Fix cuda illegal memory access in overlap mode (#2070)

python/sglang/srt/managers/schedule_batch.py

@@ -1055,9 +1055,6 @@ class ScheduleBatch:
         )
 
     def copy(self):
-        # We need a stream synchronization here. Otherwise, there will be cuda illegal memory access errors.
-        _ = self.seq_lens[0].item()
-
         # Only contain fields that will be used by process_batch_result
         return ScheduleBatch(
             reqs=self.reqs,

python/sglang/srt/managers/scheduler.py

@@ -390,6 +390,9 @@ class Scheduler:
             batch = self.get_next_batch_to_run()
             self.cur_batch = batch
             if batch:
+                # We need a stream synchronization here. Otherwise, there will be cuda illegal memory access errors.
+                _ = batch.seq_lens[0].item()
+
                 result = self.run_batch(batch)
                 result_queue.append((batch.copy(), result))