del wait in sharding for npu (#34637)

ce733495 · Baibaifan · GitHub · c91b1e03 · ce733495
隐藏空白更改
内联并排

Showing with 3 addition and 2 deletion

python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py ...e/distributed/fleet/meta_optimizers/sharding_optimizer.py +3 -2

未找到文件。
--- a/python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py
+++ b/python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py
@@ -379,8 +379,9 @@ class ShardingOptimizer(MetaOptimizerBase):
                  'w') as f:
            f.writelines(str(main_block.program))
-        # GPU and NPU need to wait server ready
+        # GPU need to wait server ready, GPU and NPU is Layered connection
-        self._wait()
+        if not core.is_compiled_with_npu():
+            self._wait()
        return optimize_ops, params_grads
    def _init_pair_comm(self, pair, ring_id):