bullet

A domain-specific ML library, generally used for training NNUE-style networks for many of the strongest chess engines in the world due to its best-in-class performance, chess-specific tooling and ease of use. Training is supported on NVIDIA (cuda), AMD (rocm) and Apple Silicon (metal) GPUs.

Usage for NNUE/Value Network Training

Before attempting to use, check out the docs which contain 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:

• simple - a basic (768 -> N)x2 -> 1 network, with example inference code, that uses ValueTrainerBuilder
• progression - from a first network to input buckets, output buckets and multiple layers
• advanced - a SOTA training example that does not use ValueTrainerBuilder, instead using bullet-trainer directly
• ataxx - a non-chess example, where a simple custom data format is defined

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" }