fix(vllm): tell the reasoning parser whether thinking was enabled

backend/python/vllm/backend.py

@@ -558,6 +558,73 @@ def _build_sampling_params(self, request):
 
         return sampling_params
 
+    def _new_reasoning_parser(self, chat_template_kwargs):
+        """Build the reasoning parser, telling it whether thinking is on.
+
+        vLLM's newer parser engines decide their *initial state* from
+        ``chat_template_kwargs``: ``Qwen3Parser`` reads
+        ``chat_template_kwargs["enable_thinking"]`` and defaults to ``True``,
+        starting in the REASONING state. Constructed without it, a completion
+        produced with thinking disabled is classified as reasoning end to end,
+        and the answer is reported in both ``reasoning_content`` and
+        ``content``.
+
+        vLLM's own OpenAI server forwards the request's chat template kwargs
+        here; this backend renders the template itself, so it forwards the
+        same dict. Older parsers do not accept the argument — fall back to the
+        plain constructor for those.
+        """
+        try:
+            return self.reasoning_parser_cls(
+                self.tokenizer, chat_template_kwargs=chat_template_kwargs or {},
+            )
+        except TypeError:
+            return self.reasoning_parser_cls(self.tokenizer)
+
+    @staticmethod
+    def _split_reasoning(rp, generated_text, prompt, reasoning, content):
+        """Decide what the reasoning parser's output actually means.
+
+        Covers the *older* parser shape, which has no initial state to set:
+        ``BaseThinkingReasoningParser.extract_reasoning`` documents its own
+        fallback — "For models that may not generate start token, assume the
+        reasoning content is always at the start." When no end token is
+        present it returns *everything* as reasoning and ``None`` as content,
+        which is right for a truncated reasoning run and wrong for a
+        completion that never contained reasoning at all.
+
+        Taking ``None`` content to mean "keep the raw text" then duplicates
+        the answer into both fields.
+
+        The prompt says which case it is. A template with thinking on leaves
+        the reasoning block open (the prompt ends with the start token); with
+        thinking off it closes the block in the prompt, so the completion is
+        plain content. Parsers that expose no token pair (the engine-based
+        adapters, which take the ``chat_template_kwargs`` route above) keep
+        the parser's verdict unchanged.
+        """
+        start = getattr(rp, "start_token", None)
+        end = getattr(rp, "end_token", None)
+
+        if end and end in generated_text:
+            # The parser split on the end token. Empty content here means the
+            # model stopped right after it, not that parsing failed.
+            return reasoning or "", content or ""
+
+        if not start:
+            # Unknown token layout — keep the previous behaviour rather than
+            # guess.
+            return reasoning or "", content if content is not None else generated_text
+
+        if not (start in generated_text or (prompt or "").rstrip().endswith(start)):
+            # No end token and the block was never open: the "reasoning starts
+            # at the beginning" fallback does not apply to this completion.
+            return "", generated_text
+
+        # Block was open and the end token never arrived — reasoning ran out of
+        # budget. It is all reasoning, and there is no answer to report.
+        return reasoning or "", content or ""
+
     async def _predict(self, request, context, streaming=False):
         # Build the sampling parameters
         sampling_params = self._build_sampling_params(request)
@@ -572,6 +639,9 @@ async def _predict(self, request, context, streaming=False):
 
         # Extract image paths and process images
         prompt = request.Prompt
+        # Kept in scope: the reasoning parser needs to know which chat
+        # template kwargs produced this prompt.
+        template_kwargs = {}
 
         image_paths = request.Images
         image_data = [self.load_image(img_path) for img_path in image_paths]
@@ -582,7 +652,7 @@ async def _predict(self, request, context, streaming=False):
         # If tokenizer template is enabled and messages are provided instead of prompt, apply the tokenizer template
         if not request.Prompt and request.UseTokenizerTemplate and request.Messages:
             messages_dicts = self._messages_to_dicts(request.Messages)
-            template_kwargs = {"tokenize": False, "add_generation_prompt": True}
+            template_kwargs.update({"tokenize": False, "add_generation_prompt": True})
 
             # Pass tools for tool calling
             if request.Tools:
@@ -757,10 +827,11 @@ async def _predict(self, request, context, streaming=False):
 
         if self.reasoning_parser_cls:
             try:
-                rp = self.reasoning_parser_cls(self.tokenizer)
+                rp = self._new_reasoning_parser(template_kwargs)
                 r, c = rp.extract_reasoning(generated_text, request=None)
-                reasoning_content = r or ""
-                content = c if c is not None else generated_text
+                reasoning_content, content = self._split_reasoning(
+                    rp, generated_text, prompt, r, c,
+                )
             except Exception as e:
                 print(f"Reasoning parser error: {e}", file=sys.stderr)
 

backend/python/vllm/test.py

@@ -549,3 +549,116 @@ def test_no_tool_parser_unchanged_per_delta_stream(self):
             intermediate, ["Hello ", "world", "!"],
             f"plain streaming changed; got {intermediate!r}",
         )
+
+
+class TestReasoningSplit(unittest.TestCase):
+    """Server-less tests for BackendServicer._split_reasoning.
+
+    vLLM's BaseThinkingReasoningParser returns the whole completion as
+    reasoning and None as content whenever the end token is missing. Taken
+    literally that duplicates a thinking-disabled answer into both fields.
+    """
+
+    class _Parser:
+        start_token = "<think>"
+        end_token = "</think>"
+
+    def _split(self, generated, prompt, reasoning, content):
+        import sys, os
+        sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+        from backend import BackendServicer
+        return BackendServicer._split_reasoning(
+            self._Parser(), generated, prompt, reasoning, content,
+        )
+
+    def test_thinking_off_is_not_duplicated_into_reasoning(self):
+        """No tags anywhere: the answer is content, and only content."""
+        r, c = self._split(
+            "391", "user: 17*23?\n<think>\n\n</think>\n\n",
+            reasoning="391", content=None,
+        )
+        self.assertEqual(r, "")
+        self.assertEqual(c, "391")
+
+    def test_prefilled_start_tag_keeps_truncated_reasoning(self):
+        """Prompt left the block open and the end token never arrived
+        (budget exhausted): that really is all reasoning."""
+        r, c = self._split(
+            "thinking and thinking", "user: hi\n<think>\n",
+            reasoning="thinking and thinking", content=None,
+        )
+        self.assertEqual(r, "thinking and thinking")
+        self.assertEqual(c, "")
+
+    def test_end_token_present_keeps_parser_split(self):
+        r, c = self._split(
+            "adding two and two</think>4", "user: hi\n<think>\n",
+            reasoning="adding two and two", content="4",
+        )
+        self.assertEqual(r, "adding two and two")
+        self.assertEqual(c, "4")
+
+    def test_stop_right_after_end_token_yields_empty_content(self):
+        """Content must not fall back to the raw text — that would put the
+        reasoning into the answer."""
+        r, c = self._split(
+            "reasoned</think>", "user: hi\n<think>\n",
+            reasoning="reasoned", content=None,
+        )
+        self.assertEqual(r, "reasoned")
+        self.assertEqual(c, "")
+
+    def test_unknown_token_layout_keeps_previous_behaviour(self):
+        class _Bare:
+            pass
+        import sys, os
+        sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+        from backend import BackendServicer
+        r, c = BackendServicer._split_reasoning(
+            _Bare(), "raw", "prompt", "raw", None,
+        )
+        self.assertEqual(r, "raw")
+        self.assertEqual(c, "raw")
+
+
+class TestReasoningParserConstruction(unittest.TestCase):
+    """The parser must learn whether thinking was on for this request.
+
+    vLLM's engine-based parsers (Qwen3Parser and friends) read
+    chat_template_kwargs["enable_thinking"] and default to True, so a parser
+    built without it treats a thinking-disabled completion as pure reasoning.
+    """
+
+    def _servicer(self):
+        import sys, os
+        sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+        from backend import BackendServicer
+        s = BackendServicer()
+        s.tokenizer = object()
+        return s
+
+    def test_chat_template_kwargs_are_forwarded(self):
+        seen = {}
+
+        class _Parser:
+            def __init__(self, tokenizer, **kwargs):
+                seen.update(kwargs)
+
+        s = self._servicer()
+        s.reasoning_parser_cls = _Parser
+        s._new_reasoning_parser({"enable_thinking": False})
+        self.assertEqual(
+            seen.get("chat_template_kwargs"), {"enable_thinking": False},
+        )
+
+    def test_parser_without_the_kwarg_still_builds(self):
+        """Older parsers take only the tokenizer — must not break them."""
+        class _Old:
+            def __init__(self, tokenizer):
+                self.tokenizer = tokenizer
+
+        s = self._servicer()
+        s.reasoning_parser_cls = _Old
+        self.assertIsInstance(
+            s._new_reasoning_parser({"enable_thinking": False}), _Old,
+        )