vllm-ascend-benchmark · diff
git:20260911.9f92063 to git:20260911.0ddd029
1 added, 1 removed. Audit A to A.
---
name: vllm-ascend-benchmark
- description: Run vLLM online-serving benchmarks on a workspace-managed remote container. Use for requests like "跑个 benchmark", "对比性能", "压测一下", "测下吞吐", or "看下有没有性能回退". Do not use for accuracy tests, nightly CI matrix runs, offline inference, or service-only lifecycle.
+ description: Measure throughput and latency of a vLLM online service with a specified workload, including request-rate and concurrency sweeps. Use for 跑 benchmark, 压测, or 测吞吐. Code baseline-versus-candidate comparisons and performance regression decisions use vllm-ascend-performance-regression; accuracy and service lifecycle have separate workflows.
---
# vllm-ascend-benchmark
Measure a vLLM service with one or several benchmark iterations and return raw and normalized metrics.
Choose input/output lengths, concurrency, request rate and endpoint for the intended workload. User choices override presets and nightly examples. Report variance and failures alongside throughput and latency.
## Agent entry
Run from the repository root using the platform's Python launcher. The workspace
selects its installed platform environment automatically.
```text
python .agents/skills/vllm-ascend-benchmark/scripts/bench_run.py --model /models/example --runs 3 --warmup-runs 1
```
Use --execution-id to measure an existing service, or let the workflow start and clean up its own service. --serve-args and --bench-args forward business options; --preset supplies reusable defaults. The managed interpreter and actual launch observations are recorded with measurements.
Use performance-regression for code comparisons: it binds actual local worktrees and handles alternating runs. Use correctness-validation for accuracy claims.
Read the relevant detail only when needed:
- [behavior](references/behavior.md)