Nvidia docker runtime setup for CUDA images

nvidia-docker-notes.org

Manual setup of nvidia-docker

To not use the pre-bundled images or .deb-s.

Runtime

• Checkout (opencontainers/runc) from hash id’d from nvidia-container-runtime/runtime/runc and the nvidia-container-runtime/runtime/Makefile.
• Patch (runc) with patch from directory matching the commit hash.
• Build runc, and cp/rename it to nvidia-container-runtime (needs BSD make: bmake).
• cp (runc) nvidia-container-runtime /usr/local/bin

Hook

• Merge (nvidia-docker) docker/daemon.conf with system’s.
• Copy config.toml.debian to /etc/nvidia-container-runtime/config.toml
• Build and cp (nvidia-container-runtime/hook/nvidia-container-runtime-hook/) nvidia-container-runtime-hook /usr/local/bin

libnvidia-container

• Make install on libnvidia-container (needed libseccomp-dev)

Notes

My machine is too old for tensorflow/tensorflow:latest-gpu, :1.5.0-gpu appears to work. libtensorflow_framework.so hits an illegal instruction on newer ones.

Repos and req.

• https://github.com/opencontainers/runc
• https://github.com/NVIDIA/nvidia-container-runtime
• https://github.com/NVIDIA/nvidia-docker
• https://github.com/NVIDIA/libnvidia-container
• Golang 1.7 appears to be enough.