fix: report stop_sequence stop_reason in Anthropic Messages API (#45807)

Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
Yufeng He
2026-08-20 03:23:43 +00:00
committed by GitHub
co-authored by Chauncey
parent f85e060ec2
commit e85dd21497
2 changed files with 122 additions and 5 deletions
@@ -975,6 +975,7 @@ def _make_stream_chunk(
*,
delta: DeltaMessage | None = None,
finish_reason: str | None = None,
stop_reason: int | str | None = None,
choices: list[ChatCompletionResponseStreamChoice] | None = None,
usage: UsageInfo | None = None,
) -> str:
@@ -984,6 +985,7 @@ def _make_stream_chunk(
index=0,
delta=delta or DeltaMessage(),
finish_reason=finish_reason,
stop_reason=stop_reason,
)
]
chunk = ChatCompletionStreamResponse(
@@ -1463,6 +1465,103 @@ class TestCacheSalt:
handler.create_messages.assert_not_awaited()
class TestStopSequenceReason:
"""When generation stops because a configured stop string matched, the
Anthropic Messages API must report ``stop_reason="stop_sequence"`` and echo
the matched string in ``stop_sequence``. vLLM surfaces the matched string in
the OpenAI choice's ``stop_reason`` field (a str) while ``finish_reason``
stays ``"stop"``. A natural EOS (stop_reason None) or a stop token id (int)
must still map to ``end_turn``.
"""
def test_non_streaming_stop_string_maps_to_stop_sequence(self):
converter = _make_full_converter()
response = ChatCompletionResponse(
id="chatcmpl-test",
model="test-model",
choices=[
ChatCompletionResponseChoice(
index=0,
message=ChatMessage(role="assistant", content="hello"),
finish_reason="stop",
stop_reason="</tool>",
)
],
usage=UsageInfo(prompt_tokens=5, total_tokens=8, completion_tokens=3),
)
result = converter.messages_full_converter(response)
assert result.stop_reason == "stop_sequence"
assert result.stop_sequence == "</tool>"
def test_non_streaming_natural_eos_maps_to_end_turn(self):
converter = _make_full_converter()
response = ChatCompletionResponse(
id="chatcmpl-test",
model="test-model",
choices=[
ChatCompletionResponseChoice(
index=0,
message=ChatMessage(role="assistant", content="hello"),
finish_reason="stop",
stop_reason=None,
)
],
usage=UsageInfo(prompt_tokens=5, total_tokens=8, completion_tokens=3),
)
result = converter.messages_full_converter(response)
assert result.stop_reason == "end_turn"
assert result.stop_sequence is None
def test_non_streaming_stop_token_id_maps_to_end_turn(self):
converter = _make_full_converter()
response = ChatCompletionResponse(
id="chatcmpl-test",
model="test-model",
choices=[
ChatCompletionResponseChoice(
index=0,
message=ChatMessage(role="assistant", content="hello"),
finish_reason="stop",
stop_reason=128009,
)
],
usage=UsageInfo(prompt_tokens=5, total_tokens=8, completion_tokens=3),
)
result = converter.messages_full_converter(response)
assert result.stop_reason == "end_turn"
assert result.stop_sequence is None
@pytest.mark.asyncio
async def test_streaming_stop_string_maps_to_stop_sequence(self):
async def sse_input():
yield _make_stream_chunk(
delta=DeltaMessage(content="hi"),
usage=UsageInfo(prompt_tokens=5, total_tokens=5, completion_tokens=0),
)
yield _make_stream_chunk(finish_reason="stop", stop_reason="</tool>")
yield _make_stream_chunk(
choices=[],
usage=UsageInfo(prompt_tokens=5, total_tokens=8, completion_tokens=3),
)
yield "data: [DONE]"
converter = _make_stream_converter()
output = []
async for event in converter.message_stream_converter(sse_input()):
output.append(event)
events = _parse_sse_events(output)
msg_deltas = [data for ev_type, data in events if ev_type == "message_delta"]
assert msg_deltas[0]["delta"]["stop_reason"] == "stop_sequence"
assert msg_deltas[0]["delta"]["stop_sequence"] == "</tool>"
# ======================================================================
# Client-caused errors are 4xx, not 500 (Issue #52088)
# ======================================================================
+23 -5
View File
@@ -631,7 +631,13 @@ class AnthropicServingMessages(OpenAIServingChat):
)
choice = generator.choices[0]
if choice.finish_reason == "stop":
result.stop_reason = "end_turn"
# vLLM reports the matched stop string in stop_reason (a str);
# an int stop-token-id or None (natural EOS) maps to end_turn.
if isinstance(choice.stop_reason, str):
result.stop_reason = "stop_sequence"
result.stop_sequence = choice.stop_reason
else:
result.stop_reason = "end_turn"
elif choice.finish_reason == "length":
result.stop_reason = "max_tokens"
elif choice.finish_reason == "tool_calls":
@@ -715,6 +721,9 @@ class AnthropicServingMessages(OpenAIServingChat):
first_item = True
finish_reason = None
# Matched stop string, when generation stopped on one (a str);
# int stop-token-id / None are not stop sequences.
stop_sequence: int | str | None = None
state = _ActiveBlockState()
# Map from tool call index to tool_use_id
tool_index_to_id: dict[int, str] = {}
@@ -824,12 +833,20 @@ class AnthropicServingMessages(OpenAIServingChat):
if len(origin_chunk.choices) == 0:
for event in stop_and_flush():
yield event
stop_reason = self.stop_reason_map.get(
finish_reason or "stop"
)
if isinstance(stop_sequence, str):
stop_delta = AnthropicDelta(
stop_reason="stop_sequence",
stop_sequence=stop_sequence,
)
else:
stop_delta = AnthropicDelta(
stop_reason=self.stop_reason_map.get(
finish_reason or "stop"
)
)
chunk = AnthropicStreamEvent(
type="message_delta",
delta=AnthropicDelta(stop_reason=stop_reason),
delta=stop_delta,
usage=_build_anthropic_usage(origin_chunk.usage),
)
data = chunk.model_dump_json(exclude_unset=True)
@@ -838,6 +855,7 @@ class AnthropicServingMessages(OpenAIServingChat):
if origin_chunk.choices[0].finish_reason is not None:
finish_reason = origin_chunk.choices[0].finish_reason
stop_sequence = origin_chunk.choices[0].stop_reason
# continue
# thinking / text content