Reproduce “SQuAT: Spherical Lattice Vector Quantization-Aware Training” 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
Sebastian Loeschcke, Ivan Koryakovskiy, Tycho F. A. van der Ouderaa, Riccardo Del Chiaro, Mart van Baalen, Paul N. Whatmough, Markus Nagel
SQuAT improves low-bit language-model quantization-aware training by separating each weight block’s normalized direction from its magnitude, quantizing them independently, and applying separate gradient updates. Its exact-assignment implementation reduces E8S training time from 62.5 to 4.9 seconds per iteration. On instruction-tuned LLaMA-3.2-1B at 2.25 bits per weight, it improves WikiText2 perplexity and accuracy scores over Standard VQ-QAT; gains are not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.