Fix model_accepts_loss_kwargs for timm model

[qubvel]

What does this PR do?

TimmWrapper model training is broken due to the num_items_in_batch param passed to the model forward. TimmWrapper model has **kwargs in the transformers model forward, but it goes straight to the timm model and not to the loss. This PR introduces a fix to avoid passing loss_kwargs even if **kwargs are in model forward.

Relevant to:

• Fix GA loss bugs and add unit test #35121
• [trainer] fix the GA model_accepts_loss_kwargs #34915

cc @ArthurZucker

[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.

[ArthurZucker]

Okay, I think we might want to document accepts_loss_kwargs as it's scope right now is super small.
This can't be set to True by default, we can just add a property that checks the forward signature, unless _accept_loss_kwargs is set to false.

[qubvel]

@ArthurZucker Do you mean to move it to the PretrainedModel level?

class PreTrainedModel:

    _accept_loss_kwargs = None
    
    @property
    def accepts_loss_kwargs(self):
        if self._accept_loss_kwargs is not None:
            return self._accept_loss_kwargs
        # otherwise check for signature here
        ...
        return accept_loss_kwargs

And then in modeling:

class MyModel(PreTrainedModel):
    _accept_loss_kwargs = False

[ArthurZucker]

No for now let's merge as is! Not sure we'll need the abstraction for more models!