h2oai/h2o4gpu

What it solves

H2O4GPU provides GPU-accelerated versions of common machine learning algorithms to speed up model training and inference. It is designed as a drop-in replacement for scikit-learn, allowing users to leverage GPU hardware without changing their existing Python or R codebases.

How it works

The project implements GPU solvers for specific algorithms and integrates them into a scikit-learn-compatible API. When a GPU-enabled algorithm is used, the library leverages NVIDIA GPUs via CUDA. If a specific GPU implementation is unavailable or does not support a required option, the system automatically falls back to the standard scikit-learn CPU algorithms (or faster CPU alternatives like Intel DAAL).

Who it’s for

Data scientists and machine learning engineers who use scikit-learn or R and want to accelerate their ML workflows using NVIDIA GPUs with minimal code changes.

Highlights

• Scikit-learn Compatibility: Acts as a drop-in replacement (import h2o4gpu as sklearn).
• Multi-Language Support: Provides APIs for both Python and R.
• Broad Algorithm Support: Includes GPU solvers for GLM (Lasso, Ridge, Logistic Regression, Elastic Net), KMeans, GBM (via XGBoost), SVD, and PCA.
• GOAI Integration: Aims to support operations directly on GPU memory pointers to reduce data transfer overhead between host and device.