Reproduce “Quantizing Recurrent Language Models Is Easier Than You Think” 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
Philipp Nazari, Neehal Tumma, Daniela Rus, T. Konstantin Rusch
LoReS quantizes recurrent-model state during decoding by grouping scales along value rows, assigning more bits to slowly decaying components, and using stochastic rounding to limit systematic error reinforcement. On one H100, Nemotron-H at batch size 256 measured 17.32 ms TPOT with INT8 state versus 29.35 ms with 32-bit state, in this specified serving test.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.