Open
Description
The issue is reported by compiler team (@zou3519, @oulgen) in which the compilation time seems to have a higher variance across runs. This happens on both PT2 (no compiler cache) and CacheBench dashboard, which seems to indicate an underlying problem with the runner.
One potential explanation is that the H100/A100 where these benchmarks are running is multi-tenancy. So, other jobs running in parallel on the same runner could cause this.
cc @seemethere @malfet @pytorch/pytorch-dev-infra @ZainRizvi @clee2000
Metadata
Metadata
Assignees
Labels
Type
Projects
Status
Done