GPTQ reference implementation (IST-DASLab/gptq GitHub repository README)

Code repository. Confirms GPTQ is a weight-only method for the OPT and BLOOM families at 2/3/4 bits with weight grouping, ships a 3-bit matrix-vector CUDA kernel, and quantizes columns in order of decreasing activation size (--act-order). Nothing in the README or paper quantizes activations. Wiki2 PPL table for LLaMa-7B: FP16 5.68, 4bit-GPTQ 6.09, 3bit-GPTQ 8.07, 3bit group-128 GPTQ 6.61 - i.e. group size is what makes 3-bit usable, not the algorithm alone. | SOURCE AUDIT 2026-10-03: url was https://arxiv.org/abs/github-ist-daslab-gptq, an unexpanded template that arXiv returns 404 for; kind was 'paper' although the document is a code repository. Repointed to https://github.com/ist-daslab/gptq (confirmed live 2026-10-03 via gh api: 'Code for the ICLR 2023 paper GPTQ...', 2381 stars) and kind set to repo.