If Nvidia GPUs are available on a node, they can be configured as follows.
controller: docker: nodes: gpu_node1: base_url: "tcp://192.168.0.101:2376" tls: verify: "/home/cc/.docker/machine/machines/gpu_node1/ca.pem" client_cert: - "/home/cc/.docker/machine/machines/gpu_node1/cert.pem" - "/home/cc/.docker/machine/machines/gpu_node1/key.pem" assert_hostname: False hardware: gpus: - id: 0 vram: 1024 - id: 1 vram: 1024
This configuration means that two GPUs are present on a node “gpu_node1”. Each GPU has 1024 MB VRAM. Currently only Nvidia-GPUs are supported. To make the GPUs accessible for docker, Nvidia-Docker has to be installed on each GPU node.
The IDs shown in the configuration, are the nvidia-device-IDs, which can be identified with
nvidia-smi for example