[Frontend] Make finish_reason and Responses item status truthful (#59324)

Signed-off-by: Flora Feng <4florafeng@gmail.com>
Signed-off-by: sfeng33 <4florafeng@gmail.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
Flora Feng
2026-10-15 10:12:02 +01:00
committed by GitHub
co-authored by Chauncey
parent 427915bed2
commit 80dd705c1d
14 changed files with 421 additions and 62 deletions
@@ -42,6 +42,7 @@ from vllm.entrypoints.openai.models.serving import (
)
from vllm.entrypoints.openai.parser.harmony_utils import get_encoding
from vllm.entrypoints.serve.engine.protocol import ErrorResponse
from vllm.entrypoints.serve.utils.tool_calls_utils import resolve_finish_reason
from vllm.exceptions import QueueOverflowError, VLLMValidationError
from vllm.inputs import TokensPrompt
from vllm.logprobs import Logprob
@@ -3051,3 +3052,137 @@ def test_chat_kv_transfer_prompt_token_ids_allows_text_only_parts():
kv_transfer_params={"prompt_token_ids": [10, 20, 30]},
)
assert request.kv_transfer_params == {"prompt_token_ids": [10, 20, 30]}
_TOOL_FINISH_REASON_TOOLS = [
{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get weather",
"parameters": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"],
},
},
}
]
_HERMES_TOOL_CALL_TEXT = (
'<tool_call>\n{"name": "get_weather", "arguments": {"city": "Tokyo"}}\n</tool_call>'
)
_NAMED_WEATHER_TOOL_CHOICE = {"type": "function", "function": {"name": "get_weather"}}
@pytest.mark.parametrize(
("tool_choice", "engine_finish_reason", "tool_calls_made", "expected"),
[
("auto", "stop", True, "tool_calls"),
("auto", "length", True, "length"),
("auto", "stop", False, "stop"),
("required", "stop", True, "tool_calls"),
("required", "stop", False, "stop"),
("required", "length", True, "length"),
(_NAMED_WEATHER_TOOL_CHOICE, "stop", True, "stop"),
(_NAMED_WEATHER_TOOL_CHOICE, "length", True, "length"),
("none", "stop", False, "stop"),
(None, None, False, "stop"),
],
)
def test_resolve_finish_reason(
tool_choice, engine_finish_reason, tool_calls_made, expected
):
"""``tool_calls`` needs a real call, a natural stop and no forced tool;
``length`` is never hidden behind a parsed call."""
request = ChatCompletionRequest(
model=MODEL_NAME,
messages=[{"role": "user", "content": "test"}],
tools=_TOOL_FINISH_REASON_TOOLS if tool_choice else None,
tool_choice=tool_choice,
)
assert (
resolve_finish_reason(engine_finish_reason, request, tool_calls_made)
== expected
)
@pytest.mark.asyncio
@pytest.mark.parametrize(
("tool_choice", "text", "engine_finish_reason", "expected_finish_reason"),
[
("auto", _HERMES_TOOL_CALL_TEXT, "stop", "tool_calls"),
("auto", _HERMES_TOOL_CALL_TEXT, "length", "length"),
("required", "Hello, how are you?", "stop", "stop"),
(_NAMED_WEATHER_TOOL_CHOICE, _HERMES_TOOL_CALL_TEXT, "stop", "stop"),
],
)
async def test_non_streaming_finish_reason_matches_tool_calls(
monkeypatch: pytest.MonkeyPatch,
tool_choice,
text: str,
engine_finish_reason: str,
expected_finish_reason: str,
):
"""Non-streaming ``finish_reason`` must agree with ``message.tool_calls``
and never hide a ``length`` truncation (#53255, #58228)."""
serving_chat = _build_serving_chat(
_build_mock_engine(), tool_parser="hermes", enable_auto_tools=True
)
# Parse required/named calls from the native <tool_call> format regardless
# of VLLM_ENFORCE_STRICT_TOOL_CALLING, so one text serves every case.
monkeypatch.setattr(
serving_chat.parser_cls.tool_parser_cls, "supports_required_and_named", False
)
tokenizer = get_tokenizer(MODEL_NAME)
request = ChatCompletionRequest(
model=MODEL_NAME,
messages=[{"role": "user", "content": "test"}],
tools=_TOOL_FINISH_REASON_TOOLS,
tool_choice=tool_choice,
)
parser = serving_chat.parser_cls(
tokenizer,
request.tools,
chat_template_kwargs=serving_chat._effective_chat_template_kwargs(request),
model_config=serving_chat.model_config,
)
request_output = RequestOutput(
request_id="test-req",
prompt="test",
prompt_token_ids=[1, 2, 3],
prompt_logprobs=None,
outputs=[
CompletionOutput(
index=0,
text=text,
token_ids=tokenizer.encode(text, add_special_tokens=False),
cumulative_logprob=0.0,
logprobs=None,
finish_reason=engine_finish_reason,
)
],
finished=True,
)
response = await serving_chat.chat_completion_full_generator(
request=request,
result_generator=_single_request_output(request_output),
request_id="test-req",
model_name=MODEL_NAME,
conversation=[],
tokenizer=tokenizer,
request_metadata=RequestResponseMetadata(
request_id="test-req", model_name=MODEL_NAME
),
parser=parser,
)
assert isinstance(response, ChatCompletionResponse)
choice = response.choices[0]
assert choice.finish_reason == expected_finish_reason
assert bool(choice.message.tool_calls) == (text != "Hello, how are you?"), (
"finish_reason must reflect whether tool calls were emitted"
)
@@ -456,3 +456,27 @@ def test_num_init_messages_offset():
items = ctx.make_response_output_items()
assert len(items) == 1
assert items[0].type == "message"
def test_length_truncation_marks_last_item_incomplete():
"""A turn cut by max_output_tokens ends in an ``incomplete`` item, both
with a parser (tool call) and without one (plain message)."""
request = _make_request(
tool_choice="auto",
tools=[
{
"type": "function",
"name": "get_weather",
"parameters": {"type": "object", "properties": {}},
}
],
)
ctx = _make_context(_ToolCallingParser, request=request, enable_auto_tools=True)
ctx.append_output(_make_request_output(text="calling tool", finish_reason="length"))
assert ctx.response_messages[-1].type == "function_call"
assert ctx.response_messages[-1].status == "incomplete"
ctx = _make_context(None)
ctx.append_output(_make_request_output(text="cut mid sen", finish_reason="length"))
assert ctx.response_messages[-1].type == "message"
assert ctx.response_messages[-1].status == "incomplete"
@@ -991,3 +991,28 @@ class TestConstructInputMessagesInstructionsLeak:
assert len(msgs) == 2
assert msgs[0] == {"role": "system", "content": "be helpful"}
assert msgs[1] == {"role": "user", "content": "hello"}
class TestBuildResponseOutputItemsStatus:
"""Item ``status`` must follow the finish reason of the generation
(#57998): a ``length`` cut marks the item that was cut ``incomplete``."""
_CALLS = [
FunctionCall(name="get_weather", arguments='{"city": "Paris"}'),
FunctionCall(name="get_weather", arguments='{"city": "Tok'),
]
def test_length_marks_truncated_message_incomplete(self):
(item,) = build_response_output_items(
reasoning=None, content="cut mid sen", tool_calls=[], finish_reason="length"
)
assert item.type == "message"
assert item.status == "incomplete"
def test_length_marks_only_last_tool_call_incomplete(self):
message, first, last = build_response_output_items(
reasoning=None, content="hi", tool_calls=self._CALLS, finish_reason="length"
)
assert message.status == "completed"
assert first.status == "completed"
assert last.status == "incomplete"
@@ -1,6 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import asyncio
from collections import deque
from contextlib import AsyncExitStack, asynccontextmanager
from types import SimpleNamespace
from unittest.mock import AsyncMock, MagicMock
@@ -51,6 +53,7 @@ from vllm.entrypoints.openai.responses.context import (
from vllm.entrypoints.openai.responses.protocol import (
ResponseCompletedEvent,
ResponseCreatedEvent,
ResponseIncompleteEvent,
ResponseRawMessageAndToken,
ResponsesRequest,
ResponsesResponse,
@@ -1414,6 +1417,7 @@ def _make_request_output(
text,
token_ids,
metrics: RequestStateStats | None = None,
finish_reason: str | None = None,
):
completion = CompletionOutput(
index=0,
@@ -1421,7 +1425,7 @@ def _make_request_output(
token_ids=token_ids,
cumulative_logprob=0.0,
logprobs=None,
finish_reason=None,
finish_reason=finish_reason,
stop_reason=None,
)
return RequestOutput(
@@ -1441,10 +1445,11 @@ def _make_simple_context_with_output(
token_ids,
response_parser=None,
metrics: RequestStateStats | None = None,
finish_reason: str | None = None,
):
"""Create a SimpleContext with a RequestOutput containing the given text."""
ctx = SimpleContext(response_parser=response_parser)
req_output = _make_request_output(text, token_ids, metrics)
req_output = _make_request_output(text, token_ids, metrics, finish_reason)
ctx.append_output(req_output)
ctx.request_metrics = metrics
return ctx
@@ -1561,6 +1566,66 @@ async def test_responses_streaming_metrics_only_on_completed_event():
assert events[-1].response.metrics.time_to_first_token_ms == pytest.approx(500.0)
async def _stream_truncated_by_length(serving: OpenAIServingResponses):
request = ResponsesRequest(input="hi", tools=[], stream=True, store=False)
context = _make_simple_context_with_output(
"cut mid sen", [10, 20], finish_reason="length"
)
async def one_context():
yield context
return [
event
async for event in serving.responses_stream_generator(
request=request,
sampling_params=SamplingParams(max_tokens=2),
result_generator=one_context(),
context=context,
model_name="test-model",
tokenizer=MagicMock(),
request_metadata=RequestResponseMetadata(request_id="req"),
)
]
@pytest.mark.asyncio
async def test_responses_stream_truncated_by_length_reports_incomplete():
"""A max_output_tokens cut must end the stream with ``response.incomplete``
and mark the cut item ``incomplete`` in both ``output_item.done`` and the
final response (#57998)."""
events = await _stream_truncated_by_length(_make_serving_instance())
item_done = [e for e in events if e.type == "response.output_item.done"]
assert [e.item.status for e in item_done] == ["incomplete"]
final = events[-1]
assert isinstance(final, ResponseIncompleteEvent)
assert final.response.status == "incomplete"
assert final.response.incomplete_details.reason == "max_output_tokens"
assert [item.status for item in final.response.output] == ["incomplete"]
@pytest.mark.asyncio
async def test_background_stream_replay_stops_on_response_incomplete():
"""Replaying a stored background stream must end at ``response.incomplete``
as it does at ``response.completed``, instead of waiting for more events."""
serving = _make_serving_instance()
events = await _stream_truncated_by_length(serving)
assert isinstance(events[-1], ResponseIncompleteEvent)
serving.event_store["resp_bg"] = (deque(events), asyncio.Event())
async def replay(starting_after=None):
return [
event
async for event in serving.responses_background_stream_generator(
"resp_bg", starting_after=starting_after
)
]
assert await asyncio.wait_for(replay(), timeout=5) == events
assert await asyncio.wait_for(replay(starting_after=0), timeout=5) == events[1:]
@pytest.mark.asyncio
async def test_responses_metrics_suppressed_for_multiple_generation_turns():
response = await _make_full_metrics_response(
@@ -3,11 +3,18 @@
"""Tests for the /derender endpoints (postprocessing counterpart to /render)."""
from unittest.mock import MagicMock
import httpx
import pytest
import pytest_asyncio
from tests.utils import RemoteLaunchRenderServer
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
GenerateTokensResponse,
)
from vllm.renderers.online_derenderer import OnlineDerenderer
from vllm.tokenizers import get_tokenizer
MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
@@ -1432,3 +1439,55 @@ async def test_e2e_harmony_reasoning(harmony_client, harmony_tokenizer):
assert msg["reasoning"] is not None
assert reasoning_text in msg["reasoning"]
assert answer_text in (msg["content"] or "")
_HERMES_TOOL_CALL_TEXT = (
'<tool_call>\n{"name": "get_weather", "arguments": {"city": "Paris"}}\n</tool_call>'
)
@pytest.mark.parametrize(
("tool_choice", "text", "engine_finish_reason", "expected_finish_reason"),
[
("auto", _HERMES_TOOL_CALL_TEXT, "stop", "tool_calls"),
("auto", _HERMES_TOOL_CALL_TEXT, "length", "length"),
("required", "Hello there", "stop", "stop"),
],
)
def test_derender_chat_finish_reason_matches_chat_completion(
tool_choice, text, engine_finish_reason, expected_finish_reason
):
"""/derender applies the same finish_reason rule as the coupled chat
path: ``tool_calls`` only for a real call at a natural stop."""
tokenizer = get_tokenizer(MODEL_NAME)
renderer = MagicMock()
renderer.get_tokenizer.return_value = tokenizer
model_config = MagicMock()
model_config.model = MODEL_NAME
model_config.hf_config.model_type = "llama"
derenderer = OnlineDerenderer(
model_config,
renderer,
request_logger=None,
chat_template="{{ messages }}",
chat_template_content_format="auto",
enable_auto_tools=True,
tool_parser="hermes",
)
request = ChatCompletionRequest(
model=MODEL_NAME,
messages=[{"role": "user", "content": "Weather in Paris?"}],
tools=_E2E_TOOLS,
tool_choice=tool_choice,
)
response = GenerateTokensResponse.model_validate(
_make_generate_response(
tokenizer.encode(text, add_special_tokens=False),
finish_reason=engine_finish_reason,
)
)
(choice,) = derenderer._derender_chat(response, request)
assert choice.finish_reason == expected_finish_reason
assert bool(choice.message.tool_calls) == (text == _HERMES_TOOL_CALL_TEXT)
@@ -1289,7 +1289,15 @@ class TestDerenderChatStreamParsed:
assert state.last_tool_call_ids == [first_id]
@pytest.mark.asyncio
async def test_finish_reason_rewritten_to_tool_calls(self, parsed_derenderer):
@pytest.mark.parametrize(
("engine_finish_reason", "expected_finish_reason"),
[("stop", "tool_calls"), ("length", "length")],
)
async def test_finish_reason_rewritten_to_tool_calls(
self, parsed_derenderer, engine_finish_reason, expected_finish_reason
):
"""A streamed tool call is reported as ``tool_calls`` only at a
natural stop; a ``length`` cut must stay visible."""
chat_request = _chat_request(
tools=[{"type": "function", "function": {"name": "get_weather"}}],
tool_choice="auto",
@@ -1297,11 +1305,12 @@ class TestDerenderChatStreamParsed:
chunk, _ = await parsed_derenderer.derender_chat_stream(
model=MODEL_NAME,
generate_chunk=_make_stream_chunk(
[_FakeParser.TOOL_START, _FakeParser.TOOL_ARG], finish_reason="stop"
[_FakeParser.TOOL_START, _FakeParser.TOOL_ARG],
finish_reason=engine_finish_reason,
),
chat_request=chat_request,
)
assert chunk.choices[0].finish_reason == "tool_calls"
assert chunk.choices[0].finish_reason == expected_finish_reason
@pytest.mark.asyncio
async def test_finish_reason_stop_for_named_tool_choice(self, parsed_derenderer):