ENH: stats.circ___: add array-API support

[mdhaber]

Reference issue

gh-20544

What does this implement/fix?

Makes circmean, circstd, and circvar array-API compatible.

[mdhaber]

Some questions and comments.

[mdhaber]

This is getting a little out of hand with lint complaining about UP031 in CI but not locally. I fix one thing and it starts complaining about another : (

[lucascolley]

Do you have the same ruff version locally (0.4.2)?

[mdhaber]

Probably not. I thought we would have configured a rule set, though. I guess as ruff has added these, and we don't exclude them.

[lucascolley]

We enabled all UP rules in gh-19516. UP031 should be in any compatible version of ruff that you have locally, as our minimum is 0.0.292, and the rule was added in 0.0.229. I would guess that it is due to detection differences over versions of ruff due to bug fixes.

[tylerjereddy]

I think the regular CI is already picking some of them up, but here are fails I see locally with GPU:

FAILED scipy/stats/tests/test_axis_nan_policy.py::test_axis_nan_policy_decorated_positional_args - AssertionError: Regex pattern did not match.
FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_nan_propagate_array[circmean-expected0-torch] - AssertionError: Scalars are not close!
FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_circmean_axis[circvar-torch] - AssertionError: Tensor-likes are not close!
FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_circmean_axis[circstd-torch] - AssertionError: Tensor-likes are not close!
FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_circfuncs_array_like[circmean-0.167690146-torch] - AssertionError: Scalars are not close!
FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_circmean_axis[circmean-torch] - AssertionError: Tensor-likes are not close!

[mdhaber]

Interesting. No, I don't think it is picking up in CI. I am running with CuPy locally, and I wasn't getting those.
Upgrading ruff seems to have broken my in-place build, and pip uninstall scipy is erroring out... I'll test again when I get things fixed.

[mdhaber]

Thanks @tylerjereddy. Can you give the full failure information (e.g. actual relative error)?

Also, can you reproduce these two:

FAILED scipy/stats/tests/test_morestats.py::TestCircFuncs::test_circfuncs_small[torch] - TypeError: var() received an invalid combination of arguments - got (Tensor, tuple, correction=float), but expected one of:
FAILED scipy/special/tests/test_support_alternative_backends.py::test_dispatch_to_unrecognize_library - TypeError: Cannot interpret 'array_api_strict.float64' as a data type

I can't reproduce them on this machine.

[tylerjereddy]

I can't reproduce those two either. The PR seems fine on CPU, I'll try to add some inline comments re: the failure details on GPU.

[mdhaber]

OK, I'm happy with this now. aside from the comments here and the one failure, which I can't reproduce.

=================================== FAILURES ===================================
__________________ TestCircFuncs.test_circfuncs_small[torch] ___________________
scipy/stats/tests/test_morestats.py:2479: in test_circfuncs_small
    V1 = xp_test.var(x*xp.pi/180, correction=xp.asarray(0.0))
        M1         = tensor(19.9571, dtype=torch.float64)
        M2         = tensor(19.9571, dtype=torch.float64)
        self       = <scipy.stats.tests.test_morestats.TestCircFuncs object at 0x7f625bdd31d0>
        x          = tensor([20.0000, 21.0000, 22.0000, 18.0000, 19.0000, 20.5000, 19.2000],
       dtype=torch.float64)
        xp         = <module 'torch' from '/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/torch/__init__.py'>
        xp_test    = <module 'scipy._lib.array_api_compat.torch' from '/home/runner/work/scipy/scipy/build-install/lib/python3.11/site-packages/scipy/_lib/array_api_compat/torch/__init__.py'>
scipy/_lib/array_api_compat/torch/_aliases.py:425: in var
    res = torch.var(x, tuple(range(x.ndim)), correction=correction, **kwargs)
E   TypeError: var() received an invalid combination of arguments - got (Tensor, tuple, correction=Tensor), but expected one of:
E    * (Tensor input, tuple of ints dim, bool unbiased, bool keepdim, *, Tensor out)
E    * (Tensor input, tuple of ints dim, *, int correction, bool keepdim, Tensor out)
E    * (Tensor input, bool unbiased)
E    * (Tensor input, tuple of names dim, bool unbiased, bool keepdim, *, Tensor out)
E    * (Tensor input, tuple of names dim, *, int correction, bool keepdim, Tensor out)
        axis       = None
        correction = tensor(0.)
        keepdims   = False
        x          = tensor([0.3491, 0.3665, 0.3840, 0.3142, 0.3316, 0.3578, 0.3351],
       dtype=torch.float64)

The whole point of calling array_namespace is to get an array-API xp_test so that it has var. Clearly x is a tensor. Technically, correction can just be an int or float, but I think I've tried that, in addition to leaving it at its default value. I would think it's a bug in array_api_compat, but this is using SciPy's vendored version, right?

Well, the message suggests that it should be an int, whereas the default is a 0.0, so let's try that.

Yeah we must have an older version of torch that isn't perfectly compatible with our array-api-compat.

[mdhaber]

@tylerjereddy please run these tests once more with TF GPU when you have a moment. Thanks!

[j-bowhay]

This is looking good, some minor comments, most of which are for my own understanding

[tylerjereddy]

please run these tests once more with TF GPU when you have a moment. Thanks!

@mdhaber

SCIPY_DEVICE=cuda python dev.py test -j 32 -b all

• this branch (230f383): 3 failed, 51023 passed, 11078 skipped, 155 xfailed, 13 xpassed in 55.71s
• main (4ee0de3): 4 failed, 50958 passed, 11090 skipped, 155 xfailed, 13 xpassed in 60.08s (0:01:00)

So, you're at parity with main in my hands, which seems good (I always get a few thread safety issues with fft locally on GPU). I didn't check the diff myself again, just reporting this result.

[lucascolley]

I always get a few thread safety issues with fft locally on GPU

Gentle reminder that these should be gone with gh-19900

[mdhaber]

OK If tests pass, I think I'm happy with this, and I think comments have been addressed.

[j-bowhay]

One query but otherwise this is good to go

[j-bowhay]

Thanks @mdhaber LGTM