GPTAPI类型模型，fastapi返回消息报错：缺少必要字段“input”

[idontwantagirlfriend]

症状描述

在frontend中正常输入搜索关键词并触发搜索后，卡住不动。fastapi报错：

{"object":"error","message":"[{'type': 'missing', 'loc': ('body', 'messages'), 'msg': 'Field required', 'input': {'model': 'qwen2-vl-7b-instruct', 'input': {'messages': [{'role': 'system', 'content': 'The current date is 2024-12-04.'}, {'role': 'system', 'content': \"## Character Profile\\nYou are a programmer capable of Python programming in a Jupyter environment. You can utilize the provided API to construct a Web Search Graph, ultimately generating and executing code.\\n\\n## API Description\\n\\nBelow is the API documentation for the WebSearchGraph class, including detailed attribute descriptions:\\n\\n### Class: WebSearchGraph\\n\\nThis class manages nodes and edges of a web search graph and conducts searches via a web proxy.\\n\\n#### Initialization Method\\n\\nInitializes an instance of WebSearchGraph.\\n\\n**Attributes:**\\n\\n- nodes (Dict[str, Dict[str, str]]): A dictionary storing all nodes in the graph. Each node is indexed by its name and contains content, type, and other related information.\\n- adjacency_list (Dict[str, List[str]]): An adjacency list storing the connections between all nodes in the graph. Each node is indexed by its name and contains a list of adjacent node names.\\n\\n#### Method: add_root_node\\n\\nAdds the initial question as the root node.\\n**Parameters:**\\n\\n- node_content (str): The user's question.\\n- node_name (str, optional): The node name, default is 'root'.\\n\\n#### Method: add_node\\n\\nAdds a sub-question node and returns search results.\\n**Parameters:**\\n\\n- node_name (str): The node name.\\n- node_content (str): The sub-question content.\\n\\n**Returns:**\\n\\n- str: Returns the search results.\\n\\n#### Method: add_response_node\\n\\nAdds a response node when the current information satisfies the question's requirements.\\n\\n**Parameters:**\\n\\n- node_name (str, optional): The node name, default is 'response'.\\n\\n#### Method: add_edge\\n\\nAdds an edge.\\n\\n**Parameters:**\\n\\n- start_node (str): The starting node name.\\n- end_node (str): The ending node name.\\n\\n#### Method: reset\\n\\nResets nodes and edges.\\n\\n#### Method: node\\n\\nGet node information.\\n\\npython\\ndef node(self, node_name: str) -> str\\n\\n**Parameters:**\\n\\n- node_name (str): The node name.\\n\\n**Returns:**\\n\\n- str: Returns a dictionary containing the node's information, including content, type, thought process (if any), and list of predecessor nodes.\\n\\n## Task Description\\nBy breaking down a question into sub-questions that can be answered through searches (unrelated questions can be searched concurrently), each search query should be a single question focusing on a specific person, event, object, specific time point, location, or knowledge point. It should not be a compound question (e.g., a time period). Step by step, build the search graph to finally answer the question.\\n\\n## Considerations\\n\\n1. Each search node's content must be a single question; do not include multiple questions (e.g., do not ask multiple knowledge points or compare and filter multiple things simultaneously, like asking for differences between A, B, and C, or price ranges -> query each separately).\\n2. Do not fabricate search results; wait for the code to return results.\\n3. Do not repeat the same question; continue asking based on existing questions.\\n4. When adding a response node, add it separately; do not add a response node and other nodes simultaneously.\\n5. In a single output, do not include multiple code blocks; only one code block per output.\\n6. Each code block should be placed within a code block marker, and after generating the code, add an <|action_end|> tag as shown below:\\n    <|action_start|><|interpreter|>\\n    ```python\\n    # Your code block (Note that the 'Get new added node information' logic must be added at the end of the code block, such as 'graph.node('...')')\\n    ```<|action_end|>\\n7. The final response should add a response node with node_name 'response', and no other nodes should be added.\\n\", 'name': 'interpreter'}, {'role': 'user', 'content': '如果想要更丝滑的体验，请在本地搭建-Mi'}]}, 'parameters': {'top_p': 0.9, 'temperature': 0.2, 'stream': True, 'max_tokens': 4096, 'stop': ['<|im_end|>'], 'repetition_penalty': 1.1, 'result_format': 'message'}}}, {'type': 'extra_forbidden', 'loc': ('body', 'input'), 'msg': 'Extra inputs are not permitted', 'input': {'messages': [{'role': 'system', 'content': 'The current date is 2024-12-04.'}, {'role': 'system', 'content': \"## Character Profile\\nYou are a programmer capable of Python programming in a Jupyter environment. You can utilize the provided API to construct a Web Search Graph, ultimately generating and executing code.\\n\\n## API Description\\n\\nBelow is the API documentation for the WebSearchGraph class, including detailed attribute descriptions:\\n\\n### Class: WebSearchGraph\\n\\nThis class manages nodes and edges of a web search graph and conducts searches via a web proxy.\\n\\n#### Initialization Method\\n\\nInitializes an instance of WebSearchGraph.\\n\\n**Attributes:**\\n\\n- nodes (Dict[str, Dict[str, str]]): A dictionary storing all nodes in the graph. Each node is indexed by its name and contains content, type, and other related information.\\n- adjacency_list (Dict[str, List[str]]): An adjacency list storing the connections between all nodes in the graph. Each node is indexed by its name and contains a list of adjacent node names.\\n\\n#### Method: add_root_node\\n\\nAdds the initial question as the root node.\\n**Parameters:**\\n\\n- node_content (str): The user's question.\\n- node_name (str, optional): The node name, default is 'root'.\\n\\n#### Method: add_node\\n\\nAdds a sub-question node and returns search results.\\n**Parameters:**\\n\\n- node_name (str): The node name.\\n- node_content (str): The sub-question content.\\n\\n**Returns:**\\n\\n- str: Returns the search results.\\n\\n#### Method: add_response_node\\n\\nAdds a response node when the current information satisfies the question's requirements.\\n\\n**Parameters:**\\n\\n- node_name (str, optional): The node name, default is 'response'.\\n\\n#### Method: add_edge\\n\\nAdds an edge.\\n\\n**Parameters:**\\n\\n- start_node (str): The starting node name.\\n- end_node (str): The ending node name.\\n\\n#### Method: reset\\n\\nResets nodes and edges.\\n\\n#### Method: node\\n\\nGet node information.\\n\\npython\\ndef node(self, node_name: str) -> str\\n\\n**Parameters:**\\n\\n- node_name (str): The node name.\\n\\n**Returns:**\\n\\n- str: Returns a dictionary containing the node's information, including content, type, thought process (if any), and list of predecessor nodes.\\n\\n## Task Description\\nBy breaking down a question into sub-questions that can be answered through searches (unrelated questions can be searched concurrently), each search query should be a single question focusing on a specific person, event, object, specific time point, location, or knowledge point. It should not be a compound question (e.g., a time period). Step by step, build the search graph to finally answer the question.\\n\\n## Considerations\\n\\n1. Each search node's content must be a single question; do not include multiple questions (e.g., do not ask multiple knowledge points or compare and filter multiple things simultaneously, like asking for differences between A, B, and C, or price ranges -> query each separately).\\n2. Do not fabricate search results; wait for the code to return results.\\n3. Do not repeat the same question; continue asking based on existing questions.\\n4. When adding a response node, add it separately; do not add a response node and other nodes simultaneously.\\n5. In a single output, do not include multiple code blocks; only one code block per output.\\n6. Each code block should be placed within a code block marker, and after generating the code, add an <|action_end|> tag as shown below:\\n    <|action_start|><|interpreter|>\\n    ```python\\n    # Your code block (Note that the 'Get new added node information' logic must be added at the end of the code block, such as 'graph.node('...')')\\n    ```<|action_end|>\\n7. The final response should add a response node with node_name 'response', and no other nodes should be added.\\n\", 'name': 'interpreter'}, {'role': 'user', 'content': '如果想要更丝滑的体验，请在本地搭建-Mi'}]}}, {'type': 'extra_forbidden', 'loc': ('body', 'parameters'), 'msg': 'Extra inputs are not permitted', 'input': {'top_p': 0.9, 'temperature': 0.2, 'stream': True, 'max_tokens': 4096, 'stop': ['<|im_end|>'], 'repetition_penalty': 1.1, 'result_format': 'message'}}]","type":"BadRequestError","param":null,"code":400}

可以从信息中发现request字典多了一层外面的模型名，应该是里面那层input。

vLLM后端报错：

"POST /v1/chat/completions HTTP/1.1" 400 BAD REQUEST

如何复现

（models.py要加一行load_dotenv，否则还加载不出来api和key）
设置以下环境变量：

VLLM_BASE_URL="http://192.168.1.9:8008/v1"
VLLM_API_KEY=[MY_API_KEY]
VLLM_MODEL_NAME="qwen2-vl-7b-instruct"

使用以下模型设置：

vllm = dict(
    type=GPTAPI,
    model_type=os.environ.get("VLLM_MODEL_NAME"),
    key=os.environ.get("VLLM_API_KEY"),
    api_base= f"{os.environ.get("VLLM_BASE_URL")}/chat/completions",
    meta_template=[
        dict(role="system", api_role="system"),
        dict(role="user", api_role="user"),
        dict(role="assistant", api_role="assistant"),
        dict(role="environment", api_role="system"),
    ],
    top_p=0.9,
    # top_k=1,
    temperature=0.2,
    max_new_tokens=4096,
    repetition_penalty=1.1,
    stop_words=["<|im_end|>"],
)

启动fastapi

python -m mindsearch.app --lang en --model_format vllm --search_engine GoogleSearch --asy

在前端正常输入搜索就会报错。

但是把模型名字改成gpt-smart，就能输出一部分内容了，虽然也是gibberish，但是至少输出了。（vllm显示名，实际模型和之前一样）

真是令人费解！

是否影响使用

由于这个问题而无法使用。