Bump vLLM version to 0.3.0

[lapp0]

• pin new version
• Fix serving with vLLM, LLMEngine.tokenizer uses a new class now: TokenizerGroup
• Update docs regarding pydantic

Fixes #576

Fixes #504

vLLM main is incompatible with outlines. Do not merge until vllm==0.2.8 is released, leaving as draft until then.

[jc-louis]

vLLM 0.3.0 was just released

(you should also be able to remove this line as vLLM 0.3.0 did migrate to pydantic 2.0)

[jc-louis]

vLLM skipped 0.2.8

Suggested change

	"vllm>=0.2.8",
	"vllm>=0.3.0",

[rlouf]

Let's add vllm>=0.3.0 to pyproject.toml as well. I'll cut a new release as soon as all of this is ready.

[rlouf]

@lapp0 I made a few changes on your branch. Can you confirm that it works on your end, I don't have access to my desktop today.

[lapp0]

@remic33 yes, I will smoke test in a few hours and report the results.

[lapp0]

Change-set looks good after a simple smoke test.

On this branch with vllm==0.3.0:

root@C.9153721:~/outlines$ pip freeze | grep -P "(outlines|vllm)"
UNKNOWN @ file:///root/outlines
vllm==0.3.0
root@C.9153721:~/outlines$ git status
On branch fix-vllm-new-tokenizer
Your branch is up to date with 'origin/fix-vllm-new-tokenizer'.

curl http://127.0.0.1:8000/generate \
    -d '{
        "prompt": "What is Pi? Give me the first 15 digits: ",
        "regex": "(-)?(0|[1-9][0-9]*)(\\.[0-9]+)?([eE][+-][0-9]+)?"
        }'

{"text":["What is Pi? Give me the first 15 digits: 3.14159265358979"]}

On main with vllm==0.3.0:

AttributeError: 'TokenizerGroup' object has no attribute 'get_vocab'.

[rlouf]

All good then, merging