Raise if fit and predict use different array API namespaces or devices (continued) - #33076

Reference Issues/PRs

Continuation of #29313

What does this implement/fix? Explain your changes.

This PR adds _array_api.check_same_namespace to raise an error if the array namespace used during fit and predict/transform do not match, and _array_api.convert_estimator to move the array attributes of an estimator to the namespace and device of a reference array. Eventually we should put convert_estimator in a public part of the API, because users should use it. For now I think it is Ok to have it in sklearn.utils._array_api as things are still experimental.

A change I've made compared to #29313 is that none of our library code turns on sklearn.set_config(array_api_dispatch=True) for its run (original version in #29313). I think this makes sense because users should opt-in themselves. But it does open up the question of what exactly should happen if you use convert_estimator on an estimator without turning on array API support in scikit-learn. My thinking is there should be an exception to let people know they are doing something weird.

>>> from sklearn.linear_model import Ridge
>>> import sklearn
>>> import torch
>>> from sklearn import datasets

>>> sklearn.set_config(array_api_dispatch=True)

>>> X, y = datasets.make_regression()
>>> X_torch = torch.asarray(X, device='cpu')
>>> y_torch = torch.asarray(y, device='cpu')
>>> ridge = Ridge().fit(X_torch, y_torch)
>>> type(ridge.predict(X_torch))
<class 'torch.Tensor'>
>>> ridge.predict(X)
Traceback (most recent call last):
    ...
ValueError: Inputs passed to Ridge.predict() must use the same array library and the same device as those passed to fit(). Array namespaces used during fit (sklearn.externals.array_api_compat.torch) and predict (sklearn.externals.array_api_compat.numpy) differ. You can convert the estimator to the same library and device as X with: 'from sklearn.utils._array_api import convert_estimator; estimator = convert_estimator(estimator, X)'
>>> from sklearn.utils._array_api import convert_estimator
>>> ridge = convert_estimator(ridge, X)
>>> type(ridge.predict(X))
<class 'numpy.ndarray'>
>>> ridge = convert_estimator(ridge, X_torch)
>>> type(ridge.predict(X_torch))
<class 'torch.Tensor'>

AI usage disclosure

There were quite a few merge conflicts which Cursor managed to resolve and it managed to figure out how to change from the old "what's new" system to the new one once I explained that there is a README to read.

I used AI assistance for:

• Code generation (e.g., when writing an implementation or fixing a bug)
• Test/benchmark generation
• Documentation (including examples)
• Research and understanding