| 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_11 = const()[name = tensor<string, []>("op_11"), 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_33 = const()[name = tensor<string, []>("op_33"), 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_20_to_fp16 = const()[name = tensor<string, []>("op_20_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 64, 1, 1001]> input_1_cast_fp16 = transpose(perm = var_33, 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_20_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_40 = const()[name = tensor<string, []>("op_40"), 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, [192, 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, [192, 1, 16, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832)))]; |
| tensor<fp16, [192]> m_encoder_patch_embed_proj_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_patch_embed_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99200)))]; |
| tensor<fp16, [1, 1, 64, 1001]> x_5_cast_fp16 = transpose(perm = var_40, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_25")]; |
| tensor<fp16, [1, 192, 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, 192, 1, 62]> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, [1, 192, 1, 62]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99648)))]; |
| tensor<fp16, [1, 192, 4, 62]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = var_54_to_fp16)[name = tensor<string, []>("x_9_cast_fp16")]; |
| tensor<fp16, [1, 192, 4, 1]> m_encoder_freq_pos_embed_to_fp16 = const()[name = tensor<string, []>("m_encoder_freq_pos_embed_to_fp16"), val = tensor<fp16, [1, 192, 4, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123520)))]; |
| tensor<fp16, [1, 192, 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, 192, 248])]; |
| tensor<fp16, [1, 192, 248]> var_61_cast_fp16 = reshape(shape = concat_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_61_cast_fp16")]; |
| tensor<int32, [3]> var_62 = const()[name = tensor<string, []>("op_62"), 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, [192]> m_encoder_blocks_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125120)))]; |
| tensor<fp16, [192]> m_encoder_blocks_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125568)))]; |
| tensor<fp16, []> var_14_to_fp16 = const()[name = tensor<string, []>("op_14_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)]; |
| tensor<fp16, [1, 248, 192]> input_3_cast_fp16 = transpose(perm = var_62, x = var_61_cast_fp16)[name = tensor<string, []>("transpose_24")]; |
| tensor<fp16, [1, 248, 192]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = m_encoder_blocks_0_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126016)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347264)))]; |
| tensor<fp16, [1, 248, 576]> 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_96 = const()[name = tensor<string, []>("op_96"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_97_cast_fp16 = reshape(shape = var_96, x = linear_0_cast_fp16)[name = tensor<string, []>("op_97_cast_fp16")]; |
| tensor<int32, [5]> var_98 = const()[name = tensor<string, []>("op_98"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_100_split_sizes_0 = const()[name = tensor<string, []>("op_100_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_100_axis_0 = const()[name = tensor<string, []>("op_100_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_1_cast_fp16 = transpose(perm = var_98, x = var_97_cast_fp16)[name = tensor<string, []>("transpose_23")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_100_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_100_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_100_cast_fp16_2 = split(axis = var_100_axis_0, split_sizes = var_100_split_sizes_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("op_100_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, 3, 248, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_100_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, 3, 248, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_100_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, 3, 248, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_100_cast_fp16_2)[name = tensor<string, []>("squeeze_2_cast_fp16")]; |
| tensor<bool, []> var_105_transpose_x_1 = const()[name = tensor<string, []>("op_105_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_105_transpose_y_1 = const()[name = tensor<string, []>("op_105_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_105_cast_fp16 = matmul(transpose_x = var_105_transpose_x_1, transpose_y = var_105_transpose_y_1, x = squeeze_0_cast_fp16, y = squeeze_1_cast_fp16)[name = tensor<string, []>("op_105_cast_fp16")]; |
| tensor<fp16, []> var_106_to_fp16 = const()[name = tensor<string, []>("op_106_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_1_cast_fp16 = mul(x = var_105_cast_fp16, y = var_106_to_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_7_cast_fp16 = softmax(axis = var_11, x = attn_1_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| tensor<bool, []> var_110_transpose_x_0 = const()[name = tensor<string, []>("op_110_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_110_transpose_y_0 = const()[name = tensor<string, []>("op_110_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_110_cast_fp16 = matmul(transpose_x = var_110_transpose_x_0, transpose_y = var_110_transpose_y_0, x = input_7_cast_fp16, y = squeeze_2_cast_fp16)[name = tensor<string, []>("op_110_cast_fp16")]; |
| tensor<int32, [4]> var_111_perm_0 = const()[name = tensor<string, []>("op_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_112 = const()[name = tensor<string, []>("op_112"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_111_cast_fp16 = transpose(perm = var_111_perm_0, x = var_110_cast_fp16)[name = tensor<string, []>("transpose_22")]; |
| tensor<fp16, [1, 248, 192]> input_9_cast_fp16 = reshape(shape = var_112, x = var_111_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348480)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422272)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422720)))]; |
| tensor<fp16, [192]> m_encoder_blocks_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_0_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423168)))]; |
| tensor<fp16, [1, 248, 192]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = m_encoder_blocks_0_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423616)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(718592)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_2_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720192)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1015168)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1015616)))]; |
| tensor<fp16, [192]> m_encoder_blocks_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1016064)))]; |
| tensor<fp16, [1, 248, 192]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = m_encoder_blocks_1_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1016512)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1237760)))]; |
| tensor<fp16, [1, 248, 576]> 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_154 = const()[name = tensor<string, []>("op_154"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_155_cast_fp16 = reshape(shape = var_154, x = linear_4_cast_fp16)[name = tensor<string, []>("op_155_cast_fp16")]; |
| tensor<int32, [5]> var_156 = const()[name = tensor<string, []>("op_156"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_158_split_sizes_0 = const()[name = tensor<string, []>("op_158_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_158_axis_0 = const()[name = tensor<string, []>("op_158_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_3_cast_fp16 = transpose(perm = var_156, x = var_155_cast_fp16)[name = tensor<string, []>("transpose_21")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_158_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_158_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_158_cast_fp16_2 = split(axis = var_158_axis_0, split_sizes = var_158_split_sizes_0, x = qkv_3_cast_fp16)[name = tensor<string, []>("op_158_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, 3, 248, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_158_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, 3, 248, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_158_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, 3, 248, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_158_cast_fp16_2)[name = tensor<string, []>("squeeze_5_cast_fp16")]; |
| tensor<bool, []> var_163_transpose_x_1 = const()[name = tensor<string, []>("op_163_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_163_transpose_y_1 = const()[name = tensor<string, []>("op_163_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_163_cast_fp16 = matmul(transpose_x = var_163_transpose_x_1, transpose_y = var_163_transpose_y_1, x = squeeze_3_cast_fp16, y = squeeze_4_cast_fp16)[name = tensor<string, []>("op_163_cast_fp16")]; |
| tensor<fp16, []> var_164_to_fp16 = const()[name = tensor<string, []>("op_164_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_5_cast_fp16 = mul(x = var_163_cast_fp16, y = var_164_to_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_27_cast_fp16 = softmax(axis = var_11, x = attn_5_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")]; |
| tensor<bool, []> var_168_transpose_x_0 = const()[name = tensor<string, []>("op_168_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_168_transpose_y_0 = const()[name = tensor<string, []>("op_168_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_168_cast_fp16 = matmul(transpose_x = var_168_transpose_x_0, transpose_y = var_168_transpose_y_0, x = input_27_cast_fp16, y = squeeze_5_cast_fp16)[name = tensor<string, []>("op_168_cast_fp16")]; |
| tensor<int32, [4]> var_169_perm_0 = const()[name = tensor<string, []>("op_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_170 = const()[name = tensor<string, []>("op_170"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_169_cast_fp16 = transpose(perm = var_169_perm_0, x = var_168_cast_fp16)[name = tensor<string, []>("transpose_20")]; |
| tensor<fp16, [1, 248, 192]> input_29_cast_fp16 = reshape(shape = var_170, x = var_169_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1238976)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1312768)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1313216)))]; |
| tensor<fp16, [192]> m_encoder_blocks_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_1_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1313664)))]; |
| tensor<fp16, [1, 248, 192]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = m_encoder_blocks_1_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1314112)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1609088)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = linear_6_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1610688)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1905664)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_2_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1906112)))]; |
| tensor<fp16, [192]> m_encoder_blocks_2_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1906560)))]; |
| tensor<fp16, [1, 248, 192]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = m_encoder_blocks_2_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1907008)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2128256)))]; |
| tensor<fp16, [1, 248, 576]> 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_212 = const()[name = tensor<string, []>("op_212"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_213_cast_fp16 = reshape(shape = var_212, x = linear_8_cast_fp16)[name = tensor<string, []>("op_213_cast_fp16")]; |
| tensor<int32, [5]> var_214 = const()[name = tensor<string, []>("op_214"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_216_split_sizes_0 = const()[name = tensor<string, []>("op_216_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_216_axis_0 = const()[name = tensor<string, []>("op_216_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_5_cast_fp16 = transpose(perm = var_214, x = var_213_cast_fp16)[name = tensor<string, []>("transpose_19")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_216_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_216_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_216_cast_fp16_2 = split(axis = var_216_axis_0, split_sizes = var_216_split_sizes_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("op_216_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, 3, 248, 64]> squeeze_6_cast_fp16 = squeeze(axes = squeeze_6_axes_0, x = var_216_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, 3, 248, 64]> squeeze_7_cast_fp16 = squeeze(axes = squeeze_7_axes_0, x = var_216_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, 3, 248, 64]> squeeze_8_cast_fp16 = squeeze(axes = squeeze_8_axes_0, x = var_216_cast_fp16_2)[name = tensor<string, []>("squeeze_8_cast_fp16")]; |
| tensor<bool, []> var_221_transpose_x_1 = const()[name = tensor<string, []>("op_221_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_221_transpose_y_1 = const()[name = tensor<string, []>("op_221_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_221_cast_fp16 = matmul(transpose_x = var_221_transpose_x_1, transpose_y = var_221_transpose_y_1, x = squeeze_6_cast_fp16, y = squeeze_7_cast_fp16)[name = tensor<string, []>("op_221_cast_fp16")]; |
| tensor<fp16, []> var_222_to_fp16 = const()[name = tensor<string, []>("op_222_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_9_cast_fp16 = mul(x = var_221_cast_fp16, y = var_222_to_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_47_cast_fp16 = softmax(axis = var_11, x = attn_9_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")]; |
| tensor<bool, []> var_226_transpose_x_0 = const()[name = tensor<string, []>("op_226_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_226_transpose_y_0 = const()[name = tensor<string, []>("op_226_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_226_cast_fp16 = matmul(transpose_x = var_226_transpose_x_0, transpose_y = var_226_transpose_y_0, x = input_47_cast_fp16, y = squeeze_8_cast_fp16)[name = tensor<string, []>("op_226_cast_fp16")]; |
| tensor<int32, [4]> var_227_perm_0 = const()[name = tensor<string, []>("op_227_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_228 = const()[name = tensor<string, []>("op_228"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_227_cast_fp16 = transpose(perm = var_227_perm_0, x = var_226_cast_fp16)[name = tensor<string, []>("transpose_18")]; |
| tensor<fp16, [1, 248, 192]> input_49_cast_fp16 = reshape(shape = var_228, x = var_227_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2129472)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2203264)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_2_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2203712)))]; |
| tensor<fp16, [192]> m_encoder_blocks_2_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_2_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2204160)))]; |
| tensor<fp16, [1, 248, 192]> input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = m_encoder_blocks_2_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2204608)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2499584)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2501184)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2796160)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_3_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2796608)))]; |
| tensor<fp16, [192]> m_encoder_blocks_3_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2797056)))]; |
| tensor<fp16, [1, 248, 192]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = m_encoder_blocks_3_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2797504)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3018752)))]; |
| tensor<fp16, [1, 248, 576]> 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_270 = const()[name = tensor<string, []>("op_270"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_271_cast_fp16 = reshape(shape = var_270, x = linear_12_cast_fp16)[name = tensor<string, []>("op_271_cast_fp16")]; |
| tensor<int32, [5]> var_272 = const()[name = tensor<string, []>("op_272"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_274_split_sizes_0 = const()[name = tensor<string, []>("op_274_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_274_axis_0 = const()[name = tensor<string, []>("op_274_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_7_cast_fp16 = transpose(perm = var_272, x = var_271_cast_fp16)[name = tensor<string, []>("transpose_17")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_274_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_274_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_274_cast_fp16_2 = split(axis = var_274_axis_0, split_sizes = var_274_split_sizes_0, x = qkv_7_cast_fp16)[name = tensor<string, []>("op_274_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, 3, 248, 64]> squeeze_9_cast_fp16 = squeeze(axes = squeeze_9_axes_0, x = var_274_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, 3, 248, 64]> squeeze_10_cast_fp16 = squeeze(axes = squeeze_10_axes_0, x = var_274_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, 3, 248, 64]> squeeze_11_cast_fp16 = squeeze(axes = squeeze_11_axes_0, x = var_274_cast_fp16_2)[name = tensor<string, []>("squeeze_11_cast_fp16")]; |
| tensor<bool, []> var_279_transpose_x_1 = const()[name = tensor<string, []>("op_279_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_279_transpose_y_1 = const()[name = tensor<string, []>("op_279_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_279_cast_fp16 = matmul(transpose_x = var_279_transpose_x_1, transpose_y = var_279_transpose_y_1, x = squeeze_9_cast_fp16, y = squeeze_10_cast_fp16)[name = tensor<string, []>("op_279_cast_fp16")]; |
| tensor<fp16, []> var_280_to_fp16 = const()[name = tensor<string, []>("op_280_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_13_cast_fp16 = mul(x = var_279_cast_fp16, y = var_280_to_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_67_cast_fp16 = softmax(axis = var_11, x = attn_13_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")]; |
| tensor<bool, []> var_284_transpose_x_0 = const()[name = tensor<string, []>("op_284_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_284_transpose_y_0 = const()[name = tensor<string, []>("op_284_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_284_cast_fp16 = matmul(transpose_x = var_284_transpose_x_0, transpose_y = var_284_transpose_y_0, x = input_67_cast_fp16, y = squeeze_11_cast_fp16)[name = tensor<string, []>("op_284_cast_fp16")]; |
| tensor<int32, [4]> var_285_perm_0 = const()[name = tensor<string, []>("op_285_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_286 = const()[name = tensor<string, []>("op_286"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_285_cast_fp16 = transpose(perm = var_285_perm_0, x = var_284_cast_fp16)[name = tensor<string, []>("transpose_16")]; |
| tensor<fp16, [1, 248, 192]> input_69_cast_fp16 = reshape(shape = var_286, x = var_285_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3019968)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3093760)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_3_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3094208)))]; |
| tensor<fp16, [192]> m_encoder_blocks_3_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_3_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3094656)))]; |
| tensor<fp16, [1, 248, 192]> input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = m_encoder_blocks_3_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3095104)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3390080)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_14_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3391680)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3686656)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_4_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3687104)))]; |
| tensor<fp16, [192]> m_encoder_blocks_4_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3687552)))]; |
| tensor<fp16, [1, 248, 192]> x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = m_encoder_blocks_4_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3688000)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3909248)))]; |
| tensor<fp16, [1, 248, 576]> 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_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_329_cast_fp16 = reshape(shape = var_328, x = linear_16_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")]; |
| tensor<int32, [5]> var_330 = const()[name = tensor<string, []>("op_330"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_332_split_sizes_0 = const()[name = tensor<string, []>("op_332_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_332_axis_0 = const()[name = tensor<string, []>("op_332_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_9_cast_fp16 = transpose(perm = var_330, x = var_329_cast_fp16)[name = tensor<string, []>("transpose_15")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_332_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_332_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_332_cast_fp16_2 = split(axis = var_332_axis_0, split_sizes = var_332_split_sizes_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("op_332_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, 3, 248, 64]> squeeze_12_cast_fp16 = squeeze(axes = squeeze_12_axes_0, x = var_332_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, 3, 248, 64]> squeeze_13_cast_fp16 = squeeze(axes = squeeze_13_axes_0, x = var_332_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, 3, 248, 64]> squeeze_14_cast_fp16 = squeeze(axes = squeeze_14_axes_0, x = var_332_cast_fp16_2)[name = tensor<string, []>("squeeze_14_cast_fp16")]; |
| tensor<bool, []> var_337_transpose_x_1 = const()[name = tensor<string, []>("op_337_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_337_transpose_y_1 = const()[name = tensor<string, []>("op_337_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_337_cast_fp16 = matmul(transpose_x = var_337_transpose_x_1, transpose_y = var_337_transpose_y_1, x = squeeze_12_cast_fp16, y = squeeze_13_cast_fp16)[name = tensor<string, []>("op_337_cast_fp16")]; |
| tensor<fp16, []> var_338_to_fp16 = const()[name = tensor<string, []>("op_338_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_17_cast_fp16 = mul(x = var_337_cast_fp16, y = var_338_to_fp16)[name = tensor<string, []>("attn_17_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_87_cast_fp16 = softmax(axis = var_11, x = attn_17_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")]; |
| tensor<bool, []> var_342_transpose_x_0 = const()[name = tensor<string, []>("op_342_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_342_transpose_y_0 = const()[name = tensor<string, []>("op_342_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_342_cast_fp16 = matmul(transpose_x = var_342_transpose_x_0, transpose_y = var_342_transpose_y_0, x = input_87_cast_fp16, y = squeeze_14_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")]; |
| tensor<int32, [4]> var_343_perm_0 = const()[name = tensor<string, []>("op_343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_343_cast_fp16 = transpose(perm = var_343_perm_0, x = var_342_cast_fp16)[name = tensor<string, []>("transpose_14")]; |
| tensor<fp16, [1, 248, 192]> input_89_cast_fp16 = reshape(shape = var_344, x = var_343_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3910464)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3984256)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_4_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3984704)))]; |
| tensor<fp16, [192]> m_encoder_blocks_4_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_4_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3985152)))]; |
| tensor<fp16, [1, 248, 192]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = m_encoder_blocks_4_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3985600)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4280576)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_18_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4282176)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4577152)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_5_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4577600)))]; |
| tensor<fp16, [192]> m_encoder_blocks_5_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4578048)))]; |
| tensor<fp16, [1, 248, 192]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = m_encoder_blocks_5_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4578496)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4799744)))]; |
| tensor<fp16, [1, 248, 576]> 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_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_387_cast_fp16 = reshape(shape = var_386, x = linear_20_cast_fp16)[name = tensor<string, []>("op_387_cast_fp16")]; |
| tensor<int32, [5]> var_388 = const()[name = tensor<string, []>("op_388"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_390_split_sizes_0 = const()[name = tensor<string, []>("op_390_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_390_axis_0 = const()[name = tensor<string, []>("op_390_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_11_cast_fp16 = transpose(perm = var_388, x = var_387_cast_fp16)[name = tensor<string, []>("transpose_13")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_390_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_390_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_390_cast_fp16_2 = split(axis = var_390_axis_0, split_sizes = var_390_split_sizes_0, x = qkv_11_cast_fp16)[name = tensor<string, []>("op_390_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, 3, 248, 64]> squeeze_15_cast_fp16 = squeeze(axes = squeeze_15_axes_0, x = var_390_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, 3, 248, 64]> squeeze_16_cast_fp16 = squeeze(axes = squeeze_16_axes_0, x = var_390_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, 3, 248, 64]> squeeze_17_cast_fp16 = squeeze(axes = squeeze_17_axes_0, x = var_390_cast_fp16_2)[name = tensor<string, []>("squeeze_17_cast_fp16")]; |
| tensor<bool, []> var_395_transpose_x_1 = const()[name = tensor<string, []>("op_395_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_395_transpose_y_1 = const()[name = tensor<string, []>("op_395_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_395_cast_fp16 = matmul(transpose_x = var_395_transpose_x_1, transpose_y = var_395_transpose_y_1, x = squeeze_15_cast_fp16, y = squeeze_16_cast_fp16)[name = tensor<string, []>("op_395_cast_fp16")]; |
| tensor<fp16, []> var_396_to_fp16 = const()[name = tensor<string, []>("op_396_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_21_cast_fp16 = mul(x = var_395_cast_fp16, y = var_396_to_fp16)[name = tensor<string, []>("attn_21_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_107_cast_fp16 = softmax(axis = var_11, x = attn_21_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")]; |
| tensor<bool, []> var_400_transpose_x_0 = const()[name = tensor<string, []>("op_400_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_400_transpose_y_0 = const()[name = tensor<string, []>("op_400_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = input_107_cast_fp16, y = squeeze_17_cast_fp16)[name = tensor<string, []>("op_400_cast_fp16")]; |
| tensor<int32, [4]> var_401_perm_0 = const()[name = tensor<string, []>("op_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_402 = const()[name = tensor<string, []>("op_402"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_401_cast_fp16 = transpose(perm = var_401_perm_0, x = var_400_cast_fp16)[name = tensor<string, []>("transpose_12")]; |
| tensor<fp16, [1, 248, 192]> input_109_cast_fp16 = reshape(shape = var_402, x = var_401_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4800960)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4874752)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_5_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4875200)))]; |
| tensor<fp16, [192]> m_encoder_blocks_5_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_5_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4875648)))]; |
| tensor<fp16, [1, 248, 192]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = m_encoder_blocks_5_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4876096)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5171072)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5172672)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5467648)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_6_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5468096)))]; |
| tensor<fp16, [192]> m_encoder_blocks_6_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5468544)))]; |
| tensor<fp16, [1, 248, 192]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = m_encoder_blocks_6_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5468992)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5690240)))]; |
| tensor<fp16, [1, 248, 576]> 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_444 = const()[name = tensor<string, []>("op_444"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_445_cast_fp16 = reshape(shape = var_444, x = linear_24_cast_fp16)[name = tensor<string, []>("op_445_cast_fp16")]; |
| tensor<int32, [5]> var_446 = const()[name = tensor<string, []>("op_446"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_448_split_sizes_0 = const()[name = tensor<string, []>("op_448_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_448_axis_0 = const()[name = tensor<string, []>("op_448_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_13_cast_fp16 = transpose(perm = var_446, x = var_445_cast_fp16)[name = tensor<string, []>("transpose_11")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_448_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_448_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_448_cast_fp16_2 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("op_448_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, 3, 248, 64]> squeeze_18_cast_fp16 = squeeze(axes = squeeze_18_axes_0, x = var_448_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, 3, 248, 64]> squeeze_19_cast_fp16 = squeeze(axes = squeeze_19_axes_0, x = var_448_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, 3, 248, 64]> squeeze_20_cast_fp16 = squeeze(axes = squeeze_20_axes_0, x = var_448_cast_fp16_2)[name = tensor<string, []>("squeeze_20_cast_fp16")]; |
| tensor<bool, []> var_453_transpose_x_1 = const()[name = tensor<string, []>("op_453_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_453_transpose_y_1 = const()[name = tensor<string, []>("op_453_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_453_cast_fp16 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = squeeze_18_cast_fp16, y = squeeze_19_cast_fp16)[name = tensor<string, []>("op_453_cast_fp16")]; |
| tensor<fp16, []> var_454_to_fp16 = const()[name = tensor<string, []>("op_454_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_25_cast_fp16 = mul(x = var_453_cast_fp16, y = var_454_to_fp16)[name = tensor<string, []>("attn_25_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_127_cast_fp16 = softmax(axis = var_11, x = attn_25_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")]; |
| tensor<bool, []> var_458_transpose_x_0 = const()[name = tensor<string, []>("op_458_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_458_transpose_y_0 = const()[name = tensor<string, []>("op_458_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_458_cast_fp16 = matmul(transpose_x = var_458_transpose_x_0, transpose_y = var_458_transpose_y_0, x = input_127_cast_fp16, y = squeeze_20_cast_fp16)[name = tensor<string, []>("op_458_cast_fp16")]; |
| tensor<int32, [4]> var_459_perm_0 = const()[name = tensor<string, []>("op_459_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_460 = const()[name = tensor<string, []>("op_460"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_459_cast_fp16 = transpose(perm = var_459_perm_0, x = var_458_cast_fp16)[name = tensor<string, []>("transpose_10")]; |
| tensor<fp16, [1, 248, 192]> input_129_cast_fp16 = reshape(shape = var_460, x = var_459_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5691456)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5765248)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_6_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5765696)))]; |
| tensor<fp16, [192]> m_encoder_blocks_6_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_6_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5766144)))]; |
| tensor<fp16, [1, 248, 192]> input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = m_encoder_blocks_6_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5766592)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6061568)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_26_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6063168)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6358144)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_7_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6358592)))]; |
| tensor<fp16, [192]> m_encoder_blocks_7_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6359040)))]; |
| tensor<fp16, [1, 248, 192]> x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = m_encoder_blocks_7_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6359488)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6580736)))]; |
| tensor<fp16, [1, 248, 576]> 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_502 = const()[name = tensor<string, []>("op_502"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_503_cast_fp16 = reshape(shape = var_502, x = linear_28_cast_fp16)[name = tensor<string, []>("op_503_cast_fp16")]; |
| tensor<int32, [5]> var_504 = const()[name = tensor<string, []>("op_504"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_506_split_sizes_0 = const()[name = tensor<string, []>("op_506_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_506_axis_0 = const()[name = tensor<string, []>("op_506_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_15_cast_fp16 = transpose(perm = var_504, x = var_503_cast_fp16)[name = tensor<string, []>("transpose_9")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_506_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_506_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_506_cast_fp16_2 = split(axis = var_506_axis_0, split_sizes = var_506_split_sizes_0, x = qkv_15_cast_fp16)[name = tensor<string, []>("op_506_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, 3, 248, 64]> squeeze_21_cast_fp16 = squeeze(axes = squeeze_21_axes_0, x = var_506_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, 3, 248, 64]> squeeze_22_cast_fp16 = squeeze(axes = squeeze_22_axes_0, x = var_506_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, 3, 248, 64]> squeeze_23_cast_fp16 = squeeze(axes = squeeze_23_axes_0, x = var_506_cast_fp16_2)[name = tensor<string, []>("squeeze_23_cast_fp16")]; |
| tensor<bool, []> var_511_transpose_x_1 = const()[name = tensor<string, []>("op_511_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_511_transpose_y_1 = const()[name = tensor<string, []>("op_511_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_511_cast_fp16 = matmul(transpose_x = var_511_transpose_x_1, transpose_y = var_511_transpose_y_1, x = squeeze_21_cast_fp16, y = squeeze_22_cast_fp16)[name = tensor<string, []>("op_511_cast_fp16")]; |
| tensor<fp16, []> var_512_to_fp16 = const()[name = tensor<string, []>("op_512_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_29_cast_fp16 = mul(x = var_511_cast_fp16, y = var_512_to_fp16)[name = tensor<string, []>("attn_29_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_147_cast_fp16 = softmax(axis = var_11, x = attn_29_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")]; |
| tensor<bool, []> var_516_transpose_x_0 = const()[name = tensor<string, []>("op_516_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_516_transpose_y_0 = const()[name = tensor<string, []>("op_516_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_516_cast_fp16 = matmul(transpose_x = var_516_transpose_x_0, transpose_y = var_516_transpose_y_0, x = input_147_cast_fp16, y = squeeze_23_cast_fp16)[name = tensor<string, []>("op_516_cast_fp16")]; |
| tensor<int32, [4]> var_517_perm_0 = const()[name = tensor<string, []>("op_517_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_518 = const()[name = tensor<string, []>("op_518"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_517_cast_fp16 = transpose(perm = var_517_perm_0, x = var_516_cast_fp16)[name = tensor<string, []>("transpose_8")]; |
| tensor<fp16, [1, 248, 192]> input_149_cast_fp16 = reshape(shape = var_518, x = var_517_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6581952)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6655744)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_7_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6656192)))]; |
| tensor<fp16, [192]> m_encoder_blocks_7_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_7_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6656640)))]; |
| tensor<fp16, [1, 248, 192]> input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = m_encoder_blocks_7_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6657088)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6952064)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = linear_30_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6953664)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7248640)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_8_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7249088)))]; |
| tensor<fp16, [192]> m_encoder_blocks_8_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7249536)))]; |
| tensor<fp16, [1, 248, 192]> x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = m_encoder_blocks_8_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7249984)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7471232)))]; |
| tensor<fp16, [1, 248, 576]> 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_560 = const()[name = tensor<string, []>("op_560"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_561_cast_fp16 = reshape(shape = var_560, x = linear_32_cast_fp16)[name = tensor<string, []>("op_561_cast_fp16")]; |
| tensor<int32, [5]> var_562 = const()[name = tensor<string, []>("op_562"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_564_split_sizes_0 = const()[name = tensor<string, []>("op_564_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_564_axis_0 = const()[name = tensor<string, []>("op_564_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_17_cast_fp16 = transpose(perm = var_562, x = var_561_cast_fp16)[name = tensor<string, []>("transpose_7")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_564_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_564_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_564_cast_fp16_2 = split(axis = var_564_axis_0, split_sizes = var_564_split_sizes_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("op_564_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, 3, 248, 64]> squeeze_24_cast_fp16 = squeeze(axes = squeeze_24_axes_0, x = var_564_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, 3, 248, 64]> squeeze_25_cast_fp16 = squeeze(axes = squeeze_25_axes_0, x = var_564_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, 3, 248, 64]> squeeze_26_cast_fp16 = squeeze(axes = squeeze_26_axes_0, x = var_564_cast_fp16_2)[name = tensor<string, []>("squeeze_26_cast_fp16")]; |
| tensor<bool, []> var_569_transpose_x_1 = const()[name = tensor<string, []>("op_569_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_569_transpose_y_1 = const()[name = tensor<string, []>("op_569_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_569_cast_fp16 = matmul(transpose_x = var_569_transpose_x_1, transpose_y = var_569_transpose_y_1, x = squeeze_24_cast_fp16, y = squeeze_25_cast_fp16)[name = tensor<string, []>("op_569_cast_fp16")]; |
| tensor<fp16, []> var_570_to_fp16 = const()[name = tensor<string, []>("op_570_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_33_cast_fp16 = mul(x = var_569_cast_fp16, y = var_570_to_fp16)[name = tensor<string, []>("attn_33_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_167_cast_fp16 = softmax(axis = var_11, x = attn_33_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")]; |
| tensor<bool, []> var_574_transpose_x_0 = const()[name = tensor<string, []>("op_574_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_574_transpose_y_0 = const()[name = tensor<string, []>("op_574_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_574_cast_fp16 = matmul(transpose_x = var_574_transpose_x_0, transpose_y = var_574_transpose_y_0, x = input_167_cast_fp16, y = squeeze_26_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")]; |
| tensor<int32, [4]> var_575_perm_0 = const()[name = tensor<string, []>("op_575_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_576 = const()[name = tensor<string, []>("op_576"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_575_cast_fp16 = transpose(perm = var_575_perm_0, x = var_574_cast_fp16)[name = tensor<string, []>("transpose_6")]; |
| tensor<fp16, [1, 248, 192]> input_169_cast_fp16 = reshape(shape = var_576, x = var_575_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7472448)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7546240)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_8_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7546688)))]; |
| tensor<fp16, [192]> m_encoder_blocks_8_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_8_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7547136)))]; |
| tensor<fp16, [1, 248, 192]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = m_encoder_blocks_8_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7547584)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7842560)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7844160)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8139136)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_9_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8139584)))]; |
| tensor<fp16, [192]> m_encoder_blocks_9_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8140032)))]; |
| tensor<fp16, [1, 248, 192]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = m_encoder_blocks_9_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8140480)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8361728)))]; |
| tensor<fp16, [1, 248, 576]> 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_618 = const()[name = tensor<string, []>("op_618"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_619_cast_fp16 = reshape(shape = var_618, x = linear_36_cast_fp16)[name = tensor<string, []>("op_619_cast_fp16")]; |
| tensor<int32, [5]> var_620 = const()[name = tensor<string, []>("op_620"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_622_split_sizes_0 = const()[name = tensor<string, []>("op_622_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_622_axis_0 = const()[name = tensor<string, []>("op_622_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_19_cast_fp16 = transpose(perm = var_620, x = var_619_cast_fp16)[name = tensor<string, []>("transpose_5")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_622_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_622_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_622_cast_fp16_2 = split(axis = var_622_axis_0, split_sizes = var_622_split_sizes_0, x = qkv_19_cast_fp16)[name = tensor<string, []>("op_622_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, 3, 248, 64]> squeeze_27_cast_fp16 = squeeze(axes = squeeze_27_axes_0, x = var_622_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, 3, 248, 64]> squeeze_28_cast_fp16 = squeeze(axes = squeeze_28_axes_0, x = var_622_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, 3, 248, 64]> squeeze_29_cast_fp16 = squeeze(axes = squeeze_29_axes_0, x = var_622_cast_fp16_2)[name = tensor<string, []>("squeeze_29_cast_fp16")]; |
| tensor<bool, []> var_627_transpose_x_1 = const()[name = tensor<string, []>("op_627_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_627_transpose_y_1 = const()[name = tensor<string, []>("op_627_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_627_cast_fp16 = matmul(transpose_x = var_627_transpose_x_1, transpose_y = var_627_transpose_y_1, x = squeeze_27_cast_fp16, y = squeeze_28_cast_fp16)[name = tensor<string, []>("op_627_cast_fp16")]; |
| tensor<fp16, []> var_628_to_fp16 = const()[name = tensor<string, []>("op_628_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_37_cast_fp16 = mul(x = var_627_cast_fp16, y = var_628_to_fp16)[name = tensor<string, []>("attn_37_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_187_cast_fp16 = softmax(axis = var_11, x = attn_37_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")]; |
| tensor<bool, []> var_632_transpose_x_0 = const()[name = tensor<string, []>("op_632_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_632_transpose_y_0 = const()[name = tensor<string, []>("op_632_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_632_cast_fp16 = matmul(transpose_x = var_632_transpose_x_0, transpose_y = var_632_transpose_y_0, x = input_187_cast_fp16, y = squeeze_29_cast_fp16)[name = tensor<string, []>("op_632_cast_fp16")]; |
| tensor<int32, [4]> var_633_perm_0 = const()[name = tensor<string, []>("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_634 = const()[name = tensor<string, []>("op_634"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = var_632_cast_fp16)[name = tensor<string, []>("transpose_4")]; |
| tensor<fp16, [1, 248, 192]> input_189_cast_fp16 = reshape(shape = var_634, x = var_633_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8362944)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8436736)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_9_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8437184)))]; |
| tensor<fp16, [192]> m_encoder_blocks_9_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_9_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8437632)))]; |
| tensor<fp16, [1, 248, 192]> input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = m_encoder_blocks_9_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8438080)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8733056)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_38_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8734656)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9029632)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_10_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9030080)))]; |
| tensor<fp16, [192]> m_encoder_blocks_10_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9030528)))]; |
| tensor<fp16, [1, 248, 192]> x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = m_encoder_blocks_10_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9030976)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9252224)))]; |
| tensor<fp16, [1, 248, 576]> 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_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_677_cast_fp16 = reshape(shape = var_676, x = linear_40_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")]; |
| tensor<int32, [5]> var_678 = const()[name = tensor<string, []>("op_678"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_680_split_sizes_0 = const()[name = tensor<string, []>("op_680_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_680_axis_0 = const()[name = tensor<string, []>("op_680_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_21_cast_fp16 = transpose(perm = var_678, x = var_677_cast_fp16)[name = tensor<string, []>("transpose_3")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_680_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_680_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_680_cast_fp16_2 = split(axis = var_680_axis_0, split_sizes = var_680_split_sizes_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("op_680_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, 3, 248, 64]> squeeze_30_cast_fp16 = squeeze(axes = squeeze_30_axes_0, x = var_680_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, 3, 248, 64]> squeeze_31_cast_fp16 = squeeze(axes = squeeze_31_axes_0, x = var_680_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, 3, 248, 64]> squeeze_32_cast_fp16 = squeeze(axes = squeeze_32_axes_0, x = var_680_cast_fp16_2)[name = tensor<string, []>("squeeze_32_cast_fp16")]; |
| tensor<bool, []> var_685_transpose_x_1 = const()[name = tensor<string, []>("op_685_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_685_transpose_y_1 = const()[name = tensor<string, []>("op_685_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_685_cast_fp16 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = squeeze_30_cast_fp16, y = squeeze_31_cast_fp16)[name = tensor<string, []>("op_685_cast_fp16")]; |
| tensor<fp16, []> var_686_to_fp16 = const()[name = tensor<string, []>("op_686_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_41_cast_fp16 = mul(x = var_685_cast_fp16, y = var_686_to_fp16)[name = tensor<string, []>("attn_41_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_207_cast_fp16 = softmax(axis = var_11, x = attn_41_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")]; |
| tensor<bool, []> var_690_transpose_x_0 = const()[name = tensor<string, []>("op_690_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_690_transpose_y_0 = const()[name = tensor<string, []>("op_690_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_690_cast_fp16 = matmul(transpose_x = var_690_transpose_x_0, transpose_y = var_690_transpose_y_0, x = input_207_cast_fp16, y = squeeze_32_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")]; |
| tensor<int32, [4]> var_691_perm_0 = const()[name = tensor<string, []>("op_691_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_692 = const()[name = tensor<string, []>("op_692"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_691_cast_fp16 = transpose(perm = var_691_perm_0, x = var_690_cast_fp16)[name = tensor<string, []>("transpose_2")]; |
| tensor<fp16, [1, 248, 192]> input_209_cast_fp16 = reshape(shape = var_692, x = var_691_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9253440)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9327232)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_10_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9327680)))]; |
| tensor<fp16, [192]> m_encoder_blocks_10_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_10_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9328128)))]; |
| tensor<fp16, [1, 248, 192]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = m_encoder_blocks_10_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9328576)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9623552)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = linear_42_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9625152)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9920128)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_11_norm1_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm1_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9920576)))]; |
| tensor<fp16, [192]> m_encoder_blocks_11_norm1_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm1_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9921024)))]; |
| tensor<fp16, [1, 248, 192]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = m_encoder_blocks_11_norm1_bias_to_fp16, epsilon = var_14_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, [576, 192]> 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, [576, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9921472)))]; |
| tensor<fp16, [576]> 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, [576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10142720)))]; |
| tensor<fp16, [1, 248, 576]> 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_734 = const()[name = tensor<string, []>("op_734"), val = tensor<int32, [5]>([1, 248, 3, 3, 64])]; |
| tensor<fp16, [1, 248, 3, 3, 64]> var_735_cast_fp16 = reshape(shape = var_734, x = linear_44_cast_fp16)[name = tensor<string, []>("op_735_cast_fp16")]; |
| tensor<int32, [5]> var_736 = const()[name = tensor<string, []>("op_736"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [3]> var_738_split_sizes_0 = const()[name = tensor<string, []>("op_738_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<int32, []> var_738_axis_0 = const()[name = tensor<string, []>("op_738_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<fp16, [3, 1, 3, 248, 64]> qkv_23_cast_fp16 = transpose(perm = var_736, x = var_735_cast_fp16)[name = tensor<string, []>("transpose_1")]; |
| tensor<fp16, [1, 1, 3, 248, 64]> var_738_cast_fp16_0, tensor<fp16, [1, 1, 3, 248, 64]> var_738_cast_fp16_1, tensor<fp16, [1, 1, 3, 248, 64]> var_738_cast_fp16_2 = split(axis = var_738_axis_0, split_sizes = var_738_split_sizes_0, x = qkv_23_cast_fp16)[name = tensor<string, []>("op_738_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, 3, 248, 64]> squeeze_33_cast_fp16 = squeeze(axes = squeeze_33_axes_0, x = var_738_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, 3, 248, 64]> squeeze_34_cast_fp16 = squeeze(axes = squeeze_34_axes_0, x = var_738_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, 3, 248, 64]> squeeze_35_cast_fp16 = squeeze(axes = squeeze_35_axes_0, x = var_738_cast_fp16_2)[name = tensor<string, []>("squeeze_35_cast_fp16")]; |
| tensor<bool, []> var_743_transpose_x_1 = const()[name = tensor<string, []>("op_743_transpose_x_1"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_743_transpose_y_1 = const()[name = tensor<string, []>("op_743_transpose_y_1"), val = tensor<bool, []>(true)]; |
| tensor<fp16, [1, 3, 248, 248]> var_743_cast_fp16 = matmul(transpose_x = var_743_transpose_x_1, transpose_y = var_743_transpose_y_1, x = squeeze_33_cast_fp16, y = squeeze_34_cast_fp16)[name = tensor<string, []>("op_743_cast_fp16")]; |
| tensor<fp16, []> var_744_to_fp16 = const()[name = tensor<string, []>("op_744_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 3, 248, 248]> attn_45_cast_fp16 = mul(x = var_743_cast_fp16, y = var_744_to_fp16)[name = tensor<string, []>("attn_45_cast_fp16")]; |
| tensor<fp16, [1, 3, 248, 248]> input_227_cast_fp16 = softmax(axis = var_11, x = attn_45_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")]; |
| tensor<bool, []> var_748_transpose_x_0 = const()[name = tensor<string, []>("op_748_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> var_748_transpose_y_0 = const()[name = tensor<string, []>("op_748_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 3, 248, 64]> var_748_cast_fp16 = matmul(transpose_x = var_748_transpose_x_0, transpose_y = var_748_transpose_y_0, x = input_227_cast_fp16, y = squeeze_35_cast_fp16)[name = tensor<string, []>("op_748_cast_fp16")]; |
| tensor<int32, [4]> var_749_perm_0 = const()[name = tensor<string, []>("op_749_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_750 = const()[name = tensor<string, []>("op_750"), val = tensor<int32, [3]>([1, 248, 192])]; |
| tensor<fp16, [1, 248, 3, 64]> var_749_cast_fp16 = transpose(perm = var_749_perm_0, x = var_748_cast_fp16)[name = tensor<string, []>("transpose_0")]; |
| tensor<fp16, [1, 248, 192]> input_229_cast_fp16 = reshape(shape = var_750, x = var_749_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")]; |
| tensor<fp16, [192, 192]> 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, [192, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10143936)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10217728)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_blocks_11_norm2_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm2_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10218176)))]; |
| tensor<fp16, [192]> m_encoder_blocks_11_norm2_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_blocks_11_norm2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10218624)))]; |
| tensor<fp16, [1, 248, 192]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = m_encoder_blocks_11_norm2_bias_to_fp16, epsilon = var_14_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, [768, 192]> 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, [768, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10219072)))]; |
| tensor<fp16, [768]> 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, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10514048)))]; |
| tensor<fp16, [1, 248, 768]> 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, 768]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")]; |
| tensor<fp16, [192, 768]> 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, [192, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10515648)))]; |
| tensor<fp16, [192]> 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, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10810624)))]; |
| tensor<fp16, [1, 248, 192]> 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, 192]> 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, [192]> m_encoder_norm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_norm_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10811072)))]; |
| tensor<fp16, [192]> m_encoder_norm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_norm_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10811520)))]; |
| tensor<fp16, [1, 248, 192]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = m_encoder_norm_bias_to_fp16, epsilon = var_14_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, 192]> 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, [192]> m_outputlayer_0_weight_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_0_weight_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10811968)))]; |
| tensor<fp16, [192]> m_outputlayer_0_bias_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_0_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10812416)))]; |
| tensor<fp16, []> var_790_to_fp16 = const()[name = tensor<string, []>("op_790_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 192]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = m_outputlayer_0_bias_to_fp16, epsilon = var_790_to_fp16, gamma = m_outputlayer_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; |
| tensor<fp16, [527, 192]> m_outputlayer_1_weight_to_fp16 = const()[name = tensor<string, []>("m_outputlayer_1_weight_to_fp16"), val = tensor<fp16, [527, 192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10812864)))]; |
| 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, []>(11015296)))]; |
| 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); |
| } |
