✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: d46840b. Link to the linter CI: here}

@lesteve Just one test failing and that has to do with array api strict on device and float32. Maybe we need to increase the tolerance further for this specific scenario.

My honest impression is that these tests are fragile on float32 data but I don't really know if there is much we can do to improve the situation ...

Even for array-api-strict the results are different because of the difference between scipy.linalg.choleksy and numpy.linalg.cholesky and between scipy.linalg.triangular_solve and numpy.linalg.solve.

On a GPU VM I also saw some test failures (a few more than in the CI actually) and raised the atol and rtol a bit to get them to pass locally. I trigger another run of the CUDA CI, let's see what happens 🤞.

I don't think we can do much with trying to improve array-api-strict tests for float32 especially with respect to accuracy. As long as array-api-strict works generally I think that should be sufficient.

Thanks for the reviews @OmarManzoor and @ogrisel!

One of the remaining question in the old and long TODO list: should we implement __sklearn_tags__ to tell that GaussianMixture has array_api_support?

PCA ___sklearn_tags__ does this currently and always sets array_api = True although array API support is implemented for some values of the parameters, not sure whether this is expected or not:

I am guessing the array_api tags is only used for the common tests right now, right?

Good questions:

• indeed, we could make PCA only return the tags.array_api_support = True when the solver supports array API inputs.
• similarly for GaussianMixture (depending on the choice of the init).