[Bugfix] [Frontend] Cleanup gpt-oss non-streaming chat tool calls (#25514)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
This commit is contained in:
Ben Browning
2025-09-24 03:20:38 +00:00
committed by GitHub
parent d747c2ef18
commit 5caaeb714c
4 changed files with 102 additions and 12 deletions
+5 -8
View File
@@ -1186,6 +1186,10 @@ class OpenAIServingChat(OpenAIServing):
logprobs = None
if self.use_harmony:
reasoning_content, content, _ = parse_chat_output(token_ids)
if not request.include_reasoning:
reasoning_content = None
if self.tool_parser is not None:
tool_parser = self.tool_parser(tokenizer)
# NOTE: We use token_ids for openai tool parser
@@ -1194,10 +1198,7 @@ class OpenAIServingChat(OpenAIServing):
request=request,
token_ids=token_ids, # type: ignore
)
reasoning_content, content = None, tool_call_info.content
if request.include_reasoning:
reasoning_content, content, _ = parse_chat_output(
token_ids)
content = tool_call_info.content
message = ChatMessage(
role=role,
reasoning_content=reasoning_content,
@@ -1205,10 +1206,6 @@ class OpenAIServingChat(OpenAIServing):
tool_calls=tool_call_info.tool_calls,
)
else:
reasoning_content, content, _ = parse_chat_output(
token_ids)
if not request.include_reasoning:
reasoning_content = None
message = ChatMessage(
role=role,
reasoning_content=reasoning_content,
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import json
from collections.abc import Sequence
from typing import TYPE_CHECKING
@@ -12,10 +13,13 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
FunctionCall, ToolCall)
from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
ToolParser, ToolParserManager)
from vllm.logger import init_logger
if TYPE_CHECKING:
from vllm.transformers_utils.tokenizer import AnyTokenizer
logger = init_logger(__name__)
@ToolParserManager.register_module("openai")
class OpenAIToolParser(ToolParser):
@@ -40,17 +44,33 @@ class OpenAIToolParser(ToolParser):
if len(parser.messages) > 0:
for msg in parser.messages:
if len(msg.content) < 1:
continue
msg_text = msg.content[0].text
if msg.recipient and msg.recipient.startswith("functions."):
# If no content-type is given assume JSON, as that's the
# most common case with gpt-oss models.
if not msg.content_type or "json" in msg.content_type:
# load and dump the JSON text to check validity and
# remove any extra newlines or other odd formatting
try:
tool_args = json.dumps(json.loads(msg_text))
except json.JSONDecodeError:
logger.exception(
"Error decoding JSON tool call from response.")
tool_args = msg_text
else:
tool_args = msg_text
tool_calls.append(
ToolCall(
type="function",
function=FunctionCall(
name=msg.recipient.split("functions.")[1],
arguments=msg.content[0].text,
arguments=tool_args,
),
))
elif msg.channel == "final":
final_content = msg.content[0].text
final_content = msg_text
return ExtractedToolCallInformation(
tools_called=len(tool_calls) > 0,