[mattip]

See numpy/numpy#33074

I ran uv lock but had to trick it to not use "7 days" since the version of scipy-openblas was just released.

• remove libquadmath
• adjust the license to point out libgfortran is statically linked

[mattip]

Ahh, this can only run after the numpy/numpy PR is merged.

[andyfaff]

Ahh, this can only run after the numpy/numpy PR is merged.

The workflow is: once the numpy/numpy PR has been merged run tools/update_lock.sh from this repo. This updates the pyproject.toml and uv.lock files in this repo.

How did you "trick uv lock"? Did you adjust update_lock.sh to remove the --exclude-newer flag? I think it should be possible to do something like tools/update_lock.sh --upgrade --exclude-newer-package=scipy_openblas64 because the update lock script has a $@ in it for adding extra flags.

I was a bit surprised that the uv lock --check --exclude-newer "7 days" gateway check doesn't fail in wheels.yml. Perhaps it assumes that once pins are present in the uv.lock file then they're regarded as being ok (i.e. uv lock --check doesn't retrieve all publishing dates when it runs).

[charris]

I there any reason not to merge this?

[andyfaff]

No reason not to

[andyfaff]

@mattip, do we need to add a check to see if libquadmath is not installed. We can add a flag to the auditwheel steps to specifically exclude libquadmath.

[rgommers]

Probably useful, but maybe as a follow-up PR, since main is currently failing on checking the lock file?

[mattip]

I have a check in openblas-libs that neither libquadmath.so nor libgfortran.so are shipped in the posix wheel, is that enough? For windows, we have statically linked without libquadmath for a number of years now.

[mattip]

Not sure where to put this, but here is a comparison of the PyPI package sizes for the before and after

platform	0.3.34.237.0	0.3.34.237.2	delta
manylinux_2_27_x86_64	8.6 MB	7.6 MB	-1.0 MB (-11.7%)
musllinux_1_2_x86_64	9.3 MB	7.8 MB	-1.4 MB (-15.3%)
manylinux_2_27_aarch64	9.2 MB	8.8 MB	-0.5 MB (-4.9%)
musllinux_1_2_aarch64	9.5 MB	8.9 MB	-0.6 MB (-6.3%)
macosx_11_0_arm64	6.8 MB	5.4 MB	-1.4 MB (-20.4%)
macosx_10_9_x86_64	10.3 MB	7.7 MB	-2.6 MB (-25.0%)
manylinux_2_27_ppc64le	9.0 MB	8.3 MB	-0.7 MB (-8.0%)
manylinux_2_27_s390x	6.7 MB	6.2 MB	-0.5 MB (-7.4%)
manylinux_2_38_riscv64	5.9 MB	5.4 MB	-0.6 MB (-9.6%)
win_amd64	7.0 MB	7.0 MB	+0.0 MB (+0.0%)
win_arm64	5.0 MB	5.0 MB	+0.0 MB (+0.0%)

Edit: that is the package size of scipy-openblas64, not NumPy.