[BugFix][Frontend] Pass reasoning_ended through render → generate (#60059) (#60062)

Signed-off-by: Yida Weng <84164537+YidaWeng@users.noreply.github.com>
Signed-off-by: Flora Feng <4florafeng@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
This commit is contained in:
Yida Weng
2026-10-05 20:39:25 +00:00
committed by GitHub
co-authored by Flora Feng
parent 74c5cbcd7b
commit 18c8a65edd
11 changed files with 249 additions and 22 deletions
@@ -4,7 +4,7 @@
# Adapted from
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
import time
from typing import Annotated, Any, ClassVar, Literal
from typing import TYPE_CHECKING, Annotated, Any, ClassVar, Literal
from openai.types.chat.chat_completion_audio import (
ChatCompletionAudio as OpenAIChatCompletionAudio,
@@ -53,6 +53,9 @@ from vllm.sampling_params import (
)
from vllm.utils import random_uuid
if TYPE_CHECKING:
from vllm.parser.abstract_parser import Parser
logger = init_logger(__name__)
@@ -566,6 +569,25 @@ class ChatCompletionRequest(OpenAIBaseModel):
_grammar_from_parser: bool = PrivateAttr(default=False)
"""CAUTION: Should only be set by the parser-engine adapter's adjust_request."""
def resolve_reasoning_ended(
self, parser: "Parser | None", prompt_token_ids: list[int]
) -> bool | None:
"""Resolve the engine's `reasoning_ended` for structured outputs.
Call after `adjust_request`, which may set `_grammar_from_parser`.
`True` constrains from the first generated token; `None` lets the
engine check the prompt with its own reasoning parser.
"""
if not self.include_reasoning:
return True
if self._grammar_from_parser:
# The Mistral grammar already includes an optional `think?`
# rule that handles both reasoning and non-reasoning outputs.
return True
if parser is not None and parser.reasoning_parser is not None:
return parser.is_reasoning_end(prompt_token_ids)
return None
def build_chat_params(
self,
default_template: str | None,
@@ -364,17 +364,9 @@ class OpenAIServingChat(GenerateBaseServing):
session_id=session_id,
)
else:
if not request.include_reasoning:
reasoning_ended = True
elif request._grammar_from_parser:
# The Mistral grammar already includes an optional
# `think?` rule that handles both reasoning and
# non-reasoning outputs.
reasoning_ended = True
elif parser is not None and parser.reasoning_parser is not None:
reasoning_ended = parser.is_reasoning_end(prompt_token_ids or [])
else:
reasoning_ended = None
reasoning_ended = request.resolve_reasoning_ended(
parser, prompt_token_ids or []
)
generator = self.engine_client.generate(
engine_input,