alora issues

[jakelorocco]

For hf models, setting the adapter causes generation requests to the base model to be modified and sets requires_grad=True. We need to:

• unset the adapter after calling the alora
• disable requires_grad (I believe)
• implement a lock for huggingface aloras so that async non-alora generation requests don't accidentally use the alora
• have aloras use model_options?

We should also make sure that the VLLM server doesn't have similar issues with the active adapter needing to be set/unset or managed between requests.

[jakelorocco]

@guicho271828, Nathan asked me to assign you. Let me know if you have any questions.