Y
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.