메뉴
Find slowest PyTorch modules in recent steps
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
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
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
SOC 직업 분류 기준
| name | module_bottleneck |
| description | Find slowest PyTorch modules in recent steps |
| category | performance |
| tables | ["python.torch_trace"] |
| tags | ["slow","bottleneck","module","torch","慢","瓶颈"] |
| keywords | {"en":["slow","bottleneck","hotspot","which module","slowdown"],"zh":["慢","瓶颈","哪个模块","变慢","hotspot","热点"]} |
| parameters | {"recent_steps":{"type":"integer","default":10},"stage_filter":{"type":"string","default":"post forward"}} |
基于 python.torch_trace 的 post-hook duration，找出最近 step 中最耗时的模块。 与 torch.profiler 互补：这是长期采样、模块级、低开销视图。
recent_steps (integer, default 10): Analyze the last N training stepsstage_filter (string, default post forward): Hook stage to aggregate (post forward | post step)