This is possible to achieve with NVIDIA cards using nvidia PRIME (it’s just a few environment variable to set to use the GPU for discrete rendering), it’s also possible with virtualGL Seamless GPU passthrough on Qubes OS with VirtualGL (I didn’t test it)
Performance are impacted (~30% less) in this mode as the bandwidth is almost halved because the GPU has to send the picture back (it’s in the magnitude of order of Gb/s on pci express bus or thunderbolt)