Hummingbird compiles traditional machine learning models (scikit-learn, LightGBM, XGBoost) into tensor computations, enabling faster inference through neural network framework optimizations and hardware acceleration. It targets ML engineers and researchers seeking to unify their ML model serving infrastructure without re-engineering existing models.

How It Works

Hummingbird reconfigures model operators into tensor computations, primarily using matrix multiplications (GEMM) for decision tree traversal. This approach allows for vectorized and GPU-accelerated execution by transforming algorithmic steps into regular tensor operations, abstracting away the underlying hardware specifics.

Quick Start & Requirements

• Install: python -m pip install hummingbird-ml or python -m pip install hummingbird-ml[extra] for LightGBM/XGBoost.
• Requirements: Python 3.9-3.11, PyTorch >= 1.6.0. TVM backend requires Python 3.10. GPU acceleration requires PyTorch with CUDA support.
• Documentation: API Documentation, Blog Post

Highlighted Details

• Converts models to PyTorch, TorchScript, ONNX, and TVM.
• Provides a uniform inference API consistent with scikit-learn.
• Enables serving converted models via TorchServe.
• Supports scikit-learn Decision Trees, Random Forests, LightGBM, and XGBoost classifiers/regressors.

Maintenance & Community

• Key contributors include Supun Nakandala, Matteo Interlandi, and Karla Saur.
• Roadmap and contribution guides are available.

Licensing & Compatibility

• License: MIT License.
• Compatible with commercial and closed-source applications.

Limitations & Caveats

The TVM backend is restricted to Python 3.10. Support for additional backends and model types is planned for future releases.