Fix CI container name collision for parallel matrix jobs

[sunway513]

Summary

• Multiple matrix jobs sharing the same runner used a hardcoded container name atom_test, causing cleanup steps to kill sibling jobs' containers and leading to intermittent GPU OOM errors.
• Use CONTAINER_NAME=atom_test_${{ strategy.job-index }} so each job operates on its own container without interference.
• GPU isolation is already handled by gha-render-devices — the container name collision was the sole root cause.

Changes

• Added CONTAINER_NAME env var using strategy.job-index for unique per-job container names
• Replaced all 12 hardcoded atom_test container references with $CONTAINER_NAME
• Used exact-match filter name="^${CONTAINER_NAME}$" in cleanup to avoid interfering with sibling jobs
• Docker image tag atom_test:ci left unchanged (shared image is fine)

Test plan

• Verify CI passes with parallel jobs on shared runners (linux-atom-mi355-1)
• Confirm no container name collisions in job logs
• Verify cleanup step only stops/removes its own container

[sunway513]

@gyohuangxin do you think this change is relevant?

[gyohuangxin]

@gyohuangxin do you think this change is relevant?

I think it makes sense, could you convert this PR to ready to let CI verify it?

[ChuanLi1101]

LGTM, thanks.

[sunway513]

updated, plz review again @ChuanLi1101

[Copilot]

Pull request overview

This PR attempts to fix container name collisions when multiple matrix jobs run in parallel on the same GitHub Actions runner by introducing unique per-job container names. The approach uses strategy.job-index to differentiate containers, replacing 12 hardcoded references to atom_test with a dynamic $CONTAINER_NAME variable.

Changes:

• Added CONTAINER_NAME environment variable using strategy.job-index for unique container names
• Updated all container lifecycle commands (stop, rm, exec) to use $CONTAINER_NAME
• Modified cleanup filters to use exact-match patterns for container isolation

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Docker's --filter name= does not support regex anchors like ^ and $. The filter performs a substring match by default. To prevent matching containers like atom_test_1_backup when looking for atom_test_1, you should rely on exact matching by ensuring unique enough names. However, since the base approach using strategy.job-index is not valid, this will need to be reconsidered once the container naming strategy is corrected.

Suggested change

	(docker ps -aq -f name="^${CONTAINER_NAME}$" | xargs -r docker stop) || true
	(docker ps -aq -f name="^${CONTAINER_NAME}$" | xargs -r docker rm) || true
	(docker ps -aq -f name="${CONTAINER_NAME}" | xargs -r docker stop) || true
	(docker ps -aq -f name="${CONTAINER_NAME}" | xargs -r docker rm) || true

[Copilot]

strategy.job-index is not a valid GitHub Actions context variable. GitHub Actions does not provide a built-in job index for matrix jobs. Consider using a combination of matrix values to create unique container names instead. For example:

• Use github.run_id with a sanitized version of matrix.model_name
• Or use github.run_id with github.run_attempt and sanitized matrix values

Example: CONTAINER_NAME: atom_test_${{ github.run_id }}_${{ github.run_attempt }}_${{ matrix.model_name }}

Note: You'll need to sanitize matrix.model_name to remove special characters that are invalid in container names (e.g., replace spaces and special chars with underscores).

Suggested change

	CONTAINER_NAME: atom_test_${{ strategy.job-index }}
	CONTAINER_NAME: atom_test_${{ github.run_id }}_${{ github.run_attempt }}_${{ toLower(replace(replace(replace(replace(matrix.model_name, ' ', '_'), '/', '_'), ':', '_'), '.', '_')) }}

[ChuanLi1101]

Follow the Co-Pilot suggestion. Made a suggestion to avoid a low probability collision. I think you can merge when fix it. Approved in advance for convenience.

[ChuanLi1101]

Follow the co-pilot: Consider change CONTAINER_NAME: atom_test_${{ strategy.job-index }}
to:
CONTAINER_NAME: atom_test_${{ github.run_id }}_${{ strategy.job-index }}

The reason is strategy.job-index is only unique within a single workflow run. If two PRs trigger CI at the same time and are scheduled on the same runner, both runs may have job-index=0, resulting in both generating atom_test_0 and causing a naming collision. Adding github.run_id. Even if two workflow runs execute concurrently on the same runner, they will not collide.

[valarLip]

@gyohuangxin maybe you can help resolve the conflict? so we can merge this one