vGPU on Kubernetes

Must someone use vGPU with a guest OS(each vgpu profile has to boot into a guest OS so it can be consumed) even when its intended for kubernetes? (Should not be mistaken for MIG). Or the vGPU manager runs on the host OS and can expose the vGPU profiles to kubernetes with GPU Operator directly?

Yes, with vGPU, you still need a guest OS (VM) to consume each vGPU profile. The vGPU Manager runs on the host hypervisor, but Kubernetes only sees GPUs inside the guest VM, not directly from the host.

If you want Kubernetes to run directly on bare-metal with GPU partitioning, you should be looking at MIG, not vGPU.