[PyTorchModelHubMixin] Fix saving model with shared tensors

[NielsRogge]

This PR uses save_model instead of save_file in order to properly save shared tensors for the PyTorchModelHubMixin.

In the future, this may be replaced by save_file again when we support saving sharded checkpoints, which deduplicates shared tensors as in the Transformers library.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 6 lines in your changes are missing coverage. Please review.

Project coverage is 82.92%. Comparing base (0ab9391) to head (3cc9f71).

❗ Current head 3cc9f71 differs from pull request most recent head eb0111f. Consider uploading reports for the commit eb0111f to get more accurate results

Files	Patch %	Lines
src/huggingface_hub/hub_mixin.py	0.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2086      +/-   ##
==========================================
+ Coverage   82.90%   82.92%   +0.02%     
==========================================
  Files         102      102              
  Lines        9480     9477       -3     
==========================================
  Hits         7859     7859              
+ Misses       1621     1618       -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Wauplin]

Thanks for the PR @NielsRogge. I've added a test + used safetensors.torch.load_model to correctly load models with shared tensors. Everything should be fine now.

Regarding load_model, it does not support the device parameter at the moment. I have open a PR to fix this in safetensors directly: huggingface/safetensors#449. Will add back proper support once merged. Current workaround is to load on cpu first and then move to gpu (=> increase loading time but 🤷)

[Wauplin]

Let's get this merged :)