Migrate RAPIDS to `numba-cuda>=0.22.1,<0.23.0`

[brandon-b-miller]

Issue tracking the task of upgrading RAPIDS to a newer version of numba-cuda.

• RMM Update to numba-cuda >=0.22.1,<0.23.0 rmm#2168
• cuDF Update to numba-cuda >=0.22.1,<0.23.0 cudf#20750
• dask-cuda Update to numba-cuda >=0.22.1,<0.23.0 dask-cuda#1591
• ucxx Update to numba-cuda >=0.22.1,<0.23.0 ucxx#554
• cuML Update to numba-cuda >=0.22.1,<0.23.0 cuml#7549

Merge order: RMM -> cuDF -> Everything else

[jameslamb]

Thanks for opening this! I assigned you and set the target release to v26.02, so it'll show up in the roadmap here: https://github.com/orgs/rapidsai/projects/116/views/15

[bdice]

I think we're going to want NVIDIA/numba-cuda#621, which is coming in release 0.21.3.

[brandon-b-miller]

I think we're going to want NVIDIA/numba-cuda#621, which is coming in release 0.21.3.

Correct, thanks @bdice . This mirrors what I saw locally earlier.

[brandon-b-miller]

Quick daily update here. Version 0.21.3 resolves the deferred typing context initialization issue.

Currently working through the steps of converting cuDF's numba extension over to the numba.cuda namespace, this is yielding at least one bug (either in numba-cuda, or cuDF, not sure which yet). This will need to be resolved before moving forward and I'm currently looking into it.

[brandon-b-miller]

Update here

RMM
The failing tests were due to a breaking change where the NUMBA_CUDA_USE_NVIDIA_BINDING variable no longer exists causing rmm to take the legacy ctypes codepath where the desired numba version no longer provides ctypes bindings. This issue is fixed in the latest commit and the tests now pass.

cuDF
Changes in the latest commit rapidsai/cudf@eb5a77d pass tests locally, with the exception of a few groupby tests that exposed a numba issue being fixed in NVIDIA/numba-cuda#634.

dask-cuda
Failing tests are due to requiring the rmm changes outlined above.

ucxx and cuml
These libraries generally pass if cuDF passes, but I have not yet done local testing to verify.

Plan
Merge the outstanding numba-cuda PR putting cuDF in a position where it can get a numba release that passes all cuDF tests when used. At that point, the RMM, cuDF, and dask-cuda tests should should all be in a state that locally pass tests with the most recently released version of numba-cuda.

We'll need to merge the RMM PR first and publish packages, then admin merge the dask-cuda PR and publish packages again. From there, the cuDF CI will hopefully run clean. If anything pops up, we'll evaluate if it can be fixed before broken CI starts to have a nontrivial impact, and if not, we can roll back the RMM and dask-cuda updates to unblock CI while we investigate.