[Feature]: Enable debian12 support

[m3co-code]

Suggestion Description

Hi. Is it possible to enable driver installation for debian12? 🤞 The ROCm drivers should be ready according to the matrix. I didn't find one for the device plugin though, would be keen to learn if you'd know of any concerns re debian12 support on the plugin side.

Thank you so much in advance :)

---

1 controller.go:341] "Reconciler error" err="failed to handle build ConfigMap for DeviceConfig kube-amd-gpu/gpu-operator: invalid node worker-3, err: OS: debian gnu/linux 12 (bookworm) not supported. Should be one of [ubuntu coreos rhel red hat redhat]" logger="amd-gpu" controller="DriverAndPluginReconciler" controllerGroup="amd.com" controllerKind="DeviceConfig" DeviceConfig="kube-amd-gpu/gpu-operator" namespace="kube-amd-gpu" name="gpu-operator" reconcileID="fd2a95ec-3f38-4fe0-8faf-afdd4ea06c18"

  driver:
    enable: true
    blacklist: true
  devicePlugin:
    devicePluginImage: rocm/k8s-device-plugin:v1.31.0.6
    nodeLabellerImage: rocm/k8s-device-plugin:labeller-v1.31.0.6

Operating System

Debian 12

GPU

Mi300x

ROCm Component

Operator code I think or device plugin

[yansun1996]

Hi @m3co-code thanks for trying AMD GPU Operator.

1. I saw that you are configuring to ask GPU Operator to install out-of-tree kernel module on your worker node

  driver:
    enable: true
    blacklist: true

for this functionality, it requires operator to build the driver then dispatch the compiled driver to worker node for loading kernel module. We haven't fully validated this process for Debian 12 yet so it is not in the supported OS list. We've received request to add more Linux distros OS support, we are not there yet to validate all kinds of ROCm supported Linux distros.

2. Workaround
   There is a workaround: You can manually install specific version out-of-tree driver on your Debian 12 worker node, or use your worker node's existing inbox amdgpu driver if it has (check modinfo amdgpu and lsmod amdgpu). After installing the driver you can set

  driver:
    enable: false

then the operator will skip the driver installation and directly bring up the operands (device plugin, node labeller, metrics exporter, test runner, etc)

[m3co-code]

Thank you for the swift support and feedback! I ended up pre-installing the drivers beforehand and used the device plugin than standalone. Was hoping to leverage the driver part from the operator mostly.

Should we keep the issue open for the mean-time? I can at least help giving another data point and testing things when debian12 made it through your validation.

[yansun1996]

sure @m3co-code let's remain this issue open. Let us know if you have any other questions or suggestions.

We're actively working on AMD GPU Operator and more features and support are coming.