add seed params support from vllm_worker

[BuxianChen]

trafficstars

Add seed parameter support for openai api protocol.

I tested vllm_worker:

python3 -m fastchat.serve.controller
CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.vllm_worker --model-path facebook/opt-125m --model-names opt-125m --controller http://localhost:21001 --port 31001 --worker-address http://localhost:31001
python3 -m fastchat.serve.openai_api_server --host 0.0.0.0 --controller-address http://127.0.0.1:21001 --port 8000

The following code will produce same output

from openai import OpenAI
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="token-abc123",
)
completion = client.chat.completions.create(
        model="opt-125m",
        messages=[
            {"role": "user", "content": "Hello!"}
        ],
        seed=1,
        max_tokens=200,
    )
print(completion.choices[0].message.content)

Note that other types of worker may not reproduce same output.

Maybe related to these issues: issue 3215, issue 3216

[surak]

@infwinston what do you think? Makes perfect sense to me!