Back to homepage

Blog

Engineering, reflections, and reviews.

RSS
Writing activity
4 posts across 2 active months · last 12 months
  1. August 2025: 0 posts
  2. September 2025: 0 posts
  3. October 2025: 0 posts
  4. November 2025: 0 posts
  5. December 2025: 0 posts
  6. January 2026: 0 posts
  7. February 2026: 0 posts
  8. March 2026: 0 posts
  9. April 2026: 0 posts
  10. May 2026: 0 posts
  11. June 2026: 1 post
  12. July 2026: 3 posts
Less More
Engineering 10 min read
DeepSeek Sparse Attention:从 SWA 到稀疏 Attention 计算

从 SWA 的固定窗口出发,解释 Lightning Indexer 如何为完整前缀打分并选出 top-k、融合式 Sparse Attention 算子如何按索引读取 KV,以及 DSA 如何降低 Attention 的计算量与显存读取量。

DeepSeek Sparse AttentionSWAAttentionTransformer