| name | vllm-take-capture |
| description | Capture a short live vLLM profile from an already-running workload. Use when the user says "take a capture", "capture vLLM", "profile vLLM", or asks to grab a GPU+CPU vLLM trace from the local OpenAI server at 127.0.0.1:8000, automatically bounded to 4 scheduler/engine steps, then copy rank 0 to lukea@orion.local:~/Desktop without probing or health-checking the instance. |
vLLM Take Capture
Default Workflow
Run the exact helper command immediately:
python skills/vllm-take-capture/scripts/take_vllm_capture.py --profile-dir /tmp/vllm-ds4-decode
Do not do preflight work first. That means no test -f, find, rg, ls,
server probing, cleanup, endpoint discovery, /health, /v1/models, metrics,
or other validation before the capture command. Assume the helper path exists,
the profile directory is /tmp/vllm-ds4-decode, the workload is already
running, and the server is http://127.0.0.1:8000.
The helper:
- Posts directly to
/start_profile with CPU+CUDA and 4-step hints.
- Does not send inference traffic.
- Relies on the running vLLM profiler config to auto-stop after 4 iterations.
- Finds the newest rank-0 trace emitted after the start request.
- Runs
scp -r <rank0 trace> lukea@orion.local:~/Desktop.
If Defaults Fail
Only debug after the direct helper command fails. If it reports a concrete
file-location problem, then adjust the profile directory:
python skills/vllm-take-capture/scripts/take_vllm_capture.py --profile-dir /absolute/profile/dir
Use --wait-timeout only if the four-step capture needs more time to flush.
Do not manually call /stop_profile unless the user explicitly asks for that
override; the intended path is the profiler's configured max_iterations=4
auto-stop.
Notes
vLLM's public OpenAI server profiling route starts/stops the profiler over HTTP; CPU+CUDA activity selection and step bounds are normally part of the server's profiler config. The helper still sends those values in the start payload because some local builds accept them, but do not spend time verifying the server's route shape before taking the capture.