let's see if deep-ep can be ported to 4-gpu nodes.
Alexander Doria@Dorialexander
too hot to get out, time to rewrite kernels
7:23 AM · Jun 27, 2026 · 937 Views
PyTorch torchtitan pull request ports Deep-EP to 4-GPU nodes for dropless Mixture-of-Experts dispatching
The CUDA-graphable implementation is optimized for full recomputation.
012081.1K
Sentiment
Sentiment building, check back later.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS195BOOKMARKS7LIKES6
Edward Z. Yang@ezyang
@Dorialexander Just leaving this here https://github.com/pytorch/torchtitan/pull/3561
Alexander Doria@Dorialexander
let's see if deep-ep can be ported to 4-gpu nodes.
4hViews 195Likes 6Bookmarks 7