Trainer / Core : Do not change init signature order

[younesbelkada]

What does this PR do?

Currently the TRL CI is broken on transformers main due to the addition of image_processors in the middle of trainer's init signature 🤯

On TRL (and probably other libs such as axolotl), we do subclass the trainer and call:

        super().__init__(
            model,
            args,
            data_collator,
            train_dataset,
            eval_dataset,
            tokenizer,
            model_init,
            compute_metrics,
            callbacks,
            optimizers,
            preprocess_logits_for_metrics,
        )

Without explicit pos arguments

#29896

To be on the safe zone I propose to simply put image_processors at the very end of trainer's init

cc @amyeroberts @NielsRogge

[amyeroberts]

We can add this quick patch - which will be eventually be subsumed by #30102

So that we can make sure things don't break in the future - is there a reason the init is called with positional arguments rather than kwargs?

[younesbelkada]

good point, I will also update that in TRL ! actually there is no strong reason to not use kwargs :/
I think other libs might possibly do the same thing as what we did in TRL

[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.