TorchKM brings GPU acceleration to kernel learning and model selection

TorchKM is a new GPU-accelerated library for kernel methods, targeting classic machine learning algorithms like support vector machines (SVM), kernel logistic regression, and kernel quantile regression. Built with GPU-friendly linear algebra, it exploits intelligent reuse of matrix operations across both training and cross-validation sweeps, dramatically reducing runtime compared to CPU-based scikit-learn implementations. The library offers a scikit-learn-style API, making it easy for existing users to adopt without rewriting pipelines. Benchmarks confirm competitive predictive accuracy alongside significant speedups, though exact figures vary by dataset and kernel choice.

The library is open-source, available on PyPI, and includes documentation and code on GitHub. The accompanying arXiv paper (2606.06742) details the design and experiments across 14 pages. By targeting the full model-selection pipeline rather than just training, TorchKM addresses a key bottleneck in applied kernel learning. This could make kernel methods more practical for large-scale problems where GPUs are now standard, potentially reviving interest in classical approaches for modern ML workflows.