Add support for LoRA finetuning for LLaMA-65B

[YingHH1]

Since there is no Vicuna-65B, is it possible to use LoRA to finetune the LLaMA-65B base model whilst simultaneously full-finetuning the Medusa heads?

[ctlllll]

Hi,

Definitely! It's a good idea to train such a model with native Medusa heads, we list it in our Roadmap, welcome to submit a PR if you'd like to help with this :)