fix glm

Author: shell-nlp

gpt_server/model_handler/prompts.py

@@ -15,18 +15,16 @@ def __init__(
         separator="\n",
         tools="""\n\n你可以使用以下工具提供适当的答复和支持。\n\n# 可用工具\n\n在<tools></tools> XML标签中提供了function的签名(即函数的结构信息):\n<tools>""",
         eotools="""\n</tools>
-Use the following format:
+## 如果使用工具，你可以在回复中插入零次、一次或多次以下命令以调用工具：
 
-Question: the input question you must answer
-Thought: you should always think about what to do
 Action: the action to take, should be one of [{tool_names}]
 Action Input: the input to the action
-Observation: the result of the action
-... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
-Thought: I now know the final answer
-Final Answer: the final answer to the original input question
 
-Begin!      
+形如：
+Action: get_result
+Action Input: {{"a": "1","b": "2"}}
+
+如果回答问题已经不需要再继续使用工具，则不需要再使用Action、Action Input格式，可直接回答。 
 """,
         stop_words=["<|user|>", "<|endoftext|>", "<|observation|>"],
         meta_instruction="你是一个名为 GLM-4 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。",
@@ -102,15 +100,15 @@ def messages2prompt_base(self, messages, sequence_start=True, tools=None, **kwar
             for tool in tools:
                 tool_names.append(tool["function"]["name"])
             tool_names = ",".join(tool_names)
-            self.eotools = self.eotools.format(tool_names=tool_names)
+            eotools = self.eotools.format(tool_names=tool_names)
             for tool in tools:
                 tool_prompt += self.separator
                 tool_prompt += f'{{"type": "function", "function": {json.dumps(tool, ensure_ascii=False)}}}'
             if len(messages) and messages[0]["role"] == "system":
-                ret += f"{self.system}{messages[0]['content']}{self.tools}{tool_prompt}{self.eotools}{self.eosys}"
+                ret += f"{self.system}{messages[0]['content']}{self.tools}{tool_prompt}{eotools}{self.eosys}"
                 messages.pop(0)
             else:
-                ret += f"{self.system}{self.meta_instruction}{self.tools}{tool_prompt}{self.eotools}{self.eosys}"
+                ret += f"{self.system}{self.meta_instruction}{self.tools}{tool_prompt}{eotools}{self.eosys}"
 
         for message in messages:
             role = message["role"]

gpt_server/model_handler/tool_parser.py

@@ -63,7 +63,8 @@ def extract_tool_calls(
     ) -> ExtractedToolCallInformation:
         text = model_output
         try:
-
+            if "Action:" not in text:
+                raise Exception
             i = text.rfind("Action:")
             j = text.rfind("Action Input:")
             name = text[i + len("Action:") : j].strip().strip(".")
@@ -324,7 +325,10 @@ def tool_parser(full_text: str, tool_parser: ToolParser, tools, ret):
         ret["finish_reason"] = "tool_calls"
         return json.dumps(ret).encode() + b"\0"
     else:
+        logger.info(f"工具解析失败, tool_calls: {tool_calls}")
         ret["text"] = ""
+        ret["tool_calls"] = tool_calls
+        ret["finish_reason"] = "tool_calls"
         return json.dumps(ret).encode() + b"\0"
 
 

gpt_server/model_worker/chatglm.py

@@ -50,9 +50,7 @@ def __init__(
         )
         self.stop_words_ids = []
 
-        self.stop = [
-            self.tokenizer.decode(skip_word) for skip_word in self.stop_words_ids
-        ]
+        self.stop = ["Observation:"]
         logger.warning(f"{model_names[0]} 停用词: {self.stop}")
 
     async def generate_stream_gate(self, params):

tests/test_openai_completion_tool_calling.py

@@ -5,7 +5,7 @@
 client = OpenAI(api_key="EMPTY", base_url="http://localhost:8082/v1")
 
 
-def get_weather(location: str, unit: str):
+def get_weather(location: str, unit: str = "celsius"):
     return f"Getting the weather for {location} in {unit}..."
 
 
@@ -37,8 +37,10 @@ def get_weather(location: str, unit: str):
     tools=tools,
     tool_choice="auto",
 )
-tool_call = response.choices[0].message.tool_calls[0].function
+
+print("message", response.choices[0].message)
 print(response.choices[0].message.tool_calls)
+tool_call = response.choices[0].message.tool_calls[0].function
 print(f"Function called: {tool_call.name}")
 print(f"Arguments: {tool_call.arguments}")
 print(f"Result: {get_weather(**json.loads(tool_call.arguments))}")