Reproduce “PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers” 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
Yanlong Chen, Yining Chen, Song Zhang, Amirhossein Habibian, Yawei Li
PrismQuant reduces activation quantization error by rotating dominant activation directions into group-constant levels represented by an asymmetric INT4 offset, using directions estimated from all calibration tokens’ uncentered second moment. The rotation optimizes energy alignment, not quantization error. On Llama-3.1-70B at W4A4KV4 and k=8, it reports WikiText-2 perplexity 3.85 and eight-task mean accuracy 72.46%, versus 4.22 and 71.36% for Hadamard; this protocol-specific result does not establish equivalence to bf16.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.