[responsesAPI] fix incomplete_messages for simple/parsable context (#31836)

Signed-off-by: Andrew Xia <axia@fb.com>
Co-authored-by: Andrew Xia <axia@fb.com>
This commit is contained in:
Andrew Xia
2026-01-09 21:00:57 +00:00
committed by GitHub
co-authored by Andrew Xia
parent 0a0aa07747
commit 1f8b7c536b
4 changed files with 44 additions and 0 deletions
@@ -51,7 +51,13 @@ class ResponsesParser:
if tool_parser_cls is not None:
self.tool_parser_instance = tool_parser_cls(tokenizer)
# Store the last finish_reason to determine response status
self.finish_reason: str | None = None
def process(self, output: CompletionOutput) -> "ResponsesParser":
# Store the finish_reason from the output
self.finish_reason = output.finish_reason
reasoning_content, content = self.reasoning_parser_instance.extract_reasoning(
output.text, request=self.request
)
@@ -692,6 +692,10 @@ class OpenAIServingResponses(OpenAIServing):
# TODO: Calculate usage.
# assert final_res.prompt_token_ids is not None
num_tool_output_tokens = 0
# Check finish reason from the parser
if context.parser.finish_reason == "length":
status = "incomplete"
else:
assert isinstance(context, SimpleContext)
# Use final_output which has accumulated text/token_ids/logprobs
@@ -703,6 +707,10 @@ class OpenAIServingResponses(OpenAIServing):
# finish_reason='error' indicates retryable internal error
self._raise_if_error(final_output.finish_reason, request.request_id)
# Check if generation was stopped due to max_tokens
if final_output.finish_reason == "length":
status = "incomplete"
output = self._make_response_output_items(request, final_output, tokenizer)
if request.enable_response_messages: