[Perf] Eliminate full-history reasoning scans for structured outputs (#55223)

Signed-off-by: sfeng33 <4florafeng@gmail.com>
Signed-off-by: Flora Feng <4florafeng@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
Flora Feng
2026-09-08 14:07:16 -07:00
committed by GitHub
co-authored by Wentao Ye
parent 73f61d12aa
commit 29f46ac475
8 changed files with 475 additions and 42 deletions
+15
View File
@@ -219,6 +219,21 @@ class Parser:
"""
return self.is_reasoning_end(input_ids)
def find_reasoning_end_offset(self, token_ids: Sequence[int]) -> int | None:
"""
Locate the token that ends reasoning within one window of tokens.
Args:
token_ids: The tokens to examine.
Returns:
The offset within ``token_ids`` of the last reasoning token, or
``None`` when reasoning does not end in this window. Parsers that
cannot answer from a window alone always return ``None``; callers
fall back to :meth:`is_reasoning_end_streaming`.
"""
return None
@abstractmethod
def extract_content_ids(self, input_ids: list[int]) -> list[int]:
"""