Hacker News new | ask | show | jobs
by verytrivial 131 days ago
Tldr: 5% - 17% speedup due to removing a bottleneck by juggling where on a GPU/compute core a computation is done during Flash attention.