Confirmation of understanding that sharing is per-Node, not per-GPU

[chipzoller]

It isn't stated in the device plugin's README that I have found, but from inferring based on examples, issue comments here, here, and here, as well as personal experimentation, it seems like enabling sharing (Time-slicing and MPS) is on a per-node basis and not per-GPU. In other words, you either share all GPUs on a node or you share none of them. Is that understanding correct? (I understand that DRA will change the game here, but I'm asking about the current state of the device plugin.)