vllm-bench-serve — Online Benchmark Orchestrator

1. Scope & Boundaries

This skill does:

• Execute single or batch vllm bench serve online benchmarks against running inference services
• Aggregate and compare results across multiple test cases
• Auto-optimize: search for optimal concurrency/throughput given SLO constraints

This skill does NOT do (defer to other skills or decline):

• Start/deploy vLLM services → use vllm-ascend-server
• Offline batch inference throughput tests → use vllm-ascend
• Profiling / tracing (torch profiler, perfetto, NPU profiling) → out of scope
• Health checks only (just checking /v1/models) → simple curl, no skill needed
• Download/clean/convert datasets without running benchmarks → out of scope
• Analyze existing benchmark results without running new tests → out of scope