Reproduce “SinkState: On-the-Fly SVD of Recurrent States for Efficient Hybrid LLM Serving” 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
Haolin Jiang, Weihang Li, Zhongzhu Zhou, Jinwen Pan, Santosh Pandey, Hang Liu
SinkState uses adaptive low-rank factors to compress recurrent states in linear-attention models during decoding. It initializes factors with randomized SVD, buffers updates, and periodically compacts active directions while asynchronously consolidating frozen ones on CPU. The authors report average peak state memory at 0.52 of dense baselines and a 1.58× speedup on one MATH-500 test, with accuracy 0.7 points lower there.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.