vLLM support for output types and adding mistral.rs as the inference server

[torchss]

As mentioned at https://dottxt-ai.github.io/outlines/latest/features/models/:

As a result, we have limited control over text generation and some output types are not supported.

However, for vLLM, it seems like all output types are supported, in direct contrast to the other server-based models. Also surprisingly, it uses the OpenAI client while the atual OpenAI server-based model doesn't support anything but JSON Schema. As a result I have 2 questions:

1. Is the documentation accurate for Features Matrix: https://dottxt-ai.github.io/outlines/latest/features/models/#features-matrix
2. Can I host any Huggingface model that can be served by vLLM and get full support for all output types?
3. If so, what's so special about the vLLM integration that allows us to get full support for all output types compared to other server-based models
4. If I wanted to use https://github.com/EricLBuehler/mistral.rs as the inference server, depending on its OpenAI compatible endpoint, what are some changes in both mistral.rs, outlines (and anything else) that needs to be made to have it support all output types in a similar manner to vLLM (and are there PRs that I could look at for more handson guides)

RE: #3, I assume (without reading the code) that it's using extra parameters from https://docs.vllm.ai/en/stable/features/structured_outputs.html#online-serving-openai-api - it would be good to know more details and get a confirmation

[rlouf]

The vLLM server integration indeed handles all output types. I'm not sure what mistral.rs uses for structured outputs, and if they can handle every output type, so you would have to check. Happy to add an integration regardless.