Reproduce “LISA: Head-Fused Linear Indexing for Efficient Sparse Attention” 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
Zhixin Pan, Fanxu Meng, Zhaohui Wang, Taosong Fang, Muhan Zhang
LISA replaces DSA’s multi-head full-context token scoring with one fused-query score per token, using the linear term of an exact decomposition while omitting its nonlinear correction. At a 1M-token input, it reduced DeepSeek-V3.2’s offline time to first token from 383.16 seconds to 92.72 seconds, a 4.13× speedup. RULER scores were close to native baselines on the evaluated models and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.