Repository navigation
[MNT]: Which NumPy version to use for the weekly build? #26730
Description
Summary
We install both NumPy -pre-version (1.26) and 2.0, but still build MPL with 1.25:
Install Python dependencies:
Collecting numpy
Obtaining dependency information for numpy from https://files.pythonhosted.org/packages/f9/86/7b6d8e66fa9c2138516b33f6dab9854eae0736c5ab89205398c4622ed37f/numpy-1.26.0rc1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata
Downloading numpy-1.26.0rc1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (58 kB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 58.5/58.5 kB 19.2 MB/s eta 0:00:00
...
Downloading numpy-1.26.0rc1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.2 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18.2/18.2 MB 98.3 MB/s eta 0:00:00
Install the nightly dependencies
Collecting numpy
Downloading https://pypi.anaconda.org/scientific-python-nightly-wheels/simple/numpy/2.0.0.dev0/numpy-2.0.0.dev0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.3 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18.3/18.3 MB 14.8 MB/s eta 0:00:00
Install Matplotlib
Collecting numpy>=1.25
Obtaining dependency information for numpy>=1.25 from https://files.pythonhosted.org/packages/69/1f/c95b1108a9972a52d7b1b63ed8ca704[66](https://github.com/matplotlib/matplotlib/actions/runs/6129239432/job/16637131200#step:11:67)b59b8c1811bd121f1e6[67](https://github.com/matplotlib/matplotlib/actions/runs/6129239432/job/16637131200#step:11:68)cc45d8/numpy-1.25.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata
Downloading numpy-1.25.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.6 kB)
...
Downloading numpy-1.25.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.3 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18.3/18.3 MB 107.1 MB/s eta 0:00:00
Proposed fix
Wouldn't it make sense to also use one of the upcoming versions during installation?
Activity
ksunden commentedon Sep 11, 2023
This conversation may be reignited at some point, but I think I'm largely okay with using the nightly wheels for now.
There is a bit of a compatibility dance that needs to be done, but I'm not aware of further code changes that I expect to require on our end for np 2.0, and so I think testing against installing main with 2.0 is probably a good thing.
That said, there are ABI changes planned, so code compiled with np<2 may not work with np==2, though the opposite is intended to continue working, to my knowledge.
Official np guidance is to pin to
numpy<2.0for releases (something we have not done for prior releases, but have done for 3.7.3 (#26706) and I am going to do the same for the 3.8 release branch.)However, mpl is used in numpy's doc building, so they may be forced to do source installs (and potentially git installs) for that to work once the ABI changes hit.
Until quite recently, the nightly wheels, which are built from numpy main, were actually closer to 1.26 than they were to 2.0. So far it is only python deprecations/cleanups that they have merged on main, which we have addressed in our code, so the failures we are seeing are due to third parties (xarray/pandas, mostly)
oscargus commentedon Sep 12, 2023
My point is rather that we don't build MPL for testing with the nightly wheels, although we install them prior to build MPL, as if the purpose was to actually build it as well. But I may have misunderstood your reply.
QuLogic commentedon Jul 13, 2024
We always use
--no-build-isolationnow, so the nightly build will be against the nightly wheels automatically.