Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
Batch<N+1> -> PrefetchedEmbeddings<N+1>
Prefetch and route sparse embedding features for the next training iteration while simultaneously executing dense neural network layers for the current iteration.
Problem it solves
Embedding lookup and network communication stall the accelerator's compute units during large-scale recommendation training.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.