[Bug][Structured Outputs] Fix bug that leads to unconstrained generations with structural tags (#42452)

Signed-off-by: rishitdholakia13 <rishit+github@cohere.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

tests/v1/structured_output/test_reasoning_structured_output.py

@@ -10,6 +10,7 @@ import pytest
 from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
 from vllm.v1.request import Request
 from vllm.v1.structured_output import StructuredOutputManager
+from vllm.v1.structured_output.backend_types import StructuredOutputOptions
 
 
 class MockReasoner:
@@ -214,6 +215,32 @@ class TestReasoningStructuredOutput:
         )
         assert result is False
 
+    def test_should_advance_reasoning_just_ended_with_spec_decode_structural_tag(
+        self,
+        manager_with_reasoner,
+        mock_request_with_structured_output,
+    ):
+        """When reasoning ends this step, advance immediately for structural
+        tags with speculative decoding."""
+        structured_req = mock_request_with_structured_output.structured_output_request
+        structured_req.reasoning_ended = False
+        structured_req.structured_output_key = (
+            StructuredOutputOptions.STRUCTURAL_TAG,
+            "{}",
+        )
+        reasoner = MockReasoner(tokenizer=Mock())
+        reasoner.is_reasoning_end_streaming.return_value = True
+        structured_req.reasoner = reasoner
+
+        manager_with_reasoner.vllm_config.speculative_config = Mock()
+
+        result = manager_with_reasoner.should_advance(
+            mock_request_with_structured_output
+        )
+
+        assert structured_req.reasoning_ended is True
+        assert result is True
+
     def test_should_advance_reasoning_already_ended(
         self,
         manager_with_reasoner,

vllm/v1/structured_output/__init__.py

@@ -15,6 +15,7 @@ from vllm.v1.structured_output.backend_guidance import GuidanceBackend
 from vllm.v1.structured_output.backend_types import (
     StructuredOutputBackend,
     StructuredOutputGrammar,
+    StructuredOutputOptions,
 )
 from vllm.v1.structured_output.backend_xgrammar import XgrammarBackend
 
@@ -350,10 +351,22 @@ class StructuredOutputManager:
         if reasoner.is_reasoning_end_streaming(
             all_token_ids, itertools.islice(all_token_ids, start, None)
         ):
-            # Reasoning just ended, so we shouldn't advance til
-            # next pass
             structured_req.reasoning_ended = True
 
+            # Reasoning just ended this step. Defer FSM advance until the next
+            # pass (see reasoning_ended check above) for JSON/regex/choice/grammar:
+            # advancing on the closing boundary token can accept tokens that still
+            # belong to the reasoning stream. Structural tags are the only safe
+            # same-step exception: they model phased output (e.g. thinking tag ->
+            # answer tag), and speculative decoding must run grammar.validate_tokens
+            # on draft tokens produced immediately after that transition.
+            if (
+                self.vllm_config.speculative_config is not None
+                and structured_req.structured_output_key[0]
+                == StructuredOutputOptions.STRUCTURAL_TAG
+            ):
+                return True
+
         return False
 
     def clear_backend(self) -> None: