[#19229][fix] Discard a complete think tag in NemotronV3 reasoning finish()

tensorrt_llm/llmapi/reasoning_parser.py

@@ -634,10 +634,15 @@ def finish(self) -> ReasoningParserResult:
         as reasoning_content since we are still in reasoning mode.
 
         If the closing tag was already found (or reasoning was never
-        entered), flushes any remaining buffer as content."""
+        entered), flushes any remaining buffer as content.
+
+        A buffer holding exactly a complete tag is a delimiter rather than
+        model output, so it is discarded."""
+        remaining = self._buffer
+        self._buffer = ""
+        if remaining in (self.reasoning_start, self.reasoning_end):
+            remaining = ""
         if self.in_reasoning and not self._found_closing_tag:
-            remaining = self._buffer
-            self._buffer = ""
             if self._force_nonempty_content:
                 all_content = self._accumulated_reasoning + remaining
                 self._accumulated_reasoning = ""
@@ -648,8 +653,6 @@ def finish(self) -> ReasoningParserResult:
             if remaining:
                 return ReasoningParserResult(reasoning_content=remaining)
             return ReasoningParserResult()
-        remaining = self._buffer
-        self._buffer = ""
         if remaining:
             return ReasoningParserResult(content=remaining)
         return ReasoningParserResult()

tests/unittest/llmapi/test_reasoning_parser.py

@@ -795,6 +795,15 @@ def test_nano_v3_reasoning_parser_stream(delta_texts: list, content: list,
         (["reasoning", "<tool_call>data"], "", "", {
             "force_nonempty_content": True
         }),
+        # A stream that ends on a complete tag: the tag is a delimiter,
+        # not model output.
+        (["a", R1_END], "", "", None),
+        (["a", R1_END], "a", "", {
+            "force_nonempty_content": True
+        }),
+        ([R1_START], "", "", {
+            "enable_thinking": False
+        }),
     ])
 def test_nano_v3_reasoning_parser_finish(delta_texts: list, finish_content: str,
                                          finish_reasoning: str,