With normal passthrough, your host loses access to the gpu, no? So you need to have two gpus, one for the host and one for the guest. Correct me, if I am wrong, but this should make the host fully operational on a single gpu and still let vm guests have headless access to the host gpu.