Wolfi CUDA Base Image

Lightweight Docker images based on Wolfi Linux optimized for NVIDIA CUDA and deep learning frameworks. These images provide a minimal, secure base for building and deploying AI/ML applications with significantly reduced image sizes compared to official NVIDIA images.

Features

• Minimal footprint - Up to 70% smaller than official NVIDIA CUDA images
• Security-focused - Built on Wolfi Linux with minimal attack surface
• Multi-framework support - Base, PyTorch, and TensorFlow variants
• Multiple Python versions - Support for Python 3.11 and 3.12
• Automated builds - CI/CD pipeline with GitHub Actions and Dagger

Image Size Comparison

Image Type	Size
NVIDIA CUDA Base	1.4 GB
NVIDIA CUDA 12.4.1	2.5 GB
Wolfi CUDA 12.4.1 (Base)	382 MB
Wolfi CUDA 12.4.1 (TensorFlow)	1.2 GB
Wolfi CUDA 12.4.1 (PyTorch)	1.1 GB

Quick Start

Pull Pre-built Images

# Base CUDA image
docker pull ghcr.io/YOUR_USERNAME/wolfi-cuda-base-image:wolfi_python_3.11_cuda_12.4.1_base

# PyTorch image
docker pull ghcr.io/YOUR_USERNAME/wolfi-cuda-base-image:wolfi_python_3.11_cuda_12.4.1_pytorch

# TensorFlow image
docker pull ghcr.io/YOUR_USERNAME/wolfi-cuda-base-image:wolfi_python_3.11_cuda_12.4.1_tensorflow

Build Locally with Docker

# Build base image
docker build -t wolfi-cuda:base .

# Build PyTorch image
docker build --build-arg FRAMEWORK=pytorch -t wolfi-cuda:pytorch .

# Build TensorFlow image
docker build --build-arg FRAMEWORK=tensorflow -t wolfi-cuda:tensorflow .

# Build with specific Python version
docker build --build-arg PYTHON_VERSION=3.12 -t wolfi-cuda:py312 .

Run a Container

# Interactive shell
docker run --gpus all -it wolfi-cuda:pytorch /bin/sh

# Run a Python script
docker run --gpus all -v $(pwd):/app wolfi-cuda:pytorch python3 your_script.py

Build with Dagger

This project uses Dagger for reproducible builds across different environments.

Prerequisites

• Python 3.8+
• Dagger CLI
• Docker (for local builds)

Installation

# Clone the repository
git clone https://github.com/YOUR_USERNAME/wolfi-cuda-base-image.git
cd wolfi-cuda-base-image

# Install Python dependencies
pip install -r requirements.txt

Build and Publish

# Set environment variables
export USERNAME="your-github-username"
export PASSWORD="your-github-pat"  # Personal Access Token with packages:write scope
export REPOSITORY="wolfi-cuda-base-image"  # Optional, defaults to wolfi-cuda-base-image

# Run the build pipeline
python main.py

Environment Variables

Variable	Required	Description
USERNAME	Yes	GitHub username for container registry authentication
PASSWORD	Yes	GitHub Personal Access Token with packages:write scope
REPOSITORY	No	Repository name (default: wolfi-cuda-base-image)
MULTI_ARCH	No	Set to true to enable multi-architecture builds (amd64 + arm64)

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        GitHub Actions                           │
│                    (CI/CD Pipeline Trigger)                     │
└─────────────────────────┬───────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────────────┐
│                      Dagger Pipeline                            │
│                       (main.py)                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │  Build Matrix:                                           │   │
│  │  • CUDA: 12.4.1, 12.6.0                                 │   │
│  │  • Python: 3.11, 3.12                                   │   │
│  │  • Framework: base, pytorch, tensorflow                  │   │
│  └─────────────────────────────────────────────────────────┘   │
└─────────────────────────┬───────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────────────┐
│                   Wolfi Base Image                              │
│              (cgr.dev/chainguard/wolfi-base)                    │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │  + CUDA Toolkit + cuDNN + cuBLAS + Python + Framework   │   │
│  └─────────────────────────────────────────────────────────┘   │
└─────────────────────────┬───────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────────────┐
│              GitHub Container Registry (ghcr.io)                │
│                     Published Images                            │
└─────────────────────────────────────────────────────────────────┘

Available Image Tags

Images are tagged with the following format:

ghcr.io/{username}/{repository}:{os}_python_{python_version}_cuda_{cuda_version}_{framework}

Current Build Matrix

OS	Python	CUDA	Framework
wolfi	3.11, 3.12	12.4.1, 12.6.0	base, pytorch, tensorflow

Using as a Base Image

FROM ghcr.io/YOUR_USERNAME/wolfi-cuda-base-image:wolfi_python_3.11_cuda_12.4.1_pytorch

# Install additional dependencies
RUN pip install transformers datasets

# Copy your application
COPY . /app
WORKDIR /app

# Run your application
CMD ["python3", "train.py"]

Troubleshooting

GPU not detected

Ensure you have the NVIDIA Container Toolkit installed:

# Ubuntu/Debian
sudo apt-get install -y nvidia-container-toolkit
sudo systemctl restart docker

Permission denied on GHCR

Make sure your Personal Access Token has the packages:write scope and you're logged in:

echo $PASSWORD | docker login ghcr.io -u $USERNAME --password-stdin

Build fails with package not found

Wolfi packages may have different names than standard Alpine packages. Check the Wolfi package repository for available packages.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

Acknowledgments

• Chainguard for the Wolfi Linux distribution
• Dagger for the portable CI/CD engine
• NVIDIA for CUDA toolkit