Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
ExpertActivations -> DispatchedExpertExecution
Dynamically route MoE expert execution to CPU or GPU based on activation hotness and device memory layout.
Problem it solves
MoE models are too large for GPU VRAM, but executing all experts on CPU causes severe performance bottlenecks.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.