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

2024-11-17 21:29:30 -08:00
parent a9e90b4bce
commit 116685337e
3 changed files with 10 additions and 11 deletions
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/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,