New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
BUG: Test failures from BLAS-flavour testing for 1.13 in conda-forge #20472
Labels
defect
A clear bug or issue that prevents SciPy from being installed or used as expected
Comments
h-vetinari
added
the
defect
A clear bug or issue that prevents SciPy from being installed or used as expected
label
Apr 14, 2024
ev-br
added a commit
to ev-br/scipy
that referenced
this issue
Apr 14, 2024
A slight tolerance violation was reported on conda-forge in scipy#20472
ev-br
added a commit
to ev-br/scipy
that referenced
this issue
Apr 14, 2024
A small tolerance violation reported on conda-forge in scipy#20472
ev-br
added a commit
to ev-br/scipy
that referenced
this issue
Apr 14, 2024
An exact equality was reported as flaky on conda-forge in scipy#20472 The tolerance violations are of the order of fp noise (< 2e-16), and I don't think that pickling/unpickling guarantees bit-to-bit compatibility. In principle, this may invoke recalculations and those may be subject to fp noise. So I think it's OK to only require allclose(atol=eps) instead of exact equality.
#20474 has what I think are straightforward tolerance bumps in RGI and RBF. The cossin failure needs a more careful look. Off the cuff, it looks like some signs differ. So maybe it's a matter of the factorization only being unique up to consistent sign changes in the multipliers. |
ev-br
added a commit
to ev-br/scipy
that referenced
this issue
Apr 14, 2024
An exact equality was reported as flaky on conda-forge in scipy#20472 The tolerance violations are of the order of fp noise (< 2e-16), and I don't think that pickling/unpickling guarantees bit-to-bit compatibility. In principle, this may invoke recalculations and those may be subject to fp noise. So I think it's OK to only require allclose(atol=eps) instead of exact equality.
The failure in osx-64 + accelerate: minor tolerance violation in
|
tylerjereddy
pushed a commit
to tylerjereddy/scipy
that referenced
this issue
May 2, 2024
A slight tolerance violation was reported on conda-forge in scipy#20472
tylerjereddy
pushed a commit
to tylerjereddy/scipy
that referenced
this issue
May 2, 2024
A small tolerance violation reported on conda-forge in scipy#20472
tylerjereddy
pushed a commit
to tylerjereddy/scipy
that referenced
this issue
May 2, 2024
An exact equality was reported as flaky on conda-forge in scipy#20472 The tolerance violations are of the order of fp noise (< 2e-16), and I don't think that pickling/unpickling guarantees bit-to-bit compatibility. In principle, this may invoke recalculations and those may be subject to fp noise. So I think it's OK to only require allclose(atol=eps) instead of exact equality.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Collecting the various test failures (new and existing) from conda-forge/scipy-feedstock#273, I made a distinction whether something is affecting the main BLAS flavours (e.g. the conda-forge default, or other frequently used implementations like MKL), or regarding the severity of the failure (e.g. crash vs. minor tolerance violation) when opening new issues.
Here's a list of already existing or new stand-alone issues:
test_x0_equals_Mb
withbicgstab
#15533TestNNLS.test_nnls_inner_loop_case1
fails on windows with MKL #20470TestEig.test_falker
fails on windows + MKL as well as linux + aarch64 + netlib #20471A note on PyPy
On linux-64, PyPy 7.3.15 hangs completely, while on osx & win, there are merely additional test failures (compared to CPython).
Other failures
linux-64 + blis / osx + accelerate: minor tolerance violation in
test_x0_working[tfqmr]
linux-64 + MKL: single-element tolerance violation in
TestKMean.test_krandinit[numpy]
osx-64 + MKL + PyPy: in
test_cossin[True-40-12-20-float32]
Unclear why this only shows up in this combination (osx, MKL, PyPy); previously (for 1.12) also
test_cossin[False-40-12-20-float32]
failed.PyPy (everywhere):
test_dunder_ror[dok_matrix]
Fixed ✅
In #20474
{linux-64,win-64} + blis: minor (absolute) tolerance violation in
TestRegularGridInterpolator.test_derivatives
osx-64 + accelerate: tolerance violation in
TestInterpN.test_matrix_input[quintic]
osx-64 + accelerate: tolerance violation in
TestRBFInterpolatorNeighbors20.test_pickleable
(flaky)win-64 + blis: minor tolerance violation in
Test_SVDS_PROPACK.test_svds_parameter_k_which[LM-5]
The text was updated successfully, but these errors were encountered: