This repository implements the GPTQ algorithm from the ICLR 2023 paper, providing efficient quantization of generative pretrained transformers. It includes implementations for compressing OPT and BLOOM model families, custom CUDA kernels for accelerated 3-bit matrix-vector products, and evaluation tools for measuring perplexity and ZeroShot performance on quantized models. The work also supports LLaMa models with techniques like activation-ordering and true-sequential quantization.