Typed Transformer

This repository contains a simple implementation of an encoder-decoder transformer model in JAX. The primary focuses are a clean, intuition-building implementation and leveraging Python's typing facilities to clarify things like the shapes of various tensors.

There's an accompanying write-up walking through things in detail here.

Usage

The top-level file is lm.py which trains a trivial integer sequence model in lm_main.

As far as dependencies and setup, there are a variety of options:

• There's a pyproject.toml that contains Poetry declarations, but I never use poetry directly so I can't guarantee its correctness.
• I use Poetry indirectly via poetry2nix. If you're a Nix user, invoking nix develop should get you a CPU-only setup.
• For GPU support, there's a Dockerfile. A very similar Dockerfile works for Python 3.11, but I haven't actually tested the Dockerfile after bumping to Python 3.12.