Which driver I should install for pass-through now and vGPU later

Hello all
I am going to deploy a machine with two V100 SXM 32G gpus, it could be 8 in the future.
My hypervisor would be VMWare ESXi.

I know if I use the Tesla driver then I don’t need a license but grid need.
At this time, I would just use a guest machine for training, all the work can be done in ssh console, so I may not need OpenGL.
In this case, should I install NVIDIA vGPU Manager ? If I installed the NVIDIA vGPU Manager in the hypervisor, can I still use Tesla driver(or no license) to get full access to the GPUs(including NVLink and P2P).

Then when I have more GPUs, I would like to assign one or two GPU for a guest windows machine to verified my training result. So I must buy the Quadro Virtual Data Center Workstation(vDWS) license ? The other license won’t support CUDA nor Virtual Compute Server would support windows.

Besides, the subscription method for Quadro vDWS, if I want to run two VMs at the same time, I need two perpetual license and a SUMS every years ? It would sound better to deploy the windows beyond the VM and using a customer graphics card.