torch-mlir icon indicating copy to clipboard operation
torch-mlir copied to clipboard

[brainstorm] What could cause ONE e2e test to fail with packages but not locally

Open silvasean opened this issue 2 years ago • 2 comments

In IREE-Torch, there is one test that segfaults recently. The segfault is on memory access at address 0 before it falls off into ud2's in printType. Any ideas about build config/etc. that could explain the failure?

https://github.com/iree-org/iree-torch/pull/51

cc @ashay @powderluv

silvasean avatar Oct 04 '22 15:10 silvasean

This is just a shot in the dark.

Last night, I learnt the hard way that the VMs used by CI may have files from prior CI runs (or so it seems). Could this be a case of a non-hermetic build system like CMake being affected by prior builds?

ashay avatar Oct 04 '22 15:10 ashay

Thanks Ashay. In this case, I am able to reproduce locally with the installed packages, so files from prior runs are unlikely to be an issue.

silvasean avatar Oct 05 '22 13:10 silvasean