[Frontend] Add the Beam reasoning and tool parsers (#62052)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>
Co-authored-by: neilkg <33635204+neilkg@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Woosuk Kwon
2026-10-16 14:08:29 -07:00
committed by GitHub
co-authored by Bugen Zhao Yoni Gozlan neilkg Cursor Claude Opus 5.5
parent 9731f30a03
commit 153a17166e
27 changed files with 4464 additions and 92 deletions
@@ -53,6 +53,7 @@ from openai_harmony import Message as OpenAIHarmonyMessage
from pydantic import (
BeforeValidator,
Field,
PrivateAttr,
ValidationError,
field_serializer,
model_validator,
@@ -469,6 +470,14 @@ class ResponsesRequest(OpenAIBaseModel):
max_output_tokens_param="max_output_tokens",
)
_grammar_from_parser: bool = PrivateAttr(default=False)
"""CAUTION: Should only be set by the parser's adjust_request."""
_parser_structured_outputs: StructuredOutputsParams | None = PrivateAttr(
default=None
)
"""Parser-derived constraints that preserve the public request's text metadata."""
_DEFAULT_SAMPLING_PARAMS = {
"temperature": 1.0,
"top_p": 1.0,
@@ -478,6 +487,8 @@ class ResponsesRequest(OpenAIBaseModel):
def extract_structured_outputs(self) -> StructuredOutputsParams | None:
"""Normalize request constraints into ``StructuredOutputsParams``."""
if self._parser_structured_outputs is not None:
return self._parser_structured_outputs
if self.text is None or self.text.format is None:
return self.structured_outputs
@@ -526,6 +526,7 @@ class OpenAIServingResponses(GenerateBaseServing):
trace_headers=trace_headers,
session_id=session_id,
reasoning_parser_kwargs=reasoning_parser_kwargs,
reasoning_ended=True if request._grammar_from_parser else None,
)
generators.append(generator)
@@ -647,6 +648,7 @@ class OpenAIServingResponses(GenerateBaseServing):
trace_headers: Mapping[str, str] | None = None,
session_id: str | None = None,
reasoning_parser_kwargs: dict[str, Any] | None = None,
reasoning_ended: bool | None = None,
):
max_model_len = self.model_config.max_model_len
cache_salt = cast(str | None, engine_input.get("cache_salt"))
@@ -673,6 +675,7 @@ class OpenAIServingResponses(GenerateBaseServing):
priority=priority,
session_id=session_id,
reasoning_parser_kwargs=reasoning_parser_kwargs,
reasoning_ended=reasoning_ended,
)
async for res in generator: