What problem does it solve? Serving benchmarks and scaling decisions for LLM inference are routinely invalidated by saturation-ceiling runs, missing warmups, GPU-utilization autoscaling, and unrepresentative synthetic load, producing capacity numbers that collapse in production. ## Core Features & Use Cases - Benchmark design and review: Builds rate-sweep benchmarks with goodput SLOs, explicit readiness gates, warmup, noise floors, and load that matches production in length distribution, arrival shape, and prefix reuse. - Regression diagnosis: Attributes TTFT, TPOT, and throughput regressions to concrete causes such as prefix-cache collapse, KV cache pressure, preemption, or chunked-prefill budget misconfiguration. - Capacity and scaling: Derives per-replica capacity from goodput collapse points, sizes cold-start budgets, and selects queue-depth or latency-based autoscaling signals instead of GPU utilization. - Configuration evaluation: Assesses changes like speculative decoding, memory utilization, parallelism strategy, and LoRA adapters with acceptance and withdrawal rules measured at production concurrency. - Use Case: Given a benchmark report claiming 7x H100 capacity for 30 req/s, the skill identifies that --request-rate inf measured only the saturation ceiling, warmup was skipped, and the synthetic 512/512 load misrepresents production, then prescribes the correct rate-sweep methodology. ## Quick Start Review this vLLM serving benchmark report and tell me whether the capacity conclusion is safe to plan against.
