nvidia/nvidia-container-toolkit

This project is a toolkit from NVIDIA that lets you run software inside containers while giving that software access to NVIDIA graphics cards (GPUs). A container is a self-contained package of code and dependencies that runs in isolation, similar to a lightweight virtual machine. Normally, containers cannot easily talk to the physical hardware on the host computer, including the GPU. This toolkit bridges that gap. The toolkit installs a small layer that sits between the container system and the GPU hardware. When a container starts, this layer automatically detects which NVIDIA GPU is available on the host and configures the container so the software inside can use it. You do not need to install NVIDIA's full CUDA software stack on the host machine, only the basic NVIDIA driver. The main use case is running AI and machine learning workloads inside containers. Researchers and developers often package their training or inference code into containers so it can run consistently across different machines. Without a tool like this, those containers would not have GPU access and would run slowly on the CPU instead. The README is brief and points to external documentation for architecture details, installation steps, and configuration options. Setup requires a Linux system with an NVIDIA driver already installed.