bugfix: Fix multiple finish_reason chunks and tool_calls finish reason check (#8417)

test/srt/openai_server/basic/test_serving_chat.py

@@ -197,6 +197,134 @@ class ServingChatTestCase(unittest.TestCase):
             self.assertEqual(params["min_new_tokens"], 5)
             self.assertEqual(params["stop"], ["</s>"])
 
+    async def test_unstreamed_tool_args_completion(self):
+        """Test that remaining tool call arguments are sent when generation finishes."""
+
+        # Mock FunctionCallParser with detector that has partial tool call data
+        mock_parser = Mock()
+        mock_detector = Mock()
+
+        # Simulate a tool call that was partially streamed
+        mock_detector.prev_tool_call_arr = [
+            {
+                "name": "get_weather",
+                "arguments": {"location": "San Francisco", "unit": "celsius"},
+            }
+        ]
+        mock_detector.streamed_args_for_tool = [
+            '{"location": "San Francisco"'  # Partial arguments streamed so far
+        ]
+        mock_parser.detector = mock_detector
+
+        content = {
+            "meta_info": {
+                "id": "chatcmpl-test123",
+            }
+        }
+
+        request = ChatCompletionRequest(
+            model="test",
+            messages=[{"role": "user", "content": "What's the weather?"}],
+            tools=[{"type": "function", "function": {"name": "get_weather"}}],
+        )
+
+        # Test the completion method
+        result = self.chat._check_for_unstreamed_tool_args(
+            parser=mock_parser,
+            content=content,
+            request=request,
+            finish_reason_type="stop",
+            index=0,
+        )
+
+        # Should return a chunk with remaining arguments
+        self.assertIsNotNone(result, "Should return chunk with remaining arguments")
+        self.assertIn('"arguments":', result, "Should contain arguments field")
+        self.assertIn(
+            ', "unit": "celsius"}', result, "Should contain remaining arguments"
+        )
+        self.assertIn(
+            '"finish_reason":null',
+            result,
+            "Should not include finish_reason in completion chunk",
+        )
+
+    async def test_unstreamed_tool_args_no_completion_needed(self):
+        """Test that no completion chunk is sent when all arguments were already streamed."""
+
+        # Mock FunctionCallParser with detector that has complete tool call data
+        mock_parser = Mock()
+        mock_detector = Mock()
+
+        # Simulate a tool call that was completely streamed
+        mock_detector.prev_tool_call_arr = [
+            {"name": "get_weather", "arguments": {"location": "San Francisco"}}
+        ]
+        mock_detector.streamed_args_for_tool = [
+            '{"location": "San Francisco"}'  # All arguments already streamed
+        ]
+        mock_parser.detector = mock_detector
+
+        content = {
+            "meta_info": {
+                "id": "chatcmpl-test123",
+            }
+        }
+
+        request = ChatCompletionRequest(
+            model="test",
+            messages=[{"role": "user", "content": "What's the weather?"}],
+            tools=[{"type": "function", "function": {"name": "get_weather"}}],
+        )
+
+        # Test the completion method
+        result = self.chat._check_for_unstreamed_tool_args(
+            parser=mock_parser,
+            content=content,
+            request=request,
+            finish_reason_type="stop",
+            index=0,
+        )
+
+        # Should return None since no completion is needed
+        self.assertIsNone(result, "Should return None when no completion is needed")
+
+    async def test_unstreamed_tool_args_no_parser_data(self):
+        """Test that no completion chunk is sent when parser has no tool call data."""
+
+        # Mock FunctionCallParser with empty detector
+        mock_parser = Mock()
+        mock_detector = Mock()
+        mock_detector.prev_tool_call_arr = []
+        mock_detector.streamed_args_for_tool = []
+        mock_parser.detector = mock_detector
+
+        content = {
+            "meta_info": {
+                "id": "chatcmpl-test123",
+            }
+        }
+
+        request = ChatCompletionRequest(
+            model="test",
+            messages=[{"role": "user", "content": "What's the weather?"}],
+            tools=[{"type": "function", "function": {"name": "get_weather"}}],
+        )
+
+        # Test the completion method
+        result = self.chat._check_for_unstreamed_tool_args(
+            parser=mock_parser,
+            content=content,
+            request=request,
+            finish_reason_type="stop",
+            index=0,
+        )
+
+        # Should return None since there's no parser data
+        self.assertIsNone(
+            result, "Should return None when parser has no tool call data"
+        )
+
 
 if __name__ == "__main__":
     unittest.main(verbosity=2)