Update the vllm integration

[rlouf]

We need to update the vllm integration to follow the new user interface.

[justHungryMan]

Does this mean that updating vLLM will use up-to-date outlines?

[chris-aeviator]

this is not to rant, just to highlight: vllm <-> outlines integration is a pain right now and causing suffering for production usage in a lot of teams I talk to.

It would be awesome if outlines could at least state, very prominently, what version compatabilties to expect (and keep it updated) in the readme.

[rlouf]

It would be awesome if said teams could open an issue either here or on outlines-core whenever they encounter an issue. We can only help solve the issues we are aware of.

[chris-aeviator]

While I agree I also know for a fact that outlines is aware of multiple integration issues with vllm regarding

• version compatibilities between outlines and vllm (unsure how to connect two repos via issues, vllm repo is in a messy situation)
• 2-4x slower decoding
• numpy 1

[rlouf]

All 3 being either worked on with the vLLM team or solved in the latest versions. We even created vLLM tags across repos to track those. Let us know if you think there is anything else we can do.