[Fix] Simple fix: pass tokenizer stop tokens to XGrammar so constrained requests can terminate

python/sglang/srt/constrained/base_grammar_backend.py

@@ -376,8 +376,21 @@ def create_grammar_backend(
             XGrammarGrammarBackend,
         )
 
-        # Convert Set[int] to List[int] if needed
-        eos_list = list(eos_token_ids) if eos_token_ids else None
+        # XGrammar only ever unmasks stop tokens from this list, so it must
+        # match the EOS set that Req._check_token_based_finish accepts:
+        # model config EOS + tokenizer EOS + additional stop tokens. Any EOS
+        # missing here can never be emitted under a constraint, and the
+        # request runs until max_new_tokens.
+        eos_set = set(eos_token_ids) if eos_token_ids else set()
+        tokenizer_eos_token_id = getattr(tokenizer, "eos_token_id", None)
+        if isinstance(tokenizer_eos_token_id, int):
+            eos_set.add(tokenizer_eos_token_id)
+        additional_stop_token_ids = getattr(
+            tokenizer, "additional_stop_token_ids", None
+        )
+        if isinstance(additional_stop_token_ids, (list, tuple, set)):
+            eos_set.update(additional_stop_token_ids)
+        eos_list = sorted(eos_set) if eos_set else None
 
         try:
             grammar_backend = XGrammarGrammarBackend(

test/registered/unit/constrained/test_base_grammar_backend.py

@@ -339,6 +339,23 @@ def test_xgrammar_backend(self, mock_xgrammar_cls):
         )
         self.assertIs(result, mock_backend)
 
+    @patch("sglang.srt.constrained.xgrammar_backend.XGrammarGrammarBackend")
+    def test_xgrammar_stop_tokens_include_tokenizer_eos(self, mock_xgrammar_cls):
+        """Tokenizer-level stop tokens reach XGrammar even when the model
+        config's eos_token_id list omits them."""
+        mock_xgrammar_cls.return_value = MagicMock(spec=BaseGrammarBackend)
+        args = self._make_server_args("xgrammar")
+
+        class Tok:
+            eos_token_id = 151645
+            additional_stop_token_ids = [151646]
+
+        create_grammar_backend(args, Tok(), 32000, {151643})
+        self.assertEqual(
+            mock_xgrammar_cls.call_args.kwargs["model_eos_token_ids"],
+            [151643, 151645, 151646],
+        )
+
     @patch("sglang.srt.constrained.xgrammar_backend.XGrammarGrammarBackend")
     def test_xgrammar_unsupported_tokenizer_falls_back_to_none(self, mock_xgrammar_cls):
         from sglang.srt.constrained.xgrammar_backend import TokenizerNotSupportedError