Right now legate-boost pins to numpy<2 maybe just unnecessarily due to test failure.
The problem is a chain of things:
- legate-boost subclasses sklearn
sklearns score method (unintentionally) uses partial Array API related logic (specifically it checks .device). Maybe only in an effort to give a nicer error message.- On
numpy < 2 those checks are just irrelevant, but on numpy >= 2 they trigger because cupynumeric don't have array.device == "cpu"
If we can create a more minimal reproducer, this is probably actually an skimage bug, because skimage shouldn't be sensitive to this type of thing without user opt-in.
On the one hand, that may be in the future, on the other hand such details are probably also still up to change in sklearn before removing the need for opt-in.
The simplest thing is to decide that just don't expect a legate estimator to work with NumPy inputs to .score() or convert to cupynumeric first to work around it (until sklearn may have changed their behavior).
Right now legate-boost pins to
numpy<2maybe just unnecessarily due to test failure.The problem is a chain of things:
sklearnsscoremethod (unintentionally) uses partial Array API related logic (specifically it checks.device). Maybe only in an effort to give a nicer error message.numpy < 2those checks are just irrelevant, but onnumpy >= 2they trigger because cupynumeric don't havearray.device == "cpu"If we can create a more minimal reproducer, this is probably actually an skimage bug, because skimage shouldn't be sensitive to this type of thing without user opt-in.
On the one hand, that may be in the future, on the other hand such details are probably also still up to change in sklearn before removing the need for opt-in.
The simplest thing is to decide that just don't expect a legate estimator to work with NumPy inputs to
.score()or convert tocupynumericfirst to work around it (until sklearn may have changed their behavior).