Fix trainer saving safetensors: metadata is None

[hiyouga]

What does this PR do?

Fixes hiyouga/LLaMA-Factory#1959

If we use Trainer to train a model that does not belong to the PreTrainedModel class, such as the PreTrainedModelwithValuehead from the TRL library, the trainer will not save the metadata. This leads to errors in reading the metadata when using AutoModelforCausalLM.from_pretrained to load the model.

transformers/src/transformers/trainer.py

Line 2911 in 29e7a1e

safetensors.torch.save_file(state_dict, os.path.join(output_dir, SAFE_WEIGHTS_NAME))

transformers/src/transformers/modeling_utils.py

Lines 3403 to 3407 in 29e7a1e

	with safe_open(resolved_archive_file, framework="pt") as f:
	metadata = f.metadata()
	if metadata.get("format") == "pt":
	pass

Although it may sound strange to load a model that does not belong to the PreTrainedModel class using AutoModelForCausalLM.from_pretrained, this approach benefits model loading by utilizing features such as low_cpu_mem_usage if the model checkpoints share the same structure.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@muellerzr @pacman100

[pacman100]

Thank you @hiyouga for these changes which are inline with what the save_pretrained does at

transformers/src/transformers/modeling_utils.py

Line 2406 in 3cefac1

safe_save_file(shard, os.path.join(save_directory, shard_file), metadata={"format": "pt"})

LGTM!

[amyeroberts]

Thanks for fixing!