Merge branch 'main' into release/3.1

swift/llm/template/template_inputs.py

@@ -232,5 +232,10 @@ def messages_join_observation(messages: Messages, tools_prompt='react_en') -> No
                 assert isinstance(pre_content, str)
                 pre_message['content'] = pre_content + content  # assistant
                 messages.pop(i)  # remove tool
+            elif (pre_role == 'assistant' and role == 'assistant' and isinstance(pre_content, str)
+                  and isinstance(content, str)):
+                # Consecutive messages from the assistant role need to be merged to prevent errors.
+                pre_message['content'] = pre_content + content
+                messages.pop(i)
             else:
                 i += 1

swift/plugin/tools.py

@@ -29,7 +29,7 @@ def format_react_en(tool_names, tool_descs):
 
 Begin!
 """
-    tool_descs = [json.dumps(t) if not isinstance(t, str) else t for t in tool_descs]
+    tool_descs = [json.dumps(t, ensure_ascii=False) if not isinstance(t, str) else t for t in tool_descs]
     return REACT_PROMPT.format(tool_list='\n\n'.join(tool_descs), tool_names=','.join(tool_names))
 
 
@@ -49,7 +49,7 @@ def format_react_zh(tool_names, tool_descs):
 
 开始！
 """
-    tool_descs = [json.dumps(t) if not isinstance(t, str) else t for t in tool_descs]
+    tool_descs = [json.dumps(t, ensure_ascii=False) if not isinstance(t, str) else t for t in tool_descs]
     return REACT_ZH_PROMPT.format(tool_list='\n\n'.join(tool_descs), tool_names=','.join(tool_names))
 
 
@@ -59,7 +59,7 @@ def format_glm4(tool_names, tool_descs):
 # 可用工具
 
 {tool_list}"""
-    tool_descs = [json.dumps(t) if not isinstance(t, str) else t for t in tool_descs]
+    tool_descs = [json.dumps(t, ensure_ascii=False) if not isinstance(t, str) else t for t in tool_descs]
     tool_list = ''
     for name, tool in zip(tool_names, tool_descs):
         tool_list += f'## {name}\n\n{tool}\n\n'
@@ -92,7 +92,7 @@ def format_toolbench(tool_names, tool_descs):
 use function Finish->give_up_and_restart.
 2.Do not use origin tool names, use only subfunctions' names.
 Specifically, you have access to the following APIs: {tool_list}"""
-    tool_descs = [json.dumps(t) if not isinstance(t, str) else t for t in tool_descs]
+    tool_descs = [json.dumps(t, ensure_ascii=False) if not isinstance(t, str) else t for t in tool_descs]
     return TOOLBENCH_PROMPT.format(tool_list='\n\n'.join(tool_descs))
 
 
@@ -107,10 +107,20 @@ def format_qwen(tool_names, tool_descs):
 
 {tool_list}
 
-## 你可以在回复中插入以下命令以调用这些工具：
-
-{format_list}
-    '''
+## 你可以在回复中插入以下命令以并行调用N个工具：
+
+✿FUNCTION✿: 工具1的名称，必须是[{tool_names}]之一
+✿ARGS✿: 工具1的输入
+✿FUNCTION✿: 工具2的名称
+✿ARGS✿: 工具2的输入
+...
+✿FUNCTION✿: 工具N的名称
+✿ARGS✿: 工具N的输入
+✿RESULT✿: 工具1的结果
+✿RESULT✿: 工具2的结果
+...
+✿RESULT✿: 工具N的结果
+✿RETURN✿: 根据工具结果进行回复'''
     # 定义星期映射
     weekdays = {0: '星期一', 1: '星期二', 2: '星期三', 3: '星期四', 4: '星期五', 5: '星期六', 6: '星期日'}
     now = dt.datetime.now()
@@ -122,15 +132,13 @@ def format_qwen(tool_names, tool_descs):
     PROMPT = PROMPT.replace('{date}', formatted_date)
     tool_list = ''
     for name, tool in zip(tool_names, tool_descs):
-        tool_list += f'### {name} \n{name}: {tool["description"]} 输入参数: {json.dumps(tool["parameters"])}\n'
+        desc = tool.get('description', '')
+        parameters = json.dumps(params, ensure_ascii=False) if (params := tool.get('parameters')) else ''
+        tool_list += f'### {name}\n\n{name}: {desc} 输入参数: {parameters} 此工具的输入应为JSON对象。'
 
     PROMPT = PROMPT.replace('{tool_list}', tool_list)
-
-    format_list = ''
-    for i, _ in enumerate(tool_names):
-        format_list += f'✿FUNCTION✿:工具{i+1}的名称\n✿ARGS✿:工具{i + 1}的输入\n✿RESULT✿:工具{i + 1}的结果\n'
-    PROMPT = PROMPT.replace('{format_list}', format_list)
-    return PROMPT
+    PROMPT = PROMPT.replace('{tool_names}', ','.join(tool_names))
+    return PROMPT.rstrip()
 
 
 def format_custom(tool_names, tool_descs):
@@ -140,7 +148,7 @@ def format_custom(tool_names, tool_descs):
 
     {tool_list}'''
     tool_list = ''
-    tool_descs = [json.dumps(t) if not isinstance(t, str) else t for t in tool_descs]
+    tool_descs = [json.dumps(t, ensure_ascii=False) if not isinstance(t, str) else t for t in tool_descs]
     for name, tool in zip(tool_names, tool_descs):
         tool_list += f'## {name}\n\n{tool}\n\n'
     return PROMPT.format(tool_list=tool_list)

tests/llm/test_template.py

@@ -43,6 +43,52 @@ def test_template(self):
         response2 = _infer_model(pt_engine)
         assert response == response2
 
+    def test_tool_message_join(self):
+        from copy import deepcopy
+
+        from swift.llm.template.template_inputs import StdTemplateInputs
+        from swift.plugin.tools import get_tools_keyword
+
+        messages = [
+            # first round
+            {
+                'role': 'user',
+                'content': 'testing_user_message'
+            },
+            {
+                'role': 'assistant',
+                'content': ''
+            },
+            {
+                'role': 'tool',
+                'content': ''
+            },
+            # second round
+            {
+                'role': 'assistant',
+                'content': ''
+            },
+            {
+                'role': 'tool',
+                'content': ''
+            },
+        ]
+
+        # testing two template type.
+        for tool_prompt in ('react_en', 'qwen'):
+            tool_prompt = 'react_en'
+            test_messages = deepcopy(messages)
+            obs_word = get_tools_keyword(tool_prompt).get('observation')
+            test_messages[1]['content'] = f'{obs_word}'
+            test_messages[2]['content'] = 'first_round_result\n'
+            test_messages[3]['content'] = f'{obs_word}'
+            test_messages[4]['content'] = 'second_round_result\n'
+            StdTemplateInputs.messages_join_observation(test_messages, tools_prompt=tool_prompt)
+
+            # multi-round tool calling should be joined that only one assistant message left.
+            assert len(test_messages) == 2, f'Tool prompot {tool_prompt} join failed, {messages}'
+            assert test_messages[1]['content'] == f"""{obs_word}first_round_result\n{obs_word}second_round_result\n"""
+
 
 if __name__ == '__main__':
     unittest.main()