Re-introduce __array__ with a warning

[asmeurer]

#67

Copilot reviewed 2 out of 2 changed files in this pull request and generated no suggestions.

[asmeurer]

@rgommers mentioned this might still be too hard for SciPy in #67 so waiting for feedback here CC @mdhaber @lucascolley

[mdhaber]

I don't know whether silencing the warning is easier than fixing it. @lucascolley do you think we will be able to get scipy/scipy#21835 working before we need the features from array_api_strict 2.1+?

[lucascolley]

I don't think I have the capacity to push it through immediately, but maybe someone can follow through on the idea.

[asmeurer]

OK, I'll just re-enable it completely for now #92. We can continue the discussion about it at #67. If there are any changes that need to be made for either the standard or NumPy we should discuss those too.

[lucascolley]

FWIW I do think it should be removed. But giving consumer libraries some time to adapt will be helpful.

[mdhaber]

I'm happy with this being re-enabled while we figure out how this should be done. Maybe it could be controlled by an environment variable so we can test with _allow_array = False once we have converted to from_dlpack?

[asmeurer]

What would help is some clarity on what the problems are. Is there something that should be added to the standard and/or some changes that should be made to NumPy to make this easier?

I do know that for array-api-strict itself I found that I had to do some workarounds to make asarray([<list of array_api_strict arrays>]) work, because we reuse the np.asarray logic for that (I'm not keen on re-implementing the full list-of-lists logic here, and NumPy doesn't provide any useful helpers). But I don't know if this is the sort of problem that is affecting you, or if the problems are something else.

[lucascolley]

In SciPy the problem is that we punted on using DLPack by just using np.asarray where needed and adding skips for GPU backends. Now that there is a CPU backend which doesn't work with np.asarray (array-api-strict), we're seeing the consequences. We do need to move to DLPack, but it will take a bit of time.

[mdhaber]

Is there something that should be added to the standard and/or some changes that should be made to NumPy to make this easier?

We'll probably figure this out when we're making the change to from_dlpack. It just wasn't done before because we were supposed to use np.asarray to convert to NumPy according to our development documentation.

[asmeurer]

Going to discuss __array__ a little more in the meeting Thursday. For now, it sounds like a warning is no more helpful than an error, so I'm going to close this.