Skip to content

Latest commit

 

History

History
56 lines (34 loc) · 1.13 KB

File metadata and controls

56 lines (34 loc) · 1.13 KB

环境配置:

conda create -n OliVe python=3.8
conda activate OliVe

conda install pytorch=1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

cd ./olive_quantization

pip install -r requirements.txt

pip install ./quant

适配LLAMA:

配好环境后,在conda环境中更新这些包:

pip install --upgrade evaluate
pip install datasets -U
pip install --upgrade transformers==4.33
!pip install accelerate==0.20.3

运行Olive:

cd olive_quantization/llm
./scripts/run_all.sh

run_all.sh中的运行示例:可以按照实验要求手动修改

CUDA_VISIBLE_DEVICES=1 ./scripts/clm_run.sh LLAMA/llama-7b c4 realnewslike ant-int-flint 4 2 46666 outlier

其中:

LLAMA/llama-7b:是存放模型软连接的文件夹,改成opt模型的话:OPT/opt-7b

c4 realnewslike:是数据集选择,选Wikitext数据集改成:wikitext wikitext-103-raw-v1

4:是bit选择,默认是8bit,这里是4bit

2:batch_size大小

所有脚本参数的设定在 clm_run.sh 文件里 实验结果在./llm/checkpoints中 实验日志在./llm/log中