Getting Started¶
Theseus is a heterogeneous pretraining infrastructure and architecture experimentation platform.
Installation¶
First, clone the repo
For dependencies, we recommend using uv to perform installation; this depends on who your compute partner is...
- cuda13:
uv sync --group all --group cuda13 - cuda12:
uv sync --group all --group cuda12 - TPUs:
uv sync --group all --group tpu - CPU:
uv sync --group all --group cpu
Running¶
You have two options:
- Run theseus on the machine its installed in: You can install it directly onto your infrastructure, and thus sync with dependency group relating to the hardware; this enables running experiments locally
- Run theseus on a remote machine or SLURM: You can leverage the remote dispatch infrastructure to run on a remote cluster; if you do, you have to install the
cpuvariant on your head node / local machine, and use the dispatch CLI there.
For most use-cases, the first option will do just fine. For specific cases where remote infrastructure is not a normal shell (e.g., SLURM), the remote dispatch would work better.