Segmentation fault in JIT tests on Windows CPU

[pmeier]

For example: https://github.com/pytorch/vision/actions/runs/5519380703/jobs/10067038151#step:9:964

Looking at the HUD, I'm seeing the first occurrence of this around June 28 nightlies.

However, the job is flaky and passes from time to time. Chance of passing is different for the same job on different Python versions. Not sure if that actually plays a role, but it underlines the flakiness.

• A passing job from the same run as above but with a different Python version: https://github.com/pytorch/vision/actions/runs/5519380703/jobs/10067040388
• Different run than above, but same job passing there: https://github.com/pytorch/vision/actions/runs/5499611221/jobs/10021935529

cc @seemethere

[pmeier]

Commits that could have caused this are here: https://github.com/pytorch/pytorch/commits/main?after=a8b63d4d1b138aee2de1e6e9d9aaed7e5ae80bac+0

The only one that has JIT in the PR name is pytorch/pytorch#102510. But as far as I see, this is only touching tracing and thus shouldn't affect us.