Integrate CUDAGraphs with AOTAutograd

Open Chillee opened this issue 3 years ago • 2 comments

CUDAGraphs should work under NVFuser, so no reason to not do it.

Might be some tricky components about 1. setting it up so it runs after the actual compilation happens, and 2. handling situations where we incur a cuda sync (like dynamic shape operations).

May 20 '22 00:05 Chillee

cc: @eellison (this issue could also go on the TorchDynamo repo)

May 20 '22 22:05 Chillee

https://github.com/pytorch/pytorch/pull/80566 implements this

Jul 20 '22 22:07 ezyang