When the type of context in the incoming messages is text, an error occurs.

When the type of context in the incoming messages is text, an error occurs.

API: /v1/chat/completions

request

{
    "max_tokens": 0,
    "model": "qwen-72b-chat-int4",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "text": "Assistant is a large language model trained by Meta.",
                    "type": "text"
                }
            ]
        }
    ],
    "functions": null,
    "function_call": null,
    "temperature": 0,
    "top_p": 0,
    "n": 0,
    "stream": true,
    "stop": [
        "\nObservation:","Observation:"
    ],
    "presence_penalty": 0,
    "frequency_penalty": 0,
    "logit_bias": null,
    "user": "",
    "response_format": {
        "type": ""
    },
    "seed": null,
    "tools": null,
    "tool_choice": null
}

ERROR

ERROR | stderr |   File "/app/fastchat/serve/openai_api_server.py", line 420, in create_chat_completion
ERROR | stderr |     gen_params = await get_gen_params(
ERROR | stderr |   File "/app/fastchat/serve/openai_api_server.py", line 329, in get_gen_params
ERROR | stderr |     prompt = conv.get_prompt()
ERROR | stderr |   File "/app/fastchat/conversation.py", line 168, in get_prompt
ERROR | stderr |     ret += role + "\n" + message + self.sep + "\n"
ERROR | stderr | TypeError: can only concatenate str (not "tuple") to str

Start with the failing request in the issue and inspect get_gen_params in fastchat/serve/openai_api_server.py, then follow the call into conversation.py at Conversation.get_prompt. Reproduce the request and trace how the text message is represented. Done means this input no longer raises the reported TypeError while producing the expected chat-completion response.