Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
DynamicRoutingMask, Activations -> RoutedExpertOutputs
Dynamically adjust expert capacity and parallel routing configurations at runtime without performance penalties or tensor padding.
Problem it solves
Static expert capacity allocations cause tensor padding overhead or token drop when token-to-expert distributions shift dynamically.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.