[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
+5
View File
@@ -207,6 +207,11 @@ if _SDK_AVAILABLE:
if isinstance(result, ErrorResponse):
return _error_response(result, raw_request)
if (kwargs := result.reasoning_parser_kwargs) is not None:
kwargs.chat_template_kwargs = handler._engine_chat_template_kwargs(
kwargs.chat_template_kwargs
)
return JSONResponse(content=result.model_dump())
class CohereErrorEnvelopeMiddleware(BaseHTTPMiddleware):
@@ -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,
@@ -17,6 +17,7 @@ from vllm.entrypoints.scale_out.token_in_token_out.mm_features import (
from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
GenerateRequest,
MultiModalFeatures,
ReasoningParserKwargs,
)
from vllm.entrypoints.serve.engine.protocol import ErrorResponse
from vllm.entrypoints.serve.engine.serving import BaseServing
@@ -124,6 +125,20 @@ class ServingRender(BaseServing):
)
params = request.to_sampling_params(max_tokens, self.default_sampling_params)
# Resolve here what OpenAIServingChat passes to the engine:
# `_grammar_from_parser` (set by `adjust_request`) is not serialized.
reasoning_parser_kwargs = self._reasoning_parser_kwargs(request)
parser = None
if reasoning_parser_kwargs is not None:
assert self.online_renderer.parser is not None
parser = self.online_renderer.parser(
self.online_renderer.renderer.get_tokenizer(),
request.tools,
chat_template_kwargs=reasoning_parser_kwargs.chat_template_kwargs,
model_config=self.model_config,
)
reasoning_ended = request.resolve_reasoning_ended(parser, token_ids)
request_id = f"chatcmpl-{random_uuid()}"
return GenerateRequest(
@@ -132,6 +147,8 @@ class ServingRender(BaseServing):
features=self._extract_mm_features(engine_input),
sampling_params=params,
model=request.model,
reasoning_ended=reasoning_ended,
reasoning_parser_kwargs=reasoning_parser_kwargs,
stream=bool(request.stream),
stream_options=(request.stream_options if request.stream else None),
cache_salt=request.cache_salt,
@@ -263,6 +280,7 @@ class ServingRender(BaseServing):
features=self._extract_mm_features(engine_input),
sampling_params=params,
model=request.model,
reasoning_parser_kwargs=self._reasoning_parser_kwargs(request),
stream=bool(request.stream),
cache_salt=request.cache_salt,
priority=request.priority,
@@ -271,6 +289,18 @@ class ServingRender(BaseServing):
token_offsets=engine_input.get("prompt_token_offsets"),
)
def _reasoning_parser_kwargs(
self, request: ChatCompletionRequest | ResponsesRequest
) -> ReasoningParserKwargs | None:
parser_cls = self.online_renderer.parser
if parser_cls is None or parser_cls.reasoning_parser_cls is None:
return None
return ReasoningParserKwargs(
chat_template_kwargs=self.online_renderer.effective_chat_template_kwargs(
request
)
)
def _placeholder_metadata_fields(self, modality: str) -> set[str]:
"""Fields the EC consumer still requires after embeddings transfer."""
if self._placeholder_metadata_parser_failed:
@@ -4,6 +4,7 @@ from typing import Annotated, Any, Literal, TypeAlias
from pydantic import (
BaseModel,
ConfigDict,
Discriminator,
Field,
NonNegativeInt,
@@ -147,6 +148,16 @@ class MultiModalFeatures(BaseModel):
return self
class ReasoningParserKwargs(BaseModel):
"""Kwargs for the engine-side reasoning parser that gates structured
outputs. Typed so clients cannot pass arbitrary constructor kwargs."""
model_config = ConfigDict(extra="forbid")
chat_template_kwargs: dict[str, Any] = Field(default_factory=dict)
"""The effective chat template kwargs the prompt was rendered with."""
class GenerateRequest(BaseModel):
request_id: str = Field(
default_factory=lambda: f"{random_uuid()}",
@@ -212,6 +223,15 @@ class GenerateRequest(BaseModel):
model: str | None = None
reasoning_ended: bool | None = None
"""Whether reasoning has ended before the first generated token, as
resolved by /render. `True` applies structured outputs from the first
token; `None` lets the engine check the prompt with its reasoning parser."""
reasoning_parser_kwargs: ReasoningParserKwargs | None = None
"""Set by /render when it has a reasoning parser, so the engine-side
parser agrees with the frontend on flags such as `enable_thinking`."""
return_token_ids: bool | None = Field(
default=None,
description=(
@@ -349,6 +349,12 @@ class ServingTokens(GenerateBaseServing):
priority=request.priority,
data_parallel_rank=data_parallel_rank,
session_id=session_id,
reasoning_ended=request.reasoning_ended,
reasoning_parser_kwargs=(
request.reasoning_parser_kwargs.model_dump()
if request.reasoning_parser_kwargs is not None
else None
),
)
assert result_generator is not None