Closed h-vetinari closed 3 months ago
Hi! This is the friendly automated conda-forge-linting service.
I just wanted to let you know that I linted all conda-recipes in your PR (recipe
) and found it was in an excellent condition.
As expected, this essentially reproduces the state from 1.8.0rc3:
variant | blis | mkl | netlib | openblas | sum** |
---|---|---|---|---|---|
linux / x86 | ✔️ | ✔️ | ✔️ | ✔️ | - |
linux / aarch | ➖ | ➖ | ✔️ | ✔️ | - |
linux / ppc64le | ➖ | ➖ | ✖️ | ✖️ | 6F |
osx / arm | ➖ | ➖ | ✔️ | ✔️ | - |
osx / x86 | ✔️ | ✔️ | ✔️ | ✔️ | - |
win/ x86 | 💥 | 💥 | 💥 | 💥 | 12S |
sum** | 3S | 3S | 3S/3F | 3S/3F | 12S/6F |
** sum of Segfaults (S), Failures (F), resp. Timeouts (T); out of a total of 54 CI combinations being tested
This is on my radar, just swamped recently and will pick this back as soon as I can
This is on my radar, just swamped recently and will pick this back as soon as I can
No worries, there's no rush. I just wanted to open the PR for support and as a reminder for us all.
@mckib2 Given that 1.8.1 seems to be materializing, I just wanted to double-check that I haven't missed any potential updates regarding PROPACK? Guessing we retarget 1.8.2 / 1.9.0?
Hi! This is the friendly automated conda-forge-linting service.
I just wanted to let you know that I linted all conda-recipes in your PR (recipe
) and found it was in an excellent condition.
So relative to https://github.com/conda-forge/scipy-feedstock/pull/224, this has no increase of test failures, and certainly all the propack segfaults on windows are gone 🤩 🥳
Broken aarch/ppc builds have nothing to do with propack, investigating in https://github.com/conda-forge/scipy-feedstock/pull/232
OK, taking the last results of https://github.com/conda-forge/scipy-feedstock/pull/224 as a baseline (minus https://github.com/scipy/scipy/issues/18371, which got fixed in #233) and then switching on propack again, the results are as follows
variant | accelerate | blis | mkl | netlib | openblas |
---|---|---|---|---|---|
linux / x64 | ➖ | ✔️ | ✔️ | ✔️ | ✔️ |
linux / aarch64 | ➖ | ➖ | ➖ | ✔️ | ✔️ |
linux / ppc64le | ➖ | ➖ | ➖ | ✔️ | ✔️ |
osx / arm | ❔ | ➖ | ➖ | ❔ | ❔ |
osx / x64 | ✔️ | ✔️ | ❌ | ✔️ | ✔️ |
win / x64 | ➖ | ❌ | ❌ | ❌ | ❌ |
Notable:
sparse/linalg/tests/test_propack.py::test_svdp
on osx+mkl+meson+CPython (without enabling PROPACK) is tightly bounded between 11-12, now it fluctuates between 3-9 failures?!In more detail:
sparse/linalg/tests/test_propack.py::test_svdp
& sparse/linalg/tests/test_propack.py::test_examples
were shown to be due to meson in https://github.com/conda-forge/scipy-feedstock/pull/224, after backporting the latest patches we now see wildly differing errors per Python version and implementation, seemingly at random. It's also bizarre to me that only ever 15 tests fail per run, even though there's a total of ~50 that are possible sparse/linalg/tests/test_propack.py::test_svdp
& sparse/linalg/tests/test_propack.py::test_examples
only fail on CPython; PyPy only has failures for sparse/linalg/_eigen/tests/test_svds.py::Test_SVDS_PROPACK::test_svd_simple, but more of them each run.I probably know what the root cause is. It's related to the ABI wrappers, so it makes sense that it only shows up with MKL. Basically, the use of wcdotc
in PROPACK is completely wrong, and may produce arbitrary garbage. That things pass with Accelerate is just luck, I think it zero-initializes some data rather than grab random memory.
This has been fixed/re-enabled by https://github.com/scipy/scipy/pull/19855, so closing this.
The PROPACK tests were disabled in https://github.com/scipy/scipy/pull/15479 to unblock the release of 1.8.0. Still, we should fix this, so opening this PR to try to help iterate. Happy to play with any upstream commits / PRs (including submodules) here.
CC @mckib2