[Bugfix][Reasoning] Nemotron V3: surface reasoning as content when thinking is unterminated (#39091)

Signed-off-by: Andrii Skliar <askliar@nvidia.com>
Co-authored-by: Andrii Skliar <askliar@nvidia.com>
This commit is contained in:
Andrii Skliar
2026-06-10 05:58:19 -07:00
committed by GitHub
co-authored by Andrii Skliar
parent 77f42d9725
commit 4882fd7632
3 changed files with 163 additions and 10 deletions
+23 -1
View File
@@ -776,6 +776,28 @@ class DelegatingParser(Parser):
last_tc.function.arguments or ""
) + self._tool_parser.get_remaining_unstreamed_args()
def finalize_generation(
self,
delta_message: DeltaMessage | None,
request: ChatCompletionRequest | ResponsesRequest,
state: StreamState,
) -> DeltaMessage | None:
"""Finalize generation for cases where generation was incomplete.
For example, if streaming terminated before reasoning ended
"""
fallback_fn = getattr(
self._reasoning_parser, "get_streaming_fallback_content", None
)
if fallback_fn is not None and not state.reasoning_ended:
promoted = fallback_fn(state.previous_text, request)
if promoted:
if delta_message is None:
delta_message = DeltaMessage()
delta_message.content = (delta_message.content or "") + promoted
self._append_unstreamed_tool_args(delta_message)
return delta_message
def parse(
self,
model_output: str,
@@ -883,6 +905,6 @@ class DelegatingParser(Parser):
state.previous_token_ids = current_token_ids
if finished:
self._append_unstreamed_tool_args(delta_message)
delta_message = self.finalize_generation(delta_message, request, state)
return delta_message
+22 -7
View File
@@ -14,20 +14,35 @@ class NemotronV3ReasoningParser(DeepSeekR1ReasoningParser):
Reasoning parser for Nemotron V3 models.
"""
def extract_reasoning(
self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
) -> tuple[str | None, str | None]:
reasoning, final_content = super().extract_reasoning(model_output, request)
def _should_force_content(
self, request: ChatCompletionRequest | ResponsesRequest
) -> bool:
chat_template_kwargs = getattr(request, "chat_template_kwargs", None)
if (
return bool(
chat_template_kwargs
and (
chat_template_kwargs.get("enable_thinking") is False
or chat_template_kwargs.get("force_nonempty_content") is True
)
and (final_content is None or not final_content.strip())
)
def extract_reasoning(
self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
) -> tuple[str | None, str | None]:
reasoning, final_content = super().extract_reasoning(model_output, request)
if self._should_force_content(request) and (
final_content is None or not final_content.strip()
):
reasoning, final_content = final_content, reasoning
return reasoning, final_content
def get_streaming_fallback_content(
self, text: str, request: ChatCompletionRequest | ResponsesRequest
) -> str | None:
"""Reasoning to duplicate into content on the terminal streaming delta."""
if not self._should_force_content(request):
return None
reasoning, _ = super().extract_reasoning(text, request)
return reasoning