Require numba-cuda>=0.16.0#19213
Conversation
|
Seems like this change will require |
|
Here's the solution we need: conda-forge/libnvjitlink-feedstock#28 |
numba-cuda>=0.15.2numba-cuda>=0.16.0
|
Tests currently fail because |
|
@bdice is this a DetailsBut I seem to be hitting |
|
I don't see any The failures I see in CI are like |
|
Nevermind. I see that |
| - *numba-cuda-dep | ||
| - *numba-dep |
There was a problem hiding this comment.
We can eliminate these test dependencies because dask-cudf tests no longer need numba / numba-cuda. I did a very small change in dask_cudf/tests/test_distributed.py to eliminate the sole use of numba there.
@rapidsai/cudf-dask-codeowners, can you approve this?
TomAugspurger
left a comment
There was a problem hiding this comment.
Approving for the dask-cudf changes.
Note that dask-cuda still has a dependency on numba (https://github.com/rapidsai/dask-cuda/blob/7d9e28af0b08d9819c254eb70ef3a3b16dc7e448/dependencies.yaml#L145), so if you were hoping this would avoid the need for it in the test environment we aren't quite there yet unfortunately.
|
I misunderstood this at first. I thought the third-party tests were defined externally (upstream in stumpy). The source code for that third-party test is in this repo. I went ahead and fixed it in 746ede2. |
|
There's a couple pandas tests failing here still. I haven't managed to reproduce locally yet but I'm going to try with local CI next. |
|
@brandon-b-miller These pandas tests are failing on the nightly run, too: https://github.com/rapidsai/cudf/actions/runs/16187163471 This will probably block other PRs. We can xfail them, perhaps? |
|
Comparing to a recently successful job, the main difference I would suspect is |
They're XPASS'ing, so I think we should remove them from the list of xfails. |
|
That seems right. If the issue is truly between third party libraries it shouldn't block. |
Fixes nightly CI https://github.com/rapidsai/cudf/actions/runs/16187163471. xref #19213 Authors: - Matthew Murray (https://github.com/Matt711) Approvers: - Bradley Dice (https://github.com/bdice) - https://github.com/brandon-b-miller - GALI PREM SAGAR (https://github.com/galipremsagar) URL: #19341
|
NVVM now seems to be missing on the conda side. Looking in to why |
|
@bdice I think the compiler dependencies might still be needed on the conda side here. |
|
@brandon-b-miller There seems to be a bug in cuda-python packaging. Tracking it here: conda-forge/cuda-python-feedstock#129 I will push a fix shortly. |
| ) | ||
| all_gpu_devices = [ | ||
| device.id for device in cuda.list_devices() | ||
| int(device.id) for device in cuda.list_devices() |
There was a problem hiding this comment.
| int(device.id) for device in cuda.list_devices() | |
| # TODO: Revert `int(device.id)` to `device.id` once | |
| # https://github.com/NVIDIA/numba-cuda/pull/319 is released and | |
| # included in the minimum required numba-cuda version. | |
| int(device.id) for device in cuda.list_devices() |
|
/merge |
Update to numba-cuda>=0.16 [to match cuDF](rapidsai/cudf#19213), and update stream synchronization code in distributed-ucxx. Closes rapidsai/dask-upstream-testing#71. Authors: - Peter Andreas Entschev (https://github.com/pentschev) Approvers: - Bradley Dice (https://github.com/bdice) - Tom Augspurger (https://github.com/TomAugspurger) URL: #462
This reverts commit a9c395d.
- Partially reverts #19213 - Use numba-cuda>=0.15.2,<0.16 We are investigating an issue where the new CUDA bindings in numba-cuda 0.16 cause segfaults, blocking our CI. This temporarily downgrades to numba-cuda >=0.15.2,<0.16 until a fix can be made. Authors: - Bradley Dice (https://github.com/bdice) Approvers: - https://github.com/brandon-b-miller - Peter Andreas Entschev (https://github.com/pentschev) - Gil Forsyth (https://github.com/gforsyth) URL: #19413
numba-cuda0.15.0switched numba to use the NVIDIA bindings by default, and0.15.2includes a few patches that ameliorated knock-on effects. We should update things to leverage these changes and if anything is broken, fix. This PR now bumps the version to>=0.16.0to include the latest release.