5 ms·
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,
by edwardjhu 3y ago
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!