[gpt-oss] Add IncompleteDetails to ResponsesRepsonse (vllm-project#24561)

Signed-off-by: Andrew Xia <axia@meta.com>

Author: qandrew

tests/entrypoints/openai/test_response_api_with_harmony.py

@@ -74,6 +74,20 @@ async def test_basic_with_reasoning_effort(client: OpenAI, model_name: str):
     assert response.status == "completed"
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_max_tokens(client: OpenAI, model_name: str):
+    response = await client.responses.create(
+        model=model_name,
+        input="What is the first paragraph of Moby Dick?",
+        reasoning={"effort": "low"},
+        max_output_tokens=30,
+    )
+    assert response is not None
+    assert response.status == "incomplete"
+    assert response.incomplete_details.reason == "max_output_tokens"
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 async def test_chat(client: OpenAI, model_name: str):

vllm/entrypoints/context.py

@@ -112,6 +112,7 @@ def __init__(
         available_tools: list[str],
     ):
         self._messages = messages
+        self.finish_reason: Optional[str] = None
         self.available_tools = available_tools
         self._tool_sessions: dict[str, Union[ClientSession, Tool]] = {}
         self.called_tools: set[str] = set()
@@ -135,7 +136,8 @@ def _update_num_reasoning_tokens(self):
         if self.parser.current_channel in {"analysis", "commentary"}:
             self.num_reasoning_tokens += 1
 
-    def append_output(self, output) -> None:
+    def append_output(self, output: Union[RequestOutput,
+                                          list[Message]]) -> None:
         if isinstance(output, RequestOutput):
             output_token_ids = output.outputs[0].token_ids
             self.parser = get_streamable_parser_for_assistant()
@@ -150,25 +152,27 @@ def append_output(self, output) -> None:
             # Move current turn to previous turn for next turn's calculations
             self.previous_turn = self.current_turn.copy()
             output_msgs = self.parser.messages
+            # The responses finish reason is set in the last message
+            self.finish_reason = output.outputs[0].finish_reason
         else:
             # Tool output.
             output_msgs = output
         self._messages.extend(output_msgs)
 
     def _update_prefill_token_usage(self, output: RequestOutput) -> None:
         """Update token usage statistics for the prefill phase of generation.
-        
+
         The prefill phase processes the input prompt tokens. This method:
         1. Counts the prompt tokens for this turn
         2. Calculates tool output tokens for multi-turn conversations
         3. Updates cached token counts
         4. Tracks state for next turn calculations
-        
+
         Tool output tokens are calculated as:
-        current_prompt_tokens - last_turn_prompt_tokens - 
+        current_prompt_tokens - last_turn_prompt_tokens -
         last_turn_output_tokens
         This represents tokens added between turns (typically tool responses).
-        
+
         Args:
             output: The RequestOutput containing prompt token information
         """
@@ -214,18 +218,18 @@ def _update_prefill_token_usage(self, output: RequestOutput) -> None:
 
     def _update_decode_token_usage(self, output: RequestOutput) -> int:
         """Update token usage statistics for the decode phase of generation.
-        
+
         The decode phase processes the generated output tokens. This method:
         1. Counts output tokens from all completion outputs
         2. Updates the total output token count
         3. Tracks tokens generated in the current turn
-        
+
         In streaming mode, this is called for each token generated.
         In non-streaming mode, this is called once with all output tokens.
-        
+
         Args:
             output: The RequestOutput containing generated token information
-            
+
         Returns:
             int: Number of output tokens processed in this call
         """
@@ -385,7 +389,8 @@ def __init__(self, *args, **kwargs):
     def messages(self) -> list:
         return self.parser.messages
 
-    def append_output(self, output) -> None:
+    def append_output(self, output: Union[RequestOutput,
+                                          list[Message]]) -> None:
         if isinstance(output, RequestOutput):
             # append_output is called for each output token in streaming case,
             # so we only want to add the prompt tokens once for each message.

vllm/entrypoints/harmony_utils.py

@@ -387,7 +387,9 @@ def parse_remaining_state(
             id=f"msg_{random_uuid()}",
             content=[output_text],
             role="assistant",
-            status="completed",
+            # if the parser still has messages (ie if the generator got cut
+            # abruptly), this should be incomplete
+            status="incomplete",
             type="message",
         )
         return [text_item]

vllm/entrypoints/openai/protocol.py

@@ -30,7 +30,7 @@
     from openai.types.responses import (ResponseFormatTextConfig as
                                         ResponseTextConfig)
 
-from openai.types.responses.response import ToolChoice
+from openai.types.responses.response import IncompleteDetails, ToolChoice
 from openai.types.responses.tool import Tool
 from openai.types.shared import Metadata, Reasoning
 from pydantic import (BaseModel, ConfigDict, Field, TypeAdapter,
@@ -1868,7 +1868,7 @@ class ResponsesResponse(OpenAIBaseModel):
     id: str = Field(default_factory=lambda: f"resp_{random_uuid()}")
     created_at: int = Field(default_factory=lambda: int(time.time()))
     # error: Optional[ResponseError] = None
-    # incomplete_details: Optional[IncompleteDetails] = None
+    incomplete_details: Optional[IncompleteDetails] = None
     instructions: Optional[str] = None
     metadata: Optional[Metadata] = None
     model: str
@@ -1904,9 +1904,18 @@ def from_request(
         status: ResponseStatus,
         usage: Optional[ResponseUsage] = None,
     ) -> "ResponsesResponse":
+
+        incomplete_details: Optional[IncompleteDetails] = None
+        if status == 'incomplete':
+            incomplete_details = IncompleteDetails(reason='max_output_tokens')
+        # TODO: implement the other reason for incomplete_details,
+        # which is content_filter
+        # incomplete_details = IncompleteDetails(reason='content_filter')
+
         return cls(
             id=request.request_id,
             created_at=created_time,
+            incomplete_details=incomplete_details,
             instructions=request.instructions,
             metadata=request.metadata,
             model=model_name,
@@ -2109,7 +2118,7 @@ class DetokenizeResponse(OpenAIBaseModel):
 
 class TokenizerInfoResponse(OpenAIBaseModel):
     """
-    Response containing tokenizer configuration 
+    Response containing tokenizer configuration
     equivalent to tokenizer_config.json
     """
 
@@ -2199,7 +2208,7 @@ class TranscriptionRequest(OpenAIBaseModel):
     to_language: Optional[str] = None
     """The language of the output audio we transcribe to.
 
-    Please note that this is not currently used by supported models at this 
+    Please note that this is not currently used by supported models at this
     time, but it is a placeholder for future use, matching translation api.
     """
 

vllm/entrypoints/openai/serving_responses.py

@@ -27,7 +27,7 @@
                                     ResponseReasoningItem,
                                     ResponseReasoningTextDeltaEvent,
                                     ResponseReasoningTextDoneEvent,
-                                    response_text_delta_event)
+                                    ResponseStatus, response_text_delta_event)
 from openai.types.responses.response_output_text import (Logprob,
                                                          LogprobTopLogprob)
 # yapf: enable
@@ -461,10 +461,22 @@ async def responses_full_generator(
                 # TODO: Use a vllm-specific Validation Error
                 return self.create_error_response(str(e))
 
+        # NOTE: Implementation of stauts is still WIP, but for now
+        # we guarantee that if the status is not "completed", it is accurate.
+        # "completed" is implemented as the "catch-all" for now.
+        status: ResponseStatus = "completed"
+
         if self.use_harmony:
             assert isinstance(context, HarmonyContext)
             output = self._make_response_output_items_with_harmony(context)
             num_tool_output_tokens = context.num_tool_output_tokens
+            if len(output) > 0:
+                if context.finish_reason == "length":
+                    status = "incomplete"
+                elif context.finish_reason == "abort":
+                    status = "cancelled"
+            else:
+                status = "incomplete"
         else:
             assert isinstance(context, SimpleContext)
             final_res = context.last_output
@@ -501,7 +513,7 @@ async def responses_full_generator(
             model_name=model_name,
             created_time=created_time,
             output=output,
-            status="completed",
+            status=status,
             usage=usage,
         )
 
@@ -658,7 +670,7 @@ def _make_response_output_items_with_harmony(
         self,
         context: HarmonyContext,
     ) -> list[ResponseOutputItem]:
-        output_items = []
+        output_items: list[ResponseOutputItem] = []
         num_init_messages = context.num_init_messages
         for msg in context.messages[num_init_messages:]:
             output_items.extend(parse_output_message(msg))

vllm/v1/core/sched/utils.py

@@ -10,19 +10,19 @@
 
 def remove_all(lst: list, items_to_remove: set) -> list:
     """Remove all items from a list that are in the items_to_remove set.
-    
+
     This method optimizes for the common case of removing a single item,
     falling back to list comprehension for multiple items.
-    
+
     Args:
         lst: The list to remove items from
         items_to_remove: Set of items to remove
-    
+
     Returns:
         Either the modified original list (for single item removal) or
         a new list (for multiple item removal). Callers should use the
         returned value.
-    
+
     Note:
         For single item removal, this modifies the original list in-place
         and returns it. For multiple items, it creates and returns a new list.

vllm/v1/engine/output_processor.py

@@ -373,17 +373,17 @@ def process_outputs(
         1) Compute stats for logging
         2) Detokenize
         3) Create and handle RequestOutput objects:
-            * If there is a queue (for usage with AsyncLLM), 
+            * If there is a queue (for usage with AsyncLLM),
               put the RequestOutput objects into the queue for
               handling by the per-request generate() tasks.
 
-            * If there is no queue (for usage with LLMEngine), 
+            * If there is no queue (for usage with LLMEngine),
               return a list of RequestOutput objects.
 
         NOTE FOR DEVELOPERS
 
         vLLM V1 minimizes the number of python loops over the full
-        batch to ensure system overheads are minimized. This is the 
+        batch to ensure system overheads are minimized. This is the
         only function that should loop over EngineCoreOutputs.
 
         If you need to touch every element of the batch, do it from