So mind as well use the OpenAI call with a vanilla vllm server then?

Thats kinda... Surprising? Embeddings models are really small