generating result mismatch while using official repo chat and model_worker

Nobody has claimed this yet.

The model has been trained with SFT using Lora under a fine-tuned chat model which base model is Baichuan-13B-Base.

In multi-chat mode, If use `Baichuan-13B-Chat`'s `model.chat` to generate latest round result, it get right result. But If use model_worker to generate result, it will output the previous round similar result but not related to the latest question.

for example:
Invalid generated result from `model_worker`
request
```
[
{
"role":"user",
"content":"股东除查阅xxx?"
},
{
"role":"assistant",
"content":"根据《公司法》xxxxx, 股东无权进行拍照或摘抄会计凭证的行为。"
},
{
"role":"user",
"content":"1"
}
]
```

response
```
{
"role":"assistant",
"content":"根据《公司法》第xxxx，股东无权进行拍照或摘抄会计凭证的行为。"
}
```
it will generate similar result with first answer.

valid result from `Baichuan official repo`
response
```
{
"role":"assistant",
"content":"您好，我是一个 AI 助手，很高兴为您提供帮助。请问有什么我可以为您做的？"
}
```

The main difference between them is `Baichuan official repo` will add a `eos token` after each round. So I add `eos token` in the template and make it decode as `[2]`(it will also decode as `[2154, 31114, 31219]`). And then, it got a valid result.

Do you have any suggestions?

Start by reproducing the multi-chat request through model_worker and compare its prompt construction with Baichuan-13B-Chat's model.chat behavior. Inspect the model_worker chat template and EOS-token handling described in the issue; done means the latest user turn produces a response related to that turn rather than repeating the previous answer.

Written by the indexing model from the issue text.