[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
@@ -380,6 +380,40 @@ class TestDelegatingPromptDetection:
assert delta.reasoning is None
assert delta.content == "answer"
def test_prior_turn_think_end_keeps_non_streaming_reasoning(
self, mock_tokenizer, mock_request
):
parser = _Qwen3DelegatingParser(mock_tokenizer)
parser.set_prompt_token_ids([_THINK_END_ID, _IM_END_ID, _IM_START_ID, _TEXT_ID])
output_ids = [ord("a"), _THINK_END_ID, ord("b")]
reasoning, content, _ = parser.parse(
mock_tokenizer.decode(output_ids),
mock_request,
model_output_token_ids=output_ids,
)
assert reasoning == "a"
assert content == "b"
assert parser.count_reasoning_tokens(output_ids) > 0
def test_think_end_in_current_turn_parses_output_as_content(
self, mock_tokenizer, mock_request
):
parser = _Qwen3DelegatingParser(mock_tokenizer)
parser.set_prompt_token_ids([_IM_START_ID, _THINK_START_ID, _THINK_END_ID])
output_ids = [ord("a"), ord("b")]
reasoning, content, _ = parser.parse(
mock_tokenizer.decode(output_ids),
mock_request,
model_output_token_ids=output_ids,
)
assert reasoning is None
assert content == "ab"
assert parser.count_reasoning_tokens(output_ids) == 0
class TestStreaming:
def test_basic_streaming(self, parser):