DoSA sparse attention claims lossless 1.94× TTFT speedup
TL;DR
- DoSA reports up to 2.74× indexer speedup over SGLang and 1.69× over LiteTopK across English, Chinese, and code workloads.
- The authors describe DoSA as lossless because a bounds-based certificate preserves exact Top-K token selection rather than approximating it.
- End-to-end time-to-first-token speedups reach up to 1.94× and 1.50× over the two baselines on GLM-5.2 and DeepSeek-V3.2.
DoSA, a sparse-attention method from Fanxu Meng and Taosong Fang, reports up to 2.74 times indexer speedup over SGLang and 1.69 times over LiteTopK while, by construction, preserving the same Top-K token selection as exact attention. The authors describe it as "a lossless method for making this computation conditional while preserving exact token selection."
The mechanism is an early-rejection certificate. "Top-K selection requires sufficient evidence to reject a candidate, not necessarily its full score," the abstract argues. A seed set supplies a lower bound on the winning threshold; bounds on positive and negative head contributions let the method skip scoring candidates that can already be ruled out, and send full scoring only where a candidate remains eligible.
End-to-end, DoSA posts "up to 1.94 times and 1.50 times time-to-first-token speedups over the two baselines, respectively" on GLM-5.2 and DeepSeek-V3.2, across English, Chinese, and code workloads. The abstract reports no accuracy comparisons, because under the lossless claim the selected tokens should be identical to exact Top-K.
Originally reported by paper
Read the original article →Original headline: Lossless Sparse Attention DoSA Cuts TTFT 1.94× Without Approximation