theseus-cli

What problem does it solve?

This Skill streamlines the end-to-end workflow for running research training jobs with the theseus CLI by helping users generate and edit YAML configs, run experiments locally, dispatch to remote clusters, start remote REPLs, create bootstrap scripts, and manage checkpoints and restores.

Core Features & Use Cases

• Config generation: Create canonical YAML job configs from registered job schemas and merge previous configs with overrides.
• Local execution: Run jobs locally with correct output root handling and restore from checkpoints.
• Dispatch & orchestration: Submit jobs to clusters (SLURM, SSH, TPU, Volcano), start remote Jupyter REPLs, produce bootstrap scripts, and manage hardware requests and dispatch options.
• Operational guidance: Provides best practices for tokenization, sweeps, naming, and safety (always show and confirm submit commands).
• Use case: Generate a pretraining config for gpt/train/pretrain, tokenize data locally, and dispatch a multi-GPU pretraining job to an H100 cluster with checkpointing enabled.

Quick Start

Run uv run theseus configure gpt/train/pretrain train.yaml --chip h100 -n 8 to generate a config skeleton, inspect and edit the YAML as needed, then run or submit after confirming the full command.