[Bugfix][Parser] Honor a prompt-closed reasoning block in non-streaming parse and reasoning token counts (#61857)

Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>
This commit is contained in:
alberto
2026-10-16 12:54:42 -04:00
committed by GitHub
parent cffc0cfa31
commit ae8f41238c
6 changed files with 253 additions and 2 deletions
+8
View File
@@ -68,6 +68,11 @@ class ParserEngineReasoningAdapter(ReasoningParser):
def find_reasoning_end(self, token_ids: Sequence[int]) -> ReasoningEnd | None:
return self._parser_engine.find_reasoning_end(token_ids)
def set_prompt_token_ids(self, prompt_token_ids: Sequence[int]) -> None:
self._parser_engine.set_prompt_token_ids(prompt_token_ids)
if self._counting_parser_engine is not None:
self._counting_parser_engine.set_prompt_token_ids(prompt_token_ids)
def adjust_initial_state_from_prompt(self, prompt_token_ids: Sequence[int]) -> None:
self._parser_engine.adjust_initial_state_from_prompt(prompt_token_ids)
@@ -148,6 +153,9 @@ class ParserEngineReasoningAdapter(ReasoningParser):
self._counting_parser_engine = self._parser_engine_cls(
self.model_tokenizer, **self._parser_engine_kwargs
) # type: ignore[call-arg]
self._counting_parser_engine._prompt_closes_reasoning = (
self._parser_engine._prompt_closes_reasoning
)
self._counting_parser_engine._single_pass_parse(
self.model_tokenizer.decode(token_ids), token_ids
)