Limit devices visible in a Kubernetes node

[MassimoGirondi]

What is the proper way to limit GPUs visibility to a Kubernetes node?

The use case is a shared environment where we would like to run both Kubernetes and Docker, and we would like to limit the GPUs visible in the node to a subset.

I found references to CUDA_VISIBLE_DEVICES and NVIDIA_VISIBLE_DEVICES, but it's not clear (to me) how these affect the GPUs assigned and what is the suggested way to implement such a limitation.