Menu
SRE first-response triage for distributed training incidents. Automates the manual checks from PyTorch/NCCL debugging runbooks.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
NCCL proxy wait decomposition — culprit (send_gpu_wait) vs victim (recv_wait)
One-shot health check: CPU, GPU, tables, torch step progress, TorchProbe overhead, cluster nodes, NCCL profiler health
Diagnose PyTorch NCCL watchdog timeouts with Flight Recorder collective sequence alignment.
GPU memory and utilization headroom
Detect monotonic GPU memory growth across training steps
Find slowest PyTorch modules in recent steps
Based on SOC occupation classification
| name | sre_triage |
| description | SRE first-response triage for distributed training incidents. Automates the manual checks from PyTorch/NCCL debugging runbooks. |
| category | sre |
| tables | ["cluster.nodes","cpu.utilization","gpu.utilization","python.comm_collective","python.torch_nccl_flight_record","nccl.proxy_ops","rdma.mlx_hca","process.kmsg"] |
| tags | ["SRE","runbook","oncall","triage","NCCL","hang","OOM","network"] |
| keywords | {"en":["SRE","oncall","runbook","incident","triage","NCCL timeout","hang"],"zh":["SRE","值班","应急","runbook","分诊","事故","卡住","NCCL 超时"]} |
| parameters | {"use_global":{"type":"boolean","default":true},"seq_window":{"type":"integer","default":200},"sample_limit":{"type":"integer","default":20}} |
面向值班第一响应：把原本需要人工 ssh、grep、py-spy、看 NCCL/IB/GPU/内核日志的流程，尽量转成 probing 可执行证据链。
这个 skill 不追求一次性给出最终根因，而是自动回答：
incident_snapshot：当前有哪些表，决定能自动化到什么深度。cluster_nodes：rank/host 是否齐，是否有 dead/stale 节点。blocked_threads / kernel_wait_points：类似人工 py-spy/strace 的第一层替代。flight_recorder_alignment：自动化 Flight Recorder / fr_trace 的核心对齐判断。nccl_wait_summary：进入 NCCL 后的 culprit/victim 信号。rdma_network_hints / kernel_fault_hints：网络或硬件灰故障线索。watchdog_timeoutnccl_culprit_victimtraining_hangmemory_leak / gpu_pressure