aiperf-kube-sweep

Author a YAML file with a benchmark block and a sweep block defining parameter axes, then submit it with aiperf kube sweep --config sweep.yaml --image aiperf:latest. The operator creates one child AIPerfJob per variation and trial, run sequentially by a sweep-controller pod.

Set sweep.type to adaptive_search with planner bayesian or optuna, define a searchSpace with lo/hi bounds, objectives with metric and direction, and a required maxIterations between 2 and 200. The sweep-controller proposes one point per iteration and creates one child AIPerfJob each.

Credential-bearing axes such as API keys, tokens, passwords, and any endpoint.headers path are rejected before anything is created, for every sweep type. Keep credentials fixed and inject them via Secret-backed env vars like AIPERF_INJECTED_API_KEY using valueFrom.secretKeyRef in the pod template.

Sweeps are capped at 200 variations enforced by the operator at admission, and 10 trials per variation enforced as a client-side schema bound on multiRun.numRuns. Dry-run and validation pass on over-cap grids, so count the product of axis lengths before submitting.

No. Dry-run only renders the AIPerfSweep CR as JSON and returns before touching the cluster; it does not expand variations, show child counts, or check cardinality caps. Expand the sweep locally or compute the grid product yourself to preview variation counts.

Run aiperf kube cancel <sweep-name> for the whole sweep, or add -v <index> to cancel one variation. The running child drains to a terminal phase and still contributes results, remaining iterations are skipped, and aggregation runs over completed children.