Move Linux CI jobs from linux_job_v2 to linux_job_v3 (OSDC) - #22107
Draft
huydhn wants to merge 3 commits into
Draft
Move Linux CI jobs from linux_job_v2 to linux_job_v3 (OSDC)#22107huydhn wants to merge 3 commits into
huydhn wants to merge 3 commits into
Conversation
Every Linux job except ROCm moves to the OSDC reusable workflow. ROCm has no OSDC hardware and maps to itself in arc.yaml, so rocm.yml stays on v2. Most of the diff is mechanical: the v2 -> v3 rename, EC2 runner labels swapped for their OSDC equivalents per pytorch/pytorch .github/arc.yaml, and the removal of use-custom-docker-registry, which v3 ignores. Two things are not. The first is the CI image. v2 resolved `ci-image:<name>` in-job through calculate-docker-image, which also waited for docker-builds to push it. v3 hands `docker-image` straight to the runner pod's container, which the kubelet pulls before any step of the job runs, so neither the resolution nor the wait can happen inside the job any more. The new _docker-image.yml does both: it emits `git rev-parse HEAD:.ci/docker`, and on a commit that touches .ci/docker it blocks until docker-builds has pushed. Callers depend on it and spell the image out as 308535385114.dkr.ecr.us-east-1.amazonaws.com/executorch/ci-image:<name>-<hash>, the same reference calculate-docker-image derived. Reusable workflows that take a docker-image input keep taking `ci-image:<name>` unchanged, so their callers did not have to move. The second is the jobs that relied on the EC2 host. doc-build's upload-gh-pages no longer needs run-with-docker: false, because v3 checks out and runs the script in the same container, so the credentials git push needs are already in the repo's git config. The two arm backend jobs that raise fs.inotify.max_user_watches now tolerate failure, since that is a node-level setting an unprivileged pod cannot change. Review order: _docker-image.yml first, then _unittest.yml as a small worked example of the caller change, then the rest. Note that arc.yaml still maps linux.arm64.2xlarge to l-arm64g2-6-32, which ci-infra renamed to l-arm64g2-6-25 in #625; this uses the deployed name. Authored with Claude Code.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22107
Note: Links to docs will display an error until the docs builds have been completed. ❌ 7 New Failures, 15 Pending, 5 Unclassified FailuresAs of commit 61eba98 with merge base 9a2d135 ( NEW FAILURES - The following jobs have failed:
UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Aug 24, 2026
This PR needs a
|
docker-builds loses its pull_request trigger, so on a PR it only runs once the ciflow/docker label tags the commit. Without that tag there is nothing to wait for, and the poll would sit for four hours before timing out. Give the tag five minutes to appear, then fail with what to do about it. Authored with Claude Code.
Contributor
Author
|
Two follow-ups pushed:
|
arc.yaml maps linux.arm64.2xlarge to l-arm64g2-6-32 and ci-infra renamed that def to l-arm64g2-6-25, but neither name schedules: the three arm64 legs of executorch's first OSDC docker-builds run sat queued for nearly four hours and were cancelled without a runner ever picking them up. mt-l-arm64g4-16-62 is the only arm64 label with live production evidence, and it is what pytorch/pytorch's own OSDC docker-builds uses. That also folds in the one job that was mapped to mt-l-arm64g3-16-62, which is the same size and equally unattested, so all arm64 work now lands on one proven label. Authored with Claude Code.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 2 of moving ExecuTorch's Linux CI to OSDC. Independent of #22106; #22108 stacks on this.
87 call sites across 24 workflows move to
linux_job_v3. ROCm stays on v2 — no OSDC hardware.Most of the diff is mechanical. The part worth review is the CI image: v3 hands
docker-imagestraight to the pod'scontainer:, which the kubelet pulls before any step runs, so v2's in-jobcalculate-docker-image— which both resolvedci-image:<name>and waited for docker-builds to push it — has nowhere left to run._docker-image.ymlreplaces both halves, and callers name the image in full. Reusable workflows keep takingci-image:<name>unchanged, so their own callers did not move.doc-build'supload-gh-pagesno longer needsrun-with-docker: false, since v3 checks out and runs the script in one container.Review order:
_docker-image.yml, then_unittest.ymlas a worked caller example, then the rest.Draft until a CI run shows what OSDC's PyPI reachability breaks —
cuda.ymlalready neededPIP_EXTRA_INDEX_URLcleared.Authored with Claude Code.