GPU Provisiong with GCP Provider fails

What happened?
I am trying to create a simple GPU-based GCP VM (g2-standard-4) with GCP Provider (v.0.0.7 ). It fails and with the error code 400 and message “Instances with guest accelerators do not support live migration.”

What did you expect to happen instead?
Create A VM successfully with GPU. The way I found it can be bypassed is this flag. --maintenance-policy TERMINATE --restart-on-failure ( Link for troubleshooting steps: https://groups.google.com/g/gce-discussion/c/e9K3h3fQuJk/m/UxyKqskLAQAJ )

~ ❯ gcloud compute instances create vm  --machine-type=g2-standard-4 --zone=us-central1-a                G hrittik-project
ERROR: (gcloud.compute.instances.create) Could not fetch resource:
 - Instances with guest accelerators do not support live migration.

~ ❯ gcloud compute instances create vm  --machine-type=g2-standard-4 --zone=us-central1-a   --maintenance-policy TERMINATE --restart-on-failure
Created [https://www.googleapis.com/compute/v1/projects/hrittik-project/zones/us-central1-a/instances/vm].
NAME  ZONE           MACHINE_TYPE   PREEMPTIBLE  INTERNAL_IP  EXTERNAL_IP     STATUS
vm    us-central1-a  g2-standard-4               10.128.0.16  34.136.130.135  RUNNING

The purpose is to tryOlama with GPU and CUDA features. With CPU-based machines, it’s working, but I wanted to push it with a GPU during inference.

How can we reproduce the bug? (as minimally and precisely as possible)

Create g2-standard-4 VM on GCP or others with GPU.

My devcontainer.json:

{
    "name": "Python 3",
    "image": "mcr.microsoft.com/devcontainers/python:1-3.12-bullseye",
    "features": {
        "ghcr.io/devcontainers/features/docker-in-docker:2": {},
        "ghcr.io/devcontainers/features/nvidia-cuda:1": {},
        "ghcr.io/rio/features/k3d:1": {}
    },
    "customizations": {
        "vscode": {
            "extensions": [
                "ms-azuretools.vscode-docker",
                "GitHub.copilot"
            ]
        }
    }
}

Local Environment:

DevPod Version: v0.5.4
Operating System: mac
ARCH of the OS: ARM64

DevPod Provider:

Cloud Provider: google

Anything else we need to know?
Internal Chat

loft-sh / devpod

GPU Provisiong with GCP Provider fails #973