LiftQuant enables continuous bit-width quantization for LLMs

Arxiv CS.LG (machine learning)

Paper proposes continuous bit-width quantization to bridge the deployment gap where integer bit-widths (2, 3-bit) don't fit memory budgets. Method uses dimensional lifting and projection for fine-grained control.

Arxiv CS.LG (machine learning)·Liulu HeXuanAng LiuJuntao LiuTaolue FengTing LuChunsheng GanZhiyv PengYuan DuHuanrui YangYijiang LiuLi Du

7 days ago