This repository builds reusable hardware/runtime base images for downstream application images.
The public image contract is:
ghcr.io/offloadr/base/cpu-coreghcr.io/offloadr/base/amd-coreghcr.io/offloadr/base/nvidia-coreghcr.io/offloadr/base/nvidia-full
The default immutable tags built by this repository are:
cpu-core:py3.12-torch2.10.0-cpuamd-core:py3.12-torch2.10.0-rocm7.1nvidia-core:py3.12-torch2.10.0-cuda13.0.2nvidia-full:py3.12-torch2.10.0-cuda13.0.2
All published runtime images are expected to provide:
WORKDIR /workspace- system
python uv- base Python dependencies installed into the system Python environment
Downstream images that want their own virtual environment can layer one on top of the base image while still inheriting the preinstalled packages:
RUN uv venv /opt/venv --system-site-packages
ENV PATH="/opt/venv/bin:${PATH}"nvidia-core contains CUDA, Python, uv, PyTorch, and the shared Python dependencies needed by the optional accelerator wheels.
nvidia-full adds:
- xFormers
- FlashAttention 3
- SageAttention2++
- Nunchaku
Build everything locally:
docker buildx bakeBuild a specific image:
docker buildx bake nvidia-full