docker-gpu-setup

What problem does it solve?

Building and running GPU-enabled Docker containers for ML workloads is error-prone and slow without standard patterns, resulting in broken GPU passthrough, huge unreproducible images, and long rebuild times. This Skill provides clear, production-ready Dockerfile patterns, Compose snippets, and troubleshooting steps to ensure containers see GPUs, dependencies are cached, and images remain reproducible.

Core Features & Use Cases

• Multi-stage Dockerfile templates for isolating dependency installs and producing slim runtime images.
• NGC base image guidance to choose the right nvcr.io image for training, inference, dev, or runtime scenarios.
• uv-in-Docker and BuildKit cache patterns to speed dependency installs and enable reproducible builds.
• docker-compose GPU passthrough snippets and verification steps for Compose v2 deployments.
• GPU troubleshooting checklist covering NVIDIA driver, NVIDIA Container Toolkit, docker runtime config, driver/CUDA compatibility, and common runtime errors.
• Use Case: Build a multi-stage PyTorch training image that installs dependencies via uv, caches builds with BuildKit, runs with GPU passthrough, and verifies GPU visibility with nvidia-smi.

Quick Start

Create a multi-stage Dockerfile using an appropriate NGC base, copy the uv binary from the official image, build with BuildKit enabled, run the container with GPU passthrough, and verify GPU access by running nvidia-smi inside the container.