Use numba-cuda>=0.0.13 #16474
Use numba-cuda>=0.0.13 #16474
Conversation
gmarkall
added
non-breaking
numba-cuda is the NVIDIA-maintained CUDA target for Numba, which depends on the numba package. This PR replaces the numba dependency with numba-cuda. However, similar to @bdice's comment in #648, I don't know if we strictly need this here any more due to the cudf dependency (from rapidsai/cudf#16474) ensuring numba-cuda is present.
|
This looks right from the recipe edit side. You should only have to manually edit dependencies.yaml, and all the other files should then be generated. I don't understand the failing builds. I'm having a hard time finding the error that kills the build. Are you seeing this locally? |
|
Thanks @msarahan - unfortunately I edited the PR before I saw your comment - should I revert that last change? I haven't run this locally because I'm not really set up to build / test cudf at the moment (it's not what I usually work on), but will give it a try on Monday. |
|
Just noticed the problem, I introduced a space that shouldn't have been there. |
Co-authored-by: Mike Sarahan <[email protected]>
|
Thanks for looking at this @msarahan - I've just marked as "DO NOT MERGE" for now until I speak to Ben about this later in the Numba weekly sync - I want to first check whether there should be some discussion about the implications of this with other developers - will update here later. |
|
@msarahan Following the discussion in the cudf meeting last week the consensus seemed to be that this was OK to proceed with, so I've removed the "DO NOT MERGE" label, fixed up the conflicts, and I'm just waiting on CI. |
|
@gmarkall what else do we want to see on this PR before we merge? If we're happy with it, I'd encourage moving forward ASAP so we have some decent runway to test it before the 24.10 release. CC @bdice @brandon-b-miller |
|
Thanks Vyas - I don't think there's anything else we want to see in it, I just wanted to have made sure to have discussed it with the cuDF team before merging. |
|
Cool. Going to queue this up for merge then, the dep changes look fine. |
|
/merge |
|
Let's wait to adopt this until Then let's set the minimum version to the first version available on conda-forge, to avoid cross-channel constraints from causing problems. |
|
Given the delay with the conda-forge recipe and there now being conflicts on this branch, let's push this upgrade to 24.12. That will give us more time to iron out any kinks if anything goes unexpectedly wrong with the switch (particularly if it happens downstream of cudf). |
| @@ -720,7 +720,7 @@ dependencies: | |||
| matrices: | |||
| - matrix: {dependencies: "oldest"} | |||
| packages: | |||
| - numba==0.57.* | |||
| - *numba-cuda-dep | |||
There was a problem hiding this comment.
Note that I changed this so that we now require numba-cuda even to test the oldest supported dependency. If at some point we choose to support a range of versions then we should remove the alias here and insert a real version.
|
/merge |
Testing with https://github.com/NVIDIA/numba-cuda on CI. I am not sure if edits in other repos are required (e.g. I used to have to change an "integration" repo) - any pointers appreciated! Authors: - Graham Markall (https://github.com/gmarkall) - Vyas Ramasubramani (https://github.com/vyasr) - Bradley Dice (https://github.com/bdice) Approvers: - Mike Sarahan (https://github.com/msarahan) - Vyas Ramasubramani (https://github.com/vyasr) URL: #16474
**EDIT:** This now removes the numba dependency, which comes transitively from cuDF. **Original description:** numba-cuda is the NVIDIA-maintained CUDA target for Numba, which depends on the numba package. This PR replaces the numba dependency with numba-cuda. However, similar to @bdice's comment in #648, I don't know if we strictly need this here any more due to the cudf dependency (from rapidsai/cudf#16474) ensuring numba-cuda is present. Authors: - Graham Markall (https://github.com/gmarkall) - Bradley Dice (https://github.com/bdice) Approvers: - Bradley Dice (https://github.com/bdice) URL: #715
Description
Testing with https://github.com/NVIDIA/numba-cuda on CI.
I am not sure if edits in other repos are required (e.g. I used to have to change an "integration" repo) - any pointers appreciated!
Checklist