A LLVM-based compiler for LightGBM decision trees.

lleaves converts trained LightGBM models to optimized machine code, speeding-up prediction by ≥10x.

lgbm_model = lightgbm.Booster(model_file="NYC_taxi/model.txt")
%timeit lgbm_model.predict(df)
# 12.77s

llvm_model = lleaves.Model(model_file="NYC_taxi/model.txt")
llvm_model.compile()
%timeit llvm_model.predict(df)
# 0.90s 

• Speed: Both low-latency single-row prediction and high-throughput batch-prediction.
• Drop-in replacement: The interface of lleaves.Model is a subset of LightGBM.Booster.
• Dependencies: llvmlite and numpy. LLVM comes statically linked.

conda install -c conda-forge lleaves or pip install lleaves (Linux and MacOS only).

Ran on a dedicated Intel i7-4770 Haswell, 4 cores. Stated runtime is the minimum over 20.000 runs.

mostly numerical features.

mix of categorical and numerical features.

To avoid expensive recompilation, you can call lleaves.Model.compile() and pass a cache=<filepath> argument. This will store an ELF (Linux) / Mach-O (macOS) file at the given path when the method is first called. Subsequent calls of compile(cache=<same filepath>) will skip compilation and load the stored binary file instead. For more info, see docs.

To eliminate any Python overhead during inference you can link against this generated binary. For an example of how to do this see benchmarks/c_bench/. The function signature might change between major versions.

High-level explanation of the inner workings of the lleaves compiler: link

mamba env create
conda activate lleaves
pip install -e .
pre-commit install
./benchmarks/data/setup_data.sh
pytest -k "not benchmark"

If you're using lleaves for your research, I'd appreciate if you could cite it. Use:

@software{Boehm_lleaves,
  author = {Boehm, Simon},
  title = {lleaves},
  url = {https://github.com/siboehm/lleaves},
  license = {MIT},
}