Docker on NVIDIA GPU

The recommended way to use Docker for NVIDIA hardware is described here.

You can also run PyG with CUDA 12.1 inside a docker image. This method is deprecated and we highly recommend the above mentioned official NVIDIA docker containers instead.

The creation of our dockerfile refers to the dockerfiles provided by NVIDIA and PyTorch.

1. Download the dockerfile to your host server.
2. $ docker build -t "custom image name"
3. $ docker run --rm -it --init --runtime=nvidia --ipc=host --network=host --volume=$PWD:/app -e NVIDIA_VISIBLE_DEVICES=0 "custom image name" /bin/bash

If you encounter any problems, please feel free to create a GitHub issue.

Docker on Intel GPU

You can also run PyG with Intel GPU inside a docker image. The creation of our dockerfile refers to the dockerfiles provided by Intel and the installation guidance provided by Intel® Extension for PyTorch.

1. Download the dockerfile to your host server.
2. $ docker build -f docker/Dockerfile.xpu -t "custom image name"
3. $ docker run --rm -it --ipc=host -v /dev/dri:/dev/dri --volume=$PWD:/app "custom image name" /bin/bash

Singularity

You can run PyG inside a singularity image. An example singularity file can be found in this folder.

You might have to modify the script; depending on your needs, modify the following:

• cuda version 10.1: If you need another version, change From: nvidia/cuda:10.1-cudnn7-devel-ubuntu16.04 to the corresponding tag from https://hub.docker.com/r/nvidia/cuda. Same if you want to use anything but Ubuntu 16.04. Your host has to have at least this cuda version!
• python version 3.7.2: If you need another version, change pyenv install 3.7.2 and the following lines to the corresponding version.
• pytorch version 1.3.0: If you need another version, change pip install torch==1.3.0.
• pytorch_geometric versions: This uses specific versions for each of the pytorch_geometric requirements (scatter: 1.4.0, sparse: 0.4.3, cluster: 1.4.5, geometric 1.3.2). To change these, change the corresponding git checkout lines near the bottom.
• cuda compute capability 5.0 and 6.1: If you run it on multiply systems (likely, with singularity), ensure that all compute capabilities are listed here. If you have the cuda samples installed, check it with (for example) /usr/local/cuda-10.1/extras/demo_suite/deviceQuery | grep 'CUDA Capability'; if not, check here.

Note: If your harddisk runs full after multiple builds, this is known and apparently working as intended; delete the /tmp/sbuild-XXXXXXXXX files.

Building and Using the Container

To build the container, run

sudo singularity build geometric.sif singularity

then wait. Once finished, you can run the GAT example in the folder you built the image in by calling

wget https://raw.githubusercontent.com/pyg-team/pytorch_geometric/master/examples/gat.py

(to download the sample),

then

singularity exec geometric.sif python3 gat.py

to run on the CPU, or

singularity exec --nv geometric.sif python3 gat.py

to run on the GPU.