Merge commit 'e65bb85ca347d03eee82193cccf3cc9cce47c788' into refactor/v1

tastelikefeet · tastelikefeet · commit a677559d10f1 · 2025-06-03T15:17:50.000+08:00
* commit 'e65bb85ca347d03eee82193cccf3cc9cce47c788':
  minor fix
  add deepseek, dashscope llm; merge _continue_gen
diff --git a/modelscope_agent/llm/dashscope_llm.py b/modelscope_agent/llm/dashscope_llm.py
@@ -1,30 +1,28 @@
-import inspect
-from typing import Any, Optional
+from typing import List
 
-from modelscope_agent.utils.llm_utils import retry
-from modelscope_agent.llm.llm import LLM
+from modelscope_agent.llm.utils import Message, Tool
+from modelscope_agent.llm.openai_llm import OpenAI
+from omegaconf import DictConfig
 
+class DashScope(OpenAI):
 
-class DashScope(LLM):
+    def __init__(self, config: DictConfig):
+        super().__init__(config, base_url=config.llm.dashscope_base_url, api_key=config.llm.dashscope_api_key)
 
-    def __init__(self, system):
-        self.system = system
-        self.client = OpenAI(
-            api_key=self.token,
-            base_url=self.base_url,
-        )
+    def _continue_generate(self, messages: List[Message], new_message, tools: List[Tool] = None, **kwargs):
+        # ref: https://bailian.console.aliyun.com/?tab=doc#/doc/?type=model&url=https%3A%2F%2Fhelp.aliyun.com%2Fdocument_detail%2F2862210.html&renderType=iframe
+        if messages and messages[-1].to_dict().get('partial', False):
 
-    @retry(max_attempts=5)
-    def generate(self, messages, model: Optional[str] = None, tools=None, **kwargs) -> Any:
-        _e = None
-        parameters = inspect.signature(self.client.chat.completions.create).parameters
-        kwargs = {key: value for key, value in kwargs.items() if key in parameters}
-        completion = self.client.chat.completions.create(
-            model=model,
-            messages=messages,
-            tools=tools,
-            parallel_tool_calls=False,
-            **kwargs
-        )
-        return completion
+            messages[-1].reasoning_content += new_message.reasoning_content
+            messages[-1].content += new_message.content
+            if new_message.tool_calls:
+                if messages[-1].tool_calls:
+                    messages[-1].tool_calls += new_message.tool_calls
+                else:
+                    messages[-1].tool_calls = new_message.tool_calls
+        else:
+            messages.append(new_message)
+            messages[-1].partial = True
 
+        messages = self.format_input_message(messages)
+        return self._call_llm(messages=messages, tools=tools, **kwargs)
diff --git a/modelscope_agent/llm/deepseek_llm.py b/modelscope_agent/llm/deepseek_llm.py
@@ -1,25 +1,82 @@
-import inspect
-from typing import Any
-
-
-class DeepSeek:
-
-    def __init__(self, system):
-        self.system = system
-        from openai import OpenAI
-        self.client = OpenAI(
-            api_key=self.token,
-            base_url=self.base_url,
-        )
-
-    def generate(self, messages, model, tools=None, **kwargs) -> Any:
-        parameters = inspect.signature(self.client.chat.completions.create).parameters
-        kwargs = {key: value for key, value in kwargs.items() if key in parameters}
-        completion = self.client.chat.completions.create(
-            model=model,
-            messages=messages,
-            tools=tools,
-            **kwargs
-        )
-        return completion
+from typing import List
 
+from modelscope_agent.llm.utils import Message, Tool
+from modelscope_agent.llm.openai_llm import OpenAI
+from omegaconf import DictConfig
+
+class DeepSeek(OpenAI):
+    input_msg = {'role', 'content', 'tool_calls', 'prefix'}
+
+    def __init__(self, config: DictConfig):
+        super().__init__(config, base_url=config.llm.deepseek_base_url, api_key=config.llm.deepseek_api_key)
+
+    def _continue_generate(self, messages: List[Message], new_message, tools: List[Tool] = None, **kwargs):
+        # ref: https://api-docs.deepseek.com/zh-cn/guides/chat_prefix_completion
+        if messages and messages[-1].to_dict().get('prefix', False):
+
+            messages[-1].reasoning_content += new_message.reasoning_content
+            messages[-1].content += new_message.content
+            if new_message.tool_calls:
+                if messages[-1].tool_calls:
+                    messages[-1].tool_calls += new_message.tool_calls
+                else:
+                    messages[-1].tool_calls = new_message.tool_calls
+        else:
+            messages.append(new_message)
+            messages[-1].prefix = True
+
+        messages = self.format_input_message(messages)
+        stop = kwargs.pop('stop', []).append('```')
+        return self._call_llm(messages=messages, tools=tools, stop=stop, **kwargs)
+
+
+if __name__ == '__main__':
+    import os
+    from omegaconf import OmegaConf
+
+    # 创建一个嵌套的字典结构
+    conf: DictConfig = OmegaConf.create({
+        "llm": {
+            "model": "deepseek-reasoner",
+            "deepseek_base_url": "https://api.deepseek.com/v1",
+            "deepseek_api_key": os.getenv("DEEPSEEK_API_KEY"),
+            "openai_base_url": "https://api-inference.modelscope.cn/v1",
+            "openai_api_key": os.getenv("MODELSCOPE_API_KEY"),
+            "generation_config": {
+                "stream": False,
+                "max_tokens": 500,
+            }
+        }
+    })
+
+    messages = [
+        Message(role='assistant', content='You are a helpful assistant.'),
+        # Message(role='user', content='经度：116.4074，纬度：39.9042是什么地方。用这个名字作为目录名'),
+        # Message(role='user', content='请你简单介绍杭州'),
+        Message(role='user', content='创建2个文件夹，一个叫a,一个叫b'),
+
+    ]
+
+    tools = [
+        # Tool(server_name='amap-maps', tool_name='maps_regeocode', description='将一个高德经纬度坐标转换为行政区划地址信息', parameters={'type': 'object', 'properties': {'location': {'type': 'string', 'description': '经纬度'}}, 'required': ['location']}),
+        Tool(tool_name='mkdir', description='在文件系统创建目录', parameters={'type': 'object', 'properties': {'dir_name': {'type': 'string', 'description': '目录名'}}, 'required': ['dir_name']})
+    ]
+    # tools = None
+
+
+    # 打印配置
+    print(OmegaConf.to_yaml(conf))
+
+    llm = DeepSeek(conf)
+
+    # res = llm.generate(messages=messages, tools=tools, extra_body={'enable_thinking': False})
+    # for chunk in res:
+    #     print(chunk)
+
+    # kwargs覆盖conf
+    message = llm.generate(messages=messages, tools=tools, stream=False, extra_body={'enable_thinking': False})
+    print(message)
+    messages.append(message)
+    # messages.append(Message(role='tool', content='北京市朝阳区崔各庄阿里巴巴朝阳科技园'))
+    # message = llm.generate(messages=messages, tools=tools, stream=False, extra_body={'enable_thinking': False})
+    # print(message)
diff --git a/modelscope_agent/llm/modelscope_llm.py b/modelscope_agent/llm/modelscope_llm.py
@@ -8,3 +8,4 @@ class ModelScope(OpenAI):
 
     def __init__(self, config: DictConfig):
         super().__init__(config, base_url=config.llm.modelscope_base_url, api_key=config.llm.modelscope_api_key)
+
diff --git a/modelscope_agent/llm/openai_llm.py b/modelscope_agent/llm/openai_llm.py
@@ -5,12 +5,13 @@
 from openai.types.chat.chat_completion_message_tool_call import ChatCompletionMessageToolCall, Function
 
 from modelscope_agent.llm.llm import LLM
+from modelscope_agent.utils.llm_utils import retry
 from modelscope_agent.llm.utils import Message, Tool, ToolCall
 from modelscope_agent.utils.utils import assert_package_exist
 
 
-
 class OpenAI(LLM):
+    input_msg = {'role', 'content', 'tool_calls', 'partial', 'prefix'}
 
     def __init__(self, config: DictConfig, base_url: Optional[str] = None,  api_key: Optional[str] = None):
         super().__init__(config)
@@ -23,10 +24,10 @@ def __init__(self, config: DictConfig, base_url: Optional[str] = None,  api_key:
             api_key=api_key,
             base_url=base_url,
         )
-        exclude_fields = {"model", "base_url", "api_key"}
-        self.args: Dict = {k: v for k, v in OmegaConf.to_container(getattr(config, 'generation_config', {}), resolve=True).items() if k not in exclude_fields}
+        self.args: Dict = {k: v for k, v in getattr(config.llm, 'generation_config', {}).items()}
 
-    def generate(self, messages: List[Message], model: Optional[str] = None, tools: List[Tool] = None, **kwargs) -> Message | Generator[Message, None, None]:
+    @retry(max_attempts=3)
+    def generate(self, messages: List[Message], tools: List[Tool] = None, **kwargs) -> Message | Generator[Message, None, None]:
         parameters = inspect.signature(self.client.chat.completions.create).parameters
         args = self.args.copy()
         args.update(kwargs)
@@ -45,44 +46,23 @@ def generate(self, messages: List[Message], model: Optional[str] = None, tools:
                     }
                 } for tool in tools
             ]
-        completion = self._call_llm(model or self.model, messages, tools, **args)
+        completion = self._call_llm(messages, tools, **args)
 
         # 考虑到复杂任务可能存在 单次调用llm生成不完整的情况。需要调用continue_gen判断是否应多次调用以获得完整输出
         if stream:
             return self.stream_continue_generate(messages, completion, tools, **args)
         else:
             return self.continue_generate(messages, completion, tools, **args)
 
-    def _call_llm(self, model, messages, tools, **kwargs):
+    def _call_llm(self, messages, tools, **kwargs):
         messages = self.format_input_message(messages)
         return self.client.chat.completions.create(
-            model=model,
+            model=self.model,
             messages=messages,
             tools=tools,
             **kwargs
         )
 
-    def _stream_continue_generate(self, messages: List[Message], new_message, tools: List[Tool] = None, **kwargs):
-        # 如果上一条消息也和new_message一样不完整，则进行拼接
-        if messages and messages[-1].to_dict().get('partial', False):
-            # 更新最后一条消息的内容
-            messages[-1].reasoning_content += new_message.reasoning_content
-            messages[-1].content += new_message.content
-            if new_message.tool_calls:
-                if messages[-1].tool_calls:
-                    messages[-1].tool_calls += new_message.tool_calls
-                else:
-                    messages[-1].tool_calls = new_message.tool_calls
-        else:
-            # 否则添加为新的 partial 消息
-            new_message.partial = True
-            messages.append(new_message)
-
-        messages = self.format_input_message(messages)
-
-        # 继续调用 LLM 并流式返回后续结果
-        return self._call_llm(messages, tools, **kwargs)
-
     def stream_continue_generate(self, messages: List[Message], completion, tools: List[Tool] = None, **kwargs) -> Generator[Message, None, None]:
         message = None
         for chunk in completion:
@@ -114,13 +94,13 @@ def stream_continue_generate(self, messages: List[Message], completion, tools: L
             yield message_chunk
             if chunk.choices[0].finish_reason in ['length', 'null']:
                 print(f'finish_reason: {chunk.choices[0].finish_reason}， continue generate.')
-                completion = self._stream_continue_generate(messages, message, tools, **kwargs)
+                completion = self._continue_generate(messages, message, tools, **kwargs)
                 for chunk in self.stream_continue_generate(messages, completion, tools, **kwargs):
                     yield chunk
 
     def stream_format_output_message(self, completion_chunk) -> Message:
-        content = completion_chunk.choices[0].delta.content
-        reasoning_content = completion_chunk.choices[0].delta.reasoning_content
+        content = completion_chunk.choices[0].delta.content or ''
+        reasoning_content = completion_chunk.choices[0].delta.reasoning_content or ''
         tool_calls = None
         if completion_chunk.choices[0].delta.tool_calls:
             func = completion_chunk.choices[0].delta.tool_calls
@@ -135,8 +115,8 @@ def stream_format_output_message(self, completion_chunk) -> Message:
         return Message(role='assistant', content=content, reasoning_content=reasoning_content, tool_calls=tool_calls, id=completion_chunk.id)
 
     def format_output_message(self, completion) -> Message:
-        content = completion.choices[0].message.content
-        reasoning_content = completion.choices[0].message.reasoning_content
+        content = completion.choices[0].message.content or ''
+        reasoning_content = completion.choices[0].message.reasoning_content or ''
         tool_calls = None
         if completion.choices[0].message.tool_calls:
             tool_calls = [ToolCall(
@@ -149,11 +129,10 @@ def format_output_message(self, completion) -> Message:
             ]
         return Message(role='assistant', content=content, reasoning_content=reasoning_content, tool_calls=tool_calls, id=completion.id)
 
-    def _continue_generate(self, messages: List[Message], completion, tools: List[Tool] = None, **kwargs):
+    def _continue_generate(self, messages: List[Message], new_message, tools: List[Tool] = None, **kwargs):
         # ref: https://bailian.console.aliyun.com/?tab=doc#/doc/?type=model&url=https%3A%2F%2Fhelp.aliyun.com%2Fdocument_detail%2F2862210.html&renderType=iframe
         # TODO: 移到dashscope_llm并找到真正openai的续写方式
         if messages[-1].to_dict().get('partial', False):
-            new_message = self.format_output_message(completion)
             messages[-1].reasoning_content += new_message.reasoning_content
             messages[-1].content += new_message.content
             if new_message.tool_calls:
@@ -162,22 +141,21 @@ def _continue_generate(self, messages: List[Message], completion, tools: List[To
                 else:
                     messages[-1].tool_calls = new_message.tool_calls
         else:
-            messages.append(self.format_output_message(completion))
+            messages.append(new_message)
             messages[-1].partial = True
 
         messages = self.format_input_message(messages)
         return self._call_llm(messages, tools, **kwargs)
 
     def continue_generate(self, messages: List[Message], completion, tools: List[Tool] = None, **kwargs) -> Message:
-        # finish_reason: Literal["stop", "length", "tool_calls", "content_filter", "function_call"]
-
-
+        new_message = self.format_output_message(completion)
         if completion.choices[0].finish_reason in ['length', 'null']:
+            print(f'new_message: {new_message}')
             print(f'finish_reason: {completion.choices[0].finish_reason}， continue generate.')
-            completion = self._continue_generate(messages, completion, tools, **kwargs)
+            completion = self._continue_generate(messages, new_message, tools, **kwargs)
             return self.continue_generate(messages, completion, tools, **kwargs)
         else:
-            return self.format_output_message(completion)
+            return new_message
 
     def format_input_message(self, messages: List[Message]) -> List[Dict[str, Any]]:
         openai_messages = []
@@ -200,8 +178,7 @@ def format_input_message(self, messages: List[Message]) -> List[Dict[str, Any]]:
                     tool_calls.append(tool_call)
                 message['tool_calls'] = tool_calls
 
-            input_msg = {'role', 'content', 'tool_calls', 'partial'}
-            message = {key: value for key, value in message.items() if key in input_msg and value}
+            message = {key: value for key, value in message.items() if key in self.input_msg and value}
 
             openai_messages.append(message)
 
@@ -233,7 +210,7 @@ def format_input_message(self, messages: List[Message]) -> List[Dict[str, Any]]:
 
     # tools = [
     #     Tool(server_name='amap-maps', tool_name='maps_regeocode', description='将一个高德经纬度坐标转换为行政区划地址信息', parameters={'type': 'object', 'properties': {'location': {'type': 'string', 'description': '经纬度'}}, 'required': ['location']}),
-    #     Tool(tool_name='mkdir', description='在文件系统创建目录', parameters={'type': 'object', 'properties': {'dir_name': {'type': 'string', 'description': '目录名'}}, 'required': ['location']})
+    #     Tool(tool_name='mkdir', description='在文件系统创建目录', parameters={'type': 'object', 'properties': {'dir_name': {'type': 'string', 'description': '目录名'}}, 'required': ['dir_name']})
     # ]
     tools = None
 
diff --git a/modelscope_agent/llm/utils.py b/modelscope_agent/llm/utils.py
@@ -30,7 +30,7 @@ class Tool(TypedDict, total=False):
 class Message:
     role: Required[Literal['system', 'user', 'assistant', 'tool']]
 
-    content: Required[Union[str, List[Dict[str, 'Message']]]]
+    content: str = ''
 
     tool_calls: List[ToolCall] = field(default_factory=list)
 
@@ -46,6 +46,7 @@ class Message:
 
     # 续写模式
     partial: bool = False
+    prefix: bool = False
 
     def to_dict(self):
         return asdict(self)

Original file line number	Diff line number	Diff line change
`@@ -8,3 +8,4 @@ class ModelScope(OpenAI):`
`8`	`8`
`9`	`9`	`def __init__(self, config: DictConfig):`
`10`	`10`	`super().__init__(config, base_url=config.llm.modelscope_base_url, api_key=config.llm.modelscope_api_key)`
	`11`	`+`