fix chat template with tool call (#3773)

Author: RunningLeon

lmdeploy/model.py

@@ -1965,18 +1965,20 @@ def match(cls, model_path: str) -> Optional[str]:
             return 'llama4'
 
 
+@MODELS.register_module(name='intern-s1')
 @MODELS.register_module(name='interns1')
 class InternS1(InternVL2_5):
 
     def __init__(
             self,
-            tool='# External Tools\nYou have access to these tools:\n',
-            eotool='\n# Tool Call Formatted\nYour response should consist of a reasoning step (**thought**) followed immediately by a function call in valid JSON format. Wrap each function call using the `<|action_start|><|plugin|>` and `<|action_end|>` tags.\n**Format example:**\n```\n(Your thought goes here...)\n<|action_start|><|plugin|>\n{\n    "name": "tool_name",\n    "parameters": {\n        "parameter1": "value1",\n        "parameter2": "value2"\n    }\n}\n<|action_end|>\n```',  # noqa: E501
+            tool='\n\nYour response should consist of a reasoning step (**thought**) followed immediately by a function call in valid JSON format. Wrap each function call using the `<|action_start|><|plugin|>` and `<|action_end|>` tags.\n\n**Format example:**\n\n```\n(Your thought goes here...)\n\n<|action_start|><|plugin|>\n{\n    "name": "tool_name",\n    "parameters": {\n        "parameter1": "value1",\n        "parameter2": "value2"\n    }\n}\n<|action_end|>\n```\n\n# External Tools\nYou have access to these tools:\n',  # noqa: E501
+            eotool='',
             meta_instruction='You are an expert reasoner with extensive experience in all areas. You approach problems through systematic thinking and rigorous reasoning. Your response should reflect deep understanding and precise logical thinking, making your solution path and reasoning clear to others. Please put your thinking process within <think>...</think> tags.',  # noqa: E501
             **kwargs):
+        super(InternVL2_5, self).__init__(meta_instruction=meta_instruction, **kwargs)
+
         self.tool = tool or ''
         self.eotool = eotool or ''
-        super(InternVL2_5, self).__init__(meta_instruction=meta_instruction, **kwargs)
 
     def messages2prompt(self, messages, sequence_start=True, tools=None, enable_thinking=None, **kwargs):
         """Return the prompt that is concatenated with other elements in the
@@ -2000,24 +2002,40 @@ def messages2prompt(self, messages, sequence_start=True, tools=None, enable_thin
                        environment=self.eoenv,
                        tool=self.eoenv)
         name_map = dict(plugin=self.plugin, interpreter=self.interpreter)
+
         ret = ''
-        if self.meta_instruction is not None and sequence_start:
-            if len(messages):
-                if messages[0]['role'] != 'system' and enable_thinking is not False:
-                    ret += f'{self.system}{self.meta_instruction}{eox_map["system"]}'
 
         if tools:
             tools_prompt = dict(
                 role='system',
                 name='plugin',  # only support internlm2
-                content=f'{self.tool}{json.dumps(tools, ensure_ascii=False)}{self.eotool}')
-            insert_index = 0
+                content=f'{self.tool}{json.dumps(tools, ensure_ascii=False, indent=2)}{self.eotool}')
+
             if messages[0]['role'] == 'system':
-                insert_index = 1
-            messages.insert(insert_index, tools_prompt)
-        for message in messages:
+                tools_prompt['content'] = messages[0]['content'] + tools_prompt['content']
+                messages[0] = tools_prompt
+            else:
+                if self.meta_instruction is not None and sequence_start and enable_thinking is not False:
+                    tools_prompt['content'] = self.meta_instruction + tools_prompt['content']
+                else:
+                    tools_prompt['content'] = tools_prompt['content'].lstrip('\n')
+                messages.insert(0, tools_prompt)
+        elif self.meta_instruction is not None and sequence_start:
+            if len(messages):
+                if messages[0]['role'] != 'system' and enable_thinking is not False:
+                    ret += f'{self.system}{self.meta_instruction}{eox_map["system"]}'
+        # find index of last user input section
+        last_user_idx = -1
+        for idx in range(len(messages) - 1, -1, -1):
+            if messages[idx]['role'] == 'user':
+                last_user_idx = idx
+                break
+
+        for idx, message in enumerate(messages):
             role = message['role']
             content = get_text(message['content'])
+            if last_user_idx != -1 and idx > last_user_idx and message.get('reasoning_content', None) is not None:
+                content = f'<think>\n{message["reasoning_content"]}\n</think>\n{content}'
             if role == 'assistant' and message.get('tool_calls', None) is not None:
                 for tool_call in message['tool_calls']:
                     function = tool_call.get('function', {})
@@ -2026,9 +2044,14 @@ def messages2prompt(self, messages, sequence_start=True, tools=None, enable_thin
                     function.pop('arguments')
                     if isinstance(function['parameters'], str):
                         function['parameters'] = json.loads(function['parameters'])
-                    content += f'<|action_start|><|plugin|>\n{json.dumps(function, ensure_ascii=False)}<|action_end|>'
-            if 'name' in message and message['name'] in name_map:
-                begin = box_map[role].strip() + f" name={name_map[message['name']]}\n"
+                    content += f'<|action_start|><|plugin|>\n{json.dumps(function, ensure_ascii=False)}\n<|action_end|>'
+
+            if 'name' in message:
+                begin = box_map[role].strip()
+                if message['name'] in name_map:
+                    begin = begin + f" name={name_map[message['name']]}\n"
+                elif role == 'tool':
+                    begin = begin + f" name={name_map['plugin']}\n"
             else:
                 begin = box_map[role]
             ret += f'{begin}{content}{eox_map[role]}'
@@ -2048,8 +2071,8 @@ def match(cls, model_path: str) -> Optional[str]:
             model_path (str): the model path used for matching.
         """
         path = model_path.lower()
-        if 'interns1' in path:
-            return 'interns1'
+        if 'intern-s1' in path or 'interns1' in path:
+            return 'intern-s1'
 
 
 def best_match_model(query: str) -> Optional[str]:

tests/test_lmdeploy/test_model.py

@@ -1008,13 +1008,14 @@ def test_qwen3(model_path, enable_thinking):
 
 
 @pytest.mark.parametrize('model_path', ['internlm/Intern-S1'])
-@pytest.mark.parametrize('enable_thinking', [True, False, None])
-def test_interns1(model_path, enable_thinking):
+@pytest.mark.parametrize('enable_thinking', [None, True, False])
+@pytest.mark.parametrize('has_user_sys', [True, False])
+def test_interns1(model_path, enable_thinking, has_user_sys):
     from transformers import AutoTokenizer
     try:
         tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
     except OSError:
-        pytest.skip(reason='internlm/Intern-S1 not exists')
+        pytest.skip(reason=f'{model_path} not exists')
 
     chat_template_name = best_match_model(model_path)
     chat_template = MODELS.get(chat_template_name)()
@@ -1032,6 +1033,9 @@ def test_interns1(model_path, enable_thinking):
         'role': 'user',
         'content': 'AGI is?'
     }]
+    if not has_user_sys:
+        messages = messages[1:]
+
     if enable_thinking is None:
         ref = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     else:
@@ -1041,3 +1045,120 @@ def test_interns1(model_path, enable_thinking):
                                             enable_thinking=enable_thinking)
     lm_res = chat_template.messages2prompt(messages, enable_thinking=enable_thinking)
     assert ref == lm_res
+
+
+@pytest.mark.parametrize('model_path', ['internlm/Intern-S1'])
+@pytest.mark.parametrize('enable_thinking', [None, True, False])
+@pytest.mark.parametrize('has_user_sys', [True, False])
+def test_interns1_tools(model_path, enable_thinking, has_user_sys):
+    from transformers import AutoTokenizer
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    except OSError:
+        pytest.skip(reason=f'{model_path} not exists')
+
+    chat_template_name = best_match_model(model_path)
+    chat_template = MODELS.get(chat_template_name)()
+
+    tools = [
+        {
+            'type': 'function',
+            'function': {
+                'name': 'find_user_id_by_name_zip',
+                'description':
+                'Find user id by first name, last name, and zip code. If the user is not found, the function will return an error message. By default, find user id by email, and only call this function if the user is not found by email or cannot remember email.',  # noqa: E501
+                'parameters': {
+                    'type': 'object',
+                    'properties': {
+                        'first_name': {
+                            'type': 'string',
+                            'description': "The first name of the customer, such as 'John'."
+                        },
+                        'last_name': {
+                            'type': 'string',
+                            'description': "The last name of the customer, such as 'Doe'."
+                        },
+                        'zip': {
+                            'type': 'string',
+                            'description': "The zip code of the customer, such as '12345'."
+                        }
+                    },
+                    'required': ['first_name', 'last_name', 'zip']
+                }
+            }
+        },
+        {
+            'type': 'function',
+            'function': {
+                'name': 'get_order_details',
+                'description': 'Get the status and details of an order.',
+                'parameters': {
+                    'type': 'object',
+                    'properties': {
+                        'order_id': {
+                            'type':
+                            'string',
+                            'description':
+                            "The order id, such as '#W0000000'. Be careful there is a '#' symbol at the beginning of the order id."  # noqa: E501
+                        }
+                    },
+                    'required': ['order_id']
+                }
+            }
+        }
+    ]
+    messages = [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant'
+        },
+        {
+            'role': 'user',
+            'content': "Hi there! I'm looking to return a couple of items from a recent order."
+        },
+        {
+            'role':
+            'assistant',
+            'content':
+            'Could you please provide your email address associated with the account, or share your first name, last name, and zip code?',  # noqa: E501
+            'reasoning_content':
+            'Okay, the user wants to return some items from a recent order. Let me start by authenticating their identity...'  # noqa: E501
+        },
+        {
+            'role': 'user',
+            'content': 'Sure, my name is Omar Anderson and my zip code is 19031.'
+        },
+        {
+            'role':
+            'assistant',
+            'content':
+            '<content>',
+            'reasoning_content':
+            "Since he didn't provide an email, I should use the find_user_id_by_name_zip function. Let me...",  # noqa: E501
+            'tool_calls': [{
+                'function': {
+                    'arguments': '{"first_name": "Omar", "last_name": "Anderson", "zip": "19031"}',
+                    'name': 'find_user_id_by_name_zip'
+                },
+                'id': 'chatcmpl-tool-a9f439084bfc4af29fee2e5105050a38',
+                'type': 'function'
+            }]
+        },
+        {
+            'content': 'omar_anderson_3203',
+            'name': 'find_user_id_by_name_zip',
+            'role': 'tool'
+        }
+    ]
+    if not has_user_sys:
+        messages = messages[1:]
+    if enable_thinking is None:
+        ref = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, tools=tools)
+    else:
+        ref = tokenizer.apply_chat_template(messages,
+                                            tokenize=False,
+                                            add_generation_prompt=True,
+                                            tools=tools,
+                                            enable_thinking=enable_thinking)
+    lm_res = chat_template.messages2prompt(messages, enable_thinking=enable_thinking, tools=tools)
+    assert ref == lm_res