feat: Add ak.sort() for CUDA backend
#3750
Conversation
Codecov Report❌ Patch coverage is
Additional details and impacted files
🚀 New features to boost your workflow:
|
shwina
changed the title
ak.sort() for CUDA backendak.sort() for CUDA backend
|
The documentation preview is ready to be viewed at http://preview.awkward-array.org.s3-website.us-east-1.amazonaws.com/PR3750 |
| strategy: | ||
| fail-fast: false | ||
| matrix: | ||
| cuda-version: | ||
| - 11 | ||
| - 12 | ||
| - 13 | ||
|
|
||
| steps: | ||
| - name: Clean the workspace and mamba | ||
| run: | | ||
| rm -rf * .[!.]* || echo "Nothing to clean" | ||
| rm -rf ~/micromamba* || echo "Nothing to clean" | ||
| - uses: actions/checkout@v6 | ||
| with: | ||
| submodules: true | ||
|
|
||
| - name: Get micromamba | ||
| uses: mamba-org/setup-micromamba@v2 | ||
| with: | ||
| environment-name: test-env | ||
| init-shell: bash | ||
| create-args: >- | ||
| -c rapidsai | ||
| -c nvidia | ||
| python=3.13 | ||
| cccl-python | ||
| cudf | ||
| cupy | ||
| cuda-version=${{ matrix.cuda-version }} | ||
| cuda-toolkit | ||
| cxx-compiler |
There was a problem hiding this comment.
| strategy: | |
| fail-fast: false | |
| matrix: | |
| cuda-version: [11, 12, 13] | |
| steps: | |
| - name: Clean the workspace and mamba | |
| run: | | |
| rm -rf * .[!.]* || echo "Nothing to clean" | |
| rm -rf ~/micromamba* || echo "Nothing to clean" | |
| - uses: actions/checkout@v6 | |
| with: | |
| submodules: true | |
| - name: Get micromamba (CUDA <= 11) | |
| if: ${{ matrix.cuda-version < 12 }} | |
| uses: mamba-org/setup-micromamba@v2 | |
| with: | |
| environment-name: test-env | |
| init-shell: bash | |
| create-args: >- | |
| -c rapidsai | |
| -c nvidia | |
| python=3.13 | |
| cudf | |
| cupy | |
| cuda-version=${{ matrix.cuda-version }} | |
| cuda-toolkit | |
| cxx-compiler | |
| - name: Get micromamba (CUDA >= 12, add cccl-python) | |
| if: ${{ matrix.cuda-version >= 12 }} | |
| uses: mamba-org/setup-micromamba@v2 | |
| with: | |
| environment-name: test-env | |
| init-shell: bash | |
| create-args: >- | |
| -c rapidsai | |
| -c nvidia | |
| python=3.13 | |
| cccl-python | |
| cudf | |
| cupy | |
| cuda-version=${{ matrix.cuda-version }} | |
| cuda-toolkit | |
| cxx-compiler |
There was a problem hiding this comment.
Here is a hopefully correct fixed matrix that installs cccl-python only when CUDA ≥ 12, and avoids the solver error when CUDA = 11.
There was a problem hiding this comment.
@ariostas - given that we still want to use CUDA 11.
There was a problem hiding this comment.
I'm guessing CUDA 11 is EOL? I couldn't find any specific timeline, but @shwina would know. If that's the case let's just remove it and move on with 12 and 13.
There was a problem hiding this comment.
It looks like it is EOL - looking at https://docs.nvidia.com/datacenter/tesla/drivers/supported-drivers-and-cuda-toolkit-versions.html and specifically the Note: All other previous driver branches not listed in the table above (such as R575, R565, R560, R550, R525, R515, R510, R495, R465, R460, R455, R450, R440, R418, R410) are end of life.
There was a problem hiding this comment.
Would it be preferable to open a separate PR to remove CUDA 11 from the build/test matrix?
There was a problem hiding this comment.
I opened #3758. If/when that is merged, I can revert some of the CUDA 11 handling introduced in this PR.
shwina
force-pushed
the
add-cuda-backend-sort
branch
from
173b048 to
8fbd26f
Compare
shwina
force-pushed
the
add-cuda-backend-sort
branch
from
8fbd26f to
f434568
Compare
5 participants
Closes #3749.
Following the study #3748, this PR adds support for
ak.sort()for the CUDA backend, using https://nvidia.github.io/cccl/python/compute.html.