MPS control daemon wrong pod selector

[radepajic]

I’ve installed the latest version of the nvidia-device-plugin (0.16.2) using Helm. Alongside the device plugin, the MPS control daemon is also being installed. The problem is that in the MPS control daemon set, the pod selector is the same as in the device plugin daemon set. As a result, the device plugin pod starts first, and both controllers attempt to manage the same pod, preventing MPS from ever starting.

[chipzoller]

Does indeed look like an issue. This isn't apparent when deploying the device plugin with the GPU operator because the templated labels applied to the device plugin and MPS control DaemonSets (via the app key) are equal to the respective names of those DaemonSets which naturally differ.