--conv-template parameter has no effect。

Nobody has claimed this yet.

I am using the qwen 72B model, and the specified --conv-template does not take effect. If the stop parameter is not specified when calling, the conversation will never end.

启动命令
```
CUDA_VISIBLE_DEVICES=2,3 nohup python -m fastchat.serve.vllm_worker --conv-template qwen-7b-chat --model-path ./finetune_all_weight/checkpoint-6 --trust-remote-code --tensor-parallel-size 2 --dtype bfloat16 --model-names Qwen-72B-ft --gpu-memory-utilization 1 --port 31002 --worker-address http://localhost:31002 > nohup.out.Qwen-72B-ft &
```

调用代码
```
content = 'xxx'
client = OpenAI(
api_key=openai.api_key,
base_url=openai.base_url,
)

chat_response = client.chat.completions.create(
model=model_name,
messages=messages + [
{"role": "user", "content": content},
],
**# stop=['<|im_end|>']**
)
print("Chat response:", chat_response)
content = chat_response.choices[0].message.content
return content
```

Start at the fastchat.serve.vllm_worker entry point and reproduce the Qwen-72B-ft launch command with --conv-template qwen-7b-chat. Trace how that option reaches the chat completion call and verify whether the selected template is applied; done means the conversation ends correctly without requiring an explicit stop parameter.

Written by the indexing model from the issue text.