Simple working example of an Argo Workflow configured to run on the GPU nodes

example-gpu-argo-workflow.yaml

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: cuda-vector-add-
spec:
  entrypoint: main
  templates:
  - name: main
    # requires this pod to be run on an nvidia.com/gpu labeled node
    nodeSelector:
      nvidia.com/gpu: "true"
    # allows this pod to be run on an nvidia.com/gpu tainted node
    tolerations:
    - key: nvidia.com/gpu
      operator: Exists
      effect: NoSchedule
    - key: gpu
      value: "true"
      operator: Equal
      effect: NoSchedule
    container:
      # https://github.com/kubernetes/kubernetes/blob/v1.7.11/test/images/nvidia-cuda/Dockerfile
      image: "k8s.gcr.io/cuda-vector-add:v0.1"
      resources:
        # don't use requests for GPUs: https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/
        limits:
          nvidia.com/gpu: 1 # requesting 1 GPU