Skip to content

[Bug]: k8s-device-plugin fails to detect Jetson platform #1963

Description

@shahizat

Hello

Environment:

  • Hardware: NVIDIA Jetson AGX Thor Developer Kit (aarch64, iGPU "NVIDIA Thor")
  • JetPack 7.2, driver 595.78, CUDA 13.2
  • MicroK8s 1.35 (containerd 2.1.6), Helm chart nvidia-device-plugin 0.9.0
  • Device-plugin image: nvcr.io/nvidia/k8s-device-plugin:v0.19.3

Installed via:

helm upgrade -i nvidia-device-plugin \
  nvidia/nvidia-device-plugin \
  -n gpu-operator \
  --version 0.9.0 \
  --set image.tag=v0.19.3 \
  --set deviceIDStrategy=index

On Nvidia Jetson Thor the v0.19.3 device-plugin pod starts and immediately dies with Exit Code: 1 in a crash loop. The pod log shows the plugin no longer detects the Jetson platform and falls back to NVML discovery, which does not exist in the Jetson container image:

I0814 09:35:39.851565       1 main.go:369] Retrieving plugins.
I0814 09:35:39.851780       1 plugin-manager.go:100] Detected platform: unknown
W0814 09:35:39.851796       1 plugin-manager.go:107] Unsupported platform detected; defaulting to nvml
I0814 09:35:39.851804       1 plugin-manager.go:50] Using device discovery strategy: nvml
E0814 09:35:39.851839       1 factory.go:99] Failed to initialize NVML: ERROR_LIBRARY_NOT_FOUND.
E0814 09:35:39.851845       1 factory.go:100] If this is a GPU node, did you set the docker default runtime to `nvidia`?
E0814 09:35:39.851849       1 factory.go:101] You can check the prerequisites at: https://github.com/NVIDIA/k8s-device-plugin#prerequisites
E0814 09:35:39.851853       1 factory.go:102] You can learn how to set the runtime at: https://github.com/NVIDIA/k8s-device-plugin#quick-start
E0814 09:35:39.851857       1 factory.go:103] If this is not a GPU node, you should set up a toleration or nodeSelector to only deploy this plugin on GPU nodes
E0814 09:35:39.852018       1 main.go:185] error starting plugins: error getting plugins: unable to create plugins: failed to construct resource managers: nvml init failed: ERROR_LIBRARY_NOT_FOUND

The same Helm chart with image.tag=v0.16.2 (and deviceIDStrategy=index) runs fine on this board, registers successfully, and is allocatable:

nvidia.com/gpu: 1
Registered device plugin for 'nvidia.com/gpu' with Kubelet

Metadata

Metadata

Assignees

Labels

bugIssue/PR to expose/discuss/fix a bugneeds-triageissue or PR has not been assigned a priority-px label

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions