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

2025-07-27 19:01:02 -07:00
parent bb81daefb8
commit dd487e5553
3 changed files with 21 additions and 8 deletions
--- a/python/sglang/srt/managers/scheduler.py
+++ b/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