Add streaming reasoning extraction with content promotion

super_v3_reasoning_parser.py

@@ -1,9 +1,23 @@
+from collections.abc import Sequence
+
+from vllm.entrypoints.openai.engine.protocol import DeltaMessage
 from vllm.reasoning.abs_reasoning_parsers import ReasoningParserManager
 from vllm.reasoning.deepseek_r1_reasoning_parser import DeepSeekR1ReasoningParser
 
 
 @ReasoningParserManager.register_module("super_v3")
 class SuperV3ReasoningParser(DeepSeekR1ReasoningParser):
+
+    def __init__(self, tokenizer, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+        self._promote_reasoning_to_content: bool = False
+        chat_template_kwargs = kwargs.get("chat_template_kwargs") or {}
+        if isinstance(chat_template_kwargs, dict):
+            self._promote_reasoning_to_content = (
+                chat_template_kwargs.get("enable_thinking") is False
+                or chat_template_kwargs.get("force_nonempty_content") is True
+            )
+
     def extract_reasoning(self, model_output, request):
         reasoning_content, final_content = super().extract_reasoning(
             model_output, request
@@ -25,4 +39,41 @@ class SuperV3ReasoningParser(DeepSeekR1ReasoningParser):
             # Put all nonempty content into the content, rather than return content
             reasoning_content, final_content = None, reasoning_content
 
-        return reasoning_content, final_content
+        return reasoning_content, final_content
+
+    def extract_reasoning_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+    ) -> DeltaMessage | None:
+        delta_message = super().extract_reasoning_streaming(
+            previous_text,
+            current_text,
+            delta_text,
+            previous_token_ids,
+            current_token_ids,
+            delta_token_ids,
+        )
+        if not self._promote_reasoning_to_content:
+            return delta_message
+        if delta_message is None or not delta_message.reasoning:
+            return delta_message
+        # Skip the <think> tag token itself — its text is the raw tag, not content.
+        if self.start_token_id in delta_token_ids and self.end_token_id not in delta_token_ids:
+            return delta_message
+        end_pos = current_text.find(self.end_token)
+        content_after_think = (
+            current_text[end_pos + len(self.end_token):]
+            if end_pos != -1 else ""
+        )
+        if content_after_think:
+            return delta_message
+        # Duplicate mode: keep reasoning channel, also populate content so
+        # clients that only read content get a non-null value on every delta.
+        if not delta_message.content:
+            delta_message.content = delta_message.reasoning
+        return delta_message