transformers

ecf79eb2 - [generate] Completely stop relying on `cache_position` to prepare inputs (#44130)

[generate] Completely stop relying on `cache_position` to prepare inputs (#44130) * generalize * better doc * don't mixup the logic everywhere * doc * fix all * no unsqueeze * more fixes * add position_ids * simplify all a lot * doc * small oupsi * more small fixes * more fixes * fix assisted decoding * slice * fix * remove outdated test * fix openai * fix mamba2 * fix xlm * fixes * fix * fix rag * can restart from sliced inputs with a full mask * fix test * skip whisper test for now * remove tests for deprecated hub-specific generation generations * fix llava test - not related to PR but was failing... * style