adaseq踩坑记录 unpicklable

[[老环境配置记录modelscope]]

modelscope :/usr/local/bin/python是默认的python路径

2025-09-02 17:35:55,329 - modelscope - WARNING - task token-classification input definition is missing 2025-09-02 17:35:55,509 - modelscope - WARNING - task token-classification output keys are missing

3/47 周赛1 # 2025天池极客挑战赛——用户购买行为预测 全国第三

adaseq感觉已经两年没有维护了，安全加载模型没有配置方法

from adaseq.commands.train import build_trainer_from_partial_objects

work_dir = 'experiments/transformer_crf'
os.makedirs(work_dir, exist_ok=True)

trainer = build_trainer_from_partial_objects(
    config,
    work_dir=work_dir,
    seed=42,
    device='cuda:0'
)

trainer.train()
trainer.test()

Weights only load failed. This file can still be loaded, to do so you have two options, do those steps only if you trust the source of the checkpoint. 
	(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
	(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
	WeightsUnpickler error: Unsupported global: GLOBAL numpy.core.multiarray._reconstruct was not an allowed global by default. Please use `torch.serialization.add_safe_globals([_reconstruct])` or the `torch.serialization.safe_globals([_reconstruct])` context manager to allowlist this global if you trust this class/function.

echo ‘export PATH=”/mnt/workspace/miniconda3/bin:$PATH”‘ >> ~/.bashrc source ~/.bashrc
conda install ipykernel
python -m ipykernel install --user --name py311 --display-name "Python 3.11 (py311)"

解决方法 work！

!pip install torch==2.3.1 transformers==4.48.3 modelscope==1.29.1
!pip install "modelscope[nlp]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html

mego backbone架构：

modelscope/models/multi_modal/mgeo/backbone.py
编码器768，

MGeo(
  (text_encoder): BertModel(
    (embeddings): BertEmbeddings(
      (word_embeddings): Embedding(21128, 768, padding_idx=0)
      (position_embeddings): Embedding(512, 768)
      (token_type_embeddings): Embedding(2, 768)
      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (encoder): BertEncoder(
      (layer): ModuleList(
        (0-11): 12 x BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
            (intermediate_act_fn): GELUActivation()
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
    )
  )
)

modelscope/models/multi_modal/mgeo/token_classification.py
MGeoForTokenClassification 类是一个用于执行词元分类任务的 PyTorch 模型，它继承自 MGeoPreTrainedModel 初始化backbone，并使用一个专门的、基于 Transformer 的主干网络 (MGeo) 来处理输入的文本。

---

核心组件与功能分析

__init__ 方法

这个构造函数用于构建模型的架构：

• 它根据配置初始化标签数量 (self.num_labels)。
• 它实例化了 MGeo 主干模型，该模型负责核心的文本编码工作。
• 它添加了一个 dropout 层 (self.dropout) 来防止过拟合。
• 它在 MGeo 主干网络的顶部添加了一个线性分类器头 (self.classifier)。这个分类器将主干网络的输出映射到所需的标签数量。

forward 方法

forward 方法定义了模型如何处理输入并生成输出。它接收多个参数，包括 input_ids、attention_mask 和 labels。其处理流程如下：

1. 输入处理：该方法首先将输入（如 input_ids 和 attention_mask）传递给 self.bert（即 MGeo 主干网络）以获得输入序列的编码表示。
2. Dropout 与分类：它对主干网络输出的序列应用 dropout，然后将其输入线性分类器，计算出 logits（即每个词元的原始、非标准化的预测分数）。
3. 损失计算：如果提供了 labels，模型会计算**交叉熵损失 (Cross-Entropy Loss)**。它巧妙地利用 attention_mask 来确保损失只针对非填充词元进行计算，这对于模型的准确训练至关重要。
4. 输出：最后，该方法返回结果。如果 return_dict 为 True，它会返回一个详细的 AttentionTokenClassificationModelOutput 对象，其中包含计算出的 loss、logits 以及主干网络的其他输出。否则，它返回一个简单的元组。

800个词->40G

解决方法 NOT WORK

1. make sure torch>=2.6.0

   !pip install --upgrade torch==2.6.0 torchvision

2. import torch.serialization
   # Add numpy.core.multiarray._reconstruct to safe globals
   from numpy.core.multiarray import _reconstruct,ndarray
   torch.serialization.add_safe_globals([_reconstruct])
   torch.serialization.add_safe_globals([ndarray])

# initialize a trainer
import os
from adaseq.commands.train import build_trainer_from_partial_objects

work_dir = 'experiments/transformer_crf'
os.makedirs(work_dir, exist_ok=True)

trainer = build_trainer_from_partial_objects(
    config,
    work_dir=work_dir,
    seed=42,
    device='cuda:0'
)

# do training
trainer.train()

# do testing
trainer.test()

Model revision not specified, use revision: v1.0.0

ValueError: DefaultTrainer: SequenceLabelingModel: TransformerEmbedder: Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors.
See the vulnerability report here https://nvd.nist.gov/vuln/detail/CVE-2025-32434

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:211, in build_from_cfg(cfg, registry, group_key, default_args)
    210     else:
--> 211         return obj_cls(**args)
    212 except Exception as e:
    213     # Normal TypeError does not print class name.

File /usr/local/lib/python3.11/site-packages/adaseq/modules/embedders/transformer_embedder.py:84, in TransformerEmbedder.__init__(self, model_name_or_path, drop_special_tokens, sub_module, train_parameters, eval_mode, load_weights, scalar_mix, gradient_checkpointing, transformer_kwargs, sub_token_mode)
     82 self.sub_token_mode = sub_token_mode
---> 84 self.transformer_model, self.from_hf = get_transformer(
     85 model_name_or_path,
     86 load_weights=load_weights,
     87 **(transformer_kwargs or {}),
     88 )
     90 if self.from_hf:

File /usr/local/lib/python3.11/site-packages/adaseq/modules/embedders/transformer_embedder.py:363, in get_transformer(model_name_or_path, load_weights, source, **kwargs)
    362 try:
--> 363     return get_ms_transformer(model_name_or_path, **kwargs), False
    364 except HTTPError as e:

File /usr/local/lib/python3.11/site-packages/adaseq/modules/embedders/transformer_embedder.py:414, in get_ms_transformer(model_name_or_path, **kwargs)
    413 try:
--> 414     transformer = MsModel.from_pretrained(model_name_or_path, task='backbone', **kwargs)
    415 except KeyError:

File /usr/local/lib/python3.11/site-packages/modelscope/models/base/base_model.py:178, in Model.from_pretrained(cls, model_name_or_path, revision, cfg_dict, device, trust_remote_code, **kwargs)
    177 if use_hf in {True, None}:
--> 178     model = try_to_load_hf_model(local_model_dir, task_name, use_hf,
    179 **kwargs)
    180 if model is not None:

File /usr/local/lib/python3.11/site-packages/modelscope/utils/automodel_utils.py:125, in try_to_load_hf_model(model_dir, task_name, use_hf, **kwargs)
    123 if automodel_class is not None:
    124     # use hf
--> 125     model = automodel_class.from_pretrained(model_dir, **kwargs)
    126 return model

File /usr/local/lib/python3.11/site-packages/modelscope/utils/hf_util/patcher.py:281, in _patch_pretrained_class.<locals>.get_wrapped_class.<locals>.ClassWrapper.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
    278     model_dir = get_model_dir(pretrained_model_name_or_path,
    279                               **kwargs)
--> 281 module_obj = module_class.from_pretrained(
    282 model_dir, *model_args, **kwargs)
    284 if module_class.__name__.startswith('AutoModel'):

File /usr/local/lib/python3.11/site-packages/transformers/models/auto/auto_factory.py:600, in _BaseAutoModelClass.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
    599         config = config.get_text_config()
--> 600     return model_class.from_pretrained(
    601 pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
    602 )
    603 raise ValueError(
    604     f"Unrecognized configuration class {config.__class__} for this kind of AutoModel: {cls.__name__}.\n"
    605     f"Model type should be one of {', '.join(c.__name__ for c in cls._model_mapping)}."
    606 )

File /usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py:317, in restore_default_torch_dtype.<locals>._wrapper(*args, **kwargs)
    316 try:
--> 317     return func(*args, **kwargs)
    318 finally:

File /usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py:5069, in PreTrainedModel.from_pretrained(cls, pretrained_model_name_or_path, config, cache_dir, ignore_mismatched_sizes, force_download, local_files_only, token, revision, use_safetensors, weights_only, *model_args, **kwargs)
   5060         torch.set_default_dtype(dtype_orig)
   5062     (
   5063         model,
   5064         missing_keys,
   5065         unexpected_keys,
   5066         mismatched_keys,
   5067         offload_index,
   5068         error_msgs,
-> 5069     ) = cls._load_pretrained_model(
   5070 model,
   5071 state_dict,
   5072 checkpoint_files,
   5073 pretrained_model_name_or_path,
   5074 ignore_mismatched_sizes=ignore_mismatched_sizes,
   5075 sharded_metadata=sharded_metadata,
   5076 device_map=device_map,
   5077 disk_offload_folder=offload_folder,
   5078 offload_state_dict=offload_state_dict,
   5079 dtype=torch_dtype,
   5080 hf_quantizer=hf_quantizer,
   5081 keep_in_fp32_regex=keep_in_fp32_regex,
   5082 device_mesh=device_mesh,
   5083 key_mapping=key_mapping,
   5084 weights_only=weights_only,
   5085 )
   5086 # make sure token embedding weights are still tied if needed

File /usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py:5335, in PreTrainedModel._load_pretrained_model(cls, model, state_dict, checkpoint_files, pretrained_model_name_or_path, ignore_mismatched_sizes, sharded_metadata, device_map, disk_offload_folder, offload_state_dict, dtype, hf_quantizer, keep_in_fp32_regex, device_mesh, key_mapping, weights_only)
   5333 else:
   5334     original_checkpoint_keys = list(
-> 5335         load_state_dict(checkpoint_files[0], map_location="meta", weights_only=weights_only).keys()
   5336     )
   5338 # Check if we are in a special state, i.e. loading from a state dict coming from a different architecture

File /usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py:562, in load_state_dict(checkpoint_file, is_quantized, map_location, weights_only)
    561 if weights_only:
--> 562     check_torch_load_is_safe()
    563 try:

File /usr/local/lib/python3.11/site-packages/transformers/utils/import_utils.py:1622, in check_torch_load_is_safe()
   1621 if not is_torch_greater_or_equal("2.6"):
-> 1622     raise ValueError(
   1623         "Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users "
   1624         "to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply "
   1625         "when loading files with safetensors."
   1626         "\nSee the vulnerability report here [https://nvd.nist.gov/vuln/detail/CVE-2025-32434](https://nvd.nist.gov/vuln/detail/CVE-2025-32434)"
   1627     )

ValueError: Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors.
See the vulnerability report here [https://nvd.nist.gov/vuln/detail/CVE-2025-32434](https://nvd.nist.gov/vuln/detail/CVE-2025-32434)

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:209, in build_from_cfg(cfg, registry, group_key, default_args)
    208 if hasattr(obj_cls, '_instantiate'):
--> 209     return obj_cls._instantiate(**args)
    210 else:

File /usr/local/lib/python3.11/site-packages/modelscope/models/base/base_model.py:85, in Model._instantiate(cls, **kwargs)
     80 """ Define the instantiation method of a model,default method is by
     81     calling the constructor. Note that in the case of no loading model
     82     process in constructor of a task model, a load_model method is
     83     added, and thus this method is overloaded
     84 """
---> 85 return cls(**kwargs)

File /usr/local/lib/python3.11/site-packages/adaseq/models/base.py:40, in Model.__init_subclass__.<locals>.new_init(self, init, *args, **kwargs)
     39 def new_init(self, init=cls.__init__, *args, **kwargs):
---> 40     init(self, *args, **kwargs)
     41     self.post_init()

File /usr/local/lib/python3.11/site-packages/adaseq/models/sequence_labeling_model.py:64, in SequenceLabelingModel.__init__(self, id_to_label, embedder, encoder, dropout, word_dropout, use_crf, multiview, temperature, mv_loss_type, mv_interpolation, partial, chunk, **kwargs)
     63 else:
---> 64     self.embedder = Embedder.from_config(embedder)
     65 hidden_size = self.embedder.get_output_dim()

File /usr/local/lib/python3.11/site-packages/adaseq/modules/embedders/base.py:64, in Embedder.from_config(cls, cfg_dict_or_path, **kwargs)
     63 if cfg['type'] is not None and cfg['type'] in EMBEDDERS.modules['default']:
---> 64     return build_embedder(cfg, default_args=kwargs)
     65 else:

File /usr/local/lib/python3.11/site-packages/adaseq/modules/embedders/base.py:24, in build_embedder(cfg, default_args)
     15 """Build embedder from config dict
     16 
     17 Args:
   (...)     22     embedder (:obj:`Embedder`): an embedder instance
     23 """
---> 24 return build_from_cfg(cfg, EMBEDDERS, group_key='default', default_args=default_args)

File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:214, in build_from_cfg(cfg, registry, group_key, default_args)
    212 except Exception as e:
    213     # Normal TypeError does not print class name.
--> 214     raise type(e)(f'{obj_cls.__name__}: {e}') from e

ValueError: TransformerEmbedder: Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors.
See the vulnerability report here [https://nvd.nist.gov/vuln/detail/CVE-2025-32434](https://nvd.nist.gov/vuln/detail/CVE-2025-32434)

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:211, in build_from_cfg(cfg, registry, group_key, default_args)
    210     else:
--> 211         return obj_cls(**args)
    212 except Exception as e:
    213     # Normal TypeError does not print class name.

File /usr/local/lib/python3.11/site-packages/adaseq/training/default_trainer.py:67, in DefaultTrainer.__init__(self, cfg_file, work_dir, dataset_manager, data_collator, preprocessor, seed, device, **kwargs)
     56 def __init__(
     57     self,
     58     cfg_file: str,
   (...)     65     **kwargs,
     66 ) -> None:
---> 67     super().__init__(
     68 model=None,
     69 cfg_file=cfg_file,
     70 cfg_modify_fn=None,
     71 data_collator=data_collator,
     72 train_dataset=dataset_manager.train,
     73 eval_dataset=dataset_manager.valid,
     74 preprocessor=preprocessor,
     75 work_dir=work_dir,
     76 seed=seed,
     77 device=device,
     78 **kwargs,
     79 )
     81     # Setup testset if there is one

File /usr/local/lib/python3.11/site-packages/modelscope/trainers/trainer.py:175, in EpochBasedTrainer.__init__(self, model, cfg_file, cfg_modify_fn, arg_parse_fn, data_collator, train_dataset, eval_dataset, preprocessor, optimizers, model_revision, seed, callbacks, samplers, efficient_tuners, **kwargs)
    174 else:
--> 175     self.model = self.build_model()
    177 if self._compile:
    178     # Compile the model with torch 2.0

File /usr/local/lib/python3.11/site-packages/adaseq/training/default_trainer.py:102, in DefaultTrainer.build_model(self)
     99 """
    100 Override this func to build adaseq `Model`.
    101 """
--> 102 return Model.from_config(self.cfg)

File /usr/local/lib/python3.11/site-packages/adaseq/models/base.py:118, in Model.from_config(cls, cfg_dict_or_path, **kwargs)
    114     raise ValueError(
    115         'Please pass a correct cfg dict, which should be a reachable file or a dict.'
    116     )
--> 118 model = build_model(model_config, task_name=task, default_args=kwargs)
    119 cfg['framework'] = 'pytorch'

File /usr/local/lib/python3.11/site-packages/modelscope/models/builder.py:35, in build_model(cfg, task_name, default_args)
     34 try:
---> 35     model = build_from_cfg(
     36 cfg, MODELS, group_key=task_name, default_args=default_args)
     37 except KeyError as e:
     38     # Handle subtask with a backbone model that hasn't been registered
     39     # All the subtask with a parent task should have a task model, otherwise it is not a
     40     # valid subtask

File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:214, in build_from_cfg(cfg, registry, group_key, default_args)
    212 except Exception as e:
    213     # Normal TypeError does not print class name.
--> 214     raise type(e)(f'{obj_cls.__name__}: {e}') from e

ValueError: SequenceLabelingModel: TransformerEmbedder: Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors.
See the vulnerability report here [https://nvd.nist.gov/vuln/detail/CVE-2025-32434](https://nvd.nist.gov/vuln/detail/CVE-2025-32434)

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In[3], line 8
      5 work_dir = 'experiments/transformer_crf'
      6 os.makedirs(work_dir, exist_ok=True)
----> 8 trainer = build_trainer_from_partial_objects(
      9 config,
     10 work_dir=work_dir,
     11 seed=42,
     12 device='cuda:0'
     13 )
     15 # do training
     16 trainer.train()

File /usr/local/lib/python3.11/site-packages/adaseq/commands/train.py:218, in build_trainer_from_partial_objects(config, work_dir, **kwargs)
    215     collator_config = dict(type=collator_config)
    216 data_collator = build_data_collator(preprocessor.tokenizer, collator_config)
--> 218 trainer = build_trainer(
    219 config.safe_get('train.trainer', Trainers.default_trainer),
    220 cfg_file=new_config_path,
    221 work_dir=work_dir,
    222 dataset_manager=dm,
    223 data_collator=data_collator,
    224 preprocessor=preprocessor,
    225 **kwargs,
    226 )
    227 return trainer

File /usr/local/lib/python3.11/site-packages/adaseq/training/default_trainer.py:176, in build_trainer(name, **kwargs)
    173 if 'WORLD_SIZE' in os.environ and int(os.environ['WORLD_SIZE']) > 1:
    174     kwargs.update(launcher='pytorch', device='gpu')
--> 176 trainer = ms_build_trainer(name, kwargs)
    177 return trainer

File /usr/local/lib/python3.11/site-packages/modelscope/trainers/builder.py:39, in build_trainer(name, default_args)
     36         register_plugins_repo(configuration.safe_get('plugins'))
     37         register_modelhub_repo(model_dir,
     38                                configuration.get('allow_remote', False))
---> 39 return build_from_cfg(cfg, TRAINERS, default_args=default_args)

File /usr/local/lib/python3.11/site-packages/modelscope/utils/registry.py:214, in build_from_cfg(cfg, registry, group_key, default_args)
    211         return obj_cls(**args)
    212 except Exception as e:
    213     # Normal TypeError does not print class name.
--> 214     raise type(e)(f'{obj_cls.__name__}: {e}') from e

ValueError: DefaultTrainer: SequenceLabelingModel: TransformerEmbedder: Due to a serious vulnerability issue in `torch.load`, even with `weights_only=True`, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors.
See the vulnerability report here [https://nvd.nist.gov/vuln/detail/CVE-2025-32434](https://nvd.nist.gov/vuln/detail/CVE-2025-32434)