bullet

A domain-specific ML library, generally used for training NNUE-style networks for some of the strongest chess engines in the world.

Crates

The main crate is bullet_core, which is an ML framework that is generic over backends:

• A network graph is constructed using GraphBuilder
• This is then lowered to GraphIR and optimisation passes performed on it
• The GraphIR is then compiled into a Graph<D: Device>, for a specific backend device
  • Upon which forwards and backwards passes, editing weights/inputs, etc, may be performed
  • A small set of (composable) optimisers are included that ingest a graph and provide update methods for it
• A token single-threaded CPU backend is included for verifying correctness of the crate and other backend implementations

Backend Crates:

• These implement specific backends
• bullet_hip_backend currently contains both the HIP/ROCm and CUDA backends
• bullet_cuda_backend is an incomplete CUDA backend rewrite waiting on ongoing refactors in the cudarc crate

The crate bullet_lib provides a high-level wrapper around these crates specifically for training networks to do with chess (and other games e.g. Ataxx) easily. There are also various misc utilities mostly to do with handling chess data in bullet-utils.

Usage for NNUE/Value Network Training

Before attempting to use, check out the docs. They contain all the main information about building bullet, managing training data and the network output format.

Most people simply clone the repo and edit one of the examples to their taste. If you want to create your own example file to ease pulling from upstream, you need to add the example to bullet_lib's Cargo.toml.

Alternatively, import the bullet_lib crate with

bullet = { git = "https://github.com/jw1912/bullet", package = "bullet_lib" }

Specific API documentation is covered by Rust's docstrings.

Help/Feedback

• Please open an issue to file any bug reports/feature requests.
• Feel free to use the dedicated #bullet channel in the Engine Programming discord server if you run into any issues.
• For general training discussion the Engine Programming non-#bullet channels are appropriate, or #engines-dev in the Stockfish discord.