fix(schedule_batch): trim stop string when EOS matches in the same step (#28802)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

python/sglang/srt/managers/schedule_batch.py

@@ -1423,15 +1423,19 @@ class Req(ReqDllmMixin):
 
         new_accepted_tokens = self.output_ids[-new_accepted_len:]
 
-        if self._check_token_based_finish(new_accepted_tokens):
-            return
-
+        # Sanitize out-of-range / NaN token ids before any decode.
         if self._check_vocab_boundary_finish(new_accepted_tokens):
             return
 
+        # Stop string beats EOS/stop-token matched in the same step (speculative
+        # decoding can accept >1 token): token-based would trim only the last
+        # token and leak the stop string.
         if self._check_str_based_finish(new_accepted_len):
             return
 
+        if self._check_token_based_finish(new_accepted_tokens):
+            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.