Huy Do

cc @voznesenskym because I suspect that the flaky behavior is coming from https://github.com/pytorch/pytorch/pull/91821 The failures all look like this ``` 2023-01-14T18:15:41.2170660Z ====================================================================== 2023-01-14T18:15:41.2170810Z ERROR [0.008s]: test_torch_size_dynamic_shapes (torch._dynamo.testing.make_test_cls_with_patches..DummyTestClass) 2023-01-14T18:15:41.2170950Z ---------------------------------------------------------------------- 2023-01-14T18:15:41.2171020Z...

I'm a bit closer to debug this. The test itself `dynamo/test_dynamic_shapes` is not flaky and all passes when I try to run it locally. However, this test is set to...

This also fails on Linux sometimes https://hud.pytorch.org/failure/test_basicmodule2_dynamic_shapes%2Ctorch._dynamo.testing.make_test_cls_with_patches.%3Clocals%3E.DummyTestClass

As @clee2000 points out, the test is also flaky when running in `mem_leak_check` mode which runs everything sequentially, i.e. https://hud.pytorch.org/pytorch/pytorch/commit/5e0d3458eb58d21081f64d6a2347c5462453c2da, so https://github.com/pytorch/pytorch/pull/92215 is probably a red-herring and won't fix the...

I can finally reproduce the issue reliably now with the follow snippet based on `test_dynamic_shapes` ``` from torch import package from torch._dynamo.testing import make_test_cls_with_patches try: from . import ( test_misc,...

> So why exactly does that make it flaky? To my best knowledge, it should not be flaky, but always fails when the line `self.scope["__builtins__"].__dict__[name] = package_module` is run. It...

@pytorchbot revert -m 'Sorry for reverting your PR. We reverted https://github.com/pytorch/pytorch/pull/70988 in https://hud.pytorch.org/pytorch/pytorch/commit/acdd462b1a070790799ce4623ce8ecc83e197e81 and this test starts to fail. There is probably a dependency between the twos' -c weird

Please do a rebase to master and rerun the tests

Per my findings in https://github.com/pytorch/pytorch/pull/93020, I think the issue is fixed by running dynamo tests on Python 3.8, now available after https://github.com/pytorch/pytorch/pull/92928 (thank @malfet), so I will revert https://hud.pytorch.org/pytorch/pytorch/commit/0fc2f9febb8147183bcf8321ea80ab8e48ced875 to...

@pytorchbot rebase -b master