Nit about model_accepts_loss_kwargs for loss

[ArthurZucker]

What does this PR do?

There was a typo in #34915, as tests were passing, did not pay attention to it. Thanks to @techkang for reporting

[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]

from transformers import Wav2Vec2Model
import inspect

sif = inspect.signature(Wav2Vec2Model.forward)
any(k.kind == inspect.Parameter.VAR_KEYWORD for k in sif.parameters.values())

Returns false. But when in the example torch, you have unwrapped_model.forward seems to have loss_kwargs
cc @muellerzr

[techkang]

This PR still confuses me as I mentioned in previes PR.
There are two ways to use num_items_in_batch to fix GA loss issue.

1. Use num_items_in_batch in loss function defined by model. In this case, model_accepts_loss_kwargs is True.
2. The model doesn't have loss function or user has self-defined loss function, which is compute_loss_func.
   However, this PR will make the second method invalid.

[ArthurZucker]

Hey!

Use num_items_in_batch in loss function defined by model. In this case, model_accepts_loss_kwargs is True.

No, model_accepts_loss_kwargs only depends on the forward_pass of the model, not the los function defined by the model. If the model accepts kwargs, we are gonna pass num_items_in_batch. Else we cannot.

The model doesn't have loss function or user has self-defined loss function, which is compute_loss_func.
However, this PR will make the second method invalid.

Even if user has a self defined loss function, if the forward pass does not support num_items_in_batch, we ought not to pass kwargs. I might be confusing with the trainer's compute_loss_func, and in that case will leave @muellerzr have a look

[muellerzr]

@ArthurZucker I disagree with 2, it was explicitly designed that way to catch this in case the model doesn't have this figured out yet and the user can/should manually do it themselves instead