Trigger weights_only=True by default for all compatible objects

[muellerzr]

What does this PR do?

Sets weights_only=True by default when using torch.save() for all compatible objects:

• Model weights
• Optimizer states
• DataLoader states
• Scheduler states

The only one that we can't do right now are the random states, since the numpy random states will raise an exception if we add the right classes for safe loading:

Traceback (most recent call last):
  File "/home/zach/work/accelerate/test.py", line 24, in <module>
    _ = load(p / "r2.pkl", weights_only=True)
  File "/home/zach/work/accelerate/src/accelerate/utils/other.py", line 245, in load
    loaded_obj = torch.load(f, map_location=map_location, **kwargs)
  File "/home/zach/miniconda3/envs/accelerate/lib/python3.10/site-packages/torch/serialization.py", line 1096, in load
    raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
 Please file an issue with the following so that we can make `weights_only=True` compatible with your use case: WeightsUnpickler error: Can only build Tensor, parameter or OrderedDict objects, but got <class 'numpy.dtypes.UInt32DType'>

Fixes # (issue)

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?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@BenjaminBossan

[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.

[BenjaminBossan]

Thanks for tackling this.

Honestly, I'm a bit surprised that by just allowing these few numpy types, everything can be loaded fine (except for the rng). I think it's super important to ensure that this will not lead to breakage.

[mikaylagawarecki]

I believe to load numpy arrays these are the GLOBALS needed (probably only a subset of the dtypes would be needed though) https://github.com/pytorch/pytorch/pull/124763/files#diff-0a602e09e0dd231bf8cf8813744ca89b6ebbbb92ef54d356379d85e51f0113f1R119-R168,

we made an explicit decision not to allowlist globals required to rebuild numpy arrays by default in torch however

[BenjaminBossan]

Thanks for the updates. A bunch of tests are failing. I added a suggestion that should fix some, but I'm not sure if it'll fix all.

Generally, if the tests indicate that this switch will work fine, I'm okay with merging. Still I'm wondering if there should not at least be a fail-safe for users to be able to switch to weights_only=False, as otherwise there is no way they can circumvent this, short of downgrading accelerate.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[BenjaminBossan]

Change LGTM, thanks.

There is a CI issue caused by a numpy import. It might be worth checking a few numpy versions (especially < 2.0 and >= 2.0) to be sure that all the numpy imports work.