fix cuda mem in sharding parallel (#55653)

bbdc1683 · pangengzheng · GitHub · 07e788f1 · bbdc1683
隐藏空白更改
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
@@ -187,10 +187,11 @@ class HybridCommunicateGroup:
            "data"
        )

-        (
-            self.sharding_check_group,
-            self.sharding_check_comm_group,
-        ) = self._set_check_group("sharding")
+        if self._sharding_degree > 1:
+            (
+                self.sharding_check_group,
+                self.sharding_check_comm_group,
+            ) = self._set_check_group("sharding")

        # create p2p group
        self.is_first_stage = self.stage_id == 0