[edwardjhu]

Yes, the plan is to keep the original weights in VRAM and merge/unmerge LoRA weights on the fly. You can even cache a large library of LoRA ckpts in RAM.

Yup, I am!