Skip to content

Move V100/MI210/CPU release image builds to GitHub Actions#134

Merged
fluidnumericsJoe merged 1 commit into
mainfrom
feature/gha-v100-mi210-cpu-publish
Jun 14, 2026
Merged

Move V100/MI210/CPU release image builds to GitHub Actions#134
fluidnumericsJoe merged 1 commit into
mainfrom
feature/gha-v100-mi210-cpu-publish

Conversation

@fluidnumericsJoe

Copy link
Copy Markdown
Member

Summary

Consolidates all release-image publishing into the GitHub Actions docker-publish workflow and retires the buildkite main-branch release pipeline.

  • Added to the docker-publish matrix (double- + single-precision each):
  • Dockerfile fixes to support the new single-precision variants:
    • Wired SELF_DOUBLE_PRECISION into docker/x86_sm70/Dockerfile and docker/x86/Dockerfile.
    • Fixed docker/x86/Dockerfile's stale default base latest-x86latest-x86-none (the actual published selfish CPU tag; latest-x86 404s).
  • Buildkite: removed the if: build.branch == "main" Release & Publish trigger from pipeline.yml and deleted the now-superseded release-and-publish.yml. Buildkite still runs the on-hardware GPU coverage tests on pull-request branches.

Tradeoff to note

The buildkite release ran on-hardware ctest (real V100/MI210 GPUs) before pushing. GitHub Actions has no GPU runners, so release images are no longer GPU-tested at publish time — the same tradeoff already accepted for the Blackwell (sm100/sm103) targets. PR-time GPU coverage tests are unaffected and remain on buildkite.

Notes

  • All 12 matrix variants validated as parseable YAML; all referenced selfish base tags confirmed to exist on Docker Hub.
  • No source/numerics changes.

🤖 Generated with Claude Code

Publish the MI210 (gfx90a), V100 (sm70), and x86 CPU release images from
the GitHub Actions docker-publish workflow alongside the existing Blackwell
(sm100/sm103) and MI300X (gfx942) targets, in both double- and
single-precision variants. MI210 was already migrated in #130; this adds
V100 (sm70) and x86 CPU.

Wire the SELF_DOUBLE_PRECISION build arg into the x86_sm70 and x86
Dockerfiles so their single-precision variants build correctly, and fix the
x86 (CPU) Dockerfile's stale default base image (latest-x86 -> the existing
latest-x86-none selfish tag).

Disable the buildkite main-branch Release & Publish pipeline and remove the
now-superseded release-and-publish.yml; buildkite continues to run the
on-hardware GPU coverage tests on pull-request branches. Note that release
images are no longer ctested on real hardware before publishing, matching
the existing behavior for the Blackwell targets.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@fluidnumericsJoe fluidnumericsJoe force-pushed the feature/gha-v100-mi210-cpu-publish branch from a19cbf4 to eec5901 Compare June 14, 2026 06:43
@fluidnumericsJoe fluidnumericsJoe merged commit b9b5ff6 into main Jun 14, 2026
18 of 19 checks passed
@fluidnumericsJoe fluidnumericsJoe deleted the feature/gha-v100-mi210-cpu-publish branch June 14, 2026 06:44
@codecov

codecov Bot commented Jun 14, 2026

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant