name: benchmarking-profiling
description: "Plan and analyze SGLang benchmarks, throughput/latency runs,
profiling, request datasets, and reproducible performance triage."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
Benchmarking Profiling
Use this sub-skill for bench_serving, bench_one_batch, auto_benchmark, profiler/NVTX workflows, concurrency/request-rate experiments, dataset selection, warmup policy, GPU metrics, and regression triage. Keep model, backend, commit, GPU, prompt lengths, output lengths, and traffic shape fixed before comparing results.
Cross-Links
- Root routing:
../../SKILL.md
- Workflow map:
../../references/workflow-map.md
- Cross-cutting troubleshooting:
../../references/troubleshooting.md