A CUDA/CPU NN Trainer, used to train NNUE-style networks for akimbo.
Also used by many other engines, including:
- Chess
- Ataxx
Raise an issue for support of a new game.
Import the crate with
bullet = { package = "bullet_lib", version = "1.0.0", features = ["cuda"] }
Check out the wiki and examples to see how to use the crate.
You can build bullet-utils
with cargo b -r --package bullet-utils
, to do the following:
- Convert Data
- Interleave Multiple Data Files
- Shuffle Data Files
- Validate Data Files
Use ./target/release/bullet-utils[.exe] help
to see specific usage.
Reference CPU backend. It is suitable for training small networks or various utilities, such as loading nets to requantise them or test their output on specific positions.
Warning
Not intended for serious training use. If you need to train on CPU, use the legacy branch.
The "first class" supported backend. To compile to target CUDA you need to enable the cuda
feature,
as demonstrated in the wiki. You will need to install the CUDA Toolkit, and have an available C++ compiler.
Note
If you are on Windows, it is recommended to use clang, direct from LLVM github releases, for the C++ compiler.
Mainly directed toward users with AMD GPUs. To compile to target HIP you need to enable the hip
feature,
as demonstrated in the wiki. You will need to install the HIP SDK.
Note
If you are on Windows, you must also add %HIP_PATH%\bin\
to the PATH variable in your system environment variables.
Warning
Due to what appears to be a bug in RoCM, some tests will sometimes fail due to missed synchronisation between device and host in a multithreaded context. As the trainer only calls kernels from one thread, this should not be an issue in training.
Warning
The HIP backend is not officially supported on Linux (due to unresolved issues with annoying platform dependent stuff), but it has been made to work by a couple of users with some minor edits.