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
MAINT, TST: test_svds_parameter_tol failures #20157
Comments
Not sure why I'd be different from the CI though. Best guess would be OpenBLAS version maybe, I often pull that from
|
Thanks for pointing this out. In the linked PR, I enabled this test by deleting the following: def test_svds_parameter_tol(self):
- if self.solver == 'propack':
- if not has_propack:
- pytest.skip("PROPACK not available")
- return # TODO: needs work, disabling for now This test was essentially skipped before my PR was merged, which explains why no failures were previously observed. Seems the comment I deleted still applies even though the test passes in CI. |
In case something on my end could be fixed/avoided, here are more detailed OpenBLAS specs (both version reproduce):
|
I reproduced locally with
So, can't be my specific OpenBLAS config I don't think. How strange. |
I'll note that the test says:
this is quite unhealthy, since |
I thought about this as well and tested it by hard-coding the |
I think we kinda already knew that - but either way, that's for putting in an issue rather than keeping in the test suite. We have a hard rule for that, tests should use seed random number state so the tests are deterministic. Otherwise we get these random failures in unrelated PRs or elsewhere. With tens of thousands of tests in the test suite, that isn't maintainable. I'd suggest removing the randomness, and then choosing a seed at which the more strict tolerances don't fail. And opening an issue with another seed where the accuracy is very bad. |
Building and testing against NumPy
1.26.4
ormain
I sporadically see two different types of test failures inTest_SVDS_LOBPCG::test_svds_parameter_tol
(see below the fold) on x86_64 Linux. Looks like this was just touched in gh-19855, so maybe some simple explanation related to some of the tol changes already made there. I've got thePROPACK
submodule update pulled in, and it looks like if I go back to a commit that precedes that update I can't reproduce even withpytest-repeat
and a bunch of replicates onTest_SVDS_LOBPCG
.The text was updated successfully, but these errors were encountered: