| name | areno-debug-runtime |
| description | Diagnose failed, hung, slow, OOM, NaN, illegal-memory-access, NCCL, compilation, rollout, or training runs in AReno. Use when runtime evidence must identify the first causal stage. Do not use for routine capacity planning without a failure. |
Debug AReno Runtime
Do not modify parameters or code until the lifecycle stage and first causal error are identified.
When diagnosis requires a code fix, implement and commit it locally, then pull
the branch on the remote reproduction host. Do not hot-patch the remote source.
Primitives
python .agents/skills/areno-debug-runtime/scripts/summarize_traceback.py run.log
python .agents/skills/areno-debug-runtime/scripts/process_snapshot.py --pid <pid>
python .agents/skills/areno-debug-runtime/scripts/inspect_core.py <core-file> --executable <python>
Use py-spy dump -p <pid> for a Python-side stall. Use Nsight Systems only after a bounded steady-state workload exists. Read references/error-taxonomy.md and references/nan-triage.md as applicable.
Workflow
- Preserve exact command, commit, config, earliest logs, and worker exit data.
- Classify config/load, prefill, decode, reward, scoring, train forward, backward, optimizer, save, or distributed teardown.
- For multi-rank output, group signatures and prioritize the earliest distinct exception over secondary NCCL watchdog failures.
- Distinguish compilation/autotune work from deadlock using elapsed time and stacks.
- Reproduce with the same semantic workload at the smallest topology that still fails.
- Correct the owning layer. Required kernels must fail loudly; do not introduce silent fallback.
- Re-run the reproduction and the original bounded path.
Report evidence, root cause, changed ownership boundary, and verification. A process exit without the first error is incomplete diagnosis.
