[CI] Dask OneHotEncoder CuPy test fails during task serialization on Python 3.14

[csadorf]

Summary

Failing test/component: test_dask_one_hot_encoder.py::test_onehot_random_inputs[10-sparse-cupy-None]

Failure observed in:

• https://github.com/NVIDIA/cuml/actions/runs/32109006318/job/95624279626

Environment

• OS: ubuntu26.04, arm64
• Python version: 3.14.7
• CUDA version: 13.3.0
• GPU: GB300
• Dependencies: latest-deps; cuml-cu13==26.10.0a70.post260818062854, dask==2026.7.1, distributed==2026.7.1, pytest==9.1.1

Test Details

• Test file: python/cuml/cuml/tests/dask/test_dask_one_hot_encoder.py
• Test name: test_onehot_random_inputs[10-sparse-cupy-None]
• Error message:

  ExceptionGroup: multiple unraisable exception warnings (5 sub-exceptions)
  
  _pickle.PicklingError: Can't pickle <function _to_cudf ...>:
  it's not the same object as cuml.dask.common.dask_arr_utils._to_cudf
  when serializing dask._task_spec.Task object
  
  RecursionError: Stack overflow (used 8097 kB)
  RuntimeError: Failed to process unraisable exception
  

The job completed with 1 failed, 850 passed, 3685 skipped, 4 xfailed.

Root Cause Analysis

[jcrist]

I've self assigned this - we've had a few of these errors appear in the dask test suite randomly over the last several months. They shouldn't happen based on how dask's serialization routines are supposed to work, but 🤷 as to why they're showing up now. The best theory I have has to do with some annoying blub details around how dask.delayed is implemented and how that interacts with some caching magic in cloudpickle. For now I plan to go through our whole cuml.dask codebase and unify how we're handling helper functions so they're all the same and follow the best pattern (basically doing the one-off fixes I've been applying for this issue everywhere). Hard to prove a negative, but at least once everything is uniform if we see this issue again we'll know to look elsewhere.