---
format: "aidr-story-markdown/v1"
id: "a558c5da79a699bf9822eb8dfb3e6342a0e05fb8dbb57dfd9d0a75f6953b3b73"
canonical_url: "https://aidr.today/a558c5da?lang=en"
title: "DoSA: Dots Sparse Attention"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-14T00:00:00.000Z"
category: "Research"
topics: ["inference"]
source_urls: ["https://huggingface.co/papers/2610.13990"]
summary: "DoSA (Dots Sparse Attention) is a lossless method that makes sparse attention computation conditional by using a seed set and bounds to direct full scoring toward eligible candidates, accelerating the Indexer by up to 2.74 times over SGLang and 1.69 times over LiteTopK."
---

# DoSA: Dots Sparse Attention

> [Open the canonical story](<https://aidr.today/a558c5da?lang=en>)

**Published:** 2026-10-14T00:00:00.000Z
**Category:** Research
**Topics:** inference

## Summary

DoSA \(Dots Sparse Attention\) is a lossless method that makes sparse attention computation conditional by using a seed set and bounds to direct full scoring toward eligible candidates, accelerating the Indexer by up to 2\.74 times over SGLang and 1\.69 times over LiteTopK\.

## Sources

- [Story source](<https://huggingface.co/papers/2610.13990>)