(Meta AI) LLaMA 4bit 실행 방법

LLaMA (Large Language Model Meta AI)  4bit 실행 방법
    - text-generation-webui 이용 실행

 

//-----------------------------------------------------------------------------

원본 16bit에 비해서 4bit가 성능저하가 별로 없으면서 적은 VRAM에서 실행 가능

https://rentry.org/llama-tard-v2

 

8bit

Model	VRAM Used	Minimum Total VRAM	RAM /Swap to Load
LLaMA-7B	9.2GB	10GB	24 GB
LLaMA-13B	16.3GB	20GB	32GB
LLaMA-30B	36GB	40GB	64GB
LLaMA-65B	74GB	80GB	128GB

 

//

4bit

Model	VRAM Used	Minimum Total VRAM	RAM /Swap to Load
LLaMA-7B	3.5GB	6GB	16 GB
LLaMA-13B	6.5GB	10GB	32 GB
LLaMA-30B	15.8GB	20GB	64 GB
LLaMA-65B	31.2GB	40GB	128 GB

 

//

모델 파일 크기

	원본	LLaMA-HFv2	LLaMA-HFv2 4bit
7B	12.6	12.5	3.5
13B	24.2	36.3	6.5
30B	60.6	75.7	15.7
65B	121.6	121.6	31.2

 

 

//-----------------------------------------------------------------------------
* LLaMA 변환한 모델 다운로드(4bit 포함)
https://huggingface.co/decapoda-research

    - 모델 다운로드 받는 방법
https://aituts.com/llama/
https://rentry.org/llama-tard-v2


//-----------------------------------------------------------------------------
< 설치 - 윈도우 WSL 환경> 

    - python 가상 환경 설치
conda create -n textgen python=3.10
conda activate textgen

    - pytorch 설치
conda install cuda pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia/label/cuda-11.7.0
        - pytorch v2.0 이 설치되는데 정상 작동

    - pytorch 설치 확인 테스트
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"



    - text-generation-webui 다운로드
git clone https://github.com/oobabooga/text-generation-webui.git
cd text-generation-webui


    - GPTQ-for-LLaMa 다운로드
md repositories
cd repositories
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git

    - text-generation-webui 필요 패키지 설치
cd ..
pip install -r requirements.txt


//-------------------------------------
    - 패키지 설치
conda install ninja
pip install chardet

pip install cchardet
    - 에러 메시지
error: subprocess-exited-with-error
...
gcc: fatal error: cannot execute ‘cc1plus’: execvp: No such file or directory

    - 해결 방법
sudo apt-get install -y g++ build-essential


//-------------------------------------
    - GPTQ-for-LLaMa 설정, 설치
cd repositories/GPTQ-for-LLaMa
export DISTUTILS_USE_SDK=1

pip install -r requirements.txt
python setup_cuda.py install

    - 에러 메시지
subprocess.CalledProcessError: Command '['which', 'g++']' returned non-zero exit status 1.

    - 해결 방법
sudo apt-get install build-essential


//-------------------------------------
실행
python server.py --model llama-7b-4bit --gptq-bits 4

    - 에러 메시지
ValueError: Tokenizer class LLaMATokenizer does not exist or is not currently imported.

    - 해결 방법
\text-generation-webui\models\llama-7b-4bit\tokenizer_config.json 파일 수정
LLaMATokenizer  ->  LlamaTokenizer



//-----------------------------------------------------------------------------
< 에러 해결 >
    - 에러 메시지
GPTQ_loader.py", line 55,
TypeError: load_quant() missing 1 required positional argument: 'groupsize'

    - 해결 방법 : GPTQ-for-LLaMa의 변경사항을 text-generation-webui이 반영 안한 상태. GPTQ-for-LLaMa의  git commit 을 예전으로 변경

GPTQ-for-LLaMa 폴더로 이동
git checkout 468c47c01b4fe370616747b6d69a2d3f48bab5e4



//-----------------------------------------------------------------------------
< 참고 >
How to Run a ChatGPT Alternative on Your Local PC
https://www.tomshardware.com/news/running-your-own-chatbot-on-a-single-gpu

    - 윈도우 환경이라면
Visual Studio 2019 빌드 도구 설치
https://learn.microsoft.com/en-us/visualstudio/releases/2019/release-notes
- 다운로드후 C++ 만 설치

//-------------------------------------
https://www.reddit.com/r/LocalLLaMA/comments/11o6o3f/how_to_install_llama_8bit_and_4bit/