Quantization 
Enhancing GPTQv2 Format Support in vLLM: Analysis and Implementation
Deep technical analysis of GPTQv2 format limitations in vLLM, and implementation of CUDA kernel adaptations to enable efficient low-bit/asymmetric quantization inference.