What problem does it solve? Choosing SGLang launch parameters for a specific model, GPU setup, and workload is guesswork that wastes GPU time and often ships suboptimal or SLO-violating configurations. This Skill runs an evidence-driven optimization loop that validates feasibility, benchmarks candidates, diagnoses bottlenecks, and emits a statistically confirmed launch command. ## Core Features & Use Cases - Automated SGLang tuning: Discovers hardware topology, regenerates the live ServerArgs parameter contract, screens candidate configurations, and confirms winners with repeated A/B and Bayesian evidence gates. - Bottleneck diagnosis: Captures bounded Nsight Systems traces and classifies bottlenecks across scheduling, KV cache, communication, MoE, and kernel layers before tuning. - Safe experimentation: Enforces trial, wall-time, and GPU-hour budgets, never kills foreign processes, and rejects candidates that fail correctness or SLO gates. - Use Case: Given a local model, 8 GPUs, and a p99 TTFT SLO, run the init/doctor/plan/run/report pipeline to receive a copy-paste SGLang launch command backed by measured evidence. ## Quick Start Ask the agent to create an inference-autopilot task for your local model and GPUs, then run doctor, plan, and run to produce an optimized SGLang launch command and report.
