Made some adjust for the code in peft and gptq for llama, and make it possible for lora finetuning with a 4 bits base model. The same adjustment can be made for 2, 3 and 8 bits.
~~Still numerically unstable.~~ Resolved.
Reconstruct fp16 matrix from 4bit data and call torch.matmul largely increased the inference speed.
Added install script for windows and linux.

Requirements

gptq-for-llama: https://github.com/qwopqwop200/GPTQ-for-LLaMa
peft: https://github.com/huggingface/peft.git

Install

~~copy files from GPTQ-for-LLaMa into GPTQ-for-LLaMa path and re-compile cuda extension~~
~~copy files from peft/tuners/lora.py to peft path, replace it~~

Linux:

./install.sh

Windows:

./install.bat

Finetune

~~The same finetune script from https://github.com/tloen/alpaca-lora can be used.~~

After installation, this script can be used:

python finetune.py

Inference

After installation, this script can be used:

python inference.py