Array API 2025.12 support

[hameerabbasi]

What type of PR is this? (check all applicable)

• 💾 Refactor
• 🪄 Feature
• 🐞 Bug Fix
• 🔧 Optimization
• 📚 Documentation
• 🧪 Test
• 🛠️ Other

Related issues

• Closes array API: add broadcast_shapes #940

Checklist

• Code follows style guide
• Tests added
• Documented the changes

---

[hameerabbasi]

@prady0t __ifloordiv__ fails for some weird reason I can't figure out; all other tests (including elemwise) succeed. Maybe an issue with the test suite or a latent bug?

[codspeed-hq]

Merging this PR will improve performance by 11.63%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 1 improved benchmark
❌ 1 regressed benchmark
✅ 338 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	test_index_fancy[side=100-rank=1-format='coo']	1.1 ms	1.4 ms	-19.8%
⚡	test_index_slice[side=100-rank=2-format='gcxs']	3.5 ms	2.2 ms	+55.37%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing hameerabbasi:array-api-2025.12 (1bd2d5a) with main (c4ed43f)}

[prady0t]

I think the hypothesis issue has now been resolved. After a lot of effort, I found out that the asarray implementation in sparse does not take copy into account. See the latest commit I pushed.

Also see this example:

import numpy as np
from sparse.numba_backend import COO, asarray

x = COO.from_numpy(np.array(6.0))
y = asarray(x, copy=True)
a = asarray(x, copy=False)
z = np.asarray(x.todense(), copy=True)
print(id(x), id(y), id(z), id(a))
print(x is a)
print(x is y)
print(x is z)

The output is:

5727776080 5727776080 5770313616 5727776080
True
True
False

Which means x, y, and a are the same object, and it does not affect the object if set to True or False, whereas copy should produce different objects, just like z

After this commit, the output is:

6072479056 6072479376 6073926064 6072479056
True
False
False

Which is the correct behaviour.

[hameerabbasi]

Well done tracking that down! Can we add a test to ensure that asarray with copy=True copies for all formats?

[prady0t]

Can we add a test to ensure that asarray with copy=True copies for all formats?

Sure. In this PR itself?

[hameerabbasi]

Sure. In this PR itself?

That'd be nice.

[prady0t]

DOK had no copy(); added it in the latest commit(inspired by GCXS copy).

[hameerabbasi]

Still need to remove some xfails. I'll do that later today.

[lucascolley]

I found out that the asarray implementation in sparse does not take copy into account.

I wonder why we didn't catch this earlier (and more easily) via the array-api-tests suite? Is there something we should add over there?

[hameerabbasi]

I wonder why we didn't catch this earlier (and more easily) via the array-api-tests suite? Is there something we should add over there?

You probably do; but by default the arrays are also immutable (don't implement setitem) in PyData/Sparse; which is why the test probably got skipped.