Hummingbird

Hummingbird is a Microsoft library that compiles trained traditional ML models into tensor computations, so models built with scikit-learn, XGBoost or LightGBM can run inside PyTorch, TensorFlow, ONNX Runtime or TVM with the speed of neural-network inference. Instead of maintaining separate serving code for tree ensembles and pipelines, you convert once and deploy on GPU or edge runtimes. The project has around 3,550 stars on GitHub and is MIT licensed.

Overview

Traditional ML models are usually served with their training library’s runtime, which complicates deployment next to deep learning models. Hummingbird translates model structures — decision trees become tensor operations, featurisers become tensor ops — into an equivalent computation graph in your target framework. This gives classical models access to GPU acceleration, batching and optimised runtimes, and lets mixed pipelines live in a single serving stack. A large matrix of models is supported, from random forests and gradient boosting to linear models, clustering and anomaly detection, and converted models are validated against the originals for numerical parity.

Key Features

• Converts scikit-learn, XGBoost, LightGBM and Spark ML models to tensors
• Targets PyTorch, TensorFlow, ONNX Runtime and Apache TVM
• GPU acceleration and batching for classical ML inference
• Broad model coverage: trees, linear models, clustering, outliers
• Numerical parity checks against the original model
• Single serving stack for classical and deep models
• MIT licensed

💳 Pricing

Hummingbird is completely free and open source under the MIT license, courtesy of Microsoft’s open-source program.

Who Is It For

ML engineers with production scikit-learn or boosting models who want faster inference and simpler deployment — especially teams standardising on PyTorch or ONNX serving who do not want to rewrite proven classical models.