program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<fp32, [1, 64, 1001]> mel) {
            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> mel_to_fp16_dtype_0 = const()[name = tensor<string, []>("mel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 64, 1001]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor<string, []>("cast_27")];
            tensor<fp16, [1, 1, 64, 1001]> x_1_cast_fp16 = expand_dims(axes = x_1_axes_0, x = mel_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<int32, [4]> var_34 = const()[name = tensor<string, []>("op_34"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [64]> m_encoder_init_bn_running_mean_to_fp16 = const()[name = tensor<string, []>("m_encoder_init_bn_running_mean_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [64]> m_encoder_init_bn_running_var_to_fp16 = const()[name = tensor<string, []>("m_encoder_init_bn_running_var_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256)))];
            tensor<fp16, [64]> m_encoder_init_bn_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_init_bn_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448)))];
            tensor<fp16, [64]> m_encoder_init_bn_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_init_bn_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640)))];
            tensor<fp16, []> var_21_to_fp16 = const()[name = tensor<string, []>("op_21_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 64, 1, 1001]> input_1_cast_fp16 = transpose(perm = var_34, x = x_1_cast_fp16)[name = tensor<string, []>("transpose_26")];
            tensor<fp16, [1, 64, 1, 1001]> x_3_cast_fp16 = batch_norm(beta = m_encoder_init_bn_bias_to_fp16, epsilon = var_21_to_fp16, gamma = m_encoder_init_bn_weight_to_fp16, mean = m_encoder_init_bn_running_mean_to_fp16, variance = m_encoder_init_bn_running_var_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<int32, [4]> var_41 = const()[name = tensor<string, []>("op_41"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<string, []> x_7_pad_type_0 = const()[name = tensor<string, []>("x_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_7_strides_0 = const()[name = tensor<string, []>("x_7_strides_0"), val = tensor<int32, [2]>([16, 16])];
            tensor<int32, [4]> x_7_pad_0 = const()[name = tensor<string, []>("x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_7_dilations_0 = const()[name = tensor<string, []>("x_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_7_groups_0 = const()[name = tensor<string, []>("x_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [768, 1, 16, 16]> m_encoder_patch_embed_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_patch_embed_proj_weight_to_fp16"), val = tensor<fp16, [768, 1, 16, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832)))];
            tensor<fp16, [768]> m_encoder_patch_embed_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_patch_embed_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394112)))];
            tensor<fp16, [1, 1, 64, 1001]> x_5_cast_fp16 = transpose(perm = var_41, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_25")];
            tensor<fp16, [1, 768, 4, 62]> x_7_cast_fp16 = conv(bias = m_encoder_patch_embed_proj_bias_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = m_encoder_patch_embed_proj_weight_to_fp16, x = x_5_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<fp16, [1, 768, 1, 62]> var_55_to_fp16 = const()[name = tensor<string, []>("op_55_to_fp16"), val = tensor<fp16, [1, 768, 1, 62]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395712)))];
            tensor<fp16, [1, 768, 4, 62]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = var_55_to_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<fp16, [1, 768, 4, 1]> m_encoder_freq_pos_embed_to_fp16 = const()[name = tensor<string, []>("m_encoder_freq_pos_embed_to_fp16"), val = tensor<fp16, [1, 768, 4, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(491008)))];
            tensor<fp16, [1, 768, 4, 62]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = m_encoder_freq_pos_embed_to_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [3]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [3]>([1, 768, 248])];
            tensor<fp16, [1, 768, 248]> var_62_cast_fp16 = reshape(shape = concat_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_62_cast_fp16")];
            tensor<int32, [3]> var_63 = const()[name = tensor<string, []>("op_63"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497216)))];
            tensor<fp16, [768]> m_encoder_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498816)))];
            tensor<fp16, []> var_15_to_fp16 = const()[name = tensor<string, []>("op_15_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 248, 768]> input_3_cast_fp16 = transpose(perm = var_63, x = var_62_cast_fp16)[name = tensor<string, []>("transpose_24")];
            tensor<fp16, [1, 248, 768]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = m_encoder_blocks_0_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_0_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500416)))];
            tensor<fp16, [2304]> m_encoder_blocks_0_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4039424)))];
            tensor<fp16, [1, 248, 2304]> linear_0_cast_fp16 = linear(bias = m_encoder_blocks_0_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_0_attn_qkv_weight_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [5]> var_97 = const()[name = tensor<string, []>("op_97"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_98_cast_fp16 = reshape(shape = var_97, x = linear_0_cast_fp16)[name = tensor<string, []>("op_98_cast_fp16")];
            tensor<int32, [5]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_101_split_sizes_0 = const()[name = tensor<string, []>("op_101_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_101_axis_0 = const()[name = tensor<string, []>("op_101_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_1_cast_fp16 = transpose(perm = var_99, x = var_98_cast_fp16)[name = tensor<string, []>("transpose_23")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_101_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_101_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_101_cast_fp16_2 = split(axis = var_101_axis_0, split_sizes = var_101_split_sizes_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("op_101_cast_fp16")];
            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = tensor<string, []>("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_101_cast_fp16_0)[name = tensor<string, []>("squeeze_0_cast_fp16")];
            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = tensor<string, []>("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_101_cast_fp16_1)[name = tensor<string, []>("squeeze_1_cast_fp16")];
            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = tensor<string, []>("squeeze_2_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_101_cast_fp16_2)[name = tensor<string, []>("squeeze_2_cast_fp16")];
            tensor<bool, []> var_106_transpose_x_1 = const()[name = tensor<string, []>("op_106_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_106_transpose_y_1 = const()[name = tensor<string, []>("op_106_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_106_cast_fp16 = matmul(transpose_x = var_106_transpose_x_1, transpose_y = var_106_transpose_y_1, x = squeeze_0_cast_fp16, y = squeeze_1_cast_fp16)[name = tensor<string, []>("op_106_cast_fp16")];
            tensor<fp16, []> var_107_to_fp16 = const()[name = tensor<string, []>("op_107_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_1_cast_fp16 = mul(x = var_106_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_7_cast_fp16 = softmax(axis = var_12, x = attn_1_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<bool, []> var_111_transpose_x_0 = const()[name = tensor<string, []>("op_111_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_111_transpose_y_0 = const()[name = tensor<string, []>("op_111_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_111_cast_fp16 = matmul(transpose_x = var_111_transpose_x_0, transpose_y = var_111_transpose_y_0, x = input_7_cast_fp16, y = squeeze_2_cast_fp16)[name = tensor<string, []>("op_111_cast_fp16")];
            tensor<int32, [4]> var_112_perm_0 = const()[name = tensor<string, []>("op_112_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_113 = const()[name = tensor<string, []>("op_113"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_112_cast_fp16 = transpose(perm = var_112_perm_0, x = var_111_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 248, 768]> input_9_cast_fp16 = reshape(shape = var_113, x = var_112_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_0_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4044096)))];
            tensor<fp16, [768]> m_encoder_blocks_0_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5223808)))];
            tensor<fp16, [1, 248, 768]> linear_1_cast_fp16 = linear(bias = m_encoder_blocks_0_attn_proj_bias_to_fp16, weight = m_encoder_blocks_0_attn_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_13_cast_fp16 = add(x = input_3_cast_fp16, y = linear_1_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<int32, [1]> input_15_axes_0 = const()[name = tensor<string, []>("input_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5225408)))];
            tensor<fp16, [768]> m_encoder_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5227008)))];
            tensor<fp16, [1, 248, 768]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = m_encoder_blocks_0_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_0_norm2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_0_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5228608)))];
            tensor<fp16, [3072]> m_encoder_blocks_0_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9947264)))];
            tensor<fp16, [1, 248, 3072]> linear_2_cast_fp16 = linear(bias = m_encoder_blocks_0_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_0_mlp_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_2_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_0_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9953472)))];
            tensor<fp16, [768]> m_encoder_blocks_0_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14672128)))];
            tensor<fp16, [1, 248, 768]> linear_3_cast_fp16 = linear(bias = m_encoder_blocks_0_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_0_mlp_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_25_cast_fp16 = add(x = input_13_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<int32, [1]> x_15_axes_0 = const()[name = tensor<string, []>("x_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14673728)))];
            tensor<fp16, [768]> m_encoder_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14675328)))];
            tensor<fp16, [1, 248, 768]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = m_encoder_blocks_1_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_1_norm1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_1_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14676928)))];
            tensor<fp16, [2304]> m_encoder_blocks_1_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18215936)))];
            tensor<fp16, [1, 248, 2304]> linear_4_cast_fp16 = linear(bias = m_encoder_blocks_1_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_1_attn_qkv_weight_to_fp16, x = x_15_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<int32, [5]> var_155 = const()[name = tensor<string, []>("op_155"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_156_cast_fp16 = reshape(shape = var_155, x = linear_4_cast_fp16)[name = tensor<string, []>("op_156_cast_fp16")];
            tensor<int32, [5]> var_157 = const()[name = tensor<string, []>("op_157"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_159_split_sizes_0 = const()[name = tensor<string, []>("op_159_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_159_axis_0 = const()[name = tensor<string, []>("op_159_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_3_cast_fp16 = transpose(perm = var_157, x = var_156_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_159_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_159_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_159_cast_fp16_2 = split(axis = var_159_axis_0, split_sizes = var_159_split_sizes_0, x = qkv_3_cast_fp16)[name = tensor<string, []>("op_159_cast_fp16")];
            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = tensor<string, []>("squeeze_3_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_159_cast_fp16_0)[name = tensor<string, []>("squeeze_3_cast_fp16")];
            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = tensor<string, []>("squeeze_4_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_159_cast_fp16_1)[name = tensor<string, []>("squeeze_4_cast_fp16")];
            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = tensor<string, []>("squeeze_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_159_cast_fp16_2)[name = tensor<string, []>("squeeze_5_cast_fp16")];
            tensor<bool, []> var_164_transpose_x_1 = const()[name = tensor<string, []>("op_164_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_164_transpose_y_1 = const()[name = tensor<string, []>("op_164_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_164_cast_fp16 = matmul(transpose_x = var_164_transpose_x_1, transpose_y = var_164_transpose_y_1, x = squeeze_3_cast_fp16, y = squeeze_4_cast_fp16)[name = tensor<string, []>("op_164_cast_fp16")];
            tensor<fp16, []> var_165_to_fp16 = const()[name = tensor<string, []>("op_165_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_5_cast_fp16 = mul(x = var_164_cast_fp16, y = var_165_to_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_27_cast_fp16 = softmax(axis = var_12, x = attn_5_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<bool, []> var_169_transpose_x_0 = const()[name = tensor<string, []>("op_169_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_169_transpose_y_0 = const()[name = tensor<string, []>("op_169_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_169_cast_fp16 = matmul(transpose_x = var_169_transpose_x_0, transpose_y = var_169_transpose_y_0, x = input_27_cast_fp16, y = squeeze_5_cast_fp16)[name = tensor<string, []>("op_169_cast_fp16")];
            tensor<int32, [4]> var_170_perm_0 = const()[name = tensor<string, []>("op_170_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_170_cast_fp16 = transpose(perm = var_170_perm_0, x = var_169_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 248, 768]> input_29_cast_fp16 = reshape(shape = var_171, x = var_170_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_1_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18220608)))];
            tensor<fp16, [768]> m_encoder_blocks_1_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19400320)))];
            tensor<fp16, [1, 248, 768]> linear_5_cast_fp16 = linear(bias = m_encoder_blocks_1_attn_proj_bias_to_fp16, weight = m_encoder_blocks_1_attn_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_33_cast_fp16 = add(x = input_25_cast_fp16, y = linear_5_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = tensor<string, []>("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19401920)))];
            tensor<fp16, [768]> m_encoder_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19403520)))];
            tensor<fp16, [1, 248, 768]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = m_encoder_blocks_1_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_1_norm2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_1_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19405120)))];
            tensor<fp16, [3072]> m_encoder_blocks_1_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24123776)))];
            tensor<fp16, [1, 248, 3072]> linear_6_cast_fp16 = linear(bias = m_encoder_blocks_1_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_1_mlp_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = linear_6_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_1_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24129984)))];
            tensor<fp16, [768]> m_encoder_blocks_1_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28848640)))];
            tensor<fp16, [1, 248, 768]> linear_7_cast_fp16 = linear(bias = m_encoder_blocks_1_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_1_mlp_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_45_cast_fp16 = add(x = input_33_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28850240)))];
            tensor<fp16, [768]> m_encoder_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28851840)))];
            tensor<fp16, [1, 248, 768]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = m_encoder_blocks_2_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_2_norm1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_2_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28853440)))];
            tensor<fp16, [2304]> m_encoder_blocks_2_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32392448)))];
            tensor<fp16, [1, 248, 2304]> linear_8_cast_fp16 = linear(bias = m_encoder_blocks_2_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_2_attn_qkv_weight_to_fp16, x = x_17_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<int32, [5]> var_213 = const()[name = tensor<string, []>("op_213"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_214_cast_fp16 = reshape(shape = var_213, x = linear_8_cast_fp16)[name = tensor<string, []>("op_214_cast_fp16")];
            tensor<int32, [5]> var_215 = const()[name = tensor<string, []>("op_215"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_217_split_sizes_0 = const()[name = tensor<string, []>("op_217_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_217_axis_0 = const()[name = tensor<string, []>("op_217_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_5_cast_fp16 = transpose(perm = var_215, x = var_214_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_217_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_217_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_217_cast_fp16_2 = split(axis = var_217_axis_0, split_sizes = var_217_split_sizes_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("op_217_cast_fp16")];
            tensor<int32, [1]> squeeze_6_axes_0 = const()[name = tensor<string, []>("squeeze_6_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_217_cast_fp16_0)[name = tensor<string, []>("squeeze_6_cast_fp16")];
            tensor<int32, [1]> squeeze_7_axes_0 = const()[name = tensor<string, []>("squeeze_7_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_217_cast_fp16_1)[name = tensor<string, []>("squeeze_7_cast_fp16")];
            tensor<int32, [1]> squeeze_8_axes_0 = const()[name = tensor<string, []>("squeeze_8_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_217_cast_fp16_2)[name = tensor<string, []>("squeeze_8_cast_fp16")];
            tensor<bool, []> var_222_transpose_x_1 = const()[name = tensor<string, []>("op_222_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_222_transpose_y_1 = const()[name = tensor<string, []>("op_222_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_222_cast_fp16 = matmul(transpose_x = var_222_transpose_x_1, transpose_y = var_222_transpose_y_1, x = squeeze_6_cast_fp16, y = squeeze_7_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")];
            tensor<fp16, []> var_223_to_fp16 = const()[name = tensor<string, []>("op_223_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_9_cast_fp16 = mul(x = var_222_cast_fp16, y = var_223_to_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_47_cast_fp16 = softmax(axis = var_12, x = attn_9_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<bool, []> var_227_transpose_x_0 = const()[name = tensor<string, []>("op_227_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_227_transpose_y_0 = const()[name = tensor<string, []>("op_227_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_227_cast_fp16 = matmul(transpose_x = var_227_transpose_x_0, transpose_y = var_227_transpose_y_0, x = input_47_cast_fp16, y = squeeze_8_cast_fp16)[name = tensor<string, []>("op_227_cast_fp16")];
            tensor<int32, [4]> var_228_perm_0 = const()[name = tensor<string, []>("op_228_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_229 = const()[name = tensor<string, []>("op_229"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_228_cast_fp16 = transpose(perm = var_228_perm_0, x = var_227_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, 248, 768]> input_49_cast_fp16 = reshape(shape = var_229, x = var_228_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_2_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32397120)))];
            tensor<fp16, [768]> m_encoder_blocks_2_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33576832)))];
            tensor<fp16, [1, 248, 768]> linear_9_cast_fp16 = linear(bias = m_encoder_blocks_2_attn_proj_bias_to_fp16, weight = m_encoder_blocks_2_attn_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_53_cast_fp16 = add(x = input_45_cast_fp16, y = linear_9_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<int32, [1]> input_55_axes_0 = const()[name = tensor<string, []>("input_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33578432)))];
            tensor<fp16, [768]> m_encoder_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33580032)))];
            tensor<fp16, [1, 248, 768]> input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = m_encoder_blocks_2_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_2_norm2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_2_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33581632)))];
            tensor<fp16, [3072]> m_encoder_blocks_2_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38300288)))];
            tensor<fp16, [1, 248, 3072]> linear_10_cast_fp16 = linear(bias = m_encoder_blocks_2_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_2_mlp_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<string, []> input_59_mode_0 = const()[name = tensor<string, []>("input_59_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_2_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38306496)))];
            tensor<fp16, [768]> m_encoder_blocks_2_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43025152)))];
            tensor<fp16, [1, 248, 768]> linear_11_cast_fp16 = linear(bias = m_encoder_blocks_2_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_2_mlp_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_65_cast_fp16 = add(x = input_53_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43026752)))];
            tensor<fp16, [768]> m_encoder_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43028352)))];
            tensor<fp16, [1, 248, 768]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = m_encoder_blocks_3_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_3_norm1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_3_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43029952)))];
            tensor<fp16, [2304]> m_encoder_blocks_3_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46568960)))];
            tensor<fp16, [1, 248, 2304]> linear_12_cast_fp16 = linear(bias = m_encoder_blocks_3_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_3_attn_qkv_weight_to_fp16, x = x_19_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<int32, [5]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_272_cast_fp16 = reshape(shape = var_271, x = linear_12_cast_fp16)[name = tensor<string, []>("op_272_cast_fp16")];
            tensor<int32, [5]> var_273 = const()[name = tensor<string, []>("op_273"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_275_split_sizes_0 = const()[name = tensor<string, []>("op_275_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_275_axis_0 = const()[name = tensor<string, []>("op_275_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_7_cast_fp16 = transpose(perm = var_273, x = var_272_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_275_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_275_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_275_cast_fp16_2 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = qkv_7_cast_fp16)[name = tensor<string, []>("op_275_cast_fp16")];
            tensor<int32, [1]> squeeze_9_axes_0 = const()[name = tensor<string, []>("squeeze_9_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_275_cast_fp16_0)[name = tensor<string, []>("squeeze_9_cast_fp16")];
            tensor<int32, [1]> squeeze_10_axes_0 = const()[name = tensor<string, []>("squeeze_10_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_275_cast_fp16_1)[name = tensor<string, []>("squeeze_10_cast_fp16")];
            tensor<int32, [1]> squeeze_11_axes_0 = const()[name = tensor<string, []>("squeeze_11_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_275_cast_fp16_2)[name = tensor<string, []>("squeeze_11_cast_fp16")];
            tensor<bool, []> var_280_transpose_x_1 = const()[name = tensor<string, []>("op_280_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_280_transpose_y_1 = const()[name = tensor<string, []>("op_280_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_280_cast_fp16 = matmul(transpose_x = var_280_transpose_x_1, transpose_y = var_280_transpose_y_1, x = squeeze_9_cast_fp16, y = squeeze_10_cast_fp16)[name = tensor<string, []>("op_280_cast_fp16")];
            tensor<fp16, []> var_281_to_fp16 = const()[name = tensor<string, []>("op_281_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_13_cast_fp16 = mul(x = var_280_cast_fp16, y = var_281_to_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_67_cast_fp16 = softmax(axis = var_12, x = attn_13_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<bool, []> var_285_transpose_x_0 = const()[name = tensor<string, []>("op_285_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_285_transpose_y_0 = const()[name = tensor<string, []>("op_285_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_285_cast_fp16 = matmul(transpose_x = var_285_transpose_x_0, transpose_y = var_285_transpose_y_0, x = input_67_cast_fp16, y = squeeze_11_cast_fp16)[name = tensor<string, []>("op_285_cast_fp16")];
            tensor<int32, [4]> var_286_perm_0 = const()[name = tensor<string, []>("op_286_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_287 = const()[name = tensor<string, []>("op_287"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_286_cast_fp16 = transpose(perm = var_286_perm_0, x = var_285_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, 248, 768]> input_69_cast_fp16 = reshape(shape = var_287, x = var_286_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_3_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46573632)))];
            tensor<fp16, [768]> m_encoder_blocks_3_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47753344)))];
            tensor<fp16, [1, 248, 768]> linear_13_cast_fp16 = linear(bias = m_encoder_blocks_3_attn_proj_bias_to_fp16, weight = m_encoder_blocks_3_attn_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_73_cast_fp16 = add(x = input_65_cast_fp16, y = linear_13_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<int32, [1]> input_75_axes_0 = const()[name = tensor<string, []>("input_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47754944)))];
            tensor<fp16, [768]> m_encoder_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47756544)))];
            tensor<fp16, [1, 248, 768]> input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = m_encoder_blocks_3_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_3_norm2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_3_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47758144)))];
            tensor<fp16, [3072]> m_encoder_blocks_3_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52476800)))];
            tensor<fp16, [1, 248, 3072]> linear_14_cast_fp16 = linear(bias = m_encoder_blocks_3_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_3_mlp_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<string, []> input_79_mode_0 = const()[name = tensor<string, []>("input_79_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_14_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_3_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52483008)))];
            tensor<fp16, [768]> m_encoder_blocks_3_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57201664)))];
            tensor<fp16, [1, 248, 768]> linear_15_cast_fp16 = linear(bias = m_encoder_blocks_3_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_3_mlp_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_85_cast_fp16 = add(x = input_73_cast_fp16, y = linear_15_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<int32, [1]> x_21_axes_0 = const()[name = tensor<string, []>("x_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57203264)))];
            tensor<fp16, [768]> m_encoder_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57204864)))];
            tensor<fp16, [1, 248, 768]> x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = m_encoder_blocks_4_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_4_norm1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_4_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57206464)))];
            tensor<fp16, [2304]> m_encoder_blocks_4_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60745472)))];
            tensor<fp16, [1, 248, 2304]> linear_16_cast_fp16 = linear(bias = m_encoder_blocks_4_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_4_attn_qkv_weight_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<int32, [5]> var_329 = const()[name = tensor<string, []>("op_329"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_330_cast_fp16 = reshape(shape = var_329, x = linear_16_cast_fp16)[name = tensor<string, []>("op_330_cast_fp16")];
            tensor<int32, [5]> var_331 = const()[name = tensor<string, []>("op_331"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_333_split_sizes_0 = const()[name = tensor<string, []>("op_333_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_333_axis_0 = const()[name = tensor<string, []>("op_333_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_9_cast_fp16 = transpose(perm = var_331, x = var_330_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_333_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_333_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_333_cast_fp16_2 = split(axis = var_333_axis_0, split_sizes = var_333_split_sizes_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("op_333_cast_fp16")];
            tensor<int32, [1]> squeeze_12_axes_0 = const()[name = tensor<string, []>("squeeze_12_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_333_cast_fp16_0)[name = tensor<string, []>("squeeze_12_cast_fp16")];
            tensor<int32, [1]> squeeze_13_axes_0 = const()[name = tensor<string, []>("squeeze_13_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_333_cast_fp16_1)[name = tensor<string, []>("squeeze_13_cast_fp16")];
            tensor<int32, [1]> squeeze_14_axes_0 = const()[name = tensor<string, []>("squeeze_14_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_333_cast_fp16_2)[name = tensor<string, []>("squeeze_14_cast_fp16")];
            tensor<bool, []> var_338_transpose_x_1 = const()[name = tensor<string, []>("op_338_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_338_transpose_y_1 = const()[name = tensor<string, []>("op_338_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_338_cast_fp16 = matmul(transpose_x = var_338_transpose_x_1, transpose_y = var_338_transpose_y_1, x = squeeze_12_cast_fp16, y = squeeze_13_cast_fp16)[name = tensor<string, []>("op_338_cast_fp16")];
            tensor<fp16, []> var_339_to_fp16 = const()[name = tensor<string, []>("op_339_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_17_cast_fp16 = mul(x = var_338_cast_fp16, y = var_339_to_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_87_cast_fp16 = softmax(axis = var_12, x = attn_17_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<bool, []> var_343_transpose_x_0 = const()[name = tensor<string, []>("op_343_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_343_transpose_y_0 = const()[name = tensor<string, []>("op_343_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_343_cast_fp16 = matmul(transpose_x = var_343_transpose_x_0, transpose_y = var_343_transpose_y_0, x = input_87_cast_fp16, y = squeeze_14_cast_fp16)[name = tensor<string, []>("op_343_cast_fp16")];
            tensor<int32, [4]> var_344_perm_0 = const()[name = tensor<string, []>("op_344_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_345 = const()[name = tensor<string, []>("op_345"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_344_cast_fp16 = transpose(perm = var_344_perm_0, x = var_343_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, 248, 768]> input_89_cast_fp16 = reshape(shape = var_345, x = var_344_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_4_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60750144)))];
            tensor<fp16, [768]> m_encoder_blocks_4_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61929856)))];
            tensor<fp16, [1, 248, 768]> linear_17_cast_fp16 = linear(bias = m_encoder_blocks_4_attn_proj_bias_to_fp16, weight = m_encoder_blocks_4_attn_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_93_cast_fp16 = add(x = input_85_cast_fp16, y = linear_17_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61931456)))];
            tensor<fp16, [768]> m_encoder_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61933056)))];
            tensor<fp16, [1, 248, 768]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = m_encoder_blocks_4_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_4_norm2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_4_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61934656)))];
            tensor<fp16, [3072]> m_encoder_blocks_4_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66653312)))];
            tensor<fp16, [1, 248, 3072]> linear_18_cast_fp16 = linear(bias = m_encoder_blocks_4_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_4_mlp_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_18_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_4_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66659520)))];
            tensor<fp16, [768]> m_encoder_blocks_4_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71378176)))];
            tensor<fp16, [1, 248, 768]> linear_19_cast_fp16 = linear(bias = m_encoder_blocks_4_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_4_mlp_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_105_cast_fp16 = add(x = input_93_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71379776)))];
            tensor<fp16, [768]> m_encoder_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71381376)))];
            tensor<fp16, [1, 248, 768]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = m_encoder_blocks_5_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_5_norm1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_5_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71382976)))];
            tensor<fp16, [2304]> m_encoder_blocks_5_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74921984)))];
            tensor<fp16, [1, 248, 2304]> linear_20_cast_fp16 = linear(bias = m_encoder_blocks_5_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_5_attn_qkv_weight_to_fp16, x = x_23_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<int32, [5]> var_387 = const()[name = tensor<string, []>("op_387"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_388_cast_fp16 = reshape(shape = var_387, x = linear_20_cast_fp16)[name = tensor<string, []>("op_388_cast_fp16")];
            tensor<int32, [5]> var_389 = const()[name = tensor<string, []>("op_389"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_391_split_sizes_0 = const()[name = tensor<string, []>("op_391_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_391_axis_0 = const()[name = tensor<string, []>("op_391_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_11_cast_fp16 = transpose(perm = var_389, x = var_388_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_391_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_391_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_391_cast_fp16_2 = split(axis = var_391_axis_0, split_sizes = var_391_split_sizes_0, x = qkv_11_cast_fp16)[name = tensor<string, []>("op_391_cast_fp16")];
            tensor<int32, [1]> squeeze_15_axes_0 = const()[name = tensor<string, []>("squeeze_15_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_391_cast_fp16_0)[name = tensor<string, []>("squeeze_15_cast_fp16")];
            tensor<int32, [1]> squeeze_16_axes_0 = const()[name = tensor<string, []>("squeeze_16_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_391_cast_fp16_1)[name = tensor<string, []>("squeeze_16_cast_fp16")];
            tensor<int32, [1]> squeeze_17_axes_0 = const()[name = tensor<string, []>("squeeze_17_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_391_cast_fp16_2)[name = tensor<string, []>("squeeze_17_cast_fp16")];
            tensor<bool, []> var_396_transpose_x_1 = const()[name = tensor<string, []>("op_396_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_396_transpose_y_1 = const()[name = tensor<string, []>("op_396_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_396_cast_fp16 = matmul(transpose_x = var_396_transpose_x_1, transpose_y = var_396_transpose_y_1, x = squeeze_15_cast_fp16, y = squeeze_16_cast_fp16)[name = tensor<string, []>("op_396_cast_fp16")];
            tensor<fp16, []> var_397_to_fp16 = const()[name = tensor<string, []>("op_397_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_21_cast_fp16 = mul(x = var_396_cast_fp16, y = var_397_to_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_107_cast_fp16 = softmax(axis = var_12, x = attn_21_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<bool, []> var_401_transpose_x_0 = const()[name = tensor<string, []>("op_401_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_401_transpose_y_0 = const()[name = tensor<string, []>("op_401_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_401_cast_fp16 = matmul(transpose_x = var_401_transpose_x_0, transpose_y = var_401_transpose_y_0, x = input_107_cast_fp16, y = squeeze_17_cast_fp16)[name = tensor<string, []>("op_401_cast_fp16")];
            tensor<int32, [4]> var_402_perm_0 = const()[name = tensor<string, []>("op_402_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_403 = const()[name = tensor<string, []>("op_403"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_402_cast_fp16 = transpose(perm = var_402_perm_0, x = var_401_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 248, 768]> input_109_cast_fp16 = reshape(shape = var_403, x = var_402_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_5_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74926656)))];
            tensor<fp16, [768]> m_encoder_blocks_5_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76106368)))];
            tensor<fp16, [1, 248, 768]> linear_21_cast_fp16 = linear(bias = m_encoder_blocks_5_attn_proj_bias_to_fp16, weight = m_encoder_blocks_5_attn_proj_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_113_cast_fp16 = add(x = input_105_cast_fp16, y = linear_21_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<int32, [1]> input_115_axes_0 = const()[name = tensor<string, []>("input_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76107968)))];
            tensor<fp16, [768]> m_encoder_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76109568)))];
            tensor<fp16, [1, 248, 768]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = m_encoder_blocks_5_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_5_norm2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_5_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76111168)))];
            tensor<fp16, [3072]> m_encoder_blocks_5_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80829824)))];
            tensor<fp16, [1, 248, 3072]> linear_22_cast_fp16 = linear(bias = m_encoder_blocks_5_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_5_mlp_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<string, []> input_119_mode_0 = const()[name = tensor<string, []>("input_119_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_5_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80836032)))];
            tensor<fp16, [768]> m_encoder_blocks_5_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85554688)))];
            tensor<fp16, [1, 248, 768]> linear_23_cast_fp16 = linear(bias = m_encoder_blocks_5_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_5_mlp_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_125_cast_fp16 = add(x = input_113_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85556288)))];
            tensor<fp16, [768]> m_encoder_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85557888)))];
            tensor<fp16, [1, 248, 768]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = m_encoder_blocks_6_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_6_norm1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_6_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85559488)))];
            tensor<fp16, [2304]> m_encoder_blocks_6_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89098496)))];
            tensor<fp16, [1, 248, 2304]> linear_24_cast_fp16 = linear(bias = m_encoder_blocks_6_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_6_attn_qkv_weight_to_fp16, x = x_25_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<int32, [5]> var_445 = const()[name = tensor<string, []>("op_445"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_446_cast_fp16 = reshape(shape = var_445, x = linear_24_cast_fp16)[name = tensor<string, []>("op_446_cast_fp16")];
            tensor<int32, [5]> var_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_449_split_sizes_0 = const()[name = tensor<string, []>("op_449_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_449_axis_0 = const()[name = tensor<string, []>("op_449_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_13_cast_fp16 = transpose(perm = var_447, x = var_446_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_449_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_449_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_449_cast_fp16_2 = split(axis = var_449_axis_0, split_sizes = var_449_split_sizes_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("op_449_cast_fp16")];
            tensor<int32, [1]> squeeze_18_axes_0 = const()[name = tensor<string, []>("squeeze_18_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_449_cast_fp16_0)[name = tensor<string, []>("squeeze_18_cast_fp16")];
            tensor<int32, [1]> squeeze_19_axes_0 = const()[name = tensor<string, []>("squeeze_19_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_449_cast_fp16_1)[name = tensor<string, []>("squeeze_19_cast_fp16")];
            tensor<int32, [1]> squeeze_20_axes_0 = const()[name = tensor<string, []>("squeeze_20_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_449_cast_fp16_2)[name = tensor<string, []>("squeeze_20_cast_fp16")];
            tensor<bool, []> var_454_transpose_x_1 = const()[name = tensor<string, []>("op_454_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_454_transpose_y_1 = const()[name = tensor<string, []>("op_454_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_454_cast_fp16 = matmul(transpose_x = var_454_transpose_x_1, transpose_y = var_454_transpose_y_1, x = squeeze_18_cast_fp16, y = squeeze_19_cast_fp16)[name = tensor<string, []>("op_454_cast_fp16")];
            tensor<fp16, []> var_455_to_fp16 = const()[name = tensor<string, []>("op_455_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_25_cast_fp16 = mul(x = var_454_cast_fp16, y = var_455_to_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_127_cast_fp16 = softmax(axis = var_12, x = attn_25_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<bool, []> var_459_transpose_x_0 = const()[name = tensor<string, []>("op_459_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_459_transpose_y_0 = const()[name = tensor<string, []>("op_459_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_459_cast_fp16 = matmul(transpose_x = var_459_transpose_x_0, transpose_y = var_459_transpose_y_0, x = input_127_cast_fp16, y = squeeze_20_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")];
            tensor<int32, [4]> var_460_perm_0 = const()[name = tensor<string, []>("op_460_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_461 = const()[name = tensor<string, []>("op_461"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_460_cast_fp16 = transpose(perm = var_460_perm_0, x = var_459_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 248, 768]> input_129_cast_fp16 = reshape(shape = var_461, x = var_460_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_6_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89103168)))];
            tensor<fp16, [768]> m_encoder_blocks_6_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90282880)))];
            tensor<fp16, [1, 248, 768]> linear_25_cast_fp16 = linear(bias = m_encoder_blocks_6_attn_proj_bias_to_fp16, weight = m_encoder_blocks_6_attn_proj_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_133_cast_fp16 = add(x = input_125_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<int32, [1]> input_135_axes_0 = const()[name = tensor<string, []>("input_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90284480)))];
            tensor<fp16, [768]> m_encoder_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90286080)))];
            tensor<fp16, [1, 248, 768]> input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = m_encoder_blocks_6_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_6_norm2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_6_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90287680)))];
            tensor<fp16, [3072]> m_encoder_blocks_6_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95006336)))];
            tensor<fp16, [1, 248, 3072]> linear_26_cast_fp16 = linear(bias = m_encoder_blocks_6_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_6_mlp_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<string, []> input_139_mode_0 = const()[name = tensor<string, []>("input_139_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_26_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_6_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95012544)))];
            tensor<fp16, [768]> m_encoder_blocks_6_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99731200)))];
            tensor<fp16, [1, 248, 768]> linear_27_cast_fp16 = linear(bias = m_encoder_blocks_6_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_6_mlp_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_145_cast_fp16 = add(x = input_133_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<int32, [1]> x_27_axes_0 = const()[name = tensor<string, []>("x_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99732800)))];
            tensor<fp16, [768]> m_encoder_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99734400)))];
            tensor<fp16, [1, 248, 768]> x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = m_encoder_blocks_7_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_7_norm1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_7_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99736000)))];
            tensor<fp16, [2304]> m_encoder_blocks_7_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103275008)))];
            tensor<fp16, [1, 248, 2304]> linear_28_cast_fp16 = linear(bias = m_encoder_blocks_7_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_7_attn_qkv_weight_to_fp16, x = x_27_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<int32, [5]> var_503 = const()[name = tensor<string, []>("op_503"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_504_cast_fp16 = reshape(shape = var_503, x = linear_28_cast_fp16)[name = tensor<string, []>("op_504_cast_fp16")];
            tensor<int32, [5]> var_505 = const()[name = tensor<string, []>("op_505"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_507_split_sizes_0 = const()[name = tensor<string, []>("op_507_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_507_axis_0 = const()[name = tensor<string, []>("op_507_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_15_cast_fp16 = transpose(perm = var_505, x = var_504_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_507_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_507_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_507_cast_fp16_2 = split(axis = var_507_axis_0, split_sizes = var_507_split_sizes_0, x = qkv_15_cast_fp16)[name = tensor<string, []>("op_507_cast_fp16")];
            tensor<int32, [1]> squeeze_21_axes_0 = const()[name = tensor<string, []>("squeeze_21_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_507_cast_fp16_0)[name = tensor<string, []>("squeeze_21_cast_fp16")];
            tensor<int32, [1]> squeeze_22_axes_0 = const()[name = tensor<string, []>("squeeze_22_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_507_cast_fp16_1)[name = tensor<string, []>("squeeze_22_cast_fp16")];
            tensor<int32, [1]> squeeze_23_axes_0 = const()[name = tensor<string, []>("squeeze_23_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_507_cast_fp16_2)[name = tensor<string, []>("squeeze_23_cast_fp16")];
            tensor<bool, []> var_512_transpose_x_1 = const()[name = tensor<string, []>("op_512_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_512_transpose_y_1 = const()[name = tensor<string, []>("op_512_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_512_cast_fp16 = matmul(transpose_x = var_512_transpose_x_1, transpose_y = var_512_transpose_y_1, x = squeeze_21_cast_fp16, y = squeeze_22_cast_fp16)[name = tensor<string, []>("op_512_cast_fp16")];
            tensor<fp16, []> var_513_to_fp16 = const()[name = tensor<string, []>("op_513_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_29_cast_fp16 = mul(x = var_512_cast_fp16, y = var_513_to_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_147_cast_fp16 = softmax(axis = var_12, x = attn_29_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<bool, []> var_517_transpose_x_0 = const()[name = tensor<string, []>("op_517_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_517_transpose_y_0 = const()[name = tensor<string, []>("op_517_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = input_147_cast_fp16, y = squeeze_23_cast_fp16)[name = tensor<string, []>("op_517_cast_fp16")];
            tensor<int32, [4]> var_518_perm_0 = const()[name = tensor<string, []>("op_518_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_519 = const()[name = tensor<string, []>("op_519"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_518_cast_fp16 = transpose(perm = var_518_perm_0, x = var_517_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 248, 768]> input_149_cast_fp16 = reshape(shape = var_519, x = var_518_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_7_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103279680)))];
            tensor<fp16, [768]> m_encoder_blocks_7_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104459392)))];
            tensor<fp16, [1, 248, 768]> linear_29_cast_fp16 = linear(bias = m_encoder_blocks_7_attn_proj_bias_to_fp16, weight = m_encoder_blocks_7_attn_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_153_cast_fp16 = add(x = input_145_cast_fp16, y = linear_29_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<int32, [1]> input_155_axes_0 = const()[name = tensor<string, []>("input_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104460992)))];
            tensor<fp16, [768]> m_encoder_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104462592)))];
            tensor<fp16, [1, 248, 768]> input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = m_encoder_blocks_7_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_7_norm2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_7_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104464192)))];
            tensor<fp16, [3072]> m_encoder_blocks_7_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109182848)))];
            tensor<fp16, [1, 248, 3072]> linear_30_cast_fp16 = linear(bias = m_encoder_blocks_7_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_7_mlp_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<string, []> input_159_mode_0 = const()[name = tensor<string, []>("input_159_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = linear_30_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_7_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109189056)))];
            tensor<fp16, [768]> m_encoder_blocks_7_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113907712)))];
            tensor<fp16, [1, 248, 768]> linear_31_cast_fp16 = linear(bias = m_encoder_blocks_7_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_7_mlp_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_165_cast_fp16 = add(x = input_153_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<int32, [1]> x_29_axes_0 = const()[name = tensor<string, []>("x_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113909312)))];
            tensor<fp16, [768]> m_encoder_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113910912)))];
            tensor<fp16, [1, 248, 768]> x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = m_encoder_blocks_8_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_8_norm1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_8_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113912512)))];
            tensor<fp16, [2304]> m_encoder_blocks_8_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117451520)))];
            tensor<fp16, [1, 248, 2304]> linear_32_cast_fp16 = linear(bias = m_encoder_blocks_8_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_8_attn_qkv_weight_to_fp16, x = x_29_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [5]> var_561 = const()[name = tensor<string, []>("op_561"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_562_cast_fp16 = reshape(shape = var_561, x = linear_32_cast_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
            tensor<int32, [5]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_565_split_sizes_0 = const()[name = tensor<string, []>("op_565_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_565_axis_0 = const()[name = tensor<string, []>("op_565_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_17_cast_fp16 = transpose(perm = var_563, x = var_562_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_565_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_565_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_565_cast_fp16_2 = split(axis = var_565_axis_0, split_sizes = var_565_split_sizes_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("op_565_cast_fp16")];
            tensor<int32, [1]> squeeze_24_axes_0 = const()[name = tensor<string, []>("squeeze_24_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_565_cast_fp16_0)[name = tensor<string, []>("squeeze_24_cast_fp16")];
            tensor<int32, [1]> squeeze_25_axes_0 = const()[name = tensor<string, []>("squeeze_25_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_565_cast_fp16_1)[name = tensor<string, []>("squeeze_25_cast_fp16")];
            tensor<int32, [1]> squeeze_26_axes_0 = const()[name = tensor<string, []>("squeeze_26_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_565_cast_fp16_2)[name = tensor<string, []>("squeeze_26_cast_fp16")];
            tensor<bool, []> var_570_transpose_x_1 = const()[name = tensor<string, []>("op_570_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_570_transpose_y_1 = const()[name = tensor<string, []>("op_570_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_570_cast_fp16 = matmul(transpose_x = var_570_transpose_x_1, transpose_y = var_570_transpose_y_1, x = squeeze_24_cast_fp16, y = squeeze_25_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
            tensor<fp16, []> var_571_to_fp16 = const()[name = tensor<string, []>("op_571_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_33_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_167_cast_fp16 = softmax(axis = var_12, x = attn_33_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<bool, []> var_575_transpose_x_0 = const()[name = tensor<string, []>("op_575_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_575_transpose_y_0 = const()[name = tensor<string, []>("op_575_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_575_cast_fp16 = matmul(transpose_x = var_575_transpose_x_0, transpose_y = var_575_transpose_y_0, x = input_167_cast_fp16, y = squeeze_26_cast_fp16)[name = tensor<string, []>("op_575_cast_fp16")];
            tensor<int32, [4]> var_576_perm_0 = const()[name = tensor<string, []>("op_576_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_576_cast_fp16 = transpose(perm = var_576_perm_0, x = var_575_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 248, 768]> input_169_cast_fp16 = reshape(shape = var_577, x = var_576_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_8_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117456192)))];
            tensor<fp16, [768]> m_encoder_blocks_8_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118635904)))];
            tensor<fp16, [1, 248, 768]> linear_33_cast_fp16 = linear(bias = m_encoder_blocks_8_attn_proj_bias_to_fp16, weight = m_encoder_blocks_8_attn_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_173_cast_fp16 = add(x = input_165_cast_fp16, y = linear_33_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<int32, [1]> input_175_axes_0 = const()[name = tensor<string, []>("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118637504)))];
            tensor<fp16, [768]> m_encoder_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118639104)))];
            tensor<fp16, [1, 248, 768]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = m_encoder_blocks_8_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_8_norm2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_8_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118640704)))];
            tensor<fp16, [3072]> m_encoder_blocks_8_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123359360)))];
            tensor<fp16, [1, 248, 3072]> linear_34_cast_fp16 = linear(bias = m_encoder_blocks_8_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_8_mlp_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<string, []> input_179_mode_0 = const()[name = tensor<string, []>("input_179_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_8_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123365568)))];
            tensor<fp16, [768]> m_encoder_blocks_8_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128084224)))];
            tensor<fp16, [1, 248, 768]> linear_35_cast_fp16 = linear(bias = m_encoder_blocks_8_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_8_mlp_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_185_cast_fp16 = add(x = input_173_cast_fp16, y = linear_35_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [1]> x_31_axes_0 = const()[name = tensor<string, []>("x_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128085824)))];
            tensor<fp16, [768]> m_encoder_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128087424)))];
            tensor<fp16, [1, 248, 768]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = m_encoder_blocks_9_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_9_norm1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_9_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128089024)))];
            tensor<fp16, [2304]> m_encoder_blocks_9_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131628032)))];
            tensor<fp16, [1, 248, 2304]> linear_36_cast_fp16 = linear(bias = m_encoder_blocks_9_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_9_attn_qkv_weight_to_fp16, x = x_31_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<int32, [5]> var_619 = const()[name = tensor<string, []>("op_619"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_620_cast_fp16 = reshape(shape = var_619, x = linear_36_cast_fp16)[name = tensor<string, []>("op_620_cast_fp16")];
            tensor<int32, [5]> var_621 = const()[name = tensor<string, []>("op_621"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_623_split_sizes_0 = const()[name = tensor<string, []>("op_623_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_623_axis_0 = const()[name = tensor<string, []>("op_623_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_19_cast_fp16 = transpose(perm = var_621, x = var_620_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_623_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_623_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_623_cast_fp16_2 = split(axis = var_623_axis_0, split_sizes = var_623_split_sizes_0, x = qkv_19_cast_fp16)[name = tensor<string, []>("op_623_cast_fp16")];
            tensor<int32, [1]> squeeze_27_axes_0 = const()[name = tensor<string, []>("squeeze_27_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_623_cast_fp16_0)[name = tensor<string, []>("squeeze_27_cast_fp16")];
            tensor<int32, [1]> squeeze_28_axes_0 = const()[name = tensor<string, []>("squeeze_28_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_623_cast_fp16_1)[name = tensor<string, []>("squeeze_28_cast_fp16")];
            tensor<int32, [1]> squeeze_29_axes_0 = const()[name = tensor<string, []>("squeeze_29_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_623_cast_fp16_2)[name = tensor<string, []>("squeeze_29_cast_fp16")];
            tensor<bool, []> var_628_transpose_x_1 = const()[name = tensor<string, []>("op_628_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_628_transpose_y_1 = const()[name = tensor<string, []>("op_628_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_628_cast_fp16 = matmul(transpose_x = var_628_transpose_x_1, transpose_y = var_628_transpose_y_1, x = squeeze_27_cast_fp16, y = squeeze_28_cast_fp16)[name = tensor<string, []>("op_628_cast_fp16")];
            tensor<fp16, []> var_629_to_fp16 = const()[name = tensor<string, []>("op_629_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_37_cast_fp16 = mul(x = var_628_cast_fp16, y = var_629_to_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_187_cast_fp16 = softmax(axis = var_12, x = attn_37_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<bool, []> var_633_transpose_x_0 = const()[name = tensor<string, []>("op_633_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_633_transpose_y_0 = const()[name = tensor<string, []>("op_633_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_633_cast_fp16 = matmul(transpose_x = var_633_transpose_x_0, transpose_y = var_633_transpose_y_0, x = input_187_cast_fp16, y = squeeze_29_cast_fp16)[name = tensor<string, []>("op_633_cast_fp16")];
            tensor<int32, [4]> var_634_perm_0 = const()[name = tensor<string, []>("op_634_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_635 = const()[name = tensor<string, []>("op_635"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_634_cast_fp16 = transpose(perm = var_634_perm_0, x = var_633_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 248, 768]> input_189_cast_fp16 = reshape(shape = var_635, x = var_634_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_9_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131632704)))];
            tensor<fp16, [768]> m_encoder_blocks_9_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132812416)))];
            tensor<fp16, [1, 248, 768]> linear_37_cast_fp16 = linear(bias = m_encoder_blocks_9_attn_proj_bias_to_fp16, weight = m_encoder_blocks_9_attn_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_193_cast_fp16 = add(x = input_185_cast_fp16, y = linear_37_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<int32, [1]> input_195_axes_0 = const()[name = tensor<string, []>("input_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132814016)))];
            tensor<fp16, [768]> m_encoder_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132815616)))];
            tensor<fp16, [1, 248, 768]> input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = m_encoder_blocks_9_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_9_norm2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_9_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132817216)))];
            tensor<fp16, [3072]> m_encoder_blocks_9_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137535872)))];
            tensor<fp16, [1, 248, 3072]> linear_38_cast_fp16 = linear(bias = m_encoder_blocks_9_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_9_mlp_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<string, []> input_199_mode_0 = const()[name = tensor<string, []>("input_199_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_38_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_9_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137542080)))];
            tensor<fp16, [768]> m_encoder_blocks_9_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142260736)))];
            tensor<fp16, [1, 248, 768]> linear_39_cast_fp16 = linear(bias = m_encoder_blocks_9_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_9_mlp_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_205_cast_fp16 = add(x = input_193_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<int32, [1]> x_33_axes_0 = const()[name = tensor<string, []>("x_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_10_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142262336)))];
            tensor<fp16, [768]> m_encoder_blocks_10_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142263936)))];
            tensor<fp16, [1, 248, 768]> x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = m_encoder_blocks_10_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_10_norm1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_10_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142265536)))];
            tensor<fp16, [2304]> m_encoder_blocks_10_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145804544)))];
            tensor<fp16, [1, 248, 2304]> linear_40_cast_fp16 = linear(bias = m_encoder_blocks_10_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_10_attn_qkv_weight_to_fp16, x = x_33_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<int32, [5]> var_677 = const()[name = tensor<string, []>("op_677"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_678_cast_fp16 = reshape(shape = var_677, x = linear_40_cast_fp16)[name = tensor<string, []>("op_678_cast_fp16")];
            tensor<int32, [5]> var_679 = const()[name = tensor<string, []>("op_679"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_681_split_sizes_0 = const()[name = tensor<string, []>("op_681_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_681_axis_0 = const()[name = tensor<string, []>("op_681_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_21_cast_fp16 = transpose(perm = var_679, x = var_678_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_681_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_681_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_681_cast_fp16_2 = split(axis = var_681_axis_0, split_sizes = var_681_split_sizes_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("op_681_cast_fp16")];
            tensor<int32, [1]> squeeze_30_axes_0 = const()[name = tensor<string, []>("squeeze_30_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_681_cast_fp16_0)[name = tensor<string, []>("squeeze_30_cast_fp16")];
            tensor<int32, [1]> squeeze_31_axes_0 = const()[name = tensor<string, []>("squeeze_31_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_681_cast_fp16_1)[name = tensor<string, []>("squeeze_31_cast_fp16")];
            tensor<int32, [1]> squeeze_32_axes_0 = const()[name = tensor<string, []>("squeeze_32_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_681_cast_fp16_2)[name = tensor<string, []>("squeeze_32_cast_fp16")];
            tensor<bool, []> var_686_transpose_x_1 = const()[name = tensor<string, []>("op_686_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_686_transpose_y_1 = const()[name = tensor<string, []>("op_686_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_686_cast_fp16 = matmul(transpose_x = var_686_transpose_x_1, transpose_y = var_686_transpose_y_1, x = squeeze_30_cast_fp16, y = squeeze_31_cast_fp16)[name = tensor<string, []>("op_686_cast_fp16")];
            tensor<fp16, []> var_687_to_fp16 = const()[name = tensor<string, []>("op_687_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_41_cast_fp16 = mul(x = var_686_cast_fp16, y = var_687_to_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_207_cast_fp16 = softmax(axis = var_12, x = attn_41_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<bool, []> var_691_transpose_x_0 = const()[name = tensor<string, []>("op_691_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_691_transpose_y_0 = const()[name = tensor<string, []>("op_691_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_691_cast_fp16 = matmul(transpose_x = var_691_transpose_x_0, transpose_y = var_691_transpose_y_0, x = input_207_cast_fp16, y = squeeze_32_cast_fp16)[name = tensor<string, []>("op_691_cast_fp16")];
            tensor<int32, [4]> var_692_perm_0 = const()[name = tensor<string, []>("op_692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_693 = const()[name = tensor<string, []>("op_693"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = var_691_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<fp16, [1, 248, 768]> input_209_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_10_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145809216)))];
            tensor<fp16, [768]> m_encoder_blocks_10_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146988928)))];
            tensor<fp16, [1, 248, 768]> linear_41_cast_fp16 = linear(bias = m_encoder_blocks_10_attn_proj_bias_to_fp16, weight = m_encoder_blocks_10_attn_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_213_cast_fp16 = add(x = input_205_cast_fp16, y = linear_41_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<int32, [1]> input_215_axes_0 = const()[name = tensor<string, []>("input_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_10_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146990528)))];
            tensor<fp16, [768]> m_encoder_blocks_10_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146992128)))];
            tensor<fp16, [1, 248, 768]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = m_encoder_blocks_10_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_10_norm2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_10_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146993728)))];
            tensor<fp16, [3072]> m_encoder_blocks_10_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151712384)))];
            tensor<fp16, [1, 248, 3072]> linear_42_cast_fp16 = linear(bias = m_encoder_blocks_10_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_10_mlp_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<string, []> input_219_mode_0 = const()[name = tensor<string, []>("input_219_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = linear_42_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_10_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151718592)))];
            tensor<fp16, [768]> m_encoder_blocks_10_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156437248)))];
            tensor<fp16, [1, 248, 768]> linear_43_cast_fp16 = linear(bias = m_encoder_blocks_10_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_10_mlp_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_225_cast_fp16 = add(x = input_213_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<int32, [1]> x_35_axes_0 = const()[name = tensor<string, []>("x_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_11_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156438848)))];
            tensor<fp16, [768]> m_encoder_blocks_11_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156440448)))];
            tensor<fp16, [1, 248, 768]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = m_encoder_blocks_11_norm1_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_11_norm1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<fp16, [2304, 768]> m_encoder_blocks_11_attn_qkv_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_attn_qkv_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156442048)))];
            tensor<fp16, [2304]> m_encoder_blocks_11_attn_qkv_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_attn_qkv_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159981056)))];
            tensor<fp16, [1, 248, 2304]> linear_44_cast_fp16 = linear(bias = m_encoder_blocks_11_attn_qkv_bias_to_fp16, weight = m_encoder_blocks_11_attn_qkv_weight_to_fp16, x = x_35_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<int32, [5]> var_735 = const()[name = tensor<string, []>("op_735"), val = tensor<int32, [5]>([1, 248, 3, 12, 64])];
            tensor<fp16, [1, 248, 3, 12, 64]> var_736_cast_fp16 = reshape(shape = var_735, x = linear_44_cast_fp16)[name = tensor<string, []>("op_736_cast_fp16")];
            tensor<int32, [5]> var_737 = const()[name = tensor<string, []>("op_737"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
            tensor<int32, [3]> var_739_split_sizes_0 = const()[name = tensor<string, []>("op_739_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<int32, []> var_739_axis_0 = const()[name = tensor<string, []>("op_739_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [3, 1, 12, 248, 64]> qkv_23_cast_fp16 = transpose(perm = var_737, x = var_736_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp16, [1, 1, 12, 248, 64]> var_739_cast_fp16_0, tensor<fp16, [1, 1, 12, 248, 64]> var_739_cast_fp16_1, tensor<fp16, [1, 1, 12, 248, 64]> var_739_cast_fp16_2 = split(axis = var_739_axis_0, split_sizes = var_739_split_sizes_0, x = qkv_23_cast_fp16)[name = tensor<string, []>("op_739_cast_fp16")];
            tensor<int32, [1]> squeeze_33_axes_0 = const()[name = tensor<string, []>("squeeze_33_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_739_cast_fp16_0)[name = tensor<string, []>("squeeze_33_cast_fp16")];
            tensor<int32, [1]> squeeze_34_axes_0 = const()[name = tensor<string, []>("squeeze_34_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_739_cast_fp16_1)[name = tensor<string, []>("squeeze_34_cast_fp16")];
            tensor<int32, [1]> squeeze_35_axes_0 = const()[name = tensor<string, []>("squeeze_35_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 12, 248, 64]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_739_cast_fp16_2)[name = tensor<string, []>("squeeze_35_cast_fp16")];
            tensor<bool, []> var_744_transpose_x_1 = const()[name = tensor<string, []>("op_744_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_744_transpose_y_1 = const()[name = tensor<string, []>("op_744_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 248, 248]> var_744_cast_fp16 = matmul(transpose_x = var_744_transpose_x_1, transpose_y = var_744_transpose_y_1, x = squeeze_33_cast_fp16, y = squeeze_34_cast_fp16)[name = tensor<string, []>("op_744_cast_fp16")];
            tensor<fp16, []> var_745_to_fp16 = const()[name = tensor<string, []>("op_745_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 12, 248, 248]> attn_45_cast_fp16 = mul(x = var_744_cast_fp16, y = var_745_to_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<fp16, [1, 12, 248, 248]> input_227_cast_fp16 = softmax(axis = var_12, x = attn_45_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<bool, []> var_749_transpose_x_0 = const()[name = tensor<string, []>("op_749_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_749_transpose_y_0 = const()[name = tensor<string, []>("op_749_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 248, 64]> var_749_cast_fp16 = matmul(transpose_x = var_749_transpose_x_0, transpose_y = var_749_transpose_y_0, x = input_227_cast_fp16, y = squeeze_35_cast_fp16)[name = tensor<string, []>("op_749_cast_fp16")];
            tensor<int32, [4]> var_750_perm_0 = const()[name = tensor<string, []>("op_750_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_751 = const()[name = tensor<string, []>("op_751"), val = tensor<int32, [3]>([1, 248, 768])];
            tensor<fp16, [1, 248, 12, 64]> var_750_cast_fp16 = transpose(perm = var_750_perm_0, x = var_749_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 248, 768]> input_229_cast_fp16 = reshape(shape = var_751, x = var_750_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
            tensor<fp16, [768, 768]> m_encoder_blocks_11_attn_proj_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_attn_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159985728)))];
            tensor<fp16, [768]> m_encoder_blocks_11_attn_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_attn_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161165440)))];
            tensor<fp16, [1, 248, 768]> linear_45_cast_fp16 = linear(bias = m_encoder_blocks_11_attn_proj_bias_to_fp16, weight = m_encoder_blocks_11_attn_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_233_cast_fp16 = add(x = input_225_cast_fp16, y = linear_45_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_blocks_11_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm2_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161167040)))];
            tensor<fp16, [768]> m_encoder_blocks_11_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161168640)))];
            tensor<fp16, [1, 248, 768]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = m_encoder_blocks_11_norm2_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_blocks_11_norm2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<fp16, [3072, 768]> m_encoder_blocks_11_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161170240)))];
            tensor<fp16, [3072]> m_encoder_blocks_11_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165888896)))];
            tensor<fp16, [1, 248, 3072]> linear_46_cast_fp16 = linear(bias = m_encoder_blocks_11_mlp_fc1_bias_to_fp16, weight = m_encoder_blocks_11_mlp_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp16, [1, 248, 3072]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<fp16, [768, 3072]> m_encoder_blocks_11_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165895104)))];
            tensor<fp16, [768]> m_encoder_blocks_11_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170613760)))];
            tensor<fp16, [1, 248, 768]> linear_47_cast_fp16 = linear(bias = m_encoder_blocks_11_mlp_fc2_bias_to_fp16, weight = m_encoder_blocks_11_mlp_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, [1, 248, 768]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_encoder_norm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_norm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170615360)))];
            tensor<fp16, [768]> m_encoder_norm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_norm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170616960)))];
            tensor<fp16, [1, 248, 768]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = m_encoder_norm_bias_to_fp16, epsilon = var_15_to_fp16, gamma = m_encoder_norm_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = tensor<string, []>("input_247_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> input_247_keep_dims_0 = const()[name = tensor<string, []>("input_247_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 768]> input_247_cast_fp16 = reduce_mean(axes = input_247_axes_0, keep_dims = input_247_keep_dims_0, x = x_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> m_outputlayer_0_weight_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_0_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170618560)))];
            tensor<fp16, [768]> m_outputlayer_0_bias_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_0_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170620160)))];
            tensor<fp16, []> var_791_to_fp16 = const()[name = tensor<string, []>("op_791_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
            tensor<fp16, [1, 768]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = m_outputlayer_0_bias_to_fp16, epsilon = var_791_to_fp16, gamma = m_outputlayer_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<fp16, [527, 768]> m_outputlayer_1_weight_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_1_weight_to_fp16"), val = tensor<fp16, [527, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170621760)))];
            tensor<fp16, [527]> m_outputlayer_1_bias_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_1_bias_to_fp16"), val = tensor<fp16, [527]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171431296)))];
            tensor<fp16, [1, 527]> linear_48_cast_fp16 = linear(bias = m_outputlayer_1_bias_to_fp16, weight = m_outputlayer_1_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<string, []> linear_48_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_48_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 527]> logits = cast(dtype = linear_48_cast_fp16_to_fp32_dtype_0, x = linear_48_cast_fp16)[name = tensor<string, []>("cast_26")];
        } -> (logits);
}