fixed reasoning streaming with tool_choice="required" (vllm-project#24108)

Signed-off-by: CNE Pierre FICHEPOIL <pierre-1.fichepoil@gendarmerie.interieur.gouv.fr>
Signed-off-by: ExtReMLapin <3909752+ExtReMLapin@users.noreply.github.com>
Co-authored-by: CNE Pierre FICHEPOIL <pierre-1.fichepoil@gendarmerie.interieur.gouv.fr>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>

Author: 3 people

tests/entrypoints/openai/test_completion_with_function_calling.py

@@ -194,11 +194,19 @@ async def test_function_tool_use(
         )
 
         output = []
+        reasoning = []
         async for chunk in output_stream:
-            if chunk.choices and chunk.choices[0].delta.tool_calls:
-                output.extend(chunk.choices[0].delta.tool_calls)
+            if chunk.choices:
+                if enable_thinking and getattr(
+                    chunk.choices[0].delta, "reasoning_content", None
+                ):
+                    reasoning.append(chunk.choices[0].delta.reasoning_content)
+                if chunk.choices[0].delta.tool_calls:
+                    output.extend(chunk.choices[0].delta.tool_calls)
 
         assert len(output) > 0
+        if enable_thinking:
+            assert len(reasoning) > 0
 
 
 @pytest.fixture(scope="module")

vllm/entrypoints/openai/serving_chat.py

@@ -563,8 +563,6 @@ async def chat_completion_stream_generator(
             # For reasoning parser and tool call all enabled
             added_content_delta_arr = [False] * num_choices
             reasoning_end_arr = [False] * num_choices
-        elif request.tool_choice == "required":
-            all_previous_token_ids = None
         else:
             all_previous_token_ids = None
 
@@ -880,29 +878,56 @@ async def chat_completion_stream_generator(
                         previous_text = previous_texts[i]
                         current_text = previous_text + delta_text
                         fn_name_returned = function_name_returned[i]
+                        output_token_ids = as_list(output.token_ids)
 
-                        if self.reasoning_parser:
-                            _, content = reasoning_parser.extract_reasoning_content(
-                                current_text, request
+                        if (
+                            self.reasoning_parser is not None
+                            and not reasoning_end_arr[i]
+                            and res.prompt_token_ids
+                            and reasoning_parser.is_reasoning_end(res.prompt_token_ids)
+                        ):
+                            reasoning_end_arr[i] = True
+
+                        if self.reasoning_parser and not reasoning_end_arr[i]:
+                            delta_message = (
+                                reasoning_parser.extract_reasoning_content_streaming(
+                                    previous_text,
+                                    current_text,
+                                    delta_text,
+                                    previous_token_ids,
+                                    current_token_ids,
+                                    output_token_ids,
+                                )
                             )
+                            if reasoning_parser.is_reasoning_end(output_token_ids):
+                                reasoning_end_arr[i] = True
+                                if delta_message and delta_message.content:
+                                    current_text = delta_message.content
+                                    delta_message.content = None
+                                else:
+                                    # reasoning ended
+                                    current_text = ""
+
                         else:
+                            # either finished reasoning or no reasoning at all
                             content = current_text
-                        delta_message, function_name_returned[i] = (
-                            self.extract_tool_call_required_streaming(
-                                previous_text=previous_text,
-                                current_text=content,
-                                delta_text=delta_text,
-                                function_name_returned=fn_name_returned,
-                                tool_call_idx=history_tool_call_cnt,
+
+                            delta_message, function_name_returned[i] = (
+                                self.extract_tool_call_required_streaming(
+                                    previous_text=previous_text,
+                                    current_text=content,
+                                    delta_text=delta_text,
+                                    function_name_returned=fn_name_returned,
+                                    tool_call_idx=history_tool_call_cnt,
+                                )
                             )
-                        )
-                        if (
-                            delta_message
-                            and delta_message.tool_calls
-                            and delta_message.tool_calls[0].id is not None
-                        ):
-                            history_tool_call_cnt += 1
-                            tools_streamed[i] = True
+                            if (
+                                delta_message
+                                and delta_message.tool_calls
+                                and delta_message.tool_calls[0].id is not None
+                            ):
+                                history_tool_call_cnt += 1
+                                tools_streamed[i] = True
 
                     # handle streaming deltas for tools with "auto" tool choice
                     # and reasoning parser