Sweeps

Run a hyperparameter sweep and see which point won, through the engine your workspace already has.

This hub stores nothing. It reads the parameters and metrics your pipeline already declares, runs a trial per point through an installed sweep engine, and ranks what comes back. With DVC that costs no extra Python package at all — the grid and the worker pool are already part of it.

What you can do

• See what your pipeline can be swept over — the parameters it declares, with the value each one currently has.
• Try several values at once — type them per parameter and the panel says how many trials that adds up to before you start.
• Run the trials in parallel — pick how many go at a time.
• Watch them land — the list fills in while the sweep runs, and stops polling when it finishes.
• Rank by the number you care about — choose the metric and whether higher or lower is better.
• Take the winner — one click puts that point in your workspace and runs your MLOps pipeline over it.
• Stop a sweep — the workers shut down on request.

What it does not do

It does not guess which number is the point. A pipeline that records both an accuracy and a loss gives no clue about which one you are optimising or which direction is better, so the panel asks and never assumes. If your pipeline declares no metrics, trials still run and nothing is ranked, and the panel says so instead of inventing an order.

It does not cut a space down to fit. A sweep larger than 200 trials is refused with its number rather than trimmed, because a trimmed grid quietly stops varying one of your parameters.

Requirements

A sweep engine. extension-mlops-dvc 1.6.0 or newer ships the DVC one, and needs a workspace that is a git repository with params: and metrics: declared in dvc.yaml.

Getting started

1. Install this extension and a sweep engine.
2. Declare the parameters your pipeline reads and the metrics it writes.
3. Open Sweeps, type the values to try, and press Run.
4. When the trials land, click the best one to apply it and run your pipeline.

The engine contract

An engine is a set of declared capability ids. The hub finds it by reading the capability catalogue once; nothing is imported across extensions and nothing is registered by hand.

Id	Required	Meaning
mlops.<engine>.sweep	yes	queue a trial per point and start the workers
mlops.<engine>.trials	no	the space, the trials and the workers
mlops.<engine>.trialApply	no	put one trial's point in the workspace
mlops.<engine>.sweepStop	no	stop the workers
mlops.<engine>.reveal	no	the engine's own panel

The full contract is punica-editor/docs/sweep-engines.md.

Capabilities

Id	What it does
sweeps.run	Expand a space and queue one trial per point
sweeps.trials	The declared parameters and metrics, plus every trial
sweeps.apply	Apply a trial and run the pipeline over it
sweeps.stop	Stop the workers
sweeps.open	Reveal the panel

License

MIT