[BugFix] Omit empty tool_calls from OpenAI chat responses (#44105)

Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Signed-off-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
jack
2026-06-23 13:43:53 +08:00
committed by GitHub
co-authored by QwertyJack Chauncey
parent 3ce5823762
commit 6c427dd401
5 changed files with 133 additions and 4 deletions
@@ -540,8 +540,8 @@ async def test_max_tokens_with_tool_choice_required(
tool_choice=tool_choice,
)
# When `tool_choice="required"` and the tokens of `tools` exceed `max_tokens`,
# both `tool_calls` and `content` should be empty.
# `tool_calls` should be absent and `content` should be empty.
# This behavior should be consistent with OpenAI.
choice = chat_completion.choices[0]
assert choice.finish_reason == "length"
assert len(choice.message.tool_calls) == 0
assert choice.message.tool_calls is None
@@ -465,7 +465,7 @@ class TestGPTOSSChat:
)
msg = tool_choice_none.choices[0].message
assert len(msg.tool_calls) == 0
assert msg.tool_calls is None
class TestGPTOSSSpeculativeChat:
@@ -2,8 +2,25 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from openai.types.chat.chat_completion import ChatCompletion as OpenAIChatCompletion
from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
ChatCompletionResponse,
ChatCompletionResponseChoice,
ChatCompletionResponseStreamChoice,
ChatCompletionStreamResponse,
ChatMessage,
)
from vllm.entrypoints.openai.engine.protocol import (
DeltaFunctionCall,
DeltaMessage,
DeltaToolCall,
FunctionCall,
ToolCall,
UsageInfo,
)
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.parser.abstract_parser import DelegatingParser
@@ -86,3 +103,101 @@ def test_responses_parser_allows_named_tool_choice_with_none_content():
assert content is None
assert tool_calls == []
def _chat_response(message: ChatMessage) -> ChatCompletionResponse:
return ChatCompletionResponse(
model="test-model",
choices=[
ChatCompletionResponseChoice(
index=0,
message=message,
finish_reason="stop",
)
],
usage=UsageInfo(prompt_tokens=1, completion_tokens=1, total_tokens=2),
)
def test_chat_completion_response_omits_empty_tool_calls_payload():
response = _chat_response(ChatMessage(role="assistant", content="done"))
payload = response.model_dump()
payload_exclude_unset = response.model_dump(exclude_unset=True)
assert "tool_calls" not in payload["choices"][0]["message"]
assert "tool_calls" not in payload_exclude_unset["choices"][0]["message"]
parsed = OpenAIChatCompletion.model_validate(payload)
assert parsed.choices[0].message.tool_calls is None
def test_chat_completion_response_keeps_non_empty_tool_calls_payload():
response = _chat_response(
ChatMessage(
role="assistant",
content="",
tool_calls=[
ToolCall(
function=FunctionCall(
name="get_weather",
arguments='{"city": "Beijing"}',
)
)
],
)
)
message = response.model_dump()["choices"][0]["message"]
assert len(message["tool_calls"]) == 1
assert message["tool_calls"][0]["function"]["name"] == "get_weather"
def _stream_response(delta: DeltaMessage) -> ChatCompletionStreamResponse:
return ChatCompletionStreamResponse(
id="chatcmpl-test",
object="chat.completion.chunk",
created=1,
model="test-model",
choices=[
ChatCompletionResponseStreamChoice(
index=0,
delta=delta,
finish_reason=None,
)
],
)
def test_chat_completion_stream_response_omits_empty_tool_calls_payload():
response = _stream_response(DeltaMessage(content="done"))
payload = response.model_dump(exclude_unset=True)
payload_json = response.model_dump_json(exclude_unset=True)
assert "tool_calls" not in payload["choices"][0]["delta"]
parsed = ChatCompletionChunk.model_validate_json(payload_json)
assert parsed.choices[0].delta.tool_calls is None
def test_chat_completion_stream_response_keeps_non_empty_tool_calls_payload():
response = _stream_response(
DeltaMessage(
tool_calls=[
DeltaToolCall(
index=0,
id="call-test",
type="function",
function=DeltaFunctionCall(
name="get_weather",
arguments='{"city": "Beijing"}',
),
)
]
)
)
delta = response.model_dump(exclude_unset=True)["choices"][0]["delta"]
assert len(delta["tool_calls"]) == 1
assert delta["tool_calls"][0]["function"]["name"] == "get_weather"
@@ -66,6 +66,13 @@ class ChatMessage(OpenAIBaseModel):
# vLLM-specific fields that are not in OpenAI spec
reasoning: str | None = None
@model_serializer(mode="wrap")
def _serialize(self, handler):
data = handler(self)
if len(data.get("tool_calls", [])) == 0:
data.pop("tool_calls", None)
return data
class ChatCompletionLogProb(OpenAIBaseModel):
token: str
@@ -353,6 +353,13 @@ class DeltaMessage(OpenAIBaseModel):
reasoning: str | None = None
tool_calls: list[DeltaToolCall] = Field(default_factory=list)
@model_serializer(mode="wrap")
def _serialize(self, handler):
data = handler(self)
if len(data.get("tool_calls", [])) == 0:
data.pop("tool_calls", None)
return data
class GenerationError(Exception):
"""raised when finish_reason indicates internal server error (500)"""