Reproduce “A Unified Hessian-Guided Framework for Post-Training 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
Mads Toftrup, Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel
HZ-PTQ guides post-training quantization of LLMs. It approximates output-KL curvature with a Kronecker-factorized Fisher matrix, then applies the proxy in staged choices for transformations, weight ranges and rounding, and activation clipping. On Llama v3.2 1B Instruct, it reports WikiText perplexity 16.05 and six-task accuracy 54.21, outperforming FlatQuant+GPTQ in that setting; improvements are not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.