Is the use of CUDA Graph Python API discouraged?

torch.compile
1

Hi,

I’m currently using both the CUDA Graph API and torch.compile() in my source.

I’m aware that torch.compile also provides an option to automatically use CUDA graph based on this post.

Question: It is fair to say that the manual construction of CUDA Graph using the aforementioned API is discouraged then *based on this post? I couldn’t find any real world example (other than the ones in CUDA Graph API) I find it quite challenging to debug as the only debug option is to use CUDAGraph.debug_dump() and view the graph.

2

Sorry I should’ve posted this under torch.compile category. I can’t edit it anymore for some reason, though.

3

Let me move it for you.

1 Like