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

python/sglang/srt/constrained/base_grammar_backend.py

@@ -168,7 +168,10 @@ class BaseGrammarBackend:
 
 
 def create_grammar_backend(
-    server_args: ServerArgs, tokenizer, vocab_size: int
+    server_args: ServerArgs,
+    tokenizer,
+    vocab_size: int,
+    eos_token_ids: Optional[set] = None,
 ) -> Optional[BaseGrammarBackend]:
     if server_args.grammar_backend == "outlines":
         from sglang.srt.constrained.outlines_backend import OutlinesGrammarBackend
@@ -180,7 +183,12 @@ def create_grammar_backend(
     elif server_args.grammar_backend == "xgrammar":
         from sglang.srt.constrained.xgrammar_backend import XGrammarGrammarBackend
 
-        grammar_backend = XGrammarGrammarBackend(tokenizer, vocab_size=vocab_size)
+        # Convert Set[int] to List[int] if needed
+        eos_list = list(eos_token_ids) if eos_token_ids else None
+
+        grammar_backend = XGrammarGrammarBackend(
+            tokenizer, vocab_size=vocab_size, model_eos_token_ids=eos_list
+        )
     elif server_args.grammar_backend == "llguidance":
         from sglang.srt.constrained.llguidance_backend import GuidanceBackend
 

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

python/sglang/srt/managers/scheduler.py

@@ -458,7 +458,10 @@ class Scheduler(
         self.grammar_queue: List[Req] = []
         if not server_args.skip_tokenizer_init:
             self.grammar_backend = create_grammar_backend(
-                server_args, self.tokenizer, self.model_config.vocab_size
+                server_args,
+                self.tokenizer,
+                self.model_config.vocab_size,
+                self.model_config.hf_eos_token_id,
             )
         else:
             self.grammar_backend = None