[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:
@@ -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(
|
||||
|
||||
Reference in New Issue
Block a user