Training technique that splits model layers across devices and streams microbatches.