tracker: move prepare_inputs_for_generation into the generation mixin 🧹

[gante]

🧹 This is a tracker regarding the move of prepare_inputs_for_generation into the generation mixin 🧹

Why?

1. prepare_inputs_for_generation is not part of the core modeling, but rather a utility for generate
2. it should greatly reduce the need to touch modeling code, on generate changes. Fewer modeling changes -> improved model stability
3. greatly reduced number of lines of code 🙏

Tracker

Ordered list of tasks:

• Fix related slow tests before we start — all llama, generate, and cache_utils [except sink cache, broken atm] slow tests should be passing to ensure we don’t break anything (Llama: make slow tests green 🟢  #33138)
• PreTrainedModel doesn't inherit from GenerationMixin, so that can_generate() becomes independent of prepare_inputs_for_generation being overwritten or not (Generation: deprecate PreTrainedModel inheriting from GenerationMixin  #33203 )
• Move llama’s prepare_inputs_for_generation to the generation mixin. This implies moving one function that prepares the 4D mask too (the one that is called there)
• Add tests for the generalist prepare_inputs_for_generation — currently we don’t test it directly, and we should
• Address the case of synced_gpus in generate: when synced_gpus and cache_positions is out of bounds, take the latest available input_ids for dummy computations (see Fix synced GPUs #33252; should fix Multi GPU generate with llama shape error #32885, Shape mismatch when generating with multiple processes #32603, and Bugfix for generation with an early-stopping process #32641)
• Delete prepare_inputs_for_generation from as many models as possible. There may be merge conflicts here, due to the 4D mask function. Try to iron out as many trivial cases as possible
• Change prepare_inputs_for_generation to forward **kwargs from its input to its output. With minimal changes, this should enable most VLMs to use the shared function (they forward pixel_values from the input to the output)
• By this point most cases of prepare_inputs_for_generation should be removed 🤗 We would need to check the others individually, there may be further simplification patterns available!

[gante]

@ydshieh edit the tracker above as soon as you start working on a task, so we don't risk doing redundant work 🤗 (e.g. with the link to a draft PR)

I'll do the same!

[ydshieh]

Thanks