Yes we have setup the NVIDIA GPUs as passthrough to the container based on LXD's documentation: https://documentation.ubuntu.com/lxd/latest/reference/devices_gpu/ We also have nvidia.runtime set to True to pass the host NVIDIA and CUDA runtime libraries into the container: https://documentation.ubuntu.com/lxd/latest/reference/instance_options/#nvid...
With lxd VMs, have you had experience with MIG in the VMs? For example, it explicitly says that the MIG GPU type is not supported for VMs in LXD documentation, but at the same time, maybe I just need to pass in the GPUs, then enable MIG within the lxd VM.