biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Fix Responses JSON schema alias serialization (#38519)

Browse Source 
Signed-off-by: noobhappylife <aratar1991@hotmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
noobHappylife
2026-04-09 10:50:16 +08:00
committed by GitHub
parent d37b378762
commit 2a49284c8a
6 changed files with 104 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
tests/entrypoints/openai/responses/test_harmony.py
View File

@@ -999,17 +999,21 @@ async def test_mcp_tool_multi_turn(client: OpenAI, model_name: str, server):
(msg.get("recipient") or "").startswith("python")
for msg in response1.output_messages
)
parsed_output_messages = [
Message.from_dict(msg) for msg in response1.output_messages
]
tool_response_found = any(
msg.get("author", {}).get("role") == "tool"
and (msg.get("author", {}).get("name") or "").startswith("python")
for msg in response1.output_messages
(msg.author.role == "tool" and (msg.author.name or "").startswith("python"))
for msg in parsed_output_messages
)
assert tool_call_found, "MCP tool call not found in output_messages"
assert tool_response_found, "MCP tool response not found in output_messages"
# No developer messages expected for elevated tools
developer_msgs = [
msg for msg in response1.input_messages if msg["author"]["role"] == "developer"
msg
for msg in (Message.from_dict(raw) for raw in response1.input_messages)
if msg.author.role == "developer"
]
assert len(developer_msgs) == 0, "No developer message expected for elevated tools"
@@ -1119,12 +1123,10 @@ async def test_function_call_with_previous_input_messages(
num_system = 0
num_developer = 0
num_tool = 0
for msg_dict in response_2.input_messages:
# input_messages use {"author": {"role": "..."}} format,
# not the top-level {"role": "..."} that Message.from_dict
# expects.
author = msg_dict.get("author", {})
role = author.get("role") if isinstance(author, dict) else None
for message in (
Message.from_dict(msg_dict) for msg_dict in response_2.input_messages
):
role = message.author.role
if role == "system":
num_system += 1
elif role == "developer":
@@ -1183,12 +1185,8 @@ async def test_system_prompt_override_no_duplication(client: OpenAI, model_name:
assert response.output_text is not None
num_system = 0
for msg in response.input_messages:
# input_messages use {"author": {"role": "system"}} format,
# not the top-level {"role": "system"} that Message.from_dict expects.
author = msg.get("author", {})
role = author.get("role") if isinstance(author, dict) else None
if role == "system":
for message in (Message.from_dict(msg) for msg in response.input_messages):
if message.author.role == "system":
num_system += 1
assert num_system == 1, f"Expected 1 system message, got {num_system}"

12
tests/entrypoints/openai/responses/test_mcp_tools.py
View File

@@ -7,7 +7,7 @@ from __future__ import annotations
import pytest
import pytest_asyncio
from openai import OpenAI
from openai_harmony import ToolDescription, ToolNamespaceConfig
from openai_harmony import Message, ToolDescription, ToolNamespaceConfig
from tests.utils import RemoteOpenAIServer
from vllm.entrypoints.mcp.tool_server import MCPToolServer
@@ -173,10 +173,10 @@ class TestMCPEnabled:
if recipient and recipient.startswith("python"):
tool_call_found = True
assert message.get("channel") == "commentary"
author = message.get("author", {})
if author.get("role") == "tool" and (author.get("name") or "").startswith(
"python"
):
parsed_message = Message.from_dict(message)
if parsed_message.author.role == "tool" and (
parsed_message.author.name or ""
).startswith("python"):
tool_response_found = True
assert message.get("channel") == "commentary"
@@ -188,7 +188,7 @@ class TestMCPEnabled:
assert tool_response_found, "No Python tool response found"
for message in response.input_messages:
assert message.get("author", {}).get("role") != "developer"
assert Message.from_dict(message).author.role != "developer"
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])

72
tests/entrypoints/openai/responses/test_serving_responses.py
View File

@@ -11,8 +11,12 @@ from openai.types.responses import (
ResponseReasoningItem,
ResponseReasoningTextDeltaEvent,
ResponseReasoningTextDoneEvent,
ResponseTextConfig,
ResponseTextDeltaEvent,
)
from openai.types.responses.response_format_text_json_schema_config import (
ResponseFormatTextJSONSchemaConfig,
)
from openai.types.responses.tool import (
CodeInterpreterContainerCodeInterpreterToolAuto,
LocalShell,
@@ -28,7 +32,13 @@ from vllm.entrypoints.openai.engine.protocol import (
RequestResponseMetadata,
)
from vllm.entrypoints.openai.responses.context import ConversationContext, SimpleContext
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.entrypoints.openai.responses.protocol import (
ResponseCreatedEvent,
ResponseRawMessageAndToken,
ResponsesRequest,
ResponsesResponse,
serialize_message,
)
from vllm.entrypoints.openai.responses.serving import (
OpenAIServingResponses,
_extract_allowed_tools_from_mcp_requests,
@@ -73,6 +83,16 @@ class MockConversationContext(ConversationContext):
pass
def test_serialize_message_pydantic_model_returns_dict() -> None:
msg = ResponseRawMessageAndToken(message="hello", tokens=[1, 2, 3])
serialized = serialize_message(msg)
assert isinstance(serialized, dict)
assert serialized["type"] == "raw_message_tokens"
assert serialized["message"] == "hello"
@pytest.fixture
def mock_serving_responses():
"""Create a mock OpenAIServingResponses instance"""
@@ -132,6 +152,56 @@ def test_extract_tool_types(monkeypatch: pytest.MonkeyPatch) -> None:
}
@pytest.mark.skip_global_cleanup
def test_response_created_event_uses_public_json_schema_alias() -> None:
schema = {
"type": "object",
"properties": {
"event_name": {"type": "string"},
"date": {"type": "string"},
"participants": {"type": "array", "items": {"type": "string"}},
},
"required": ["event_name", "date", "participants"],
"additionalProperties": False,
}
text = ResponseTextConfig()
text.format = ResponseFormatTextJSONSchemaConfig(
type="json_schema",
name="calendar_event",
schema=schema,
description="A calendar event.",
strict=True,
)
request = ResponsesRequest(
model="test-model",
input="Alice and Bob are going to a science fair on Friday.",
text=text,
)
sampling_params = request.to_sampling_params(default_max_tokens=64)
initial_response = ResponsesResponse.from_request(
request=request,
sampling_params=sampling_params,
model_name="test-model",
created_time=0,
output=[],
status="in_progress",
usage=None,
).model_dump(mode="json", by_alias=True)
fmt = initial_response["text"]["format"]
assert fmt["schema"] == schema
assert "schema_" not in fmt
event = ResponseCreatedEvent(
type="response.created",
sequence_number=0,
response=initial_response,
)
assert event.response.text is not None
assert event.response.text.format is not None
assert event.response.text.format.model_dump(by_alias=True)["schema"] == schema
class TestInitializeToolSessions:
"""Test class for _initialize_tool_sessions method"""

18
vllm/entrypoints/openai/responses/api_router.py
View File

@@ -39,7 +39,8 @@ async def _convert_stream_to_sse_events(
event_type = getattr(event, "type", "unknown")
# https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#event_stream_format
event_data = (
f"event: {event_type}\ndata: {event.model_dump_json(indent=None)}\n\n"
f"event: {event_type}\ndata: "
f"{event.model_dump_json(indent=None, by_alias=True)}\n\n"
)
yield event_data
@@ -65,10 +66,11 @@ async def create_responses(request: ResponsesRequest, raw_request: Request):
if isinstance(generator, ErrorResponse):
return JSONResponse(
content=generator.model_dump(), status_code=generator.error.code
content=generator.model_dump(mode="json", by_alias=True),
status_code=generator.error.code,
)
elif isinstance(generator, ResponsesResponse):
return JSONResponse(content=generator.model_dump())
return JSONResponse(content=generator.model_dump(mode="json", by_alias=True))
return StreamingResponse(
content=_convert_stream_to_sse_events(generator), media_type="text/event-stream"
@@ -95,10 +97,11 @@ async def retrieve_responses(
if isinstance(response, ErrorResponse):
return JSONResponse(
content=response.model_dump(), status_code=response.error.code
content=response.model_dump(mode="json", by_alias=True),
status_code=response.error.code,
)
elif isinstance(response, ResponsesResponse):
return JSONResponse(content=response.model_dump())
return JSONResponse(content=response.model_dump(mode="json", by_alias=True))
return StreamingResponse(
content=_convert_stream_to_sse_events(response), media_type="text/event-stream"
)
@@ -115,9 +118,10 @@ async def cancel_responses(response_id: str, raw_request: Request):
if isinstance(response, ErrorResponse):
return JSONResponse(
content=response.model_dump(), status_code=response.error.code
content=response.model_dump(mode="json", by_alias=True),
status_code=response.error.code,
)
return JSONResponse(content=response.model_dump())
return JSONResponse(content=response.model_dump(mode="json", by_alias=True))
def attach_router(app: FastAPI):

2
vllm/entrypoints/openai/responses/protocol.py
View File

@@ -106,7 +106,7 @@ def serialize_message(msg):
return msg.to_dict()
else:
# fallback to pydantic dump
return msg.model_dump_json(by_alias=True)
return msg.model_dump(mode="json", by_alias=True)
def serialize_messages(msgs):

2
vllm/entrypoints/openai/responses/serving.py
View File

@@ -1909,7 +1909,7 @@ class OpenAIServingResponses(OpenAIServing):
output=[],
status="in_progress",
usage=None,
).model_dump()
).model_dump(mode="json", by_alias=True)
yield _increment_sequence_number_and_return(
ResponseCreatedEvent(
type="response.created",