Option to hard fail for cudagraphs and/or more cudagraph logging #124506
Labels
module: cuda graphs
Ability to capture and then replay streams of CUDA kernels
oncall: pt2
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
馃殌 The feature, motivation and pitch
We lack a way to know how much of the torch.compiled model has been cudagraphed since we silently fallback ("skipping cudagraphs due to"). The debugging workflow today relies on grepping through logs for cudagraph logs and matching them against graph id logs.
Something like a hard fail config for users who want to ensure their model is always fully cudagraphed, or logging a final cudagraph report can help.
Alternatives
No response
Additional context
No response
cc @mcarilli @ezyang @eellison @peterbell10 @msaroufim @bdhirsh @anijain2305 @chauhang @BoyuanFeng
The text was updated successfully, but these errors were encountered: