peft model inference so slow!!

[LIU Jiarun]

As shown, I tried set `load_in_8bit=False` or set `model = model.merge_and_unload()`, but neither work. I mean it can output result like in 2000 years later SO is there a solution yet??