TST: stats array API failure for test_skew_constant_value[torch] on GPU

[tylerjereddy]

On latest main at time of writing.

SCIPY_DEVICE=cuda python dev.py test -j 32 -t scipy/stats/tests/test_stats.py::TestSkew::test_skew_constant_value -b all

______________________________________________________________________________________________________________________________________ TestSkew.test_skew_constant_value[torch] ______________________________________________________________________________________________________________________________________
[gw1] linux -- Python 3.11.2 /home/treddy/python_venvs/py_311_scipy_dev/bin/python
scipy/stats/tests/test_stats.py:3436: in test_skew_constant_value
    assert_equal(stats.skew(a), xp.asarray(xp.nan))
E   AssertionError: 
E   Items are not equal:
E    ACTUAL: tensor(nan, device='cuda:0')
E    DESIRED: tensor(nan, device='cuda:0')
        a          = tensor([-0.2783, -0.2783, -0.2783, -0.2783, -0.2783, -0.2783, -0.2783, -0.2783,
        -0.2783, -0.2783], device='cuda:0')
        self       = <scipy.stats.tests.test_stats.TestSkew object at 0x7f1da916e550>
        xp         = <module 'torch' from '/home/treddy/python_venvs/py_311_scipy_dev/lib/python3.11/site-packages/torch/__init__.py'>

The patch below the fold switches to using the xp_assert.. flavor in a few places and restores the passing test. It is more idiomatic anyway, so we should perhaps do that? I'll leave it to Matt and Lucas though, I'm probably rusty on this stuff compared to them.

--- a/scipy/stats/tests/test_stats.py
+++ b/scipy/stats/tests/test_stats.py
@@ -3433,16 +3433,16 @@ class TestSkew(SkewKurtosisTest):
         # exact (gh-13245)
         with pytest.warns(RuntimeWarning, match="Precision loss occurred"):
             a = xp.asarray([-0.27829495]*10)  # xp.repeat not currently available
-            assert_equal(stats.skew(a), xp.asarray(xp.nan))
-            assert_equal(stats.skew(a*2.**50), xp.asarray(xp.nan))
-            assert_equal(stats.skew(a/2.**50), xp.asarray(xp.nan))
-            assert_equal(stats.skew(a, bias=False), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a*2.**50), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a/2.**50), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a, bias=False), xp.asarray(xp.nan))
 
             # # similarly, from gh-11086:
             a = xp.asarray([14.3]*7)
-            assert_equal(stats.skew(a), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a), xp.asarray(xp.nan))
             a = 1. + xp.arange(-3., 4)*1e-16
-            assert_equal(stats.skew(a), xp.asarray(xp.nan))
+            xp_assert_equal(stats.skew(a), xp.asarray(xp.nan))
 
     @array_api_compatible
     def test_precision_loss_gh15554(self, xp):

[mdhaber]

Oops! I think our eyes just missed that it was assert_equal and not xp_assert_equal. Thanks; I'll submit a PR.