Advice user to use --gpus instead of gpu-addon on docker driver

[sivanaikk]

What Happened?

I am trying to integrate GPU with Minikube, My host machine is Windows and I am using docker desktop in wsl2 and trying to provision Minikube with the support of GPU.

I was able to integrate GPU with docker,

sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi

• Minikube start command

minikube start --driver=docker --addons=ingress --addons=nvidia-gpu-device-plugin

• Pod failing with the following error

  Warning  FailedScheduling  45s   default-scheduler  0/1 nodes are available: 1 Insufficient nvidia.com/gpu. preemption: 0/1 nodes are available: 1 No preemption victims found for incoming pod.

Attach the log file

log.txt

Operating System

Windows

Driver

Docker

step I followed

Medium Link

[medyagh]

@sivanaikk that addon is mean for KVM driver and sorry that minikube did not Warn and advice correctly, this is something we could work on for a better user experience

for docker driver please use --gpus flag

I think the right thing is to Rename this flag to kvm-nvidia-gp and then still accept the people using old name (deprciated name) but for new users it wont be confusing

[sivanaikk]

Thank you @medyagh. It's working now!

minikube start --gpus all --driver=docker --addons=ingress

[medyagh]

seems like we have 3 addons,
two of them are for KVM and one is for docker, it is important that we figure out if we can merge some of them

| nvidia-device-plugin | minikube | disabled | 3rd party (NVIDIA) |
| nvidia-driver-installer | minikube | disabled | 3rd party (Nvidia) |
| nvidia-gpu-device-plugin | minikube | disabled | 3rd party (Nvidia) |

https://minikube.sigs.k8s.io/docs/tutorials/nvidia/#docker

in our tutorial seems like, for KVM we need:

• nvidia-driver-installer
• nvidia-gpu-device-plugin

for "none" we use:

• nvidia-device-plugin

so the LEAST we we should do is have a description that this is for KVM only
but ideally we could merge "nvidia-gpu-device-plugin" and "nvidia-device-plugin" so they be one

[RamBITS-AI]

/assign

[RamBITS-AI]

I'm starting work (PR - #19545) on this issue.