ci: isolate PR230 SDMA initialization failures (do not merge)

[vegetabledoww]

Diagnostic run for #230. Do not merge.

The full-log prefix-cache reproducer failed before serving became healthy: SDMA warmup returned 507034, with CANN reporting vector core task timeouts on multiple devices. The original prefill TENSOR_WAIT_TIMEOUT remains unresolved.

This revision runs model-free ChipWorker initialization probes on the eight task-submit assigned devices. Each probe uses a fresh process. All SDMA-disabled controls run first, followed by SDMA-enabled probes, sequentially per device. This tests whether initialization fails without model weights, inference kernels, or simultaneous multi-device startup. Each process has a 240-second timeout.

Artifacts preserve per-device stdout and CANN logs, JSON exit codes and durations, the warmup binary SHA256, and pinned source revisions. Pipeline failure propagation is explicit; a failed or timed-out probe fails CI.

Runtime and toolchain pins remain the same as the failing run. No model code, runtime implementation, or original skill PR files are changed. The normal model matrix is replaced only on this temporary diagnostic branch.

Validation: Python syntax, workflow YAML parsing, shell syntax, and git diff checks passed locally. Hardware probes run only on upstream CI.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}