Preferences

zhihaojia parent
You are right that CUDA graph can help reduce launch overhead but does not support overlapping computation/communication across layers, since data dependencies are described at the kernel level.

This item has no comments currently.