0])];
            string x_53_mode_0 = const()[name = string("x_53_mode_0"), val = string("constant")];
            fp16 const_25_to_fp16 = const()[name = string("const_25_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32768]> x_53_cast_fp16 = pad(constant_val = const_25_to_fp16, mode = x_53_mode_0, pad = x_53_pad_0, x = input_83_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<int32, [4]> var_534 = const()[name = string("op_534"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<fp16, [1, 4, 128, 256]> x_55_cast_fp16 = reshape(shape = var_534, x = x_53_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<int32, [4]> rel_weights_1_begin_0 = const()[name = string("rel_weights_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> rel_weights_1_end_0 = const()[name = string("rel_weights_1_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> rel_weights_1_end_mask_0 = const()[name = string("rel_weights_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 255]> rel_weights_1_cast_fp16 = slice_by_index(begin = rel_weights_1_begin_0, end = rel_weights_1_end_0, end_mask = rel_weights_1_end_mask_0, x = x_55_cast_fp16)[name = string("rel_weights_1_cast_fp16")];
            bool var_541_transpose_x_0 = const()[name = string("op_541_transpose_x_0"), val = bool(false)];
            bool var_541_transpose_y_0 = const()[name = string("op_541_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 255, 64]> var_540_to_fp16 = const()[name = string("op_540_to_fp16"), val = tensor<fp16, [1, 1, 255, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11023296)))];
            tensor<fp16, [1, 4, 128, 64]> var_541_cast_fp16 = matmul(transpose_x = var_541_transpose_x_0, transpose_y = var_541_transpose_y_0, x = rel_weights_1_cast_fp16, y = var_540_to_fp16)[name = string("op_541_cast_fp16")];
            tensor<fp16, [1, 4, 128, 64]> out_15_cast_fp16 = add(x = out_13_cast_fp16, y = var_541_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [4]> var_543_perm_0 = const()[name = string("op_543_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_544 = const()[name = string("op_544"), val = tensor<int32, [3]>([1, 256, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_543_cast_fp16 = transpose(perm = var_543_perm_0, x = out_15_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 256, 128]> input_85_cast_fp16 = reshape(shape = var_544, x = var_543_cast_fp16)[name = string("input_85_cast_fp16")];
            string y_1_pad_type_0 = const()[name = string("y_1_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> y_1_strides_0 = const()[name = string("y_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> y_1_pad_0 = const()[name = string("y_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> y_1_dilations_0 = const()[name = string("y_1_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 y_1_groups_0 = const()[name = string("y_1_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_0_attn_conv_o_weight_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_o_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11056000)))];
            tensor<fp16, [256]> m_attn_layers_0_attn_conv_o_bias_to_fp16 = const()[name = string("m_attn_layers_0_attn_conv_o_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11187136)))];
            tensor<fp16, [1, 256, 128]> y_1_cast_fp16 = conv(bias = m_attn_layers_0_attn_conv_o_bias_to_fp16, dilations = y_1_dilations_0, groups = y_1_groups_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = m_attn_layers_0_attn_conv_o_weight_to_fp16, x = input_85_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_57_cast_fp16 = add(x = x_41_cast_fp16, y = y_1_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<int32, [3]> input_87_perm_0 = const()[name = string("input_87_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_559_axes_0 = const()[name = string("op_559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_0_norm_1_norm_weight_to_fp16 = const()[name = string("m_attn_layers_0_norm_1_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11187712)))];
            tensor<fp16, [256]> m_attn_layers_0_norm_1_norm_bias_to_fp16 = const()[name = string("m_attn_layers_0_norm_1_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11188288)))];
            tensor<fp16, [1, 128, 256]> input_87_cast_fp16 = transpose(perm = input_87_perm_0, x = x_57_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 128, 256]> var_559_cast_fp16 = layer_norm(axes = var_559_axes_0, beta = m_attn_layers_0_norm_1_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_0_norm_1_norm_weight_to_fp16, x = input_87_cast_fp16)[name = string("op_559_cast_fp16")];
            tensor<int32, [3]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_559_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 256, 128]> input_89_cast_fp16 = mul(x = x_59_cast_fp16, y = text_mask_to_fp16)[name = string("input_89_cast_fp16")];
            string input_91_pad_type_0 = const()[name = string("input_91_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input_91_strides_0 = const()[name = string("input_91_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_91_pad_0 = const()[name = string("input_91_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_91_dilations_0 = const()[name = string("input_91_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input_91_groups_0 = const()[name = string("input_91_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_attn_layers_0_ffn_conv_1_weight_to_fp16 = const()[name = string("m_attn_layers_0_ffn_conv_1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11188864)))];
            tensor<fp16, [1024]> m_attn_layers_0_ffn_conv_1_bias_to_fp16 = const()[name = string("m_attn_layers_0_ffn_conv_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11713216)))];
            tensor<fp16, [1, 1024, 128]> input_91_cast_fp16 = conv(bias = m_attn_layers_0_ffn_conv_1_bias_to_fp16, dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = m_attn_layers_0_ffn_conv_1_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> h_37_cast_fp16 = relu(x = input_91_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> input_93_cast_fp16 = mul(x = h_37_cast_fp16, y = text_mask_to_fp16)[name = string("input_93_cast_fp16")];
            string h_39_pad_type_0 = const()[name = string("h_39_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_39_strides_0 = const()[name = string("h_39_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_39_pad_0 = const()[name = string("h_39_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_39_dilations_0 = const()[name = string("h_39_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_39_groups_0 = const()[name = string("h_39_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> m_attn_layers_0_ffn_conv_2_weight_to_fp16 = const()[name = string("m_attn_layers_0_ffn_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11715328)))];
            tensor<fp16, [256]> m_attn_layers_0_ffn_conv_2_bias_to_fp16 = const()[name = string("m_attn_layers_0_ffn_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12239680)))];
            tensor<fp16, [1, 256, 128]> h_39_cast_fp16 = conv(bias = m_attn_layers_0_ffn_conv_2_bias_to_fp16, dilations = h_39_dilations_0, groups = h_39_groups_0, pad = h_39_pad_0, pad_type = h_39_pad_type_0, strides = h_39_strides_0, weight = m_attn_layers_0_ffn_conv_2_weight_to_fp16, x = input_93_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1, 256, 128]> y_3_cast_fp16 = mul(x = h_39_cast_fp16, y = text_mask_to_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = y_3_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [3]> input_95_perm_0 = const()[name = string("input_95_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_587_axes_0 = const()[name = string("op_587_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_0_norm_2_norm_weight_to_fp16 = const()[name = string("m_attn_layers_0_norm_2_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12240256)))];
            tensor<fp16, [256]> m_attn_layers_0_norm_2_norm_bias_to_fp16 = const()[name = string("m_attn_layers_0_norm_2_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12240832)))];
            tensor<fp16, [1, 128, 256]> input_95_cast_fp16 = transpose(perm = input_95_perm_0, x = x_61_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 128, 256]> var_587_cast_fp16 = layer_norm(axes = var_587_axes_0, beta = m_attn_layers_0_norm_2_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_0_norm_2_norm_weight_to_fp16, x = input_95_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<int32, [3]> x_63_perm_0 = const()[name = string("x_63_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_63_cast_fp16 = transpose(perm = x_63_perm_0, x = var_587_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 256, 128]> x_65_cast_fp16 = mul(x = x_63_cast_fp16, y = text_mask_to_fp16)[name = string("x_65_cast_fp16")];
            string var_609_pad_type_0 = const()[name = string("op_609_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_609_strides_0 = const()[name = string("op_609_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_609_pad_0 = const()[name = string("op_609_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_609_dilations_0 = const()[name = string("op_609_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_609_groups_0 = const()[name = string("op_609_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_1_attn_conv_q_weight_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_q_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12241408)))];
            tensor<fp16, [256]> m_attn_layers_1_attn_conv_q_bias_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_q_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12372544)))];
            tensor<fp16, [1, 256, 128]> var_609_cast_fp16 = conv(bias = m_attn_layers_1_attn_conv_q_bias_to_fp16, dilations = var_609_dilations_0, groups = var_609_groups_0, pad = var_609_pad_0, pad_type = var_609_pad_type_0, strides = var_609_strides_0, weight = m_attn_layers_1_attn_conv_q_weight_to_fp16, x = x_65_cast_fp16)[name = string("op_609_cast_fp16")];
            tensor<int32, [4]> var_610 = const()[name = string("op_610"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_611_cast_fp16 = reshape(shape = var_610, x = var_609_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_619_pad_type_0 = const()[name = string("op_619_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_619_strides_0 = const()[name = string("op_619_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_619_pad_0 = const()[name = string("op_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_619_dilations_0 = const()[name = string("op_619_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_619_groups_0 = const()[name = string("op_619_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_1_attn_conv_k_weight_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_k_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12373120)))];
            tensor<fp16, [256]> m_attn_layers_1_attn_conv_k_bias_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_k_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12504256)))];
            tensor<fp16, [1, 256, 128]> var_619_cast_fp16 = conv(bias = m_attn_layers_1_attn_conv_k_bias_to_fp16, dilations = var_619_dilations_0, groups = var_619_groups_0, pad = var_619_pad_0, pad_type = var_619_pad_type_0, strides = var_619_strides_0, weight = m_attn_layers_1_attn_conv_k_weight_to_fp16, x = x_65_cast_fp16)[name = string("op_619_cast_fp16")];
            tensor<int32, [4]> var_620 = const()[name = string("op_620"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_621_cast_fp16 = reshape(shape = var_620, x = var_619_cast_fp16)[name = string("op_621_cast_fp16")];
            string var_629_pad_type_0 = const()[name = string("op_629_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_629_strides_0 = const()[name = string("op_629_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_629_pad_0 = const()[name = string("op_629_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_629_dilations_0 = const()[name = string("op_629_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_629_groups_0 = const()[name = string("op_629_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_1_attn_conv_v_weight_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_v_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12504832)))];
            tensor<fp16, [256]> m_attn_layers_1_attn_conv_v_bias_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_v_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12635968)))];
            tensor<fp16, [1, 256, 128]> var_629_cast_fp16 = conv(bias = m_attn_layers_1_attn_conv_v_bias_to_fp16, dilations = var_629_dilations_0, groups = var_629_groups_0, pad = var_629_pad_0, pad_type = var_629_pad_type_0, strides = var_629_strides_0, weight = m_attn_layers_1_attn_conv_v_weight_to_fp16, x = x_65_cast_fp16)[name = string("op_629_cast_fp16")];
            tensor<int32, [4]> var_630 = const()[name = string("op_630"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_631_cast_fp16 = reshape(shape = var_630, x = var_629_cast_fp16)[name = string("op_631_cast_fp16")];
            fp16 var_633_to_fp16 = const()[name = string("op_633_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 4, 128, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_611_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 4, 128, 64]> var_634_cast_fp16 = mul(x = q_3_cast_fp16, y = var_633_to_fp16)[name = string("op_634_cast_fp16")];
            bool scores_5_transpose_x_0 = const()[name = string("scores_5_transpose_x_0"), val = bool(false)];
            bool scores_5_transpose_y_0 = const()[name = string("scores_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> scores_5_cast_fp16 = matmul(transpose_x = scores_5_transpose_x_0, transpose_y = scores_5_transpose_y_0, x = var_634_cast_fp16, y = var_621_cast_fp16)[name = string("scores_5_cast_fp16")];
            bool x_67_transpose_x_0 = const()[name = string("x_67_transpose_x_0"), val = bool(false)];
            bool x_67_transpose_y_0 = const()[name = string("x_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 255]> var_658_to_fp16 = const()[name = string("op_658_to_fp16"), val = tensor<fp16, [1, 1, 64, 255]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12636544)))];
            tensor<fp16, [1, 4, 128, 255]> x_67_cast_fp16 = matmul(transpose_x = x_67_transpose_x_0, transpose_y = x_67_transpose_y_0, x = var_634_cast_fp16, y = var_658_to_fp16)[name = string("x_67_cast_fp16")];
            tensor<int32, [8]> x_69_pad_0 = const()[name = string("x_69_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("constant")];
            fp16 const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 256]> x_69_cast_fp16 = pad(constant_val = const_32_to_fp16, mode = x_69_mode_0, pad = x_69_pad_0, x = x_67_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [3]> var_670 = const()[name = string("op_670"), val = tensor<int32, [3]>([1, 4, 32768])];
            tensor<fp16, [1, 4, 32768]> input_99_cast_fp16 = reshape(shape = var_670, x = x_69_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [6]> x_71_pad_0 = const()[name = string("x_71_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 127])];
            string x_71_mode_0 = const()[name = string("x_71_mode_0"), val = string("constant")];
            fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32895]> x_71_cast_fp16 = pad(constant_val = const_33_to_fp16, mode = x_71_mode_0, pad = x_71_pad_0, x = input_99_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> var_685 = const()[name = string("op_685"), val = tensor<int32, [4]>([1, 4, 129, 255])];
            tensor<fp16, [1, 4, 129, 255]> x_73_cast_fp16 = reshape(shape = var_685, x = x_71_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 128, 255]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = x_73_cast_fp16)[name = string("op_692_cast_fp16")];
            tensor<int32, [4]> var_693_begin_0 = const()[name = string("op_693_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 127])];
            tensor<int32, [4]> var_693_end_0 = const()[name = string("op_693_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_693_end_mask_0 = const()[name = string("op_693_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_693_cast_fp16 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = var_692_cast_fp16)[name = string("op_693_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_693_cast_fp16)[name = string("scores_7_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> input_101_cast_fp16 = sub(x = scores_7_cast_fp16, y = var_490_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> p_attn_3_cast_fp16 = softmax(axis = var_29, x = input_101_cast_fp16)[name = string("p_attn_3_cast_fp16")];
            bool out_17_transpose_x_1 = const()[name = string("out_17_transpose_x_1"), val = bool(false)];
            bool out_17_transpose_y_1 = const()[name = string("out_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 64]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_1, transpose_y = out_17_transpose_y_1, x = p_attn_3_cast_fp16, y = var_631_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [8]> x_75_pad_0 = const()[name = string("x_75_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 127])];
            string x_75_mode_0 = const()[name = string("x_75_mode_0"), val = string("constant")];
            fp16 const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 255]> x_75_cast_fp16 = pad(constant_val = const_38_to_fp16, mode = x_75_mode_0, pad = x_75_pad_0, x = p_attn_3_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<int32, [3]> var_735 = const()[name = string("op_735"), val = tensor<int32, [3]>([1, 4, 32640])];
            tensor<fp16, [1, 4, 32640]> input_105_cast_fp16 = reshape(shape = var_735, x = x_75_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [6]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 128, 0])];
            string x_77_mode_0 = const()[name = string("x_77_mode_0"), val = string("constant")];
            fp16 const_39_to_fp16 = const()[name = string("const_39_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32768]> x_77_cast_fp16 = pad(constant_val = const_39_to_fp16, mode = x_77_mode_0, pad = x_77_pad_0, x = input_105_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<int32, [4]> var_742 = const()[name = string("op_742"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<fp16, [1, 4, 128, 256]> x_79_cast_fp16 = reshape(shape = var_742, x = x_77_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [4]> rel_weights_3_begin_0 = const()[name = string("rel_weights_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> rel_weights_3_end_0 = const()[name = string("rel_weights_3_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> rel_weights_3_end_mask_0 = const()[name = string("rel_weights_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 255]> rel_weights_3_cast_fp16 = slice_by_index(begin = rel_weights_3_begin_0, end = rel_weights_3_end_0, end_mask = rel_weights_3_end_mask_0, x = x_79_cast_fp16)[name = string("rel_weights_3_cast_fp16")];
            bool var_749_transpose_x_0 = const()[name = string("op_749_transpose_x_0"), val = bool(false)];
            bool var_749_transpose_y_0 = const()[name = string("op_749_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 255, 64]> var_748_to_fp16 = const()[name = string("op_748_to_fp16"), val = tensor<fp16, [1, 1, 255, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12669248)))];
            tensor<fp16, [1, 4, 128, 64]> var_749_cast_fp16 = matmul(transpose_x = var_749_transpose_x_0, transpose_y = var_749_transpose_y_0, x = rel_weights_3_cast_fp16, y = var_748_to_fp16)[name = string("op_749_cast_fp16")];
            tensor<fp16, [1, 4, 128, 64]> out_19_cast_fp16 = add(x = out_17_cast_fp16, y = var_749_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [4]> var_751_perm_0 = const()[name = string("op_751_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_752 = const()[name = string("op_752"), val = tensor<int32, [3]>([1, 256, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_751_cast_fp16 = transpose(perm = var_751_perm_0, x = out_19_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 256, 128]> input_107_cast_fp16 = reshape(shape = var_752, x = var_751_cast_fp16)[name = string("input_107_cast_fp16")];
            string y_5_pad_type_0 = const()[name = string("y_5_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> y_5_strides_0 = const()[name = string("y_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> y_5_pad_0 = const()[name = string("y_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> y_5_dilations_0 = const()[name = string("y_5_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 y_5_groups_0 = const()[name = string("y_5_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_1_attn_conv_o_weight_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_o_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12701952)))];
            tensor<fp16, [256]> m_attn_layers_1_attn_conv_o_bias_to_fp16 = const()[name = string("m_attn_layers_1_attn_conv_o_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12833088)))];
            tensor<fp16, [1, 256, 128]> y_5_cast_fp16 = conv(bias = m_attn_layers_1_attn_conv_o_bias_to_fp16, dilations = y_5_dilations_0, groups = y_5_groups_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = m_attn_layers_1_attn_conv_o_weight_to_fp16, x = input_107_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_81_cast_fp16 = add(x = x_65_cast_fp16, y = y_5_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_767_axes_0 = const()[name = string("op_767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_1_norm_1_norm_weight_to_fp16 = const()[name = string("m_attn_layers_1_norm_1_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12833664)))];
            tensor<fp16, [256]> m_attn_layers_1_norm_1_norm_bias_to_fp16 = const()[name = string("m_attn_layers_1_norm_1_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12834240)))];
            tensor<fp16, [1, 128, 256]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_81_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 128, 256]> var_767_cast_fp16 = layer_norm(axes = var_767_axes_0, beta = m_attn_layers_1_norm_1_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_1_norm_1_norm_weight_to_fp16, x = input_109_cast_fp16)[name = string("op_767_cast_fp16")];
            tensor<int32, [3]> x_83_perm_0 = const()[name = string("x_83_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_83_cast_fp16 = transpose(perm = x_83_perm_0, x = var_767_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 256, 128]> input_111_cast_fp16 = mul(x = x_83_cast_fp16, y = text_mask_to_fp16)[name = string("input_111_cast_fp16")];
            string input_113_pad_type_0 = const()[name = string("input_113_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input_113_strides_0 = const()[name = string("input_113_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_113_pad_0 = const()[name = string("input_113_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_113_dilations_0 = const()[name = string("input_113_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input_113_groups_0 = const()[name = string("input_113_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_attn_layers_1_ffn_conv_1_weight_to_fp16 = const()[name = string("m_attn_layers_1_ffn_conv_1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12834816)))];
            tensor<fp16, [1024]> m_attn_layers_1_ffn_conv_1_bias_to_fp16 = const()[name = string("m_attn_layers_1_ffn_conv_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13359168)))];
            tensor<fp16, [1, 1024, 128]> input_113_cast_fp16 = conv(bias = m_attn_layers_1_ffn_conv_1_bias_to_fp16, dilations = input_113_dilations_0, groups = input_113_groups_0, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = input_113_strides_0, weight = m_attn_layers_1_ffn_conv_1_weight_to_fp16, x = input_111_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> h_41_cast_fp16 = relu(x = input_113_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> input_115_cast_fp16 = mul(x = h_41_cast_fp16, y = text_mask_to_fp16)[name = string("input_115_cast_fp16")];
            string h_43_pad_type_0 = const()[name = string("h_43_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_43_strides_0 = const()[name = string("h_43_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_43_pad_0 = const()[name = string("h_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_43_dilations_0 = const()[name = string("h_43_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_43_groups_0 = const()[name = string("h_43_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> m_attn_layers_1_ffn_conv_2_weight_to_fp16 = const()[name = string("m_attn_layers_1_ffn_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13361280)))];
            tensor<fp16, [256]> m_attn_layers_1_ffn_conv_2_bias_to_fp16 = const()[name = string("m_attn_layers_1_ffn_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13885632)))];
            tensor<fp16, [1, 256, 128]> h_43_cast_fp16 = conv(bias = m_attn_layers_1_ffn_conv_2_bias_to_fp16, dilations = h_43_dilations_0, groups = h_43_groups_0, pad = h_43_pad_0, pad_type = h_43_pad_type_0, strides = h_43_strides_0, weight = m_attn_layers_1_ffn_conv_2_weight_to_fp16, x = input_115_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1, 256, 128]> y_7_cast_fp16 = mul(x = h_43_cast_fp16, y = text_mask_to_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = y_7_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<int32, [3]> input_117_perm_0 = const()[name = string("input_117_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_795_axes_0 = const()[name = string("op_795_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_1_norm_2_norm_weight_to_fp16 = const()[name = string("m_attn_layers_1_norm_2_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13886208)))];
            tensor<fp16, [256]> m_attn_layers_1_norm_2_norm_bias_to_fp16 = const()[name = string("m_attn_layers_1_norm_2_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13886784)))];
            tensor<fp16, [1, 128, 256]> input_117_cast_fp16 = transpose(perm = input_117_perm_0, x = x_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 128, 256]> var_795_cast_fp16 = layer_norm(axes = var_795_axes_0, beta = m_attn_layers_1_norm_2_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_1_norm_2_norm_weight_to_fp16, x = input_117_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [3]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_795_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 256, 128]> x_89_cast_fp16 = mul(x = x_87_cast_fp16, y = text_mask_to_fp16)[name = string("x_89_cast_fp16")];
            string var_817_pad_type_0 = const()[name = string("op_817_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_817_strides_0 = const()[name = string("op_817_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_817_pad_0 = const()[name = string("op_817_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_817_dilations_0 = const()[name = string("op_817_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_817_groups_0 = const()[name = string("op_817_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_2_attn_conv_q_weight_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_q_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13887360)))];
            tensor<fp16, [256]> m_attn_layers_2_attn_conv_q_bias_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_q_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14018496)))];
            tensor<fp16, [1, 256, 128]> var_817_cast_fp16 = conv(bias = m_attn_layers_2_attn_conv_q_bias_to_fp16, dilations = var_817_dilations_0, groups = var_817_groups_0, pad = var_817_pad_0, pad_type = var_817_pad_type_0, strides = var_817_strides_0, weight = m_attn_layers_2_attn_conv_q_weight_to_fp16, x = x_89_cast_fp16)[name = string("op_817_cast_fp16")];
            tensor<int32, [4]> var_818 = const()[name = string("op_818"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_819_cast_fp16 = reshape(shape = var_818, x = var_817_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_827_pad_type_0 = const()[name = string("op_827_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_827_strides_0 = const()[name = string("op_827_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_827_pad_0 = const()[name = string("op_827_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_827_dilations_0 = const()[name = string("op_827_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_827_groups_0 = const()[name = string("op_827_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_2_attn_conv_k_weight_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_k_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14019072)))];
            tensor<fp16, [256]> m_attn_layers_2_attn_conv_k_bias_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_k_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14150208)))];
            tensor<fp16, [1, 256, 128]> var_827_cast_fp16 = conv(bias = m_attn_layers_2_attn_conv_k_bias_to_fp16, dilations = var_827_dilations_0, groups = var_827_groups_0, pad = var_827_pad_0, pad_type = var_827_pad_type_0, strides = var_827_strides_0, weight = m_attn_layers_2_attn_conv_k_weight_to_fp16, x = x_89_cast_fp16)[name = string("op_827_cast_fp16")];
            tensor<int32, [4]> var_828 = const()[name = string("op_828"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_829_cast_fp16 = reshape(shape = var_828, x = var_827_cast_fp16)[name = string("op_829_cast_fp16")];
            string var_837_pad_type_0 = const()[name = string("op_837_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_837_strides_0 = const()[name = string("op_837_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_837_pad_0 = const()[name = string("op_837_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_837_dilations_0 = const()[name = string("op_837_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_837_groups_0 = const()[name = string("op_837_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_2_attn_conv_v_weight_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_v_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14150784)))];
            tensor<fp16, [256]> m_attn_layers_2_attn_conv_v_bias_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_v_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14281920)))];
            tensor<fp16, [1, 256, 128]> var_837_cast_fp16 = conv(bias = m_attn_layers_2_attn_conv_v_bias_to_fp16, dilations = var_837_dilations_0, groups = var_837_groups_0, pad = var_837_pad_0, pad_type = var_837_pad_type_0, strides = var_837_strides_0, weight = m_attn_layers_2_attn_conv_v_weight_to_fp16, x = x_89_cast_fp16)[name = string("op_837_cast_fp16")];
            tensor<int32, [4]> var_838 = const()[name = string("op_838"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_839_cast_fp16 = reshape(shape = var_838, x = var_837_cast_fp16)[name = string("op_839_cast_fp16")];
            fp16 var_841_to_fp16 = const()[name = string("op_841_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 4, 128, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_819_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 4, 128, 64]> var_842_cast_fp16 = mul(x = q_5_cast_fp16, y = var_841_to_fp16)[name = string("op_842_cast_fp16")];
            bool scores_9_transpose_x_0 = const()[name = string("scores_9_transpose_x_0"), val = bool(false)];
            bool scores_9_transpose_y_0 = const()[name = string("scores_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> scores_9_cast_fp16 = matmul(transpose_x = scores_9_transpose_x_0, transpose_y = scores_9_transpose_y_0, x = var_842_cast_fp16, y = var_829_cast_fp16)[name = string("scores_9_cast_fp16")];
            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 255]> var_866_to_fp16 = const()[name = string("op_866_to_fp16"), val = tensor<fp16, [1, 1, 64, 255]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14282496)))];
            tensor<fp16, [1, 4, 128, 255]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = var_842_cast_fp16, y = var_866_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [8]> x_93_pad_0 = const()[name = string("x_93_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            string x_93_mode_0 = const()[name = string("x_93_mode_0"), val = string("constant")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 256]> x_93_cast_fp16 = pad(constant_val = const_46_to_fp16, mode = x_93_mode_0, pad = x_93_pad_0, x = x_91_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [3]> var_878 = const()[name = string("op_878"), val = tensor<int32, [3]>([1, 4, 32768])];
            tensor<fp16, [1, 4, 32768]> input_121_cast_fp16 = reshape(shape = var_878, x = x_93_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<int32, [6]> x_95_pad_0 = const()[name = string("x_95_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 127])];
            string x_95_mode_0 = const()[name = string("x_95_mode_0"), val = string("constant")];
            fp16 const_47_to_fp16 = const()[name = string("const_47_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32895]> x_95_cast_fp16 = pad(constant_val = const_47_to_fp16, mode = x_95_mode_0, pad = x_95_pad_0, x = input_121_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<int32, [4]> var_893 = const()[name = string("op_893"), val = tensor<int32, [4]>([1, 4, 129, 255])];
            tensor<fp16, [1, 4, 129, 255]> x_97_cast_fp16 = reshape(shape = var_893, x = x_95_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<int32, [4]> var_900_begin_0 = const()[name = string("op_900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_900_end_0 = const()[name = string("op_900_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_900_end_mask_0 = const()[name = string("op_900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 128, 255]> var_900_cast_fp16 = slice_by_index(begin = var_900_begin_0, end = var_900_end_0, end_mask = var_900_end_mask_0, x = x_97_cast_fp16)[name = string("op_900_cast_fp16")];
            tensor<int32, [4]> var_901_begin_0 = const()[name = string("op_901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 127])];
            tensor<int32, [4]> var_901_end_0 = const()[name = string("op_901_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_901_end_mask_0 = const()[name = string("op_901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_901_cast_fp16 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_901_cast_fp16)[name = string("scores_11_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> input_123_cast_fp16 = sub(x = scores_11_cast_fp16, y = var_490_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> p_attn_5_cast_fp16 = softmax(axis = var_29, x = input_123_cast_fp16)[name = string("p_attn_5_cast_fp16")];
            bool out_21_transpose_x_1 = const()[name = string("out_21_transpose_x_1"), val = bool(false)];
            bool out_21_transpose_y_1 = const()[name = string("out_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 64]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_1, transpose_y = out_21_transpose_y_1, x = p_attn_5_cast_fp16, y = var_839_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [8]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 127])];
            string x_99_mode_0 = const()[name = string("x_99_mode_0"), val = string("constant")];
            fp16 const_52_to_fp16 = const()[name = string("const_52_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 255]> x_99_cast_fp16 = pad(constant_val = const_52_to_fp16, mode = x_99_mode_0, pad = x_99_pad_0, x = p_attn_5_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [3]> var_943 = const()[name = string("op_943"), val = tensor<int32, [3]>([1, 4, 32640])];
            tensor<fp16, [1, 4, 32640]> input_127_cast_fp16 = reshape(shape = var_943, x = x_99_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<int32, [6]> x_101_pad_0 = const()[name = string("x_101_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 128, 0])];
            string x_101_mode_0 = const()[name = string("x_101_mode_0"), val = string("constant")];
            fp16 const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32768]> x_101_cast_fp16 = pad(constant_val = const_53_to_fp16, mode = x_101_mode_0, pad = x_101_pad_0, x = input_127_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> var_950 = const()[name = string("op_950"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<fp16, [1, 4, 128, 256]> x_103_cast_fp16 = reshape(shape = var_950, x = x_101_cast_fp16)[name = string("x_103_cast_fp16")];
            tensor<int32, [4]> rel_weights_5_begin_0 = const()[name = string("rel_weights_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> rel_weights_5_end_0 = const()[name = string("rel_weights_5_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> rel_weights_5_end_mask_0 = const()[name = string("rel_weights_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 255]> rel_weights_5_cast_fp16 = slice_by_index(begin = rel_weights_5_begin_0, end = rel_weights_5_end_0, end_mask = rel_weights_5_end_mask_0, x = x_103_cast_fp16)[name = string("rel_weights_5_cast_fp16")];
            bool var_957_transpose_x_0 = const()[name = string("op_957_transpose_x_0"), val = bool(false)];
            bool var_957_transpose_y_0 = const()[name = string("op_957_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 255, 64]> var_956_to_fp16 = const()[name = string("op_956_to_fp16"), val = tensor<fp16, [1, 1, 255, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14315200)))];
            tensor<fp16, [1, 4, 128, 64]> var_957_cast_fp16 = matmul(transpose_x = var_957_transpose_x_0, transpose_y = var_957_transpose_y_0, x = rel_weights_5_cast_fp16, y = var_956_to_fp16)[name = string("op_957_cast_fp16")];
            tensor<fp16, [1, 4, 128, 64]> out_23_cast_fp16 = add(x = out_21_cast_fp16, y = var_957_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [4]> var_959_perm_0 = const()[name = string("op_959_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_960 = const()[name = string("op_960"), val = tensor<int32, [3]>([1, 256, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_959_cast_fp16 = transpose(perm = var_959_perm_0, x = out_23_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 256, 128]> input_129_cast_fp16 = reshape(shape = var_960, x = var_959_cast_fp16)[name = string("input_129_cast_fp16")];
            string y_9_pad_type_0 = const()[name = string("y_9_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> y_9_strides_0 = const()[name = string("y_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> y_9_pad_0 = const()[name = string("y_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> y_9_dilations_0 = const()[name = string("y_9_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 y_9_groups_0 = const()[name = string("y_9_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_2_attn_conv_o_weight_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_o_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14347904)))];
            tensor<fp16, [256]> m_attn_layers_2_attn_conv_o_bias_to_fp16 = const()[name = string("m_attn_layers_2_attn_conv_o_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14479040)))];
            tensor<fp16, [1, 256, 128]> y_9_cast_fp16 = conv(bias = m_attn_layers_2_attn_conv_o_bias_to_fp16, dilations = y_9_dilations_0, groups = y_9_groups_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = m_attn_layers_2_attn_conv_o_weight_to_fp16, x = input_129_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_105_cast_fp16 = add(x = x_89_cast_fp16, y = y_9_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [3]> input_131_perm_0 = const()[name = string("input_131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_975_axes_0 = const()[name = string("op_975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_2_norm_1_norm_weight_to_fp16 = const()[name = string("m_attn_layers_2_norm_1_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14479616)))];
            tensor<fp16, [256]> m_attn_layers_2_norm_1_norm_bias_to_fp16 = const()[name = string("m_attn_layers_2_norm_1_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14480192)))];
            tensor<fp16, [1, 128, 256]> input_131_cast_fp16 = transpose(perm = input_131_perm_0, x = x_105_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 128, 256]> var_975_cast_fp16 = layer_norm(axes = var_975_axes_0, beta = m_attn_layers_2_norm_1_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_2_norm_1_norm_weight_to_fp16, x = input_131_cast_fp16)[name = string("op_975_cast_fp16")];
            tensor<int32, [3]> x_107_perm_0 = const()[name = string("x_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_107_cast_fp16 = transpose(perm = x_107_perm_0, x = var_975_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 256, 128]> input_133_cast_fp16 = mul(x = x_107_cast_fp16, y = text_mask_to_fp16)[name = string("input_133_cast_fp16")];
            string input_135_pad_type_0 = const()[name = string("input_135_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input_135_strides_0 = const()[name = string("input_135_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_135_pad_0 = const()[name = string("input_135_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_135_dilations_0 = const()[name = string("input_135_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input_135_groups_0 = const()[name = string("input_135_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_attn_layers_2_ffn_conv_1_weight_to_fp16 = const()[name = string("m_attn_layers_2_ffn_conv_1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14480768)))];
            tensor<fp16, [1024]> m_attn_layers_2_ffn_conv_1_bias_to_fp16 = const()[name = string("m_attn_layers_2_ffn_conv_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15005120)))];
            tensor<fp16, [1, 1024, 128]> input_135_cast_fp16 = conv(bias = m_attn_layers_2_ffn_conv_1_bias_to_fp16, dilations = input_135_dilations_0, groups = input_135_groups_0, pad = input_135_pad_0, pad_type = input_135_pad_type_0, strides = input_135_strides_0, weight = m_attn_layers_2_ffn_conv_1_weight_to_fp16, x = input_133_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> h_45_cast_fp16 = relu(x = input_135_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> input_137_cast_fp16 = mul(x = h_45_cast_fp16, y = text_mask_to_fp16)[name = string("input_137_cast_fp16")];
            string h_47_pad_type_0 = const()[name = string("h_47_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_47_strides_0 = const()[name = string("h_47_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_47_pad_0 = const()[name = string("h_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_47_dilations_0 = const()[name = string("h_47_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_47_groups_0 = const()[name = string("h_47_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> m_attn_layers_2_ffn_conv_2_weight_to_fp16 = const()[name = string("m_attn_layers_2_ffn_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15007232)))];
            tensor<fp16, [256]> m_attn_layers_2_ffn_conv_2_bias_to_fp16 = const()[name = string("m_attn_layers_2_ffn_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15531584)))];
            tensor<fp16, [1, 256, 128]> h_47_cast_fp16 = conv(bias = m_attn_layers_2_ffn_conv_2_bias_to_fp16, dilations = h_47_dilations_0, groups = h_47_groups_0, pad = h_47_pad_0, pad_type = h_47_pad_type_0, strides = h_47_strides_0, weight = m_attn_layers_2_ffn_conv_2_weight_to_fp16, x = input_137_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1, 256, 128]> y_11_cast_fp16 = mul(x = h_47_cast_fp16, y = text_mask_to_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = y_11_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [3]> input_139_perm_0 = const()[name = string("input_139_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1003_axes_0 = const()[name = string("op_1003_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_2_norm_2_norm_weight_to_fp16 = const()[name = string("m_attn_layers_2_norm_2_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15532160)))];
            tensor<fp16, [256]> m_attn_layers_2_norm_2_norm_bias_to_fp16 = const()[name = string("m_attn_layers_2_norm_2_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15532736)))];
            tensor<fp16, [1, 128, 256]> input_139_cast_fp16 = transpose(perm = input_139_perm_0, x = x_109_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 128, 256]> var_1003_cast_fp16 = layer_norm(axes = var_1003_axes_0, beta = m_attn_layers_2_norm_2_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_2_norm_2_norm_weight_to_fp16, x = input_139_cast_fp16)[name = string("op_1003_cast_fp16")];
            tensor<int32, [3]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_1003_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 256, 128]> x_113_cast_fp16 = mul(x = x_111_cast_fp16, y = text_mask_to_fp16)[name = string("x_113_cast_fp16")];
            string var_1025_pad_type_0 = const()[name = string("op_1025_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_1025_strides_0 = const()[name = string("op_1025_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1025_pad_0 = const()[name = string("op_1025_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1025_dilations_0 = const()[name = string("op_1025_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_1025_groups_0 = const()[name = string("op_1025_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_3_attn_conv_q_weight_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_q_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15533312)))];
            tensor<fp16, [256]> m_attn_layers_3_attn_conv_q_bias_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_q_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15664448)))];
            tensor<fp16, [1, 256, 128]> var_1025_cast_fp16 = conv(bias = m_attn_layers_3_attn_conv_q_bias_to_fp16, dilations = var_1025_dilations_0, groups = var_1025_groups_0, pad = var_1025_pad_0, pad_type = var_1025_pad_type_0, strides = var_1025_strides_0, weight = m_attn_layers_3_attn_conv_q_weight_to_fp16, x = x_113_cast_fp16)[name = string("op_1025_cast_fp16")];
            tensor<int32, [4]> var_1026 = const()[name = string("op_1026"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_1027_cast_fp16 = reshape(shape = var_1026, x = var_1025_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1035_pad_type_0 = const()[name = string("op_1035_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_1035_strides_0 = const()[name = string("op_1035_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1035_pad_0 = const()[name = string("op_1035_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1035_dilations_0 = const()[name = string("op_1035_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_1035_groups_0 = const()[name = string("op_1035_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_3_attn_conv_k_weight_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_k_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15665024)))];
            tensor<fp16, [256]> m_attn_layers_3_attn_conv_k_bias_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_k_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15796160)))];
            tensor<fp16, [1, 256, 128]> var_1035_cast_fp16 = conv(bias = m_attn_layers_3_attn_conv_k_bias_to_fp16, dilations = var_1035_dilations_0, groups = var_1035_groups_0, pad = var_1035_pad_0, pad_type = var_1035_pad_type_0, strides = var_1035_strides_0, weight = m_attn_layers_3_attn_conv_k_weight_to_fp16, x = x_113_cast_fp16)[name = string("op_1035_cast_fp16")];
            tensor<int32, [4]> var_1036 = const()[name = string("op_1036"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_1037_cast_fp16 = reshape(shape = var_1036, x = var_1035_cast_fp16)[name = string("op_1037_cast_fp16")];
            string var_1045_pad_type_0 = const()[name = string("op_1045_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_1045_strides_0 = const()[name = string("op_1045_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1045_pad_0 = const()[name = string("op_1045_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1045_dilations_0 = const()[name = string("op_1045_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_1045_groups_0 = const()[name = string("op_1045_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_3_attn_conv_v_weight_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_v_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15796736)))];
            tensor<fp16, [256]> m_attn_layers_3_attn_conv_v_bias_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_v_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15927872)))];
            tensor<fp16, [1, 256, 128]> var_1045_cast_fp16 = conv(bias = m_attn_layers_3_attn_conv_v_bias_to_fp16, dilations = var_1045_dilations_0, groups = var_1045_groups_0, pad = var_1045_pad_0, pad_type = var_1045_pad_type_0, strides = var_1045_strides_0, weight = m_attn_layers_3_attn_conv_v_weight_to_fp16, x = x_113_cast_fp16)[name = string("op_1045_cast_fp16")];
            tensor<int32, [4]> var_1046 = const()[name = string("op_1046"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_1047_cast_fp16 = reshape(shape = var_1046, x = var_1045_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 4, 128, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_1027_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 4, 128, 64]> var_1050_cast_fp16 = mul(x = q_7_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            bool scores_13_transpose_x_0 = const()[name = string("scores_13_transpose_x_0"), val = bool(false)];
            bool scores_13_transpose_y_0 = const()[name = string("scores_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> scores_13_cast_fp16 = matmul(transpose_x = scores_13_transpose_x_0, transpose_y = scores_13_transpose_y_0, x = var_1050_cast_fp16, y = var_1037_cast_fp16)[name = string("scores_13_cast_fp16")];
            bool x_115_transpose_x_0 = const()[name = string("x_115_transpose_x_0"), val = bool(false)];
            bool x_115_transpose_y_0 = const()[name = string("x_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 255]> var_1074_to_fp16 = const()[name = string("op_1074_to_fp16"), val = tensor<fp16, [1, 1, 64, 255]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15928448)))];
            tensor<fp16, [1, 4, 128, 255]> x_115_cast_fp16 = matmul(transpose_x = x_115_transpose_x_0, transpose_y = x_115_transpose_y_0, x = var_1050_cast_fp16, y = var_1074_to_fp16)[name = string("x_115_cast_fp16")];
            tensor<int32, [8]> x_117_pad_0 = const()[name = string("x_117_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 1])];
            string x_117_mode_0 = const()[name = string("x_117_mode_0"), val = string("constant")];
            fp16 const_60_to_fp16 = const()[name = string("const_60_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 256]> x_117_cast_fp16 = pad(constant_val = const_60_to_fp16, mode = x_117_mode_0, pad = x_117_pad_0, x = x_115_cast_fp16)[name = string("x_117_cast_fp16")];
            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([1, 4, 32768])];
            tensor<fp16, [1, 4, 32768]> input_143_cast_fp16 = reshape(shape = var_1086, x = x_117_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<int32, [6]> x_119_pad_0 = const()[name = string("x_119_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 127])];
            string x_119_mode_0 = const()[name = string("x_119_mode_0"), val = string("constant")];
            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32895]> x_119_cast_fp16 = pad(constant_val = const_61_to_fp16, mode = x_119_mode_0, pad = x_119_pad_0, x = input_143_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [4]> var_1101 = const()[name = string("op_1101"), val = tensor<int32, [4]>([1, 4, 129, 255])];
            tensor<fp16, [1, 4, 129, 255]> x_121_cast_fp16 = reshape(shape = var_1101, x = x_119_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> var_1108_begin_0 = const()[name = string("op_1108_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1108_end_0 = const()[name = string("op_1108_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_1108_end_mask_0 = const()[name = string("op_1108_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 128, 255]> var_1108_cast_fp16 = slice_by_index(begin = var_1108_begin_0, end = var_1108_end_0, end_mask = var_1108_end_mask_0, x = x_121_cast_fp16)[name = string("op_1108_cast_fp16")];
            tensor<int32, [4]> var_1109_begin_0 = const()[name = string("op_1109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 127])];
            tensor<int32, [4]> var_1109_end_0 = const()[name = string("op_1109_end_0"), val = tensor<int32, [4]>([1, 4, 128, 255])];
            tensor<bool, [4]> var_1109_end_mask_0 = const()[name = string("op_1109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_1109_cast_fp16 = slice_by_index(begin = var_1109_begin_0, end = var_1109_end_0, end_mask = var_1109_end_mask_0, x = var_1108_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> scores_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1109_cast_fp16)[name = string("scores_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> input_145_cast_fp16 = sub(x = scores_cast_fp16, y = var_490_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> p_attn_cast_fp16 = softmax(axis = var_29, x = input_145_cast_fp16)[name = string("p_attn_cast_fp16")];
            bool out_25_transpose_x_1 = const()[name = string("out_25_transpose_x_1"), val = bool(false)];
            bool out_25_transpose_y_1 = const()[name = string("out_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 64]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_1, transpose_y = out_25_transpose_y_1, x = p_attn_cast_fp16, y = var_1047_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [8]> x_123_pad_0 = const()[name = string("x_123_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 0, 127])];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("constant")];
            fp16 const_66_to_fp16 = const()[name = string("const_66_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 128, 255]> x_123_cast_fp16 = pad(constant_val = const_66_to_fp16, mode = x_123_mode_0, pad = x_123_pad_0, x = p_attn_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [3]> var_1151 = const()[name = string("op_1151"), val = tensor<int32, [3]>([1, 4, 32640])];
            tensor<fp16, [1, 4, 32640]> input_149_cast_fp16 = reshape(shape = var_1151, x = x_123_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<int32, [6]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 128, 0])];
            string x_125_mode_0 = const()[name = string("x_125_mode_0"), val = string("constant")];
            fp16 const_67_to_fp16 = const()[name = string("const_67_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 4, 32768]> x_125_cast_fp16 = pad(constant_val = const_67_to_fp16, mode = x_125_mode_0, pad = x_125_pad_0, x = input_149_cast_fp16)[name = string("x_125_cast_fp16")];
            tensor<int32, [4]> var_1158 = const()[name = string("op_1158"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<fp16, [1, 4, 128, 256]> x_127_cast_fp16 = reshape(shape = var_1158, x = x_125_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<int32, [4]> rel_weights_begin_0 = const()[name = string("rel_weights_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> rel_weights_end_0 = const()[name = string("rel_weights_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> rel_weights_end_mask_0 = const()[name = string("rel_weights_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 255]> rel_weights_cast_fp16 = slice_by_index(begin = rel_weights_begin_0, end = rel_weights_end_0, end_mask = rel_weights_end_mask_0, x = x_127_cast_fp16)[name = string("rel_weights_cast_fp16")];
            bool var_1165_transpose_x_0 = const()[name = string("op_1165_transpose_x_0"), val = bool(false)];
            bool var_1165_transpose_y_0 = const()[name = string("op_1165_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 255, 64]> var_1164_to_fp16 = const()[name = string("op_1164_to_fp16"), val = tensor<fp16, [1, 1, 255, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15961152)))];
            tensor<fp16, [1, 4, 128, 64]> var_1165_cast_fp16 = matmul(transpose_x = var_1165_transpose_x_0, transpose_y = var_1165_transpose_y_0, x = rel_weights_cast_fp16, y = var_1164_to_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [1, 4, 128, 64]> out_27_cast_fp16 = add(x = out_25_cast_fp16, y = var_1165_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [4]> var_1167_perm_0 = const()[name = string("op_1167_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [3]> var_1168 = const()[name = string("op_1168"), val = tensor<int32, [3]>([1, 256, 128])];
            tensor<fp16, [1, 4, 64, 128]> var_1167_cast_fp16 = transpose(perm = var_1167_perm_0, x = out_27_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 256, 128]> input_151_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("input_151_cast_fp16")];
            string y_13_pad_type_0 = const()[name = string("y_13_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> y_13_strides_0 = const()[name = string("y_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> y_13_pad_0 = const()[name = string("y_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> y_13_dilations_0 = const()[name = string("y_13_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 y_13_groups_0 = const()[name = string("y_13_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1]> m_attn_layers_3_attn_conv_o_weight_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_o_weight_to_fp16"), val = tensor<fp16, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15993856)))];
            tensor<fp16, [256]> m_attn_layers_3_attn_conv_o_bias_to_fp16 = const()[name = string("m_attn_layers_3_attn_conv_o_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16124992)))];
            tensor<fp16, [1, 256, 128]> y_13_cast_fp16 = conv(bias = m_attn_layers_3_attn_conv_o_bias_to_fp16, dilations = y_13_dilations_0, groups = y_13_groups_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = m_attn_layers_3_attn_conv_o_weight_to_fp16, x = input_151_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_129_cast_fp16 = add(x = x_113_cast_fp16, y = y_13_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1183_axes_0 = const()[name = string("op_1183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_3_norm_1_norm_weight_to_fp16 = const()[name = string("m_attn_layers_3_norm_1_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16125568)))];
            tensor<fp16, [256]> m_attn_layers_3_norm_1_norm_bias_to_fp16 = const()[name = string("m_attn_layers_3_norm_1_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16126144)))];
            tensor<fp16, [1, 128, 256]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_129_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 128, 256]> var_1183_cast_fp16 = layer_norm(axes = var_1183_axes_0, beta = m_attn_layers_3_norm_1_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_3_norm_1_norm_weight_to_fp16, x = input_153_cast_fp16)[name = string("op_1183_cast_fp16")];
            tensor<int32, [3]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_1183_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 256, 128]> input_155_cast_fp16 = mul(x = x_131_cast_fp16, y = text_mask_to_fp16)[name = string("input_155_cast_fp16")];
            string input_157_pad_type_0 = const()[name = string("input_157_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input_157_strides_0 = const()[name = string("input_157_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input_157_pad_0 = const()[name = string("input_157_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input_157_dilations_0 = const()[name = string("input_157_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input_157_groups_0 = const()[name = string("input_157_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 256, 1]> m_attn_layers_3_ffn_conv_1_weight_to_fp16 = const()[name = string("m_attn_layers_3_ffn_conv_1_weight_to_fp16"), val = tensor<fp16, [1024, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16126720)))];
            tensor<fp16, [1024]> m_attn_layers_3_ffn_conv_1_bias_to_fp16 = const()[name = string("m_attn_layers_3_ffn_conv_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16651072)))];
            tensor<fp16, [1, 1024, 128]> input_157_cast_fp16 = conv(bias = m_attn_layers_3_ffn_conv_1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = m_attn_layers_3_ffn_conv_1_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> h_49_cast_fp16 = relu(x = input_157_cast_fp16)[name = string("h_49_cast_fp16")];
            tensor<fp16, [1, 1024, 128]> input_159_cast_fp16 = mul(x = h_49_cast_fp16, y = text_mask_to_fp16)[name = string("input_159_cast_fp16")];
            string h_pad_type_0 = const()[name = string("h_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> h_strides_0 = const()[name = string("h_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> h_pad_0 = const()[name = string("h_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> h_dilations_0 = const()[name = string("h_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 h_groups_0 = const()[name = string("h_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1024, 1]> m_attn_layers_3_ffn_conv_2_weight_to_fp16 = const()[name = string("m_attn_layers_3_ffn_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16653184)))];
            tensor<fp16, [256]> m_attn_layers_3_ffn_conv_2_bias_to_fp16 = const()[name = string("m_attn_layers_3_ffn_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17177536)))];
            tensor<fp16, [1, 256, 128]> h_cast_fp16 = conv(bias = m_attn_layers_3_ffn_conv_2_bias_to_fp16, dilations = h_dilations_0, groups = h_groups_0, pad = h_pad_0, pad_type = h_pad_type_0, strides = h_strides_0, weight = m_attn_layers_3_ffn_conv_2_weight_to_fp16, x = input_159_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [1, 256, 128]> y_cast_fp16 = mul(x = h_cast_fp16, y = text_mask_to_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 256, 128]> x_133_cast_fp16 = add(x = x_131_cast_fp16, y = y_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1211_axes_0 = const()[name = string("op_1211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_attn_layers_3_norm_2_norm_weight_to_fp16 = const()[name = string("m_attn_layers_3_norm_2_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17178112)))];
            tensor<fp16, [256]> m_attn_layers_3_norm_2_norm_bias_to_fp16 = const()[name = string("m_attn_layers_3_norm_2_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17178688)))];
            tensor<fp16, [1, 128, 256]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_133_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 128, 256]> var_1211_cast_fp16 = layer_norm(axes = var_1211_axes_0, beta = m_attn_layers_3_norm_2_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_attn_layers_3_norm_2_norm_weight_to_fp16, x = input_161_cast_fp16)[name = string("op_1211_cast_fp16")];
            tensor<int32, [3]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_1211_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 256, 128]> x_137_cast_fp16 = mul(x = x_135_cast_fp16, y = text_mask_to_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 256, 128]> var_1214_cast_fp16 = add(x = x_137_cast_fp16, y = x_39_cast_fp16)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1, 256, 128]> text_1_cast_fp16 = mul(x = var_1214_cast_fp16, y = text_mask_to_fp16)[name = string("text_1_cast_fp16")];
            tensor<int32, [3]> text_3_perm_0 = const()[name = string("text_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [256, 256]> m_speech_prompted_attention1_W_query_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention1_W_query_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17179264)))];
            tensor<fp16, [256]> m_speech_prompted_attention1_W_query_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention1_W_query_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17310400)))];
            tensor<fp16, [1, 128, 256]> text_3_cast_fp16 = transpose(perm = text_3_perm_0, x = text_1_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 128, 256]> linear_0_cast_fp16 = linear(bias = m_speech_prompted_attention1_W_query_linear_bias_to_fp16, weight = m_speech_prompted_attention1_W_query_linear_weight_to_fp16, x = text_3_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_1235 = const()[name = string("op_1235"), val = tensor<int32, [4]>([1, 128, 2, 128])];
            tensor<fp16, [1, 128, 2, 128]> var_1236_cast_fp16 = reshape(shape = var_1235, x = linear_0_cast_fp16)[name = string("op_1236_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            string style_ttl_to_fp16_dtype_0 = const()[name = string("style_ttl_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [256, 256]> m_speech_prompted_attention1_W_value_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention1_W_value_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17310976)))];
            tensor<fp16, [256]> m_speech_prompted_attention1_W_value_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention1_W_value_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17442112)))];
            tensor<fp16, [1, 50, 256]> style_ttl_to_fp16 = cast(dtype = style_ttl_to_fp16_dtype_0, x = style_ttl)[name = string("cast_47")];
            tensor<fp16, [1, 50, 256]> linear_2_cast_fp16 = linear(bias = m_speech_prompted_attention1_W_value_linear_bias_to_fp16, weight = m_speech_prompted_attention1_W_value_linear_weight_to_fp16, x = style_ttl_to_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1249 = const()[name = string("op_1249"), val = tensor<int32, [4]>([1, 50, 2, 128])];
            tensor<fp16, [1, 50, 2, 128]> var_1250_cast_fp16 = reshape(shape = var_1249, x = linear_2_cast_fp16)[name = string("op_1250_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1254_transpose_x_0 = const()[name = string("op_1254_transpose_x_0"), val = bool(false)];
            bool var_1254_transpose_y_0 = const()[name = string("op_1254_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 50]> var_1253_to_fp16 = const()[name = string("op_1253_to_fp16"), val = tensor<fp16, [1, 2, 128, 50]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17442688)))];
            tensor<fp16, [1, 2, 128, 128]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_1236_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 2, 128, 50]> var_1254_cast_fp16 = matmul(transpose_x = var_1254_transpose_x_0, transpose_y = var_1254_transpose_y_0, x = q_9_cast_fp16, y = var_1253_to_fp16)[name = string("op_1254_cast_fp16")];
            fp16 _inversed_input_163_y_0_to_fp16 = const()[name = string("_inversed_input_163_y_0_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 2, 128, 50]> _inversed_input_163_cast_fp16 = mul(x = var_1254_cast_fp16, y = _inversed_input_163_y_0_to_fp16)[name = string("_inversed_input_163_cast_fp16")];
            tensor<fp16, [1, 2, 128, 50]> attn_1_cast_fp16 = softmax(axis = var_29, x = _inversed_input_163_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<int32, [3]> var_1258_perm_0 = const()[name = string("op_1258_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> mask_q_1_axes_0 = const()[name = string("mask_q_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_1258_cast_fp16 = transpose(perm = var_1258_perm_0, x = text_mask_to_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 1, 128, 1]> mask_q_1_cast_fp16 = expand_dims(axes = mask_q_1_axes_0, x = var_1258_cast_fp16)[name = string("mask_q_1_cast_fp16")];
            tensor<fp16, [1, 2, 128, 50]> attn_3_cast_fp16 = mul(x = attn_1_cast_fp16, y = mask_q_1_cast_fp16)[name = string("attn_3_cast_fp16")];
            bool out_29_transpose_x_0 = const()[name = string("out_29_transpose_x_0"), val = bool(false)];
            bool out_29_transpose_y_0 = const()[name = string("out_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 2, 50, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_1250_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 128, 128]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = attn_3_cast_fp16, y = v_9_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [4]> var_1262_perm_0 = const()[name = string("op_1262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1263 = const()[name = string("op_1263"), val = tensor<int32, [3]>([1, 128, 256])];
            tensor<fp16, [1, 128, 2, 128]> var_1262_cast_fp16 = transpose(perm = var_1262_perm_0, x = out_29_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 128, 256]> input_165_cast_fp16 = reshape(shape = var_1263, x = var_1262_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [256, 256]> m_speech_prompted_attention1_out_fc_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention1_out_fc_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17468352)))];
            tensor<fp16, [256]> m_speech_prompted_attention1_out_fc_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention1_out_fc_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17599488)))];
            tensor<fp16, [1, 128, 256]> linear_3_cast_fp16 = linear(bias = m_speech_prompted_attention1_out_fc_linear_bias_to_fp16, weight = m_speech_prompted_attention1_out_fc_linear_weight_to_fp16, x = input_165_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 128, 256]> attn1_out_cast_fp16 = mul(x = linear_3_cast_fp16, y = var_1258_cast_fp16)[name = string("attn1_out_cast_fp16")];
            tensor<fp16, [1, 128, 256]> text_cast_fp16 = add(x = text_3_cast_fp16, y = attn1_out_cast_fp16)[name = string("text_cast_fp16")];
            tensor<fp16, [256, 256]> m_speech_prompted_attention2_W_query_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention2_W_query_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17600064)))];
            tensor<fp16, [256]> m_speech_prompted_attention2_W_query_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention2_W_query_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17731200)))];
            tensor<fp16, [1, 128, 256]> linear_4_cast_fp16 = linear(bias = m_speech_prompted_attention2_W_query_linear_bias_to_fp16, weight = m_speech_prompted_attention2_W_query_linear_weight_to_fp16, x = text_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [4]> var_1283 = const()[name = string("op_1283"), val = tensor<int32, [4]>([1, 128, 2, 128])];
            tensor<fp16, [1, 128, 2, 128]> var_1284_cast_fp16 = reshape(shape = var_1283, x = linear_4_cast_fp16)[name = string("op_1284_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 256]> m_speech_prompted_attention2_W_value_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention2_W_value_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17731776)))];
            tensor<fp16, [256]> m_speech_prompted_attention2_W_value_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention2_W_value_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17862912)))];
            tensor<fp16, [1, 50, 256]> linear_6_cast_fp16 = linear(bias = m_speech_prompted_attention2_W_value_linear_bias_to_fp16, weight = m_speech_prompted_attention2_W_value_linear_weight_to_fp16, x = style_ttl_to_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [4]> var_1297 = const()[name = string("op_1297"), val = tensor<int32, [4]>([1, 50, 2, 128])];
            tensor<fp16, [1, 50, 2, 128]> var_1298_cast_fp16 = reshape(shape = var_1297, x = linear_6_cast_fp16)[name = string("op_1298_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1302_transpose_x_0 = const()[name = string("op_1302_transpose_x_0"), val = bool(false)];
            bool var_1302_transpose_y_0 = const()[name = string("op_1302_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 50]> var_1301_to_fp16 = const()[name = string("op_1301_to_fp16"), val = tensor<fp16, [1, 2, 128, 50]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17863488)))];
            tensor<fp16, [1, 2, 128, 128]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_1284_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2, 128, 50]> var_1302_cast_fp16 = matmul(transpose_x = var_1302_transpose_x_0, transpose_y = var_1302_transpose_y_0, x = q_cast_fp16, y = var_1301_to_fp16)[name = string("op_1302_cast_fp16")];
            fp16 _inversed_input_167_y_0_to_fp16 = const()[name = string("_inversed_input_167_y_0_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 2, 128, 50]> _inversed_input_167_cast_fp16 = mul(x = var_1302_cast_fp16, y = _inversed_input_167_y_0_to_fp16)[name = string("_inversed_input_167_cast_fp16")];
            tensor<fp16, [1, 2, 128, 50]> attn_5_cast_fp16 = softmax(axis = var_29, x = _inversed_input_167_cast_fp16)[name = string("attn_5_cast_fp16")];
            tensor<fp16, [1, 2, 128, 50]> attn_7_cast_fp16 = mul(x = attn_5_cast_fp16, y = mask_q_1_cast_fp16)[name = string("attn_7_cast_fp16")];
            bool out_33_transpose_x_0 = const()[name = string("out_33_transpose_x_0"), val = bool(false)];
            bool out_33_transpose_y_0 = const()[name = string("out_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 2, 50, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_1298_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 128, 128]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = attn_7_cast_fp16, y = v_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [4]> var_1310_perm_0 = const()[name = string("op_1310_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1311 = const()[name = string("op_1311"), val = tensor<int32, [3]>([1, 128, 256])];
            tensor<fp16, [1, 128, 2, 128]> var_1310_cast_fp16 = transpose(perm = var_1310_perm_0, x = out_33_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 128, 256]> input_169_cast_fp16 = reshape(shape = var_1311, x = var_1310_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [256, 256]> m_speech_prompted_attention2_out_fc_linear_weight_to_fp16 = const()[name = string("m_speech_prompted_attention2_out_fc_linear_weight_to_fp16"), val = tensor<fp16, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17889152)))];
            tensor<fp16, [256]> m_speech_prompted_attention2_out_fc_linear_bias_to_fp16 = const()[name = string("m_speech_prompted_attention2_out_fc_linear_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18020288)))];
            tensor<fp16, [1, 128, 256]> linear_7_cast_fp16 = linear(bias = m_speech_prompted_attention2_out_fc_linear_bias_to_fp16, weight = m_speech_prompted_attention2_out_fc_linear_weight_to_fp16, x = input_169_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 128, 256]> attn2_out_cast_fp16 = mul(x = linear_7_cast_fp16, y = var_1258_cast_fp16)[name = string("attn2_out_cast_fp16")];
            tensor<fp16, [1, 128, 256]> y_final_cast_fp16 = add(x = text_3_cast_fp16, y = attn2_out_cast_fp16)[name = string("y_final_cast_fp16")];
            tensor<int32, [1]> var_1326_axes_0 = const()[name = string("op_1326_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [256]> m_speech_prompted_norm_norm_weight_to_fp16 = const()[name = string("m_speech_prompted_norm_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18020864)))];
            tensor<fp16, [256]> m_speech_prompted_norm_norm_bias_to_fp16 = const()[name = string("m_speech_prompted_norm_norm_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18021440)))];
            tensor<fp16, [1, 128, 256]> var_1326_cast_fp16 = layer_norm(axes = var_1326_axes_0, beta = m_speech_prompted_norm_norm_bias_to_fp16, epsilon = var_25_to_fp16, gamma = m_speech_prompted_norm_norm_weight_to_fp16, x = y_final_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [3]> x_perm_0 = const()[name = string("x_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 256, 128]> x_cast_fp16 = transpose(perm = x_perm_0, x = var_1326_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 256, 128]> var_1328_cast_fp16 = mul(x = x_cast_fp16, y = text_mask_to_fp16)[name = string("op_1328_cast_fp16")];
            string var_1328_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_1328_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 256, 128]> text_emb = cast(dtype = var_1328_cast_fp16_to_fp32_dtype_0, x = var_1328_cast_fp16)[name = string("cast_46")];
        } -> (text_emb);
}