Tencent Hunyuan proposes Stem sparse attention algorithm, accepted by ICML 2026
Original title:首字延迟降低3.6倍,腾讯混元提出Stem稀疏注意力算法,长文推理加速新SOTA
AI summary
The Stem algorithm, through two innovations—Token Position Decay (TPD) and Output-Aware Metric (OAM)—approaches dense attention accuracy with only 25% compute. The accompanying HPC operator is optimized for the Hopper architecture, supports FP8 quantization and vLLM's Paged KV Cache, and reduces first-token latency by 3.6x on Hunyuan Hy3 preview. Compared with MIT-BSA, HPC-BSA consistently maintains about 3x speedup, with consistent performance across sequence lengths from 8K to 256K.
Key Takeaways
- The Stem algorithm,through two innovations—Token Position Decay (TPD) and Output-Aware Metric…
- The accompanying HPC operator is optimized for the Hopper architecture,supports FP8 quantization…
- 6x on Hunyuan Hy3 preview.
- Compared with MIT-BSA,HPC-BSA consistently maintains about 3x speedup,with consistent performance…
Quick Facts
- Published
- Updated
- Source
- 公众号:腾讯混元
- Category
- Papers & Research
- AI score
- 62
把稀疏注意力从一刀切升级成按信息流分配预算,算子在Hopper上把理论加速几乎无损转化,3.6倍首字延迟下降不是灌水——做长上下文应用的值得照着开源代码试。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Tencent Hunyuan proposes Stem sparse attention algorithm, accepted by ICML 2026. June 5, 2026. https://new.nat6.net/en/news/cmq2o71l001iisl6n8ljx5jbt/