| name | mi-ni |
| description | How to use the library code provided by mi-ni. Code design patterns that abstract compute to easily scale experiments. Read to learn about the `mini` package, the `Apparatus` class, hyperparameter schedulers, notebook/visualization utils, and how to author, run, and monitor memoized experiments from the CLI. These should be used by default. |
Library structure:
src/mini/
├── apparatus.py # Base Apparatus class
├── volume.py # Base Volume class
├── local_*.py # Apparatus that uses local compute and storage
├── modal_*.py # Apparatus that uses cloud GPU compute and storage
├── temporal/ # Advanced hyperparameter scheduling based on keyframes
└── vis/ # Visualization helpers
Two ways to compute
- Interactive
Apparatus (app.map/app.arun) — a blocking call inside a notebook; dies with the process. Use for quick, light work you watch finish.
- Memoized orchestration (
Experiment(main=...), driven by the mini CLI) — detached, durable, pollable across short-lived processes. Use for sweeps, multi-step pipelines, anything slow, and anything an agent runs autonomously.
Authoring, running, monitoring
- Author a memoized experiment — the
main(ctx) DAG, repo layout, and cache-friendly design: authoring.md. The identity/evidence model behind caching: memoization.md.
- Run & monitor one from the CLI — the wake-loop, backend routing, wall-clock budget, and how to delegate/schedule a long run: running.md.
- Fix & recover — the fix/prune/retry loop, bounded hotfixes, superseded records, and partial
map failures: recovery.md.
- Store & share large outputs — return
Artifact handles instead of volume paths, share artifacts across experiments by name, and publish artifacts to a URL: storage.md.
- Publish a report to the web — externalize a report's figures and data into a bundle: reports.md; the
themed figure hook that feeds it: vis.md.
To keep cost down, delegate launching and babysitting to the experiment-monitor subagent (it escalates to experiment-doctor); see running.md.
Apparatus and Volume
mini provides experiment infrastructure via the Apparatus class. Its interface is similar to an Executor, but it abstracts compute and storage. See apparatus.md, especially if you're using it in a notebook.
Hyperparameter scheduling
mini.temporal provides advanced hyperparameter scheduling. An experiment may define the schedule using a "dopesheet" (table of keyframes); this is then interpolated to a timeline with one value per parameter per step. Useful for adjusting learning dynamics over the course of training.
Visualization helpers
mini.vis provides utilities for figure theming (base styles, light/dark support). See vis.md. Externalizing those figures and a report's data assets so the exported HTML stays light and publishes off Git LFS is a report concern: reports.md.
