NVIDIA CUDA-X Deployment Documentation for Data Science

This repository contains the source for the NVIDIA CUDA-X Deployment Documentation. It explains how to install, configure, and operate NVIDIA CUDA-X libraries for data science across local systems, GPU clusters, and managed compute services.

The documentation includes:

• Installation paths for local workstations, custom containers, and Slurm-managed HPC clusters.
• Infrastructure-specific deployment instructions for major cloud providers, including virtual machines, managed Kubernetes, and machine learning services.
• Integration guidance for compute and application platforms such as Kubernetes, Kubeflow, Databricks, Snowflake, and Google Colab.
• Practical guides and end-to-end notebook examples covering distributed data processing, machine learning, optimization, and MLOps workflows.

Repository Layout

• source/cloud/ contains provider-specific infrastructure instructions for deploying NVIDIA CUDA-X libraries for data science on cloud platforms like AWS, Azure, Google Cloud, and IBM Cloud. These pages cover services such as virtual machines, managed Kubernetes, and hosted machine learning environments.
• source/platforms/ explains how to run NVIDIA CUDA-X libraries for data science on cloud platforms such as Kubernetes, Kubeflow, KServe, Databricks, Snowflake, Google Colab, Coiled, Modal, and NVIDIA AI Workbench.
• source/guides/ contains focused, cross-platform guidance for deployment topics such as custom CUDA containers, Multi-Instance GPU, InfiniBand, Dask scheduler sizing, Kubernetes worker placement, and image caching.
• source/examples/ contains end-to-end Jupyter notebook workflows and the supporting Python scripts, Dockerfiles, environment files, and Kubernetes manifests needed to run them.
• source/hpc.md covers running NVIDIA CUDA-X libraries for data science on Slurm-managed HPC clusters, including interactive and batch jobs, environment modules, and distributed workloads.
• source/local.md is the entry point for running NVIDIA CUDA-X libraries for data science on a workstation or server using conda, pip, Docker, or WSL2.
• source/custom-docker.md describes how to build smaller, tailored container images with only the required NVIDIA CUDA-X libraries using conda or pip.

Build Locally

The site is built with Sphinx and requires Python 3.12 or newer. Dependencies are managed with uv.

uv venv
uv sync --locked
uv run make dirhtml

The generated site is written to build/dirhtml. For live previews while editing, run:

uv run sphinx-autobuild -b dirhtml source build/html

Published Documentation

• Latest documentation is published from the main branch.
• Released documentation is published from vYY.MM.PP release tags to https://docs.nvidia.com/datascience/deployment/YY.MM/, for example 26.08. Use the version switcher in the navigation bar to move between releases.

See CONTRIBUTING.md for instructions on building, writing, linting, and releasing.