NVIDIA cuML – GPU‑Accelerated Machine Learning

What it is – cuML is NVIDIA’s open‑source library that brings the familiar scikit‑learn API to NVIDIA GPUs. It implements a range of classic ML algorithms (clustering, regression, classification, dimensionality reduction, etc.) so they run directly on the GPU, giving large speed‑ups (up to ~50× on some workloads).

How you use it

• Native API – import cuml just like scikit‑learn and call fit, predict, transform. Data stays on the GPU, so you avoid costly host‑device copies.
• Accelerate existing code – drop cuml.accel into a script or Jupyter notebook; it automatically swaps compatible scikit‑learn, UMAP, or HDBSCAN calls for GPU versions without code changes.
• Multi‑GPU scaling – the cuml.dask module lets you run selected algorithms across many GPUs or nodes via Dask.

Key features mentioned in the README

• API mirrors scikit‑learn’s estimator pattern.
• cuml.accel provides seamless GPU acceleration with fallback to CPU when needed.
• Distributed execution with cuml.dask.
• Installation via conda, pip, or Docker (selector tool).
• Model serialization with pickle/joblib (with standard security warnings).
• Extensive docs, notebooks, and compatibility with scikit‑learn ≥ 1.6.

Who it’s for – Data scientists and engineers who already use scikit‑learn and want to speed up training/inference by leveraging NVIDIA GPUs, without rewriting their pipelines.

Where to start – Follow the quick DBSCAN example in the README, explore the API reference, or run python -m cuml.accel script.py on an existing scikit‑learn script.