MKL warning on Python 3.10 unittest linux CI

[NicolasHug]

Unit-tests on Linux CPU / tests (3.10) / linux-job is failing because import torchvision is causing the following warning:

2023-02-21T17:06:25.6104129Z Error: mkl-service + Intel(R) MKL: MKL_THREADING_LAYER=INTEL is incompatible with libgomp.so.1 library.
2023-02-21T17:06:25.6104506Z 	Try to import numpy first or set the threading layer accordingly. Set MKL_SERVICE_FORCE_INTEL to force it.

This is an issue likely due to the way the CI is setup as the issue is not reproducible in different settings e.g. #7299 (comment) or #7299 (comment)

[malfet]

I finally have a reproducer:

$ python -c 'import numpy, subprocess, sys; subprocess.check_call([sys.executable, "-c", "import torch"])'

But removing numpy makes everything work as expected. This happens because libtorch_cpu.so inside those packages, depends on MKL, but on libgomp.so, instead of libomp.so.

But it behaves the same for 1.13, so this is not a new problem.

Even simpler reproducer that affects all our conda builds (why numpy sets this environment variable is a bit unclear to me):

$ MKL_THREADING_LAYER=INTEL python -c "import torch"

[malfet]

Following fixes the test issues:

diff --git a/test/test_transforms.py b/test/test_transforms.py
index 03b385e..8ff61ce 100644
--- a/test/test_transforms.py
+++ b/test/test_transforms.py
@@ -2283,6 +2283,7 @@ def test_functional_deprecation_warning(import_statement, from_private):
         import_statement = import_statement.replace("functional", "_functional")
         source = f"""
         import warnings
+        import numpy
 
         with warnings.catch_warnings():
             warnings.simplefilter("error")
@@ -2291,6 +2292,7 @@ def test_functional_deprecation_warning(import_statement, from_private):
     else:
         source = f"""
         import pytest
+        import numpy
         with pytest.warns(UserWarning, match="removed in 0.17"):
             {import_statement}
         """

[NicolasHug]

But it behaves the same for 1.13, so this is not a new problem

But this is still a significant issue, right? If every single user on 3.10 gets a warning depending on whether or not they import numpy, this seems like a pretty bad UX?

[malfet]

But it behaves the same for 1.13, so this is not a new problem

But this is still a significant issue, right? If every single user on 3.10 gets a warning depending on whether or not they import numpy, this seems like a pretty bad UX?

Indeed, it should be fixed, though it would prefer to understand what is causing this as well. As in the same environment, both import numpy; import torch; and import torch; import numpy work fine