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.