Reproduce “DoSA: Dots 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
Fanxu Meng, Taosong Fang
Choosing past tokens makes long-context sparse attention costly when every key must be scored. DoSA uses a seed of already-scored tokens to set a cutoff, then bounds candidates and fully scores only those not ruled out. At 1M tokens, DeepSeek-V3.2 TTFT is 1.94× faster than SGLang and 1.50× faster than LiteTopK; this measures serving speed, not accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.