gcloud CLI: Creating a (pre-emptible) VM instance with the NVIDIA GPU Cloud (NGC) image

create_ngc_vm_on_gcp.sh

# This creates a pre-emptible VM on Google Cloud with a
# more or less NVIDIA approved configuration and drivers
# using the public NVIDIA GPU Cloud (NGC) image on GCP.
# The VM has nvidia-docker installed so anything like
# this will work out of the box:
#
#   $ docker run --runtime=nvidia --rm paperspace/fastai:cuda9_pytorch0.3.1
#
# None of the public scripts and info had working configuration anywhere.
# Even the "official" sample scripts don't have working default configuration:
# https://github.com/NVIDIA/ngc-examples/tree/master/ncsp
#
# But this is how you do find out the available images:
#   $ gcloud compute images list --project nvidia-ngc-public --no-standard-images

image=nvidia-gpu-cloud-image-20180515

# List of zones at: https://cloud.google.com/compute/docs/gpus/
ZONE=europe-west1-b

gcloud compute instances create ngc-example-vm \
    --preemptible \
    --image-project=nvidia-ngc-public \
    --image=$image \
    --accelerator type=nvidia-tesla-k80,count=1 \
    --boot-disk-size=32 \
    --machine-type=n1-standard-4 \
    --local-ssd="interface=nvme" \
    --zone=$ZONE \
    --maintenance-policy=TERMINATE \
    --metadata startup-script='#!/bin/bash
# Don't prompt for the nvcr.io API key each login:
sed -i "s/exit [0-9]*/exit 0/" /usr/bin/ngc-get-key.sh
# Install the drivers without user interaction:
yes | /usr/bin/gcp-ngc-login.sh
# While at it, upgrade system packages
apt upgrade -y
# NVIDIA drivers require reboot after installation so reboot unattended:
if ! [ -f /root/.ngc-install-done ]; then
  touch /root/.ngc-install-done
  reboot
fi'

Great stuff :) Thanks for sharing.