use 'nvidia-ml-py' instead of 'pynvml', declare 'numba-cuda' dependency pins

[jameslamb]

Contributes to rapidsai/build-planning#208 (breaking some changes off of #7128 to help with review and debugging there)

• switches to using dask-cuda[cu12] extra for wheels (added in Build and test with CUDA 13.0.0 dask-cuda#1536)
• bumps pins on some dependencies to match the rest of RAPIDS
  • cuda-python: >=12.9.2 (CUDA 12)
  • cupy: >=13.6.0
  • numba: >=0.60.0
• adds explicit runtime dependency on numba-cuda
  • cuml uses this unconditionally but does not declare runtime dependency on it today

Contributes to https://github.com/rapidsai/build-infra/issues/293

• replaces dependency on pynvml package with nvidia-ml-py package (see that issue for details)

Notes for Reviewers

These dependency pin changes should be low-risk

All of these pins and requirements are already coming through cuml's dependencies, e.g. cudf carries most of them via rapidsai/cudf#19806

So they shouldn't change much about the test environments in CI.

[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.

[jameslamb]

Most CI here is passing, just one job is failing:

FAILED test_naive_bayes.py::test_gaussian_parameters[1e-05-balanced] - ValueError: Number of priors must match number of classes.
FAILED test_naive_bayes.py::test_gaussian_parameters[1e-05-unbalanced] - ValueError: Number of priors must match number of classes.
...
FAILED test_naive_bayes.py::test_categorical_partial_fit[True-int32-float32] - assert 0.2518 <= (0.104 + 0.0001)
FAILED test_naive_bayes.py::test_categorical_partial_fit[True-int32-float64] - assert 0.2518 <= (0.104 + 0.0001)
...
FAILED test_naive_bayes.py::test_categorical_parameters[False-False-0.1-balanced] - ValueError: Number of classes must match number of priors
FAILED test_naive_bayes.py::test_categorical_parameters[False-False-0.1-unbalanced] - ValueError: Number of classes must match number of priors
...
= 40 failed, 14342 passed, 6174 skipped, 1210 xfailed, 39 xpassed, 1101 warnings in 1676.92s (0:27:56) =

(conda-python-tests-singlegpu build link)

Those test failures are only happening on 1 test configuration for that job:

• ✅ 12.2.2, 3.10, amd64, rockylinux8, l4, earliest-driver, oldest-deps
• ✅ 12.9.1, 3.13, amd64, ubuntu24.04, h100, latest-driver, latest-deps
• ❌ 12.0.1, 3.12, arm64, ubuntu22.04, a100, latest-driver, latest-deps

It looks like the same tests are failing for that configuration on an unrelated PR, so I suspect this isn't related to this PR's changes.

• https://github.com/rapidsai/cuml/actions/runs/17407122771/job/49416123636#step:10:6895 (from Prevent CUDA issues when running UMAP with scikit-learn compatibility test suite #7107)

[jameslamb]

Tracking issue for the test failures: #7152

[jameslamb]

Thanks for restarting CI here @csadorf !

If that passes, I'll merge this and get the CUDA 13 PR (#7128) ready for review.

[jameslamb]

🎉 that worked! Thanks @csadorf and @betatim !

[jameslamb]

/merge