bugfix: Fix XGrammar backend to use model's EOS tokens for constrained generation (#8422)

python/sglang/srt/constrained/xgrammar_backend.py

@@ -150,14 +150,16 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
         self,
         tokenizer,
         vocab_size: int,
+        model_eos_token_ids: Optional[List[int]] = None,
     ):
         super().__init__()
 
-        if True:
-            tokenizer_info = TokenizerInfo.from_huggingface(
-                tokenizer, vocab_size=vocab_size
-            )
-            override_stop_tokens = None
+        # Create TokenizerInfo with model's EOS tokens as the authoritative stop tokens
+        # This ensures consistency between what the model considers EOS and what XGrammar uses
+        tokenizer_info = TokenizerInfo.from_huggingface(
+            tokenizer, vocab_size=vocab_size, stop_token_ids=model_eos_token_ids
+        )
+        override_stop_tokens = None
 
         self.grammar_compiler = GrammarCompiler(tokenizer_info=tokenizer_info)
         self.vocab_size = vocab_size