Python 3.13 Support

[TheSkyentist]

A draft for Python 3.13 compatibility. Currently solving none of the bugs, but simply updating the CI matrix to test all versions of supported Python and updating the pyproject.toml to have the 3.13 tag while disallowing 3.14.

[TheSkyentist]

Moving the discussion from #304 here. The current issue is with @AbstractQuantity.from_._f.register:

unxt/src/unxt/_src/quantity/base.py

Line 876 in e5b5b06

@AbstractQuantity.from_._f.register # noqa: SLF001

unxt/src/unxt/_src/quantity/base.py

Line 902 in e5b5b06

@AbstractQuantity.from_._f.register # type: ignore[no-redef] # noqa: SLF001

unxt/src/unxt/_src/quantity/base.py

Line 928 in e5b5b06

@AbstractQuantity.from_._f.register # type: ignore[no-redef] # noqa: SLF001

unxt/src/unxt/_interop/unxt_interop_astropy/quantity.py

Line 28 in e5b5b06

@AbstractQuantity.from_._f.register # noqa: SLF001

unxt/src/unxt/_interop/unxt_interop_astropy/quantity.py

Line 48 in e5b5b06

@AbstractQuantity.from_._f.register # type: ignore[no-redef] # noqa: SLF001

[TheSkyentist]

The error, presented in full in #303, is an AttributeError: 'function' object has no attribute 'register'. Replacing @AbstractQuantity.from_.f.register with @dispatch causes interoperability problems and replacing with @AbstractQuantity.from._f.dispatch results in the same error.

[TheSkyentist]

Perhaps this is an issue with plum, but testing the package itself with 3.12 and 3.13 yields the same results, but it may be due to something more complex.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.26%. Comparing base (00d5841) to head (9da4dc5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #307   +/-   ##
=======================================
  Coverage   96.26%   96.26%           
=======================================
  Files          43       43           
  Lines        1579     1579           
  Branches       72       72           
=======================================
  Hits         1520     1520           
  Misses         47       47           
  Partials       12       12           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nstarman]

I did some quick testing and it looks like beartype/plum#200 will fix this, allowing for from._f.register -> from_.dispatch

[TheSkyentist]

I can confirm that this appears to solve most issues! But some critical tests still fail:

There seem to be issues related to unit conversions where the resultant JAX array has the weak_type flag set to true but the example doesn't expect this.
In addition, numpy operations appear to strip the unxt unit from the array.

[nstarman]

This PR can now bump plum >= 2.5.4 2 and change to from_.dispatch. Then we can see what test failures remain.

[nstarman]

Great! It looks like everything passes.

There seem to be issues related to unit conversions where the resultant JAX array has the weak_type flag set to true but the example doesn't expect this.
In addition, numpy operations appear to strip the unxt unit from the array.

Where were these issues showing up? Given the passing tests this sounds like issues for a followup PR.

[TheSkyentist]

I'm glad it's passing all of the CIs! These issues are showing up when I run pytest locally. After doing a fresh install on this branch here is the issue I still have:

import unxt as u
x = u.Quantity(1000.0, "m")

returns a

Quantity['length'](Array(1000, dtype=int32, weak_type=True), unit='m')

note the weak_type flag being set.

This causes pytest failures as the sybil checks want the internal JAX array to be strongly typed. I'm unclear if this is an actual issue, but I thought it would be worth mentioning. If it's outside the scope of this PR then we can merge and get Python 3.13 support rolling. If it's not a problem at all, then great!

The merge-base changed after approval.

[TheSkyentist]

I'm sorry for the spam! I tried to learn how to sign a commit and I forced pushed quite a few times.

[nstarman]

I'm sorry for the spam! I tried to learn how to sign a commit and I forced pushed quite a few times.

No worries!

If it's outside the scope of this PR then we can merge and get Python 3.13 support rolling. If it's not a problem at all, then great!

It is indeed. Worth investigating. The weak dtype stuff is nice in that it indicates that copies are not being made.

[nstarman]

Thanks @TheSkyentist for the nice PRs!

[TheSkyentist]

Happy to help!