Add supported kwargs to fixed_cross_entropy

[Rocketknight1]

This PR allows some specific kwargs to be passed through to fixed_cross_entropy. Fixes #43240

cc @ArthurZucker because I think you wrote the original?

[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.

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=43254&sha=f05ae7

[stas00]

After this change I'd then drop **kwargs arg, since it silently ignores them. This is not good for the user - it either should do something with kwargs and assert if the use is unexpected or it shouldn't absorb args that shouldn't be there.

There is another version of this PR here #43251

[Rocketknight1]

Hey @stas00, good to hear from you again! I suspect the **kwargs there is because the function is often called with **kwargs in other contexts, so doing it like that lets it catch any named kwargs that match while ignoring ones that don't. That makes me a little nervous about updating it, because I think those kwargs are flowing through from a lot of different possible sources.

[stas00]

I understand your reasoning, Matt

Perhaps with v5 transformers could fix the silent kwargs black holes in its code - kwargs is designed for passing forward arguments to the next function in the stack. In this situation the function isn't passing kwargs forward therefore it shouldn't have kwargs in its API.