Comment by ddren

The python implementation[1] ran some tests using the same quantization algorithm as llama.cpp (4 bit RTN).

1: https://github.com/qwopqwop200/GPTQ-for-LLaMa