Skip to content

ci: bump te-rocm-wheels artifact retention 1d -> 5d#673

Open
wenchenvincent wants to merge 1 commit into
devfrom
ci/wheel-artifact-retention
Open

ci: bump te-rocm-wheels artifact retention 1d -> 5d#673
wenchenvincent wants to merge 1 commit into
devfrom
ci/wheel-artifact-retention

Conversation

@wenchenvincent

Copy link
Copy Markdown
Collaborator

Summary

  • Bump retention-days on the te-rocm-wheels artifact upload in .github/workflows/rocm-wheels-build.yml from 1 to 5.

Why

The te-rocm-wheels build artifact is downloaded by every downstream sGPU / mGPU test job. mi35x self-hosted runners have observed queue times of 24–36 hours before a job actually starts, which races the previous 1-day retention window. Every mi35x job on recent PRs has failed at the download step with:

Unable to download artifact(s): Artifact not found for name: te-rocm-wheels

Observed on PR #667 run 29390452937:

Job Started (offset from build finish) Result
build (uploads artifact) 0 ✓ pass
mGPU Torch (mi30x) +21h35m ✓ pass
mGPU JAX (mi30x) +22h34m ✓ pass (~1h26m before expiry)
sGPU (mi30x) +23h39m ✗ artifact expired
sGPU (mi35x) +23h39m ✗ artifact expired
mGPU Torch (mi35x) +34h14m ✗ artifact expired
mGPU JAX (mi35x) +35h31m ✗ artifact expired

Bumping to 5 days matches the retention already used for test-log artifacts in rocm-ci.yml (also 5) and comfortably outlasts the mi35x queue tail.

Test plan

  • Confirm build job still uploads te-rocm-wheels successfully.
  • Confirm downstream sGPU / mGPU jobs on both mi30x and mi35x can download the artifact even after long queueing delays.

🤖 Generated with Claude Code

The te-rocm-wheels build artifact is consumed by sGPU and mGPU test
jobs downstream of the build. mi35x self-hosted runners can queue for
24-36h before starting, so the previous retention-days: 1 racing the
queue makes every mi35x job (and any slow-queueing mi30x job) fail at
the download step with "Artifact not found for name: te-rocm-wheels".

5 days matches the retention already used for test-log artifacts in
rocm-ci.yml and comfortably outlasts observed mi35x queue times.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@matthiasdiener

Copy link
Copy Markdown
Contributor

There was a discussion regarding this retention here: #591 Not sure if the concerns still apply.

@ipanfilo

Copy link
Copy Markdown
Collaborator

There was a discussion regarding this retention here: #591 Not sure if the concerns still apply.

Indeed. With QoLA+CK-JIT added resulted binary size deflated x3, so it should be OK to increase retention period to 2 or 3 days w/o extra actions. But further extending increases risk of running out of space.

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.

3 participants