Reproduce “UniSVQ: 2-bit Unified Scalar-Vector Quantization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoyu Wang, Haiyan Zhao, Xingyu Yu, Zhangyang Yao, Xu Han, Zhiyuan Liu, Maosong Sun
UniSVQ compresses language-model weights to two bits by mapping integer groups through an affine transform, avoiding an explicit vector-quantization codebook. It combines randomized Hadamard processing, Hessian-based error compensation, and data-driven grid tuning. On Llama-3-8B, it reaches 101.65 tokens per second versus FP16’s 60.38, using 3.87 versus 15.72 GB peak memory; this is a weight-only result from one custom-kernel setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.