菜单
One-shot health check: CPU, GPU, tables, torch step progress, TorchProbe overhead, cluster nodes, NCCL profiler health
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
NCCL proxy wait decomposition — culprit (send_gpu_wait) vs victim (recv_wait)
SRE first-response triage for distributed training incidents. Automates the manual checks from PyTorch/NCCL debugging runbooks.
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
基于 SOC 职业分类
| name | health_overview |
| description | One-shot health check: CPU, GPU, tables, torch step progress, TorchProbe overhead, cluster nodes, NCCL profiler health |
| category | triage |
| tables | ["cpu.utilization","gpu.utilization","python.torch_trace","python.torch_step_timing","cluster.nodes","nccl.profiler_counters"] |
| tags | ["overview","doctor","triage","健康检查"] |
| keywords | {"en":["health","overview","status","doctor","checkup"],"zh":["健康","概览","怎么样","正常吗","doctor","体检"]} |
| parameters | {"sample_limit":{"type":"integer","default":5}} |
快速扫描目标进程：有哪些表、CPU/GPU 最近采样、训练是否在推进、TorchProbe shadow 开销、集群节点与 NCCL profiler 计数器。 适合作为 agent 的默认第一步，或用户不确定从哪查起时使用。
sample_limit (integer, default 5): Recent CPU/GPU samples to showavailable_tables — information_schema 可见表host_cpu / gpu_mem — 最近 CPU/GPU 采样latest_torch_step — python.torch_trace 最新 local_steptorch_probe_overhead — python.torch_step_timing dispatch/hook tax（80-step 窗口）cluster_nodes — cluster.nodes 存活（若有）nccl_data_integrity — nccl.profiler_counters 最新一行（若有 NCCL profiler）