Open ElePT opened 3 days ago
There's no need to pin Numpy (and #12676 doesn't do that) - the warning is entirely spurious and should be explicitly ignored. If, at some point in the future, numpy removes the numpy.linalg.linalg
shim to numpy.linalg._linalg
(possible) or Python's built-in unittest.TestCase.assertWarns
changes how it works (unlikely), we can remove the warning skip, but that's about it.
Environment
What is happening?
Numpy 2 made a few new modules private, and have warnings that trigger if you try to access attributes that would have existed. Unfortunately, Python's
warnings
module adds a field called__warningregistry__
to any module that triggers a warning, andunittest.TestCase.assertWarns
then queries said fields on all existing modules. On macOS ARM, we see some (we think harmless) warnings come out ofnumpy.linalg._linalg
(a now-private module) during transpilation, which means that subsequentassertWarns
calls can spuriously trick Numpy into sending out a nonsenseDeprecationWarning
.How can we reproduce the issue?
CI
What should happen?
It works
Any suggestions?
Pin versions until fix is found