BUG: array_namespace returns compat for NumPy scalars with use_compat=None

[lucascolley]

With array-api-compat 1.8, NumPy 2.0.1

In [22]: array_namespace(np.asarray(0.0), use_compat=None)
Out[22]: <module 'numpy' from '/Users/lucascolley/mambaforge/envs/scipy-dev/lib/python3.10/site-packages/numpy/__init__.py'>

In [23]: array_namespace(np.float64(0.0), use_compat=None)
Out[23]: <module 'scipy._lib.array_api_compat.numpy' from '/Users/lucascolley/programming/scipy/build-install/lib/python3.10/site-packages/scipy/_lib/array_api_compat/numpy/__init__.py'>

This behaviour is quite problematic for SciPy given NumPy's tendency to throw NumPy scalars at you under very little pressure.

In [25]: array_namespace(np.asarray(0.0) - 32, use_compat=None)
Out[25]: <module 'scipy._lib.array_api_compat.numpy' from '/Users/lucascolley/programming/scipy/build-install/lib/python3.10/site-packages/scipy/_lib/array_api_compat/numpy/__init__.py'>

It seems practical to have NumPy scalars return the same namespace as NumPy arrays when use_compat=None. Or perhaps a different argument should have to be passed to use_compat (that would be fine for us).

[lucascolley]

discovered in scipy/scipy#21264 (comment). The real problem it causes is that we check for matching namespaces in our tests, and often compare NumPy scalars against NumPy arrays. I suppose we could add extra checks for NumPy scalars to that machinery, but I think this should be fixed here instead.

[rgommers]

It seems practical to have NumPy scalars return the same namespace as NumPy arrays when use_compat=None.

Agreed, that should be the right default.

[lucascolley]

fix at gh-165

[asmeurer]

Seems like NumPy should be fixed here too.

[rgommers]

Can you spell out what should change in NumPy? It doesn't have an array_namespace function, so I don't immediately see it.

[lucascolley]

Do you mean giving NumPy scalars a matching __array_namespace__ attribute?

[asmeurer]

Yes. This should work

>>> np.asarray(0).__array_namespace__()
<module 'numpy' from '/Users/aaronmeurer/miniconda3/envs/numpy2/lib/python3.12/site-packages/numpy/__init__.py'>
>>> np.asarray([0])[0].__array_namespace__()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'numpy.int64' object has no attribute '__array_namespace__'. Did you mean: '__array_interface__'?