Repository navigation
SciPy tests fail on 0.3.30.359 #242
Description
When trying out SciPy on the wheels produced by #241 on a ryzen x86_64 Ubuntu 24.04 machine, I see 5 failures:
=============================================================================== short test summary info ===============================================================================
FAILED scipy/interpolate/tests/test_interpolate.py::TestPPoly::test_roots_croots - AssertionError:
FAILED scipy/linalg/tests/test_solvers.py::TestSolveContinuousAre::test_solve_continuous_are[13-case13] - ValueError: Reordering of (A, B) failed because the transformed matrix pair (A, B) would be too far from generalized Schur form; the problem is very ill-conditioned. (A, B) may h...
FAILED scipy/linalg/tests/test_solvers.py::TestSolveContinuousAre::test_solve_continuous_are[17-case17] - AssertionError:
FAILED scipy/linalg/tests/test_solvers.py::TestSolveDiscreteAre::test_solve_discrete_are[17-case17] - AssertionError:
FAILED scipy/linalg/tests/test_solvers.py::TestSolveDiscreteAre::test_solve_discrete_are[18-case18] - ValueError: Reordering of (A, B) failed because the transformed matrix pair (A, B) would be too far from generalized Schur form; the problem is very ill-conditioned. (A, B) may h...
I also see these same failures when using the latest "nightly" wheels on anaconda.org:
python -m pip install -i https://pypi.anaconda.org/scientific-python-nightly-wheels/simple scipy-openblas32
spin build --with-scipy-openblas
spin test
@rgommers is it typical that thresholds need adjustment with newer OpenBLAS or is there a problem somewhere that needs to be investigated?
Activity
rgommers commentedon Dec 12, 2025
That's fairly rare. Small bumps do happen on occasion for all BLAS libraries, but we should probably investigate this since there are multiple failures.
mattip commentedon Dec 12, 2025
Maybe i should add to the scheduled build to also test it with scipy, so we can catch these failures earlier
mattip commentedon Dec 13, 2025
git bisectsays it is commit e105411460 from OpenBLAS, which is OpenMathLib/OpenBLAS#5381. This added new interfaces and may have affected accuracy. I will ask on OpenBLAS if this is to be expected.mattip commentedon Dec 15, 2025
Fixed by OpenMathLib/OpenBLAS#5571, closing since the problem was found