Performance Optimization Skills

The evidence-driven loop that benchmarks a confirmed baseline and challenges it with candidates.

以 Markdown 格式查看

These skills form the optimization workflow: capture a workload contract, benchmark a confirmed baseline, then challenge it with one candidate at a time until the Service Level Objectives (SLOs) are met or the budget runs out. The loop also uses deploy-dynamo-recipe, listed on the Deployment and Operations page, to deploy the confirmed baseline and each approved candidate.

A prompt that reaches them: “Optimize this deployment for output tokens per second per user under a 200 ms time-to-first-token SLO. Budget 8 GPU-hours and stop after three failed deployments.”

PhaseSkillUse for
Workload contractsynthesize-user-workloadCapture the workload requirements and immutable user-provided or confirmed baseline deployment.
Baseline authoringauthor-baseline-dgdDraft a baseline when no catalog recipe matches, then present it for explicit user confirmation.
Benchmark configurationconfigure-aiperf-benchmarkFreeze the AIPerf workload, objective, load policy, and Kubernetes execution manifest.
Benchmark executionrun-aiperf-benchmarkLaunch, monitor, debug, and collect one run-scoped AIPerf benchmark.
Results analysisanalyze-aiperf-resultsValidate and normalize results, evaluate SLOs, and compare compatible candidates.
Proposal selectionconsult-perf-knowledgeSelect one evidence-backed optimization proposal from the performance rules and engine guidance.
Candidate creationcreate-optimization-hypothesisApply only the selected change to create a challenger-ready deployment draft.
Adversarial reviewperform-adversarial-reviewReview a candidate for correctness, comparability, feasibility, attribution, and worthwhile GPU spend.

See the optimization loop for the full sequence and the evidence rules for benchmark validity requirements.