numpy 2.3.0 madness

[xiki-tempula]

It seems that numpy 2.3.0 gives widely different results for different python versions.
Here is the test code

u_nk.npy.zip

import numpy as np
import pymbar

u_nk = np.load('u_nk.npy')
mbar = pymbar.MBAR(
            u_nk,
            [4001]*5,
        )
out = mbar.compute_entropy_and_enthalpy()
print(out['Delta_f'][0,:])

Mac OS python 3.11 with numpy 2.2.6 gives
[0. 1.61906927 2.55799023 2.98630159 3.0411557 ]
Mac OS python 3.12 with numpy 2.2.6 gives
[0. 1.61906927 2.55799023 2.98630159 3.0411557 ]
The result is identical.

Mac OS python 3.11 with numpy 2.3.0 gives
[0. 1.61712842 2.55818601 2.98991384 3.04777187]
Mac OS python 3.12 with numpy 2.3.0 gives
[0. 1.61774589 2.55915648 2.99099155 3.04888306]

[mikemhenry]

I've noticed the same thing, tests pass with numpy 2.2.6 but fail with 2.3. Not sure why, looking at the change log, nothing is super obvious. I will update the conda-forge feedstock to pin <2.3 for now.

Maybe the histogram changes:

https://numpy.org/doc/stable/release/2.3.0-notes.html#changes-to-automatic-bin-selection-in-numpy-histogram

I haven't dug into this a ton OR checked if this has been reported as an issue upstream yet.

[mikemhenry]

This issue is a bit more important now. The changes actually have to do with numpy using the platform implementation of math functions (for example on linux it uses glibc's implementation of cos instead of their own implementation. This makes each platform and python version produce a different result.

I will make a PR here: #566 that removes the numpy <2.3 pin and we will see what tests fail and also what numpy version gets pulled in. I am also testing python 3.14 as well since the python 3.14 rebuild of numpy is what reminded me of this issue.

@ijpulidos do you have any thoughts on this?