Converted with https://github.com/qwopqwop200/GPTQ-for-LLaMa All models tested on A100-80G *Conversion may require lot of RAM, LLaMA-7b takes ~12 GB, 13b around 21 GB, 30b around 62 and 65b takes more than 120 GB of RAM.
Installation instructions as mentioned in above repo:
- Install Anaconda and create a venv with python 3.8
- Install pytorch(tested with torch-1.13-cu116)
- Install Transformers library (you'll need the latest transformers with this PR : https://github.com/huggingface/transformers/pull/21955 ).
- Install sentencepiece from pip
- Run python cuda_setup.py install in venv
- You can either convert the llama models yourself with the instructions from GPTQ-for-llama repo
- or directly use these weights by individually downloading them following these instructions (https://huggingface.co./docs/huggingface_hub/guides/download)
- Profit!
- Best results are obtained by putting a repetition_penalty(~1/0.85),temperature=0.7 in model.generate() for most LLaMA models