WIP: [NOT READY FOR REVIEW] enforce a floor on libnvjitlink, build wheels with CUDA 13.0.x, test wheels against mix of CTK versions#21671
Conversation
…wheels against mix of CTK versions
jameslamb
added
improvement
|
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. |
github-actions
bot
added
Python
|
/ok to test |
|
/ok to test |
|
/ok to test |
|
(edit) Ugh, meant to |
|
/ok to test |
| RAPIDS_PY_WHEEL_NAME="${RAPIDS_PY_CUDA_SUFFIX}" RAPIDS_PY_WHEEL_PURE="1" rapids-get-pr-artifact --pkg_name distributed-ucxx ucxx 604 python wheel | ||
| ) | ||
| LIBUCXX_WHEELHOUSE=$( | ||
| RAPIDS_PY_WHEEL_NAME="libucxx_${RAPIDS_PY_CUDA_SUFFIX}" rapids-get-pr-artifact ucxx 604 cpp wheel "${UCXX_COMMIT}" |
There was a problem hiding this comment.
😬 I hit GitHub's rate limits
Downloading and decompressing ucxx_wheel_cpp_libucxx_cu12_x86_64 from Run ID into /tmp/tmp.FFhweajGIk
error fetching artifacts: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B4E0:1E3815:1337C1F:517B113:69AF40CF and timestamp 2026-03-09 21:51:11 UTC. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https://docs.github.com/en/site-policy/github-terms/github-terms-of-service) (https://api.github.com/repositories/617172783/actions/artifacts?per_page=100&page=322)
Look at the limits printed in that job... this repo was pushing it already:
"core": {
"limit": 15000,
"used": 14483,
"remaining": 517,
"reset": 1773094676
},
Seeing that ?page=322 in the message is concerning... nothing in our artifact-fetching should be paging over that many results!
I think the problem is that the run wasn't found. Notice "from Run ID into" (empty space!).
I'll put up a fix for that.
There was a problem hiding this comment.
Fixing that over here: rapidsai/gha-tools#249
|
/ok to test |
I've been using `rapids-get-pr-artifact` a lot recently, for rapidsai/build-planning#257 Today, one of my PRs for that briefly caused `cudf` to briefly exceed its GitHub API limits: rapidsai/cudf#21671 (comment) This proposes some fixes to `rapids-get-pr-artifact` in the future: * error out of a unique GitHub Actions run ID can't be determined from the inputs - *(instead of passing an empty string to `gh run download`, which causes it to page over **all CI runs GitHub's history retains**)* * limit initial search for runs to just `pr.yaml` workflow - *(optionally overridable by env variable `RAPIDS_BUILD_WORKFLOW_NAME`)* These should make `rapids-get-pr-artifact` faster and, more importantly, use many fewer GitHub API calls. ## Notes for Reviewers ### Haven't I seen this before? Yep! Similar changes were made in `rapids-github-run-id`: * #193 * #170 `rapids-get-pr-artifact` doesn't use `rapids-github-run-id` because its behavior of "use environment variable `GITHUB_RUN_ID` if it's available" is undesirable when you're trying to workflow runs in a repo other than the one `rapids-get-pr-artifact` is currently running in. ### How I tested this Tested on rapidsai/cudf#21671, with rapidsai/cudf@c058c43 Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Gil Forsyth (https://github.com/gforsyth) URL: #249
1 participant
Description
Contributes to rapidsai/build-planning#257
nvidia-nvjitlink>={whatever-minor-version-they-were-built-against}Contributes to rapidsai/build-planning#256
Other changes
libnvjitlink>={whatever-version-they-were-built-against}Notes for Reviewers
How I tested this
This uses wheels from similar PRs from RAPIDS dependencies, at build and test time:
Checklist