Add push_to_hub: safe_serialization=False

[zhiying318]

In terms of saving the models, I have faced the same issue with safe serialization of tensors, as mentioned in #42

I also suggest changing to this:

model.push_to_hub(cfg.checkpoint_id, safe_serialization=False)
processor.push_to_hub(cfg.checkpoint_id, safe_serialization=False)

[sergiopaniego]

Thanks for opening the PR! Could you provide more details about the issue faced? 😄

[zhiying318]

Hi, so here's a more detailed description:

I meet RunTimeError if I simply run model.push_to_hub(cfg.checkpoint_id)

Traceback (most recent call last):
  File "/workspace/train.py", line 162, in <module>
    model.push_to_hub(cfg.checkpoint_id)
  File "/workspace/.venv/lib/python3.10/site-packages/transformers/modeling_utils.py", line 3994, in push_to_hub
    return super().push_to_hub(*args, **kwargs)
  File "/workspace/.venv/lib/python3.10/site-packages/transformers/utils/hub.py", line 970, in push_to_hub
    self.save_pretrained(work_dir, max_shard_size=max_shard_size, safe_serialization=safe_serialization)
  File "/workspace/.venv/lib/python3.10/site-packages/transformers/modeling_utils.py", line 3941, in save_pretrained
    safe_save_file(shard, os.path.join(save_directory, shard_file), metadata={"format": "pt"})
  File "/workspace/.venv/lib/python3.10/site-packages/safetensors/torch.py", line 286, in save_file
    serialize_file(_flatten(tensors), filename, metadata=metadata)
  File "/workspace/.venv/lib/python3.10/site-packages/safetensors/torch.py", line 488, in _flatten
    raise RuntimeError(
RuntimeError: 
            Some tensors share memory, this will lead to duplicate memory on disk and potential differences when loading them again: [{'language_model.model.embed_tokens.weight', 'language_model.lm_head.weight'}].
            A potential way to correctly save your model is to use `save_model`.
            More information at https://huggingface.co/docs/safetensors/torch_shared_tensors
            

The default case does not allow shared memory tensors to be saved. So adding safe_serialization=False can solve this problem.

[sergiopaniego]

Thanks for sharing the trace.
We'd like to save the model as default using safe_serialization=True since that would imply saving it using .safetensors which is the default format in transformers. If we follow the approach of safe_serialization=False, the saved model would be saved using .bin.

I'd try to avoid this solution and find another way to do so. What the error says is that those layers share memory. Instead, it could be approached by cloning the problematic tensor.

I've opened a PR #49 addressing this error.

[sergiopaniego]

You can also see another example where safe_serialization=True in the official Google's fine tuning tutorial.