only pci passthrough ofc, vgpu comes in the next version (there are some release candidate versions though)

using NVlink... never properly evaluated it, since it wasnt really needed for our use case. midn you pci passthrough means that your VMs get fixed to a node and the associated GPU. I guess you could do nested virtualization with kubernetes on top for load balancing or try to get the gpus to work directly on baremetal containers (there are no drivers installed in harvester ofc)