Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
(TokenTensors, ActiveRankMask) -> DispatchedTensors
Route MoE tokens only to healthy expert ranks by filtering expert-parallel dispatch mappings against a dynamic active-rank mask.
Problem it solves
Single rank failures in MoE expert-parallel environments block or crash the entire distributed inference service.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.