ENH: use native numpy.reshape for numpy>=2.1

[stanbot8]

reshape with copy=False assigns to the array shape attribute that numpy 2.5 deprecates.

Bind the numpy namespace reshape to the native numpy function from numpy 2.1. That release added the copy argument. Keep the shared workaround for an older numpy.

Leave the cupy and dask paths unchanged. Neither namespace accepts a copy argument, so both still need the shared workaround.

[ev-br]

Development in this repository is human-first, and we very much prefer to not receive automated contributions from bots.

[stanbot8]

@ev-br I am a human contributor. stanbot8 is only my username. I reviewed and tested this change and can explain or revise each part of it.

[ev-br]

Than my apologies @stanbot8 . We receive too many automated pull requests and the default shifts to "shoot first ask later". And the combination of the GH username and the PR style was too LLM-like.

So, could you then please explain (preferably in your human words, without using an LLM), what is this patch, what is the problem it aims to address and what is the solution?

[stanbot8]

No worries @ev-br, I understand how my username and the PR style could have given that impression.

I found this while testing the project with NumPy 2.5.1. The copy=False path directly assigned to y.shape, which NumPy 2.5 deprecated. My patch uses the namespace's reshape(..., copy=False) when it supports the copy argument and keeps the existing fallback otherwise.

[ev-br]

Oh I see what's going on here now, thank you!

So the copy= behavior of reshape function is effectively untested (data-apis/array-api-tests#432) and subtly broken in several respects. For one, the copy=False workaround drops **kwargs.

Now, using the native function instead of a pile of workarounds is definitely a win and is very worthy goal to work towards!

I just quickly tried running tests with bare numpy 2.4.2 and a branch of data-apis/array-api-tests#432 and it passes:

$ ARRAY_API_TESTS_MODULE=numpy pytest array_api_tests/test_manipulation_functions.py::test_reshape -v --max-examples=10_000

Therefore, how about:

• branch on a numpy version at https://github.com/data-apis/array-api-compat/blob/main/src/array_api_compat/numpy/_aliases.py#L53 : only use the existing workaround for older numpies and use the native reshape for newer numpies; this would remove the need for inspect-ing things at runtume;
• for CuPy and dask, keep the existing workaround, too (ain't broken and all) --- for CuPy we'll be able to simplify in a year or so: ENH: add explicit copy arg to reshape cupy/cupy#10109

For testing, I think it would be cleaner to test numpy and cupy instead of common, a test with shares_memory would be great.

If you are feeling brave and have cycles to spare, help fixing data-apis/array-api-tests#432 would be appreciated but is certainly not a blocker for this PR.
If you are feeling really brave, torch/_aliases.py::reshape simply raises with copy=, but that'd even more of a "not a blocker".

[stanbot8]

Rewrote per your suggestion.

The numpy version branch sends only an older numpy through the workaround, so the runtime inspect is gone. I moved the tests out of common into the numpy and cupy files, and both assert shares_memory. My machine has a CUDA GPU, so the cupy test runs instead of skipping.

[ev-br]

CI errors seem unrelated. #456 is to fix the error with -strict, not sure yet what's going on with jax and its Array API version.

[ev-br]

close/reopen to retrigger CI after gh-456

[ev-br]

LGTM now, thank you @stanbot8 ! Merging.