You could use llama.cpp rpc mode over "network" via usb4/thunderbolt connection