I'm using `OpenAILike` to talk to a Vllm

I'm using OpenAILike to talk to a Vllm instance. I need to pass a custom stop token, and currently the only way i can figure out how to do this is like this:

llm.complete(prompt, True, extra_body={"stop_token_ids":[...]})

This doesn't work with llm.predict because it interprets all remaining kwargs as prompt template expansion arguments. Is there any other way to get this KV in the outoging OpenAI-API request?