Sweeper Results | NVIDIA Dynamo Documentation

<!— Generated from aisimulate/docs/sweeper/results.md by docs/fern/scripts/sync_aisimulate_docs.py. Edit the canonical source instead of this Fern copy. —>

Experimental. Sweeper’s replay and result contracts may change without a standard deprecation period.

Sweeper materializes every optimizer suggestion in the main process. It unrolls the backend selection, asks each configured provider for its concrete adapter configuration and runtime hooks, then constructs a ReplaySpec.

Replay Specification

ReplaySpec version 1 contains:

a BackendDeploymentSpec with topology, backend version, engine arguments, and worker counts;
the validated workload and optimization goal;
concrete concurrency when KV-load search derives it;
concrete adapter configurations and their runtime hooks.

RunnerCapabilities.require_compatible checks the version, backend/topology pair, and hooks before execution. canonical_json creates deterministic strict JSON and rejects non-finite values.

Candidate Output

For a scalar goal, Sweeper.run returns feasible Candidate objects sorted best-first. Each candidate contains:

Field	Meaning
`config`	unrolled backend sample plus nested concrete adapter configuration
`used_gpus`	total GPUs assigned to the deployment
`metrics`	normalized values returned by replay
`score`	objective normalized so larger is better
`objectives`	raw per-objective values for Pareto searches; otherwise `None`

For goal.target: pareto, the result contains only non-dominated candidates and preserves each objective’s natural direction.

1 candidates = sweeper.run(config)
2 best = candidates[0]
3 print(best.config)
4 print(best.metrics)

Exact repeated suggestions reuse a result from the current run call. The cache does not persist between calls, even when the same Sweeper instance is reused.