ci: split GPU builds per-arch on native runners (fix runner disk exhaustion, keep arm64)#46
Merged
Merged
Conversation
…haustion The single-runner amd64+arm64 CUDA builds (arm64 under QEMU) ran out of disk on the large CUDA image. Build each arch on its own native runner (amd64 on ubuntu-latest, arm64 on ubuntu-24.04-arm -- no QEMU), push by digest to GHCR, then assemble one multi-arch manifest per tag and mirror it to Docker Hub. Keeps arm64 a first-class target (NVIDIA Spark, Apple Silicon) instead of dropping it. Applies to build-cuda.yml and build-cuda-spatial.yml. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
This was referenced Jun 16, 2026
Replace the cross-registry Docker Hub copy (400 Bad Request) with the pattern validated on build-ml: push each per-arch image by digest to both registries, record each registry's own digest, assemble the multi-arch manifest natively in each. Applies to build-cuda.yml and build-cuda-spatial.yml. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem
The GPU image builds run out of disk on the GitHub-hosted runner.
build-cuda.ymlandbuild-cuda-spatial.ymlbuildlinux/amd64,linux/arm64in a single job — the arm64 leg under QEMU emulation — so two full image trees of a large CUDA + R + RStudio (+ spatial) stack land on one ~14 GB runner. The minimalrm -rf android dotnet ghccleanup isn't enough, and the CUDA-Spatial build fails.Approach (keep arm64 — don't drop it)
arm64 is a first-class target here (NVIDIA Spark / Grace, Apple Silicon users). Rather than drop it, split the multi-arch build so each architecture builds on its own native runner:
linux/amd64onubuntu-latest,linux/arm64onubuntu-24.04-arm(native — no QEMU, faster, and each runner only holds one arch → fits disk). Each pushes by digest to GHCR.docker buildx imagetools createassembles a multi-arch manifest per tag on GHCR, then mirrors each tag to Docker Hub by copying the GHCR manifest. No leftover per-arch tags in the registry;:latestet al. stay multi-arch.Also bumps action versions (
build-push-action@v6,login-action@v3,checkout@v4) and adds CodeQL to the disk-free step.Validation needed before merge
Workflows can only be fully verified by running. Worth confirming on a CI run:
ubuntu-24.04-armhosted runners are enabled for the org (free for public repos; minutes otherwise).imagetools createassembles + mirrors all tags (base image fans out tolatest,cuda13.0-py3.12-rapids25.12,cuda13.0-py3.12,cuda13.0).docker buildx imagetools inspect).Same per-arch split could be applied to the CPU workflows (
build-ml*.yml) later to cut their ~1h QEMU build time — left out here to keep this focused on the failing GPU builds.