Closed
Description
As a follow-up for #29227.
The following command fails locally:
pytest -vl sklearn/metrics/tests/test_common.py -k 'api_regression_metric and mean_poisson_deviance'
see full error in #29227 (comment)
but setting SCIPY_ARRAY_API=1 works:
SCIPY_ARRAY_API=1 pytest -vl sklearn/metrics/tests/test_common.py -k 'api_regression_metric and mean_poisson_deviance'
Honestly I think this is fair to say that as a newcomer to the array API work, my feeling is that it can break way too easily in very mysterious ways. There are probably a few unrelated reasons for this but I guess since at the last bi-weekly meetings we agreed that we should try to pay attention to this kind of breakages and improve the situation, here are a few comments on this #29277:
- I think we should have some tests with scipy>=1.14 and
SCIPY_ARRAY_APIunset. This seems something that could easily happen in practice and we should have tests that make sure that nothing breaks. It's not clear to me why only array-api-strict breaks and not pytorch example, maybe as the name implies array-api-strict is even stricter than other namespaces. - I don't think this is acceptable to skip the test for scipy<1.14 since you know a reasonable user may use
array_api_dispatch=Truewith scipy<1.14. To be fair it is not clear to me what happens in this case. I guess scipy will transform the array into a numpy array, and maybe the code afterwards has issues because some arrays are not in the same namespace? Anyway, we should make sure that the user code does not break. If this is really too hard to achieve the alternative is to raise an exception saying something like "to enable array API dispatch with this particular function you need scipy 1.14 andSCIPY_ARRAY_API=1". - the warning about
SCIPY_ARRAY_APIseems too broad (as soon asset_config(array_api_dispatch=True)is called?). Could we not have warnings only in a few (at least for now) places where we know we call scipy functions?
Originally posted by @lesteve in #29227 (comment)