What's next?¶

This spiel covered a very tiny slice of theseus's features. However, the remainder of our story today will be a choose-your-own-adventure. Take a look at the how-to guide sidebar and decide what you want to poke around in :)

I'll give you some ideas here as things that I'm particularly excited about in theseus, in no particular order... here goes:

Evaluations¶

Almost as easy as adding datasets, theseus allows you to add evaluations and run them inline during training, without having to worry about such rudimentary things as KV caches (:gasp!), rollout logistics (:gasp!), or perplexity calculations (:gasp!).

Pick your strategy and wire it into a trainer with the Evaluation System guide, or add your own evaluation.

Draw pretty pictures during and after training¶

Yk the time-travel debugging API I told you about in inspecting runs? You can actually formalize those little experiments into full-fledged jobs and attach them to historical nodes, right alongside the records from training.

Learn about the @analysis decorator, checkpoint artifacts, and plots during training in Analysis System. When you're ready to write one, head to Adding an analysis job.

The configuration system¶

Ask for any field, any time... once your component declares where it belongs. Your model, dataset, optimizer, and other ingredients bring their configuration fields along with them. There's no giant central config class to keep feeding.

Read about how the configuration system fits together, or add an experiment that puts it to work.

JuiceFS integration¶

It's a bit annoying to have to copy root directories around. How about we let our machines look at the same files instead?

The JuiceFS Integration guide covers mounting a shared root, using it from a notebook, and figuring out why the filesystem is giving you the cold shoulder.

Even more boxes to dispatch to¶

I covered a very minimal example of how to dispatch a theseus job remotely. There's like three platforms we support—SSH, SLURM, and Volcano K8s—and a billion configuration options. Meet them in Remote Configuration.

If you're picking hardware, here's the big ol' chip table. If you're deciding how those chips should share the work, here's the sharding design.

Make more things your own¶

Still itching to replace another part? Adding Things points you to models, experiments, datasets, evaluations, and analysis jobs.