Reproduce “AttentionIndex: Accelerating Learned Sparse Attention with a Causal Inverted-File Index” 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
Jeongmin Brian Park, Wen-mei Hwu
AttentionIndex reduces the cost of selecting sparse-attention keys at long contexts. It groups similar indexer keys, routes queries to promising groups, and applies the original indexer score to shortlisted keys, while maintaining indexes as prefixes grow. At one million tokens, it recovers 86.6% of exhaustive top-k keys versus HiSA’s 69.6% at comparable prefill throughput, but accuracy varies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.