All credits go to this amazing project: https://github.com/FreedomIntelligence/LLMZoo

This is the chat-instruct version

Converted with python llama.py ./chimera-7b c4 --wbits 4 --true-sequential --groupsize 128 --save chimera7b-4bit-128g.pt

It uses groupsize 128. Doesn't use act-order and got quantized with the oobabooga gpt-q branch so it works there.

Anyone need a 13b version? (Edit: Can't do it right now as I only get out of memory errors while quantizing.)