[fix]update apply_chat_template (#4249)

* [fix]Modify follow-up push parameters and Modify the verification method for thinking length (#4086)

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* 续推参数  generated_token_ids 修改成 completion_token_ids;修改思考长度校验方式

* add completion_token_ids

* add logger

* fix reasoning_max_tokens ParameterError

* add unittest

* add unittest

* add unittest

* add unittest

* add unittest

* add unit test

* fix

* [fix]update apply_chat_template (#4137)

* update apply_chat_template

* fix unittest

* fix unittest

* fix

* fix

* fix unit test

* fix

* fix unit test

* add unit test

Author: luukunn

fastdeploy/engine/engine.py

@@ -222,7 +222,9 @@ def add_requests(self, task, sampling_params=None, **kwargs):
         if sampling_params is not None:
             request.sampling_params = sampling_params
         request.preprocess_start_time = time.time()
-
+        chat_template_kwargs = kwargs.get("chat_template_kwargs") or {}
+        chat_template_kwargs["chat_template"] = kwargs.get("chat_template")
+        kwargs["chat_template_kwargs"] = chat_template_kwargs
         request = self.data_processor.process_request(request, self.cfg.max_model_len, **kwargs)
         request.prompt_token_ids_len = len(request.prompt_token_ids)
         request.need_prefill_tokens = request.prompt_token_ids_len

fastdeploy/entrypoints/engine_client.py

@@ -172,6 +172,9 @@ async def add_requests(self, task):
 
         task["preprocess_start_time"] = time.time()
         try:
+            chat_template_kwargs = task.get("chat_template_kwargs", {})
+            chat_template_kwargs.update({"chat_template": task.get("chat_template"), "tools": task.get("tools")})
+            task["chat_template_kwargs"] = chat_template_kwargs
             if inspect.iscoroutinefunction(self.data_processor.process_request_dict):
                 await self.data_processor.process_request_dict(task, self.max_model_len)
             else:

fastdeploy/input/ernie4_5_processor.py

@@ -88,7 +88,6 @@ def process_request(self, request, max_model_len=None, **kwargs):
             str: error message
         """
         data_processor_logger.info(f"Start processing request: {request}")
-        request.chat_template = kwargs.get("chat_template")
         request = self._apply_default_parameters(request)
         if request.get("eos_token_ids") is None or len(request.eos_token_ids) == 0:
             request.eos_token_ids = self.eos_token_ids
@@ -127,15 +126,15 @@ def process_request(self, request, max_model_len=None, **kwargs):
                     )
             elif request.messages is not None:
                 task = request.to_dict()
-                chat_template_kwargs = kwargs.get("chat_template_kwargs")
+                chat_template_kwargs = kwargs.get("chat_template_kwargs", {})
                 if chat_template_kwargs:
                     if isinstance(chat_template_kwargs, dict):
                         for k, v in chat_template_kwargs.items():
                             if k not in task:
                                 task[k] = v
                     else:
                         raise ValueError("Invalid input: chat_template_kwargs must be a dict")
-                request.prompt_token_ids = self.messages2ids(task)
+                request.prompt_token_ids = self.messages2ids(task, **chat_template_kwargs)
             else:
                 raise ValueError(f"The request should have `prompt_token_ids`, `prompt` or `messages`: {request}.")
 
@@ -205,15 +204,15 @@ def process_request_dict(self, request, max_model_len=None):
                     req_id = request.get("request_id", None)
                     data_processor_logger.info(f"req_id:{req_id}, tokens:{tokens}, token_ids: {token_ids}")
             elif request.get("messages"):
-                chat_template_kwargs = request.get("chat_template_kwargs")
+                chat_template_kwargs = request.get("chat_template_kwargs", {})
                 if chat_template_kwargs:
                     if isinstance(chat_template_kwargs, dict):
                         for k, v in chat_template_kwargs.items():
                             if k not in request:
                                 request[k] = v
                     else:
                         raise ValueError("Invalid input: chat_template_kwargs must be a dict")
-                request["prompt_token_ids"] = self.messages2ids(request)
+                request["prompt_token_ids"] = self.messages2ids(request, **chat_template_kwargs)
             else:
                 raise ValueError(f"Request must contain 'prompt_token_ids', 'prompt', or 'messages': {request}")
 
@@ -379,7 +378,7 @@ def process_response_dict_streaming(self, response_dict, **kwargs):
                 del self.tool_parser_dict[req_id]
         return response_dict
 
-    def messages2ids(self, request_or_messages):
+    def messages2ids(self, request_or_messages, **kwargs):
         """
         Convert multi-turn messages into ID sequences.
 
@@ -397,7 +396,7 @@ def messages2ids(self, request_or_messages):
             tokenize=False,
             split_special_tokens=False,
             add_special_tokens=False,
-            chat_template=request_or_messages.get("chat_template", None),
+            **kwargs,
         )
         request_or_messages["text_after_process"] = spliced_message
         req_id = None

fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py

@@ -113,7 +113,6 @@ def set_value(req, key, value):
 
     def process_request(self, request, max_model_len=None, **kwargs):
         """process the input data"""
-        request.chat_template = kwargs.get("chat_template")
         task = request.to_dict()
         task["chat_template_kwargs"] = kwargs.get("chat_template_kwargs")
         self.process_request_dict(task, max_model_len)

fastdeploy/input/ernie4_5_vl_processor/process.py

@@ -250,8 +250,8 @@ def request2ids(
                     "video",
                 ]:
                     image_message_list.append(item)
-
-        prompt_token_ids = self.apply_chat_template(request)
+        chat_template_kwargs = request.get("chat_template_kwargs", {})
+        prompt_token_ids = self.apply_chat_template(request, **chat_template_kwargs)
         if len(prompt_token_ids) == 0:
             raise ValueError("Invalid input: prompt_token_ids must be a non-empty sequence of token IDs")
         image_start_index = 0
@@ -480,7 +480,7 @@ def _load_tokenizer(self):
                 break
         self.tokenizer = Ernie4_5Tokenizer.from_pretrained(self.model_name_or_path)
 
-    def apply_chat_template(self, request):
+    def apply_chat_template(self, request, **kwargs):
         """
         Convert multi-turn messages into ID sequences.
 
@@ -498,7 +498,7 @@ def apply_chat_template(self, request):
             request,
             tokenize=False,
             add_generation_prompt=request.get("add_generation_prompt", True),
-            chat_template=request.get("chat_template", None),
+            **kwargs,
         )
         prompt_token_str = prompt_token_template.replace("<|image@placeholder|>", "").replace(
             "<|video@placeholder|>", ""

fastdeploy/input/text_processor.py

@@ -208,7 +208,6 @@ def process_request(self, request, max_model_len=None, **kwargs):
             str: error message
         """
         data_processor_logger.info(f"Start processing request: {request}")
-        request.chat_template = kwargs.get("chat_template")
         request = self._apply_default_parameters(request)
         if request.get("eos_token_ids") is None or len(request.eos_token_ids) == 0:
             request.eos_token_ids = self.eos_token_ids
@@ -242,7 +241,7 @@ def process_request(self, request, max_model_len=None, **kwargs):
                 if self.tokenizer.chat_template is None:
                     raise ValueError("This model does not support chat_template.")
                 task = request.to_dict()
-                chat_template_kwargs = kwargs.get("chat_template_kwargs")
+                chat_template_kwargs = kwargs.get("chat_template_kwargs", {})
                 if chat_template_kwargs:
                     if isinstance(chat_template_kwargs, dict):
                         for k, v in chat_template_kwargs.items():
@@ -251,7 +250,7 @@ def process_request(self, request, max_model_len=None, **kwargs):
                     else:
                         raise ValueError("Invalid input: chat_template_kwargs must be a dict")
                 task.setdefault("enable_thinking", True)
-                request.prompt_token_ids = self.messages2ids(task)
+                request.prompt_token_ids = self.messages2ids(task, **chat_template_kwargs)
             else:
                 raise ValueError(f"The request should have `input_ids`, `text` or `messages`: {request}.")
 
@@ -316,7 +315,7 @@ def process_request_dict(self, request, max_model_len=None, **kwargs):
             elif request.get("messages"):
                 if self.tokenizer.chat_template is None:
                     raise ValueError("This model does not support chat_template.")
-                chat_template_kwargs = request.get("chat_template_kwargs")
+                chat_template_kwargs = request.get("chat_template_kwargs", {})
                 if chat_template_kwargs:
                     if isinstance(chat_template_kwargs, dict):
                         for k, v in chat_template_kwargs.items():
@@ -325,7 +324,7 @@ def process_request_dict(self, request, max_model_len=None, **kwargs):
                     else:
                         raise ValueError("Invalid input: chat_template_kwargs must be a dict")
                 request.setdefault("enable_thinking", True)
-                request["prompt_token_ids"] = self.messages2ids(request)
+                request["prompt_token_ids"] = self.messages2ids(request, **chat_template_kwargs)
             else:
                 raise ValueError(f"Request must contain 'prompt_token_ids', 'prompt', or 'messages': {request}")
 
@@ -530,7 +529,7 @@ def text2ids(self, text, max_model_len):
 
         return tokens["input_ids"][0]
 
-    def messages2ids(self, request):
+    def messages2ids(self, request, **kwargs):
         """
         Convert multi-turn messages into ID sequences.
 
@@ -547,7 +546,7 @@ def messages2ids(self, request):
             split_special_tokens=False,
             add_special_tokens=False,
             return_tensors="pd",
-            chat_template=request.get("chat_template", None),
+            **kwargs,
         )
         request["text_after_process"] = spliced_message
         req_id = None

tests/entrypoints/test_engine_client.py

@@ -0,0 +1,36 @@
+import unittest
+from unittest.mock import MagicMock, patch
+
+from fastdeploy.entrypoints.engine_client import EngineClient
+
+
+class TestEngineClient(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        # 创建 EngineClient 实例的模拟对象
+        with patch.object(EngineClient, "__init__", return_value=None) as mock_init:
+            self.engine_client = EngineClient("model_path")
+            mock_init.side_effect = lambda *args, **kwargs: print(f"__init__ called with {args}, {kwargs}")
+
+        self.engine_client.data_processor = MagicMock()
+        self.engine_client.zmq_client = MagicMock()
+        self.engine_client.max_model_len = 1024
+        self.engine_client.enable_mm = False
+
+    async def test_add_request(self):
+        request = {
+            "chat_template_kwargs": {"enable_thinking": True},
+            "prompt_token_ids": [1],
+            "chat_template": "Hello",
+            "max_tokens": 20,
+            "tools": [1],
+        }
+
+        await self.engine_client.add_requests(request)
+        assert "chat_template" in request["chat_template_kwargs"], "'chat_template' not found in 'chat_template_kwargs"
+        assert "tools" in request["chat_template_kwargs"], "'tools' not found in 'chat_template_kwargs'"
+        assert request["chat_template_kwargs"]["chat_template"] == "Hello"
+        assert request["chat_template_kwargs"]["tools"] == [1]
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/input/test_ernie_processor.py

@@ -17,13 +17,27 @@ def setUp(self):
         self.processor.decode_status = {}
         self.processor.reasoning_end_dict = {}
         self.processor.tool_parser_dict = {}
+        self.processor.generation_config = MagicMock()
+        self.processor.eos_token_ids = [1]
 
         # 模拟 ids2tokens 方法
         def mock_ids2tokens(token_ids, task_id):
             return "delta_text", [2, 3], "previous_texts"
 
         self.processor.ids2tokens = mock_ids2tokens
 
+        def mock_messages2ids(request, **kwargs):
+            if "chat_template" in kwargs:
+                return [1]
+            else:
+                return [0]
+
+        def mock_apply_default_parameters(request):
+            return request
+
+        self.processor.messages2ids = mock_messages2ids
+        self.processor._apply_default_parameters = mock_apply_default_parameters
+
         # 模拟推理解析器
         self.mock_reasoning_parser = MagicMock()
         self.mock_reasoning_parser.__class__.__name__ = "ErnieX1ReasoningParser"
@@ -49,6 +63,17 @@ def test_process_response_dict_streaming_normal_case(self):
         # 验证结果
         self.assertEqual(result["outputs"]["raw_prediction"], "delta_text")
 
+    def test_process_request_dict(self):
+        request_dict = {
+            "messages": [{"role": "user", "content": "Hello!"}],
+            "chat_template_kwargs": {"chat_template": "Hello!"},
+            "eos_token_ids": [1],
+            "temperature": 1,
+            "top_p": 1,
+        }
+        result = self.processor.process_request_dict(request_dict, 100)
+        self.assertEqual(result["prompt_token_ids"], [1])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/input/test_text_processor.py

@@ -0,0 +1,63 @@
+import unittest
+from unittest.mock import MagicMock, patch
+
+from fastdeploy.engine.request import Request
+from fastdeploy.input.text_processor import DataProcessor
+
+
+class TestDataProcessorProcess(unittest.TestCase):
+    def setUp(self):
+        # 创建 DataProcessor 实例的模拟对象
+        with patch.object(DataProcessor, "__init__", return_value=None) as mock_init:
+            self.processor = DataProcessor("model_path")
+            mock_init.side_effect = lambda *args, **kwargs: print(f"__init__ called with {args}, {kwargs}")
+
+        # 设置必要的属性
+        self.processor.tokenizer = MagicMock()
+        self.processor.tokenizer.eos_token_id = 1
+        self.processor.decode_status = {}
+        self.processor.reasoning_end_dict = {}
+        self.processor.tool_parser_dict = {}
+        self.processor.generation_config = MagicMock()
+        self.processor.eos_token_ids = [1]
+
+        def mock_messages2ids(request, **kwargs):
+            if "chat_template" in kwargs:
+                return [1]
+            else:
+                return [0]
+
+        def mock_apply_default_parameters(request):
+            return request
+
+        self.processor.messages2ids = mock_messages2ids
+        self.processor._apply_default_parameters = mock_apply_default_parameters
+
+    def test_process_request(self):
+        request = Request.from_dict(
+            {
+                "request_id": "123",
+                "messages": [{"role": "user", "content": "Hello!"}],
+                "eos_token_ids": [1],
+                "temperature": 1,
+                "top_p": 1,
+            }
+        )
+        chat_template_kwargs = {"chat_template": "Hello!"}
+        result = self.processor.process_request(request, 100, chat_template_kwargs=chat_template_kwargs)
+        self.assertEqual(result.prompt_token_ids, [1])
+
+    def test_process_request_dict(self):
+        request_dict = {
+            "messages": [{"role": "user", "content": "Hello!"}],
+            "chat_template_kwargs": {"chat_template": "Hello!"},
+            "eos_token_ids": [1],
+            "temperature": 1,
+            "top_p": 1,
+        }
+        result = self.processor.process_request_dict(request_dict, 100)
+        self.assertEqual(result["prompt_token_ids"], [1])
+
+
+if __name__ == "__main__":
+    unittest.main()