Reproduce “HADVQ: Hessian Aware Directional 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

Daniil Li, Mars Davletshin, Vadim Klyzhenko, Artem Danilishin, Valery Sandul, Daniil Ryazanov, Ekaterina Muravleva

HADVQ is a training-free, weight-only post-training quantization method that compresses pretrained weights without retraining. It separates each block’s direction and magnitude, then uses proxy-Hessian scores to preserve 1% of estimated sensitive input channels at higher precision. On Llama2-7B, it scores 70.95% on BoolQ at 3.41 effective bits per parameter, versus 71.10% for FP16; this single-task result does not establish broader retention or faster inference.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.