latest gpu operator container toolkit daemonset behavior catastrophically breaks clusters running k0s

[doctorpangloss]

Describe the bug
k0sproject/k0s#6547

The two step import you introduced, /etc/k0s/containerd.d/nvidia.toml -> /etc/containerd/conf.d/99-nvidia.toml breaks k0s clusters.

To Reproduce
Use gpu-operator on a k0s cluster.

Expected behavior
Don't be too exotic with how you put in these files.

Environment (please provide the following information):

• GPU Operator Version: v25.10.0
• OS: Ubuntu24.04
• Kernel Version: 6.14
• Container Runtime Version: containerd 1.7.22
• Kubernetes Distro and Version: k0s

Information to attach (optional if deemed irrelevant)

(see referenced issue k0sproject/k0s#6547)

[doctorpangloss]

---
apiVersion: helm.toolkit.fluxcd.io/v2beta1
kind: HelmRelease
metadata:
  name: nvidia-gpu-operator
  namespace: gpu-operator
spec:
  ...
  values:
    ...
    toolkit:
      enabled: true
      env:
        - name: RUNTIME_CONFIG
          value: "/tmp/nvidia-operator-dummy-import.toml"
        - name: RUNTIME_DROP_IN_CONFIG
          value: "/etc/k0s/containerd.d/nvidia.toml"
        - name: CONTAINERD_SOCKET
          value: "/run/k0s/containerd.sock"
        - name: RUNTIME_EXECUTABLE_PATH
          value: "/var/lib/k0s/bin/containerd"
        - name: CONTAINERD_RUNTIME_CLASS
          value: "nvidia"
        - name: CONTAINERD_SET_AS_DEFAULT
          value: "false"

since your cluster will be dead, you will need to kubectl edit in the changes to the HelmRelease / redeploy.

then, clear the problematic file and restart k0s

ssh someusername@target-host -- 'sudo rm -f /etc/k0s/containerd.d/nvidia.toml && sudo k0s stop && sudo systemctl reset-failed k0sworker.service && sudo k0s start'

works around the issue for now

[francisguillier]

sorry for this disruption @doctorpangloss

GPU operator 25.10.0 introduces a breaking change (which will be fully documented tomorrow).
Helm command to install GPU operator for K0s is:

helm install gpu-operator nvidia/gpu-operator -n nvidia-gpu-operator --create-namespace --version=v25.10.0
--set toolkit.env[0].name=CONTAINERD_CONFIG
--set toolkit.env[0].value=/run/k0s/containerd-cri.toml
--set toolkit.env[1].name=RUNTIME_DROP_IN_CONFIG
--set toolkit.env[1].value=/etc/k0s/containerd.d/nvidia.toml
--set toolkit.env[2].name=CONTAINERD_SOCKET
--set toolkit.env[2].value=/run/k0s/containerd.sock
--set toolkit.env[3].name=CONTAINERD_RUNTIME_CLASS
--set toolkit.env[3].value=nvidia

Notice the additional parameter for RUNTIME_DROP_IN_CONFIG.

We are working with Mirantis team to reflect the change in their documentations too.

[francisguillier]

Mirantis has updated their docs accordingly:
https://catalog.k0rdent.io/latest/apps/nvidia/#install

[armaneshaghi]

Should we expect a similar break down in k3s?

[doctorpangloss]

Helm command to install GPU operator for K0s is:

helm install gpu-operator nvidia/gpu-operator -n nvidia-gpu-operator --create-namespace --version=v25.10.0 --set toolkit.env[0].name=CONTAINERD_CONFIG --set toolkit.env[0].value=/run/k0s/containerd-cri.toml --set toolkit.env[1].name=RUNTIME_DROP_IN_CONFIG --set toolkit.env[1].value=/etc/k0s/containerd.d/nvidia.toml

this isn't correct, you don't want nvidia container toolkit modifying the containerd config at all, only placing the dropin file. this will still be broken. use my workaround.

[doctorpangloss]

Should we expect a similar break down in k3s?

check if it uses drop-in configuration and merges it, if so, then yes, otherwise, must test.