Accelerating scikit-learn with GPUs

Classical ML algorithms -- random forests, logistic regression, PCA, clustering -- remain the backbone of production machine learning. Yet they are often CPU-only. That’s changing. This talk presents two open-source approaches to GPU-accelerating the ML you already use: RAPIDS cuML, which rewrites classical algorithms in CUDA and can transparently accelerate scikit-learn-based libraries like BERTopic with a single import; and scikit-learn’s array API support, which lets the same scikit-learn code run on GPU inputs with no rewrites. Through live demos I’ll show what becomes practical when classical ML goes to the GPU, and offer honest guidance on when to reach for each tool. I maintain both scikit-learn and cuML; this talk reflects what I’ve learned working on both sides.