huggingface/transformers

TypeError: 'list' object is not callable || Resume from checkpoint

satpalsr opened this issue · 3 comments

satpalsr commented

System Info

Name: transformers
Version: 4.41.0.dev0

When resuming training for mistral model.
This occurs with latest transformers version.
Previous version also have some other errors.

TypeError                                 Traceback (most recent call last)
Cell In[11], line 86
     75 trainer = CustomTrainer(
     76     model=model,
     77     args=training_args,
   (...)
     82     compute_metrics=compute_metrics
     83 )
     85 # Train the model
---> 86 trainer.train(resume_from_checkpoint='/workspace/regression_model_output/checkpoint-60')
     88 # Save model and tokenizer
     89 model_id = 'regression_mistral_model'

File /usr/local/lib/python3.10/dist-packages/transformers/trainer.py:1857, in Trainer.train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
   1855 if resume_from_checkpoint is not None:
   1856     if not is_sagemaker_mp_enabled() and not self.is_deepspeed_enabled and not self.is_fsdp_enabled:
-> 1857         self._load_from_checkpoint(resume_from_checkpoint)
   1858     # In case of repeating the find_executable_batch_size, set `self._train_batch_size` properly
   1859     state = TrainerState.load_from_json(os.path.join(resume_from_checkpoint, TRAINER_STATE_NAME))

File /usr/local/lib/python3.10/dist-packages/transformers/trainer.py:2536, in Trainer._load_from_checkpoint(self, resume_from_checkpoint, model)
   2533 if os.path.exists(resume_from_checkpoint):
   2534     # For BC for older PEFT versions
   2535     if hasattr(model, "active_adapters"):
-> 2536         active_adapters = model.active_adapters()
   2537         if len(active_adapters) > 1:
   2538             logger.warning("Multiple active adapters detected will only consider the first adapter")

TypeError: 'list' object is not callable

Who can help?

@Art

Information

  • The official example scripts
  • My own modified scripts

Tasks

  • An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
  • My own task or dataset (give details below)

Reproduction

\

Expected behavior

Training should start from checkpoint as usual.

pashminacameron commented

@younesbelkada I am seeing a similar issue in _load_best_model since #30738 I have a proposed fix in this commit. Please let me know if that looks good to you, I can open a PR. Thanks!

younesbelkada commented

Hi @pashminacameron
Thanks for your message! Sorry for the confusion, that's on me, I overlooked the fact that active_adapters() is a property method in PEFT: https://github.com/huggingface/peft/blob/2558dd872dbacd29f6b60668bfee54b594bcbd7f/src/peft/tuners/tuners_utils.py#L171
Please yes , feel free to open a PR so that we can merge it ASAP and land the fix before the release

pashminacameron commented

Thank you. It's up.

  • ❤️1