Config kwargs

[zucchini-nlp]

What does this PR do?

Decouples kwargs manipulation from hub's strict decorator, and ensures that all subclasses of a PreTrainedConfig accept any kwargs which is what we supported prev.

Not all remote code has @strict or has an __init__ defined and therefore can fail to accept arbitrary kwargs. For example, this should be a valid way to define a new config

Note that the init will be wrapped only if the subclass hasn't yet defined its own __init__. In case when subclass has a custom __init__, we assume that the user wants full control and don't wrap it. Otherwise we can cause issues because the wrapper doesn't call original_init, instead it sets attributes for known dataclass fields

from transformers import LlamaConfig

class NewLlamaConfig(LlamaConfig):
    model_type = "new_llama"

config = NewLlamaConfig(foo=1, **any_kwargs) # note the foo attr
model = MyModel(config)

cc @hmellor

[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]

LGTM

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: afmoe, aimv2, albert, align, altclip, apertus, arcee, aria, audio_spectrogram_transformer, audioflamingo3, autoformer, aya_vision, bamba, bark, bart, beit

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=44953&sha=c445b9