Two vLLM servers on one GPU: the startup check that bites, and the fix

Someone asked on the vLLM tracker how to run a small embedding model next to a chat model on one GPU,...