Model load Regression due to Auto Safetensor conversion merged in #29390

[Qubitium]

System Info

Ubuntu 22.04
Torch 2.2.1
Latest git/head Transformers

Who can help?

@LysandreJik @ArthurZucker

Reproduction

Since #29390, our model loading will auto start safetensor conversion from non-safetensor model and attempt to submit a PR...and crashes. Huh?

Shouldn't there be a toggle for this feature and default to False instead of auto-enabled for all?

Exception in thread Thread-autoconversion:
Traceback (most recent call last):
  File "/root/miniconda3/lib/python3.11/threading.py", line 1038, in _bootstrap_inner
    self.run()
  File "/root/miniconda3/lib/python3.11/threading.py", line 975, in run
    self._target(*self._args, **self._kwargs)
  File "/root/miniconda3/lib/python3.11/site-packages/transformers/safetensors_conversion.py", line 89, in auto_conversion
    sha = get_conversion_pr_reference(api, pretrained_model_name_or_path, **cached_file_kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/root/miniconda3/lib/python3.11/site-packages/transformers/safetensors_conversion.py", line 82, in get_conversion_pr_reference
    sha = f"refs/pr/{pr.num}"
                     ^^^^^^
AttributeError: 'NoneType' object has no attribute 'num'

Expected behavior

1. Not crash and burn
2. Not auto start safetensor conversion

[LysandreJik]

Hey! This should not crash your runtime. I'll quickly revert and work on a fix.

Thanks for reporting the issue!

[LysandreJik]

It was reverted, feel free to update your clone. This call should only happen in the background with no effect to your runtime, the conversion happens server side.

This is to continue pushing safetensors forward as an alternative to raw .bin checkpoints as we see more bad actors with pure .bin checkpoints. Thanks again for your report!

[Qubitium]

Fix confirmed.