[BugFix] fix max streaming tokens invalid (#3789)

Author: ltd0924

fastdeploy/entrypoints/openai/serving_chat.py

@@ -183,6 +183,8 @@ async def chat_completion_stream_generator(
             else (request.metadata or {}).get("max_streaming_response_tokens", 1)
         )  # dierctly passed & passed in metadata
 
+        max_streaming_response_tokens = max(1, max_streaming_response_tokens)
+
         enable_thinking = request.chat_template_kwargs.get("enable_thinking") if request.chat_template_kwargs else None
         if enable_thinking is None:
             enable_thinking = request.metadata.get("enable_thinking") if request.metadata else None
@@ -370,11 +372,6 @@ async def chat_completion_stream_generator(
                             api_server_logger.info(f"Chat Streaming response last send: {chunk.model_dump_json()}")
                         choices = []
 
-                if choices:
-                    chunk.choices = choices
-                    yield f"data: {chunk.model_dump_json(exclude_unset=True)}\n\n"
-                    choices = []
-
             if include_usage:
                 completion_tokens = previous_num_tokens
                 usage = UsageInfo(

fastdeploy/entrypoints/openai/serving_completion.py

@@ -331,6 +331,7 @@ async def completion_stream_generator(
                 if request.max_streaming_response_tokens is not None
                 else (request.suffix or {}).get("max_streaming_response_tokens", 1)
             )  # dierctly passed & passed in suffix
+            max_streaming_response_tokens = max(1, max_streaming_response_tokens)
             choices = []
             chunk = CompletionStreamResponse(
                 id=request_id,
@@ -461,10 +462,6 @@ async def completion_stream_generator(
                             )
                             yield f"data: {usage_chunk.model_dump_json(exclude_unset=True)}\n\n"
                         api_server_logger.info(f"Completion Streaming response last send: {chunk.model_dump_json()}")
-                if choices:
-                    chunk.choices = choices
-                    yield f"data: {chunk.model_dump_json(exclude_unset=True)}\n\n"
-                    choices = []
 
         except Exception as e:
             api_server_logger.error(f"Error in completion_stream_generator: {e}, {str(traceback.format_exc())}")

tests/entrypoints/openai/test_max_streaming_tokens.py

@@ -0,0 +1,279 @@
+import json
+import unittest
+from unittest import IsolatedAsyncioTestCase
+from unittest.mock import AsyncMock, Mock, patch
+
+from fastdeploy.entrypoints.openai.protocol import (
+    ChatCompletionRequest,
+    CompletionRequest,
+)
+from fastdeploy.entrypoints.openai.serving_chat import OpenAIServingChat
+from fastdeploy.entrypoints.openai.serving_completion import OpenAIServingCompletion
+
+
+class TestMaxStreamingResponseTokens(IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.engine_client = Mock()
+        self.engine_client.connection_initialized = False
+        self.engine_client.connection_manager = AsyncMock()
+        self.engine_client.connection_manager.initialize = AsyncMock()
+        self.engine_client.connection_manager.get_connection = AsyncMock()
+        self.engine_client.connection_manager.cleanup_request = AsyncMock()
+        self.engine_client.semaphore = Mock()
+        self.engine_client.semaphore.acquire = AsyncMock()
+        self.engine_client.semaphore.release = Mock()
+        self.engine_client.data_processor = Mock()
+        self.engine_client.is_master = True
+
+        self.chat_serving = OpenAIServingChat(
+            engine_client=self.engine_client,
+            models=None,
+            pid=123,
+            ips=None,
+            max_waiting_time=30,
+            chat_template="default",
+            enable_mm_output=False,
+            tokenizer_base_url=None,
+        )
+
+        self.completion_serving = OpenAIServingCompletion(
+            engine_client=self.engine_client, models=None, pid=123, ips=None, max_waiting_time=30
+        )
+
+    def test_metadata_parameter_setting(self):
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "Hello"}],
+            stream=True,
+            metadata={"max_streaming_response_tokens": 100},
+        )
+
+        max_tokens = (
+            request.max_streaming_response_tokens
+            if request.max_streaming_response_tokens is not None
+            else (request.metadata or {}).get("max_streaming_response_tokens", 1)
+        )
+
+        self.assertEqual(max_tokens, 100)
+
+    def test_default_value(self):
+        request = ChatCompletionRequest(
+            model="test-model", messages=[{"role": "user", "content": "Hello"}], stream=True
+        )
+
+        max_tokens = (
+            request.max_streaming_response_tokens
+            if request.max_streaming_response_tokens is not None
+            else (request.metadata or {}).get("max_streaming_response_tokens", 1)
+        )
+
+        self.assertEqual(max_tokens, 1)
+
+    def test_edge_case_zero_value(self):
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "Hello"}],
+            stream=True,
+            max_streaming_response_tokens=0,
+        )
+
+        max_streaming_response_tokens = (
+            request.max_streaming_response_tokens
+            if request.max_streaming_response_tokens is not None
+            else (request.metadata or {}).get("max_streaming_response_tokens", 1)
+        )
+        max_streaming_response_tokens = max(1, max_streaming_response_tokens)
+
+        self.assertEqual(max_streaming_response_tokens, 1)
+
+    @patch("fastdeploy.entrypoints.openai.serving_chat.api_server_logger")
+    @patch("fastdeploy.entrypoints.openai.serving_chat.ChatResponseProcessor")
+    async def test_integration_with_chat_stream_generator(self, mock_processor_class, mock_logger):
+        response_data = [
+            {
+                "outputs": {"token_ids": [1], "text": "a", "top_logprobs": None},
+                "metrics": {"first_token_time": 0.1, "inference_start_time": 0.1},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [2], "text": "b", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.2, "first_token_time": None},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [3], "text": "c", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.3, "first_token_time": None},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [4], "text": "d", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.4, "first_token_time": None},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [5], "text": "e", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.5, "first_token_time": None},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [6], "text": "f", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.6, "first_token_time": None},
+                "finished": False,
+            },
+            {
+                "outputs": {"token_ids": [7], "text": "g", "top_logprobs": None},
+                "metrics": {"arrival_time": 0.7, "first_token_time": None, "request_start_time": 0.1},
+                "finished": True,
+            },
+        ]
+
+        mock_response_queue = AsyncMock()
+        mock_response_queue.get.side_effect = response_data
+
+        mock_dealer = Mock()
+        mock_dealer.write = Mock()
+
+        # Mock the connection manager call
+        self.engine_client.connection_manager.get_connection = AsyncMock(
+            return_value=(mock_dealer, mock_response_queue)
+        )
+
+        mock_processor_instance = Mock()
+
+        async def mock_process_response_chat_single(response, stream, enable_thinking, include_stop_str_in_output):
+            yield response
+
+        mock_processor_instance.process_response_chat = mock_process_response_chat_single
+        mock_processor_instance.enable_multimodal_content = Mock(return_value=False)
+        mock_processor_class.return_value = mock_processor_instance
+
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "Hello"}],
+            stream=True,
+            max_streaming_response_tokens=3,
+        )
+
+        generator = self.chat_serving.chat_completion_stream_generator(
+            request=request,
+            request_id="test-request-id",
+            model_name="test-model",
+            prompt_token_ids=[1, 2, 3],
+            text_after_process="Hello",
+        )
+
+        chunks = []
+        async for chunk in generator:
+            chunks.append(chunk)
+
+        self.assertGreater(len(chunks), 0, "No chucks!")
+
+        parsed_chunks = []
+        for i, chunk_str in enumerate(chunks):
+            if i == 0:
+                continue
+            if chunk_str.startswith("data: ") and chunk_str.endswith("\n\n"):
+                json_part = chunk_str[6:-2]
+                if json_part == "[DONE]":
+                    parsed_chunks.append({"type": "done", "raw": chunk_str})
+                    break
+                try:
+                    chunk_dict = json.loads(json_part)
+                    parsed_chunks.append(chunk_dict)
+                except json.JSONDecodeError as e:
+                    self.fail(f"Cannot parser {i+1} chunck, JSON: {e}\n origin string: {repr(chunk_str)}")
+            else:
+                self.fail(f"{i+1} chunk is unexcepted 'data: JSON\\n\\n': {repr(chunk_str)}")
+        for chunk_dict in parsed_chunks:
+            choices_list = chunk_dict["choices"]
+            if choices_list[-1].get("finish_reason") is not None:
+                break
+            else:
+                self.assertEqual(len(choices_list), 3, f"Chunk {chunk_dict} should has three choices")
+
+        found_done = any("[DONE]" in chunk for chunk in chunks)
+        self.assertTrue(found_done, "Not Receive '[DONE]'")
+
+    @patch("fastdeploy.entrypoints.openai.serving_completion.api_server_logger")
+    async def test_integration_with_completion_stream_generator(self, mock_logger):
+        response_data = [
+            [
+                {
+                    "request_id": "test-request-id-0",
+                    "outputs": {"token_ids": [1], "text": "a", "top_logprobs": None},
+                    "metrics": {"first_token_time": 0.1, "inference_start_time": 0.1},
+                    "finished": False,
+                },
+                {
+                    "request_id": "test-request-id-0",
+                    "outputs": {"token_ids": [2], "text": "b", "top_logprobs": None},
+                    "metrics": {"arrival_time": 0.2, "first_token_time": None},
+                    "finished": False,
+                },
+            ],
+            [
+                {
+                    "request_id": "test-request-id-0",
+                    "outputs": {"token_ids": [7], "text": "g", "top_logprobs": None},
+                    "metrics": {"arrival_time": 0.7, "first_token_time": None, "request_start_time": 0.1},
+                    "finished": True,
+                }
+            ],
+        ]
+
+        mock_response_queue = AsyncMock()
+        mock_response_queue.get.side_effect = response_data
+
+        mock_dealer = Mock()
+        mock_dealer.write = Mock()
+
+        # Mock the connection manager call
+        self.engine_client.connection_manager.get_connection = AsyncMock(
+            return_value=(mock_dealer, mock_response_queue)
+        )
+
+        request = CompletionRequest(model="test-model", prompt="Hello", stream=True, max_streaming_response_tokens=3)
+
+        generator = self.completion_serving.completion_stream_generator(
+            request=request,
+            num_choices=1,
+            request_id="test-request-id",
+            model_name="test-model",
+            created_time=11,
+            prompt_batched_token_ids=[[1, 2, 3]],
+            text_after_process_list=["Hello"],
+        )
+
+        chunks = []
+        async for chunk in generator:
+            chunks.append(chunk)
+
+        self.assertGreater(len(chunks), 0, "No chucks!")
+
+        parsed_chunks = []
+        for i, chunk_str in enumerate(chunks):
+            if chunk_str.startswith("data: ") and chunk_str.endswith("\n\n"):
+                json_part = chunk_str[6:-2]
+                if json_part == "[DONE]":
+                    break
+                try:
+                    chunk_dict = json.loads(json_part)
+                    parsed_chunks.append(chunk_dict)
+                except json.JSONDecodeError as e:
+                    self.fail(f"Cannot parser {i+1} chunck, JSON: {e}\n origin string: {repr(chunk_str)}")
+            else:
+                self.fail(f"{i+1} chunk is unexcepted 'data: JSON\\n\\n': {repr(chunk_str)}")
+        self.assertEqual(len(parsed_chunks), 1)
+        for chunk_dict in parsed_chunks:
+            choices_list = chunk_dict["choices"]
+            self.assertEqual(len(choices_list), 3, f"Chunk {chunk_dict} should has three choices")
+            self.assertEqual(
+                choices_list[-1].get("finish_reason"), "stop", f"Chunk {chunk_dict} should has stop reason"
+            )
+
+        found_done = any("[DONE]" in chunk for chunk in chunks)
+        self.assertTrue(found_done, "Not Receive '[DONE]'")
+
+
+if __name__ == "__main__":
+    unittest.main()