Fix the issue of incorrect finish reason in final stream response chunk...

Fix the issue of incorrect finish reason in final stream response chunk returned during tool call (#7708)
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -484,7 +484,10 @@ class OpenAIServingChat(OpenAIServingBase):
 
                 # Handle tool calls
                 if request.tool_choice != "none" and request.tools:
-                    async for chunk in self._process_tool_call_stream(
+                    async for (
+                        chunk,
+                        tool_call_finish_reason_type,
+                    ) in self._process_tool_call_stream(
                         index,
                         delta,
                         parser_dict,
@@ -492,7 +495,10 @@ class OpenAIServingChat(OpenAIServingBase):
                         request,
                         finish_reason_type,
                     ):
+                        if chunk:
                             yield chunk
+                        finish_reason_type = tool_call_finish_reason_type
+
                 else:
                     # Regular content
                     if delta or not (
@@ -865,7 +871,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 choices=[choice_data],
                 model=request.model,
             )
-            yield f"data: {chunk.model_dump_json()}\n\n"
+            yield f"data: {chunk.model_dump_json()}\n\n", finish_reason_type
 
         # Yield tool calls
         for call_item in calls:
@@ -920,4 +926,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 choices=[choice_data],
                 model=request.model,
             )
-            yield f"data: {chunk.model_dump_json()}\n\n"
+            yield f"data: {chunk.model_dump_json()}\n\n", finish_reason_type
+
+        if finish_reason_type == "stop":
+            yield None, "tool_calls"

test/srt/openai_server/function_call/test_openai_function_calling.py

@@ -159,6 +159,13 @@ class TestOpenAIServerFunctionCalling(CustomTestCase):
             "Target function name 'get_current_weather' was not found in the streaming chunks",
         )
 
+        finish_reason = chunks[-1].choices[0].finish_reason
+        self.assertEqual(
+            finish_reason,
+            "tool_calls",
+            "Final response of function calling should have finish_reason 'tool_calls'",
+        )
+
     def test_function_calling_streaming_args_parsing(self):
         """
         Test: Whether the function call arguments returned in streaming mode can be correctly concatenated into valid JSON.