awq

Cyankiwi introduced an updated version of their AWQ 4-bit quantization method that jointly optimizes scales and quantization ranges, achieving lower KL divergence than existing methods on Llama-3 models.