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: 1.12.0rc2 prep #19797
MAINT: 1.12.0rc2 prep #19797
Conversation
…py#19735) Co-authored-by: h-vetinari <h.vetinari@gmx.com>
PchipInterpolator is nonlinear in data, hence does not guarantee that pchip(x, y) == pchip(x, y.real) + 1j*pchip(x, y.imag) The test was passing by chance.
* MSVC is too non-conformant and broken... * Also harden detection of cpp standard on MSVC platforms Co-Authored-By: H. Vetinari <h.vetinari@gmx.com>
* update the SciPy `1.12.0` release notes following more backports for the RC2 release [skip cirrus] [skip circle]
Opened a PR: #19800 |
If we want compatibility with pytest 8.0 (the just released rc1 shows up in the prerelease job), we also need the first two commits from #19806 (I want to avoid squash-merging it for reasons explained in the PR). |
One more backport candidate as it fixes a bug in my opinion: #19779 |
Ok, I'll be sure to carefully make the requested additional backports. If you think it should be backported and doesn't have the label yet, it is also useful to add it. I can always remove it if there's a gruesome merge conflict or something like that. |
gh-19830 updated the copyright to 2024 |
with the new stable release of Pythran out, I'll try to do RC2 tomorrow (on PTO today) if there are no hiccups |
* MAINT: integrate.cumulative_simpson: various improvements
* BUG: stats.nbinom.logcdf: fix IndexError
For example, in the deprecation of positional use of `tol=` in sparse solvers.
By and large, these should all be filtered out in the implementation, not in the tests, see scipy#19805.
* now that Pythran stable release is out, require `0.15.0` in `pyproject.toml`/correct the version bounds (we need this newer version to pull in `setuptools` with Python `3.12`)
* update the SciPy `1.12.0` release notes following additional backports for RC2 [skip cirrus] [skip circle]
damn, the bot is playing up again. I don't have time to look into it in the next few days unfortunately. It may be because this PR is to the maintenance branch. EDIT: yeah, I think it's just because the changes haven't been backported to this branch. Not sure whether it's worth doing that. |
I'm not sure what to make of the Python That failure persisted through a restart of the job, and |
This looks pretty unrelated to pythran, or perhaps it's triggering a bug somewhere else (e.g. somehow, a race or duplication for creating scratch space for scipy-openblas).
|
It is showing up in PRs against |
* an empty commit to test wheel builds prior to `1.12.0` RC2 release process [wheel build]
@mdhaber do you recognize the MacOS ARM wheel test failures for |
Studying the MacOS ARM64 Python I did do some side-by-side log diffing against a previous success in Cirrus, and nothing immediately stood out to me re: dependency versions that were suspiciously different. Matching the versions of the deps locally on an ARM Mac still isn't reproducing for me either. |
Maybe I'll give https://github.com/cirruslabs/cirrus-cli a go in the morning, to see if I can get some traction locally. |
In >>> import numpy as np
>>> def func(x):
... if np.isnan(x).any():
... return x.astype(np.float64)
... else:
... return x.astype(np.int64)
...
>>> np.apply_along_axis(func, 0, np.array([[0, np.nan], [1, 2]])).dtype
dtype('int64')
>>> np.apply_along_axis(func, 0, np.array([[np.nan, 0], [1, 2]])).dtype
dtype('float64') Final problem is probably that the random numbers used in the test aren't stable ( Since the implementation already has been improved in |
That seeding problem is more widespread. I'll turn this into a new issue for |
Move the code for getting the reference data in test_real_transforms to a fixture. This easily ensures that the data is loaded only once (the previous caching logic did not work correctly), and it makes it possible to easily close the data files afterwards. This also fixes test flakiness on PyPy. Unlike CPython, PyPy does not aggressive GC open files. Therefore, when every test iteration opened the files again, the number of open files quickly grew to the system limit and caused further `open()` calls to fail. Fixes scipy#19553
This avoids looking for scipy-openblas with CMake (which we never want), and avoids looking for it twice (that was an oversight). As a result, we should be robust to whatever is the underlying problem of the CI failures reported in scipygh-19852 are. Closes scipygh-19852 [skip cirrus] [skip circle]
* pin the random seed in `test_nan_policy_omit_3d`; see detailed discussion at: scipy#19797 (comment)
* update the SciPy `1.12.0` release notes following more backports/fixes [wheel build]
Two more backports were added in (updated the original list above), a random seed fix for After checking that full test suite and doc build pass locally, I've proceeded straight to another test of the full wheel builds, to see where we stand now. |
CI is fully green, including regular + wheel builds, in it goes. |
FWIW, testing of rc2 in conda-forge is looking good! |
The use of `hash('a_descriptive_string')` is cute, but unfortunately hashes are not reproducible at all. This resulted in a hard to debug failure in scipygh-19797. So remove all usages of `hash()`, and also use `np.random.RandomState` for good measure since that has more stability guarantees than `np.random.default_rng`.
Backports included (so far):
py::module_local
(fix for 1.12RC) #19751nbinom.logcdf
for invalid input #19779Newton-CG
#19785TODO:
setuptools
per MAINT: need stable Pythran release for SciPy 1.12.0 RC2 #19795