fix cuda mem in sharding parallel (#55653)

bbdc1683 · pangengzheng · GitHub · 07e788f1 · bbdc1683
显示空白变更内容
内联并排

Showing with 5 addition and 4 deletion

python/paddle/distributed/fleet/base/topology.py python/paddle/distributed/fleet/base/topology.py +5 -4

未找到文件。
--- a/python/paddle/distributed/fleet/base/topology.py
+++ b/python/paddle/distributed/fleet/base/topology.py
@@ -187,6 +187,7 @@ class HybridCommunicateGroup:
            "data"
        )

+        if self._sharding_degree > 1:
            (
                self.sharding_check_group,
                self.sharding_check_comm_group,