fix: prevent stale bitmask leakage in LLGuidance grammar backend (#23653)

Signed-off-by: jellysnack <oleg.jellysnack@gmail.com>
Co-authored-by: YorkSu <45256779+YorkSu@users.noreply.github.com>
Co-authored-by: jellysnack <oleg.jellysnack@gmail.com>

python/sglang/srt/constrained/llguidance_backend.py

@@ -51,7 +51,6 @@ class GuidanceGrammar(BaseGrammarObject):
         )
         self._check_err()
 
-        self.bitmask = None
         self.eos_token = self.llguidance_tokenizer.eos_token
 
     def accept_token(self, token: int):
@@ -83,16 +82,7 @@ class GuidanceGrammar(BaseGrammarObject):
     def allocate_vocab_mask(
         self, vocab_size: int, batch_size: int, device
     ) -> torch.Tensor:
-        if self.bitmask is None or self.bitmask.shape[0] < batch_size:
-            # only create bitmask when batch gets larger
-            self.bitmask = allocate_token_bitmask(
-                batch_size, self.llguidance_tokenizer.vocab_size
-            )
-            bitmask = self.bitmask
-        else:
-            bitmask = self.bitmask[:batch_size]
-
-        return bitmask
+        return allocate_token_bitmask(batch_size, self.llguidance_tokenizer.vocab_size)
 
     @staticmethod
     def move_vocab_mask(vocab_mask: torch.Tensor, device) -> torch.Tensor: