[ Adaptation Prompt] Fix llama rotary embedding issue with transformers main

[younesbelkada]

Currently the adaptation prompt slow tests are failing on transformers main due to a recent refactor to add compatibility with static cache. huggingface/transformers#27931

That PR broke the adaptation prompt forward pass, as module.rotary_emb forward pass now have position_ids as a required argument, thus breaking the current impmelementation of llama roatary embedding as we don't pass position_ids in PEFT.

In addition, cos / sin are now directly indexed through that method, therefore we don't need to call our method that also indexes cos / sin with positon_ids. As computing the query states given cos / sin is a one-liner, I slighlty modified the logic in llama_compute_query_states to make sure the forward pass works fine with transformers main

cc @pacman100 @BenjaminBossan

[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.

[BenjaminBossan]

Thanks for fixing this, LGTM.

Too bad that the signature changed and we have to resort to inspect, but I guess there isn't really a better way.