Skip to content

[CI] Harden ARM64 package access and Docker disk usage#2949

Closed
neverhook wants to merge 2 commits into
kvcache-ai:mainfrom
neverhook:codex/ci-harden-arm64-apt
Closed

[CI] Harden ARM64 package access and Docker disk usage#2949
neverhook wants to merge 2 commits into
kvcache-ai:mainfrom
neverhook:codex/ci-harden-arm64-apt

Conversation

@neverhook

@neverhook neverhook commented Jul 16, 2026

Copy link
Copy Markdown

Summary

  • switch Ubuntu Ports entries from HTTP to HTTPS before the ARM64 CUDA install
  • force IPv4 and add apt retry/timeout options
  • make the NVIDIA keyring download use IPv4 with bounded retries
  • free hosted-runner disk before the source Docker build and report disk usage before/after
  • make the Docker build's existing Release configuration explicit and enable plain progress logs

Motivation

ARM64 package access

The ARM64 matrix fails before Mooncake is configured or built. In PR #2934 run 29479306961, Python 3.10 failed in Install CUDA Toolkit 13.0 (arm64 SBSA) and Python 3.12 was cancelled by matrix fail-fast. PR #2935 reproduced the same failure in run 29479320966. The completed logs show IPv6 is unreachable and all IPv4 connections to ports.ubuntu.com:80 time out while fetching Ubuntu packages.

Docker build capacity

PR #2934 run 29479306961 also failed while linking inside docker/mooncake.Dockerfile with /tmp ... No space left on device; pip then reported [Errno 28]. The same Docker job passed on the larger stacked #2935 old SHA, so this is hosted-runner capacity pressure rather than a source error. The cleanup removes large unused hosted-toolcache directories before Buildx starts and leaves disk diagnostics in every run.

This PR is intentionally separate from #2934 and #2935 because both failures are runner/CI reliability issues, not their feature code.

Validation

Local checks:

  • parsed .github/workflows/ci.yml with PyYAML
  • extracted the modified ARM64 and Docker run blocks and passed them through bash -n
  • git diff --check
  • typos-cli 1.30.2 .github/workflows/ci.yml

Remote ARM64 validation on commit 140d8e76 in run 29480368390:

  • build-arm64 (3.10): success, including CUDA installation, project build, wheel build, and artifact upload
  • build-arm64 (3.12): success, including CUDA installation, project build, wheel build, and artifact upload

A new full Actions run on head a06a81df will validate the added Docker cleanup and reconfirm ARM64.

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.

@neverhook neverhook changed the title [CI] Harden ARM64 CUDA package installation [CI] Harden ARM64 package access and Docker disk usage Jul 16, 2026
@neverhook

Copy link
Copy Markdown
Author

Superseded by #2951. The follow-up Docker disk hardening was pushed after the initial opened-event run, and this fork token cannot add the required run-ci label for synchronize events. #2951 carries the same commits and was opened fresh to trigger a complete validation run.

@neverhook neverhook closed this Jul 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant