pint 0.24 / numpy 2.0 / python 3.9 issues

[keewis]

numpy>=2 will include a couple of breaking changes (to be released soonish), so it would be good to have a release of pint before that (I'd also love to have #1922 in a release, as it will fix some typing issues on python>=3.12)

Most of the urgent issues have already been fixed, the only thing left is #1971.

Other than that, __array__ has gained a optional copy kwarg which we should add (the version of __array__ without copy will cause a deprecation warning).

[cmutel]

In case anyone is having weird pint errors, or getting this error:

AttributeError: module 'numpy' has no attribute 'cumproduct'. Did you mean: 'cumprod'?

It's because you installed Numpy 2.0.RC1, which isn't yet supported.

[andrewgsavage]

I'm not sure how best to handle the __array__(copy) issue, we don't handle other optional args at present.
Other than that seems like 2.0 works now - I've added it to the matrix.(maybe too a bit too many places)

[andrewgsavage]

I think #1978 and #1960 could be merged if someone has the time to review.

Otherwise I think we're ready for a new release. Are there any other PRs that could be merged?

[keewis]

I would forward copy to the wrapped array, but in any case this is not urgent, since not having that argument will warn, not raise.

#1960 is a documentation-only PR (and from a quick glance I can't see anything wrong), so I think that PR is fine?

[dopplershift]

We're officially on the clock here, as NumPy 2.0 will release 16 June 2024.

It would be good to get some kind of release out the door since currently importing pint 0.23 with numpy 2 installed results in the AttributeError noted above.

I'm going to update on conda-forge to add a constraint for numpy < 2.

[jakirkham]

Do we know what issues still remain here?

Curious if anyone has tried to run the test suite with NumPy 2. If so, would be curious to know what that reveals

[andrewgsavage]

We're officially on the clock here, as NumPy 2.0 will release 16 June 2024.

I've released a pre-release. We've said to leave a week before full release so 0.24 can be released on the 8th.

[andrewgsavage]

0.24 released

[jakirkham]

Thanks Andrew! 🙏

[takluyver]

If anyone else is still seeing the no attribute 'cumproduct' error, note that NumPy 2.0 is compatible with Python 3.9 and up, but Pint 0.24 requires Python 3.10 or above. So unfortunately, pip install numpy pint on Python 3.9 results in pint not being importable at all.

[slokhorst]

If anyone else is still seeing the no attribute 'cumproduct' error, note that NumPy 2.0 is compatible with Python 3.9 and up, but Pint 0.24 requires Python 3.10 or above. So unfortunately, pip install numpy pint on Python 3.9 results in pint not being importable at all.

This is an issue on Ubuntu 22.04.

Would it be possible to tag a new point release 0.23.1, with:
requires-python = ">=3.9" and numpy = "^1.23"?

[takluyver]

Unfortunately I don't think that a new point release with metadata saying it needs numpy 1.x really solves this, because the dependency resolver is still free to pick pint 0.23, which appears to the tools to be compatible with numpy 2.0. In some cases it might decide to downgrade numpy instead, but not reliably.

Further complicating it, numpy is an optional dependency of pint. So if you pip install pint[numpy], pip would try to get compatible versions, but pip install pint numpy will install them as separate packages, ignoring any version constraints. There's no way to specify "pint doesn't require numpy, but numpy 2.0 breaks it" in packaging metadata.

The possible fixes I'm aware of here are:

• Downstream packages drop support for Python 3.9 entirely.
• Downstream packages require Numpy 1.x for Python 3.9 (numpy <2; python_version < "3.10")
• Pint makes a new release that's compatible with both Python 3.9 and Numpy 2.0. I think this ought to be easy in principle - it's only been a few months since Pint dropped 3.9 support - but I spent a while on it, and unfortunately neither undoing the 3.9 changes nor redoing the numpy changes was straightforward.

[andrewgsavage]

• Pint makes a new release that's compatible with both Python 3.9 and Numpy 2.0. I think this ought to be easy in principle - it's only been a few months since Pint dropped 3.9 support - but I spent a while on it, and unfortunately neither undoing the 3.9 changes nor redoing the numpy changes was straightforward.

I'm not opposed to this. Was this https://github.com/takluyver/pint/tree/0.23-np-2.0 what you tried? Did you look at making 0.24 work with 3.9?

[takluyver]

Thanks, that would be great! That was the branch I was trying on, working forwards from the 0.23 tag and trying to cherry pick the necessary changes. I tried working backwards from 0.24 first, but quickly got conflicts as I reverted commits, so I switched the approach. I didn't get far enough to push that. But I now suspect that was the better plan after all. I'm away from my computer for a few days, but I can look into this again next week if no-one else tackles it.

…

On Fri, 21 Jun 2024, 08:49 andrewgsavage, ***@***.***> wrote: - Pint makes a new release that's compatible with both Python 3.9 and Numpy 2.0. I think this ought to be easy in principle - it's only been a few months since Pint dropped 3.9 support - but I spent a while on it, and unfortunately neither undoing the 3.9 changes nor redoing the numpy changes was straightforward. I'm not opposed to this. Was this https://github.com/takluyver/pint/tree/0.23-np-2.0 what you tried? Did you look at making 0.24 work with 3.9? — Reply to this email directly, view it on GitHub <#1974 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACQB5NDECIGA45YSC4GO2DZIPLHXAVCNFSM6AAAAABG6O7422VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOBSGE4TSNZZG4> . You are receiving this because you commented.Message ID: ***@***.***>

[andrewgsavage]

I've got 0.24 working with py39 #2019 so I'll make 0.24.1 available for py39.
some seemingly unrelated failures though...

[takluyver]

Thanks! Yes, looking at the diff, that was much easier than what I was trying

…

On Fri, 21 Jun 2024, 09:42 andrewgsavage, ***@***.***> wrote: I've got 0.24 working with py39 #2019 <#2019> so I'll make 0.24.1 available for py39. some seemingly unrelated failures though... — Reply to this email directly, view it on GitHub <#1974 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACQB5IRWG5VNG4SAM2JSGTZIPRNZAVCNFSM6AAAAABG6O7422VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOBSGI4TENZUHE> . You are receiving this because you commented.Message ID: ***@***.***>

[andrewgsavage]

I think the all PRs needed to fix issues for 0.24.1 are now merged. would be good to confirm the babel issue #2006

[andrewgsavage]

0.24.1 released

[takluyver]

Thanks @andrewgsavage !