Add more estimators to the compatibility test suite

[betatim]

This adds more estimators and xfails for them to the compatibility test suite.

This is step 1 for #7061

Along the way I'm creating issues for issues that arise when running the checks that are more serious (aka we can't just mark the check as xfail). They should all refer to #7061 so you can see them there

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[betatim]

/ok to test

[betatim]

IMHO we can already merge this, even though there are more estimators that could be added. The CI gods are looking favourably on this PR right now, and the diff is already large.

[csadorf]

LGTM, just one question.

[csadorf]

We need to fix some of the imports.

[betatim]

After fixing the imports and adding more xfails and skipping more tests I am now at a point where estimators like RandomForestClassifier that were in a good state fail the most basic "is this estimator clone'able" check (which you can't xfail or skip) with a "bad alloc" failure. It seems like failures in one estimator effect other estimators. In particular some failed checks lead to an increased chance of future checks failing with MemoryErrors. Skipping those tests works to some extent, but eventually you get to the point that RandomForestClassifer is now at: even the most basic check fails, but if you run the test suite only on RandomForestClassifier everything passes or is xfailed.

Seems like a bit of a dead end to work on this without fixing what ever underlying problem is causing this, because there isn't much point is skipping all the checks.

[betatim]

The memory error that you see a lot is variations on this: MemoryError: std::bad_alloc: CUDA error (failed to allocate 1280 bytes) at: /home/thead/miniforge3/envs/cuml-20250729/include/rmm/mr/device/cuda_memory_resource.hpp

You can run the tests with pytest -sv --disable-warnings --tb=auto test_sklearn_compatibility.py which will eventually crash and produce a lot of MemoryErrors. You can also only run a subset, for example for TruncatedSVD with pytest -sv --disable-warnings --tb=auto test_sklearn_compatibility.py -k TruncatedSVD - if you run only the TruncatedSVD tests everything will either pass, be skipped or xfail. If you run all the tests new failures will appear within the TruncatedSVD tests.

[csadorf]

I'd suggest to split off the RandomForestClassifier tests for now so that we can unblock this and then address those problems separately.

[betatim]

We can completely skip RandomForestClassifier and the estimators after them. However the problem isn't with those estimators per se. The problem is that at some point a check breaks some global state that leads to RMM being unable to allocate any amount of memory. As a result we end up marking many checks as xfail, even though they'd pass if you test only a specific estimator.

[betatim]

Ok, I like this solution better: we skip the naive bayes estimators as they are the ones that cause the problems. All the other estimators are tested.

[csadorf]

Pre-approving since this generally LGTM, but I'd ask that we reference relevant issues in-code wherever applicable.

[betatim]

/merge