Fix stop boundaries for grammar-constrained speculative decoding (#31738)

python/sglang/srt/managers/schedule_batch.py

@@ -1489,11 +1489,6 @@ class Req(ReqDllmMixin):
             self.finished_len = self.sampling_params.max_new_tokens
             return
 
-        if self.grammar is not None:
-            if self.grammar.is_terminated():
-                self.finished_reason = FINISH_MATCHED_TOKEN(matched=self.output_ids[-1])
-                return
-
         new_accepted_tokens = self.output_ids[-new_accepted_len:]
 
         # Sanitize out-of-range / NaN token ids before any decode.
@@ -1509,6 +1504,10 @@ class Req(ReqDllmMixin):
         if self._check_token_based_finish(new_accepted_tokens):
             return
 
+        if self.grammar is not None and self.grammar.is_terminated():
+            self.finished_reason = FINISH_MATCHED_TOKEN(matched=self.output_ids[-1])
+            return
+
     def reset_for_retract(self):
         # Increment retraction count before resetting other state. We should not reset this
         # since we are tracking the total number of retractions for each request.