[Frontend] Fix reasoning_tokens for text-based parsers in Responses API (#33513)

Signed-off-by: Jaeyeon Kim <anencore94@gmail.com>
This commit is contained in:
Jaeyeon Kim(김재연)
2026-02-18 23:16:41 -08:00
committed by GitHub
parent b6101d384d
commit 9681068cf9
7 changed files with 208 additions and 3 deletions
+3 -1
View File
@@ -280,7 +280,6 @@ class ParsableContext(ConversationContext):
self.num_prompt_tokens = 0
self.num_output_tokens = 0
self.num_cached_tokens = 0
# TODO: num_reasoning_tokens is not implemented yet.
self.num_reasoning_tokens = 0
# not implemented yet for ParsableContext
self.all_turn_metrics: list[TurnMetrics] = []
@@ -308,12 +307,15 @@ class ParsableContext(ConversationContext):
self.input_messages: list[ResponseRawMessageAndToken] = []
self.output_messages: list[ResponseRawMessageAndToken] = []
self._accumulated_token_ids: list[int] = []
def append_output(self, output: RequestOutput) -> None:
self.num_prompt_tokens = len(output.prompt_token_ids or [])
self.num_cached_tokens = output.num_cached_tokens or 0
self.num_output_tokens += len(output.outputs[0].token_ids or [])
self.parser.process(output.outputs[0])
output_token_ids = output.outputs[0].token_ids or []
self._accumulated_token_ids.extend(output_token_ids)
# only store if enable_response_messages is True, save memory
if self.request.enable_response_messages:
@@ -759,6 +759,19 @@ class OpenAIServingResponses(OpenAIServing):
num_generated_tokens = context.num_output_tokens
num_cached_tokens = context.num_cached_tokens
num_reasoning_tokens = context.num_reasoning_tokens
# For text-based reasoning parsers (e.g., <think>...</think>),
# HarmonyContext already counts reasoning tokens via channels.
# For Simple/Parsable contexts, derive reasoning_tokens from
# accumulated output token IDs using the parser if not already set.
if (
num_reasoning_tokens == 0
and self.parser is not None
and self.parser.reasoning_parser_cls is not None
and isinstance(context, (SimpleContext, ParsableContext))
):
reasoning_parser = self.parser.reasoning_parser_cls(tokenizer)
accumulated = getattr(context, "_accumulated_token_ids", []) or []
num_reasoning_tokens = reasoning_parser.count_reasoning_tokens(accumulated)
usage = ResponseUsage(
input_tokens=num_prompt_tokens,