BUG: cp.generic is an alias for np.generic

[crusaderky]

Fix bug where plain numpy objects e.g. np.int64(0) would take cupy-specific code paths.

This is a blocker for implementing at function in array-api-extra.

[crusaderky]

This PR is ready to be reviewed and merged.

• CI failures are unrelated and should be fixed by BUG: dask.array: asarray(..., copy=None) copies in dask==2024.12 #211

[ev-br]

What would be a scenario where cupy aliasing array scalars matters? I just tried running https://github.com/data-apis/array-api-compat/blob/main/test_cupy.sh locally, and it passes with or without.

[crusaderky]

What would be a scenario where cupy aliasing array scalars matters?

Not that I'm aware of.
If there are any cases where the output of is_cupy_array(cp.somefunction(...)) needs to return True, even if the output is a np.generic, then those cases will need to be looked into.

To me this looks like a copy-paste error from is_numpy_array.

[ev-br]

So if there's no use case, what are the reasons for this then?

[crusaderky]

So if there's no use case, what are the reasons for this then?

There is no use case that I know of where is_cupy_array(cupy.asarray([1])[0]) returning False would cause a problem - which is what this PR would cause.
But it is a big problem if is_cupy_array(numpy.asarray([1])[0]) returns True - which is what happens now in main.

[crusaderky]

This is ready for final review and merge.

[ev-br]

But it is a big problem if is_cupy_array(numpy.asarray([1])[0]) returns True - which is what happens now in main.

That's bad indeed if that happens, but... does it?

In [5]: is_cupy_array(np.asarray([1])[0])
Out[5]: False

This is on main, after

$ pip install -e .
Obtaining file:///home/br/repos/array-api-compat
Preparing metadata (setup.py) ... done
Installing collected packages: array_api_compat
DEPRECATION: Legacy editable install of array_api_compat==1.9.1 from file:///home/br/repos/array-api-compat (setup.py develop) is deprecated. pip 25.0 will enforce this behaviour change. A possible replacement is to add a pyproject.toml or enable --use-pep517, and use setuptools >= 64. If the resulting installation is not behaving as expected, try using --config-settings editable_mode=compat. Please consult the setuptools documentation for more information. Discussion can be found at https://github.com/pypa/pip/issues/11457
Running setup.py develop for array_api_compat
Successfully installed array_api_compat-1.9.1
(array-api-tests) br@gonzales:~/repos/array-api-compat$ git log
commit edde2e04a984b65dc299321ea5c0cda33f1f9e5d (HEAD -> main, upstream/main, upstream/HEAD)
Merge: d9df003 484f725
Author: Evgeni Burovski <evgeny.burovskiy@gmail.com>
Date: Thu Dec 12 18:58:34 2024 +0200
Merge pull request #213 from ev-br/torch_allclose
MAINT: avoid a DeprecationWarning from torch

What am I missing?

[crusaderky]

What am I missing?

You're missing cupy in sys.modules.

In [1]: fromarray_api_compatimportis_cupy_arrayIn [2]: importnumpyasnpIn [3]: g=np.asarray([0])[0]
In [4]: is_cupy_array(g)
Out[4]: FalseIn [5]: importcupyIn [6]: is_cupy_array(g)
Out[6]: True

[ev-br]

Ouf, yes indeed, is_cupy_array(....) is False unless cupy has been imported. So yes, a clear bug with a clear fix. Thank you @crusaderky