[Bugfix][Frontend] Keep length finish_reason for max_tokens-truncated streaming tool calls (#46303)

Signed-off-by: Ting Sun <suntcrick@gmail.com>

tests/entrypoints/openai/chat_completion/test_serving_chat.py

@@ -2239,7 +2239,13 @@ async def test_tool_choice_validation_without_parser():
 
 
 @pytest.mark.asyncio
-async def test_streaming_n_gt1_independent_tool_parsers():
+@pytest.mark.parametrize(
+    ("engine_finish_reason", "expected_finish_reason"),
+    [("stop", "tool_calls"), ("length", "length")],
+)
+async def test_streaming_n_gt1_independent_tool_parsers(
+    engine_finish_reason: str, expected_finish_reason: str
+):
     """n>1 streaming must use independent parser instances
     and token-id histories per choice.
     """
@@ -2344,7 +2350,7 @@ async def test_streaming_n_gt1_independent_tool_parsers():
                     token_ids=[],
                     cumulative_logprob=0.0,
                     logprobs=None,
-                    finish_reason="stop",
+                    finish_reason=engine_finish_reason,
                 )
                 for choice_idx in range(num_choices)
             ],
@@ -2408,8 +2414,8 @@ async def test_streaming_n_gt1_independent_tool_parsers():
         assert len(reasons) == 1, (
             f"Choice {choice_idx}: expected exactly 1 finish_reason, got {reasons}"
         )
-        assert reasons[0] == "tool_calls", (
-            f"Choice {choice_idx}: expected finish_reason='tool_calls', "
+        assert reasons[0] == expected_finish_reason, (
+            f"Choice {choice_idx}: expected finish_reason={expected_finish_reason!r}, "
             f"got '{reasons[0]}'"
         )
 

vllm/entrypoints/openai/chat_completion/serving.py

@@ -757,7 +757,11 @@ class OpenAIServingChat(GenerateBaseServing):
                         # finish_reason is:
                         # "tool_calls" for "auto" or "required" tool calls,
                         # and "stop" for named tool calls.
-                        if tools_streamed[i] and not tool_choice_function_name:
+                        if (
+                            tools_streamed[i]
+                            and not tool_choice_function_name
+                            and output.finish_reason == "stop"
+                        ):
                             finish_reason_ = "tool_calls"
                         else:
                             finish_reason_ = (