Agent skill · nvidia
cuml
Use when writing, running, debugging or upgrading Python code that uses NVIDIA cuML (GPU machine learning), cuml.accel (zero-code-change GPU acceleration of scikit-learn, umap-learn and hdbscan), or GPU UMAP/HDBSCAN/RandomForest/KMeans/SVC, including through libraries that wrap cuML (BERTopic, rapids-singlecell). Covers APIs removed or renamed in 2025-2026 releases, verifying code really ran on the GPU, large-data UMAP/HDBSCAN memory recipes, and exporting trained models. Not for cuDF-only dataframe work, PyTorch or XGBoost.
What it needs
About 4k tokens when loaded.
What this skill does
cuML cuML releases every two months and removes deprecated APIs, so your training data is likely to contain APIs that no longer exist. Look things up in the installed cuML instead of recalling them. Ground facts Install from pypi.org: pip install cuml-cu13 (CUDA 13 driver, >= 580) or cuml-cu12. No extra index URL is needed. pip install cuml installs an unrelated placeholder. Check the installation guide for supported GPUs, operating systems, Python versions, CUDA/driver requirements, and install commands. Linux only; on Windows use WSL2. Google Colab GPU runtimes ship cuML preinstalled: check pip list before installing anything there. cuml.accel runs unmodified scikit-learn / umap-learn / hdbscan code on the GPU and falls back to the CPU for anything unsupported. It accelerates RandomForest, LogisticRegression, Ridge, KMeans, DBSCAN, PCA, SVC, KNN, UMAP, HDBSCAN, scalers, encoders and more. Enable it with python -m cuml.accel script.py, %loadext cuml.accel (first notebook cell), CUMLACCELENABLED=1, or cuml.accel.install() before importing sklearn. Step 1: look it up in the installed cuML Check the versions and answer for them: python -c "import cuml, sklearn; print(cuml.version, sklearn.version)". Before using a cuML class, function or argument, confirm it exists and read its documentation: python -c "import inspect, cuml; print(inspect.signature(cuml.svm.SVC))", help(cuml.manifold.UMAP). Docstrings list valid values, for example the accepted buildkwds keys and outputtype values. Run the code and read the errors: removed arguments raise TypeError, removed modules ModuleNotFoundError, invalid values usually list the valid ones. Treat deprecated APIs as removed: do not recommend anything that emits a FutureWarning or is marked deprecated in its docstring, even if it still works in the installed version. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills nvidia/cuml