[Core] Set weights_only=True nicely when we use torch.load().

[sayakpaul]

What does this PR do?

See internal discussion: https://pytorch.slack.com/archives/CN963QVJB/p1722612785082449.

Can propagate this PR to examples/ and scripts/ after I have an approval.

[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.

[sayakpaul]

@yiyixuxu WDYT?

[yiyixuxu]

can we wrap the torch_load + weights_only into a function and use it everywhere?

[yiyixuxu]

actually, we already have one, no?

diffusers/src/diffusers/models/model_loading_utils.py

Line 98 in 5d476f5

def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[str] = None):

[sayakpaul]

That could work! Will update the PR.

[asomoza]

just want to bring that there some users that even though it's not recommended and risky, still want to load checkpoints with the weights_only=False option. For example: #8874 #9154

IMO it should be the default but users should be able to change it, I don't have a strong opinion about this just bringing it up to attention.

[sayakpaul]

I think better to wait for Dhruv to come back then.

[sayakpaul]

@DN6 cc I think this is relevant.

[mikaylagawarecki]

Hey, any update on this, fyi the BC-breaking default flip is coming in torch 2.6 https://dev-discuss.pytorch.org/t/bc-breaking-change-torch-load-is-being-flipped-to-use-weights-only-true-by-default-in-the-nightlies-after-137602/2573 want to make sure everything works smoothly

Is there anything I can do to help push this along? cc @sayakpaul @DN6 @yiyixuxu

[yiyixuxu]

hi @mikaylagawarecki
I think we've flipped the default long time ago #7393

cc @DN6 to confirm

[DN6]

Hi. Yes, we flipped it because of the security issue. We don't allow setting weights_only=True when loading Diffusers models anymore.

diffusers/src/diffusers/models/model_loading_utils.py

Line 144 in 76b7d86

weights_only_kwarg = {"weights_only": True} if is_torch_version(">=", "1.13") else {}

Users would have to either convert to safetensors or deserialise those checkpoints on their own