Torch.compile page here mentions a cudagraphs backend. If I understand correctly, inductor also uses cudagraphs and aotautograd: PyTorch CI HUD. What is the cudagraphs backend and how is it different from the inductor backend with cudagraphs?
Is it that inductor backend with cudagraphs emits triton code but just cudagraphs doesn’t to triton on gpus? What code does the cudagraphs backend emit? TORCH_COMPILE_DEBUG doesn’t show an output file for cudagraphs backend.
A similar question was asked here but wasn’t clarified yet.