I've been playing with the vLLM (VLLMModel) support and generated a sample.

model = VLLMModel(model_id=model_name)

The support is super nice but it could benefit from some additional flexibility.
Other model clients like HfApiModel and OpenAIServerModel already provide client_kwargs where you can add the params for the client.
It could be interesting to extend this functionality to VLLMModel. For example, it would enable configurations like:

model = VLLMModel(model_id=model_name, client_kwargs={"max_model_len": 65536})
...

I already have a fix proposal developed 😄