fix the combination bug of sharding stage1 + dp (#46631)

6e4cba14 · wuhuachaocoding · GitHub · 6512e087 · 6e4cba14
隐藏空白更改
内联并排

Showing with 2 addition and 1 deletion

python/paddle/distributed/fleet/model.py python/paddle/distributed/fleet/model.py +2 -1

未找到文件。
--- a/python/paddle/distributed/fleet/model.py
+++ b/python/paddle/distributed/fleet/model.py
@@ -139,7 +139,8 @@ def distributed_model(model):
            model,
            comm_buffer_size=strategy.fuse_grad_size_in_MB,
            last_comm_buffer_size=strategy.last_comm_group_size_MB,
-            find_unused_parameters=strategy.find_unused_parameters)
+            find_unused_parameters=strategy.find_unused_parameters,
+            group=fleet_env._hcg.get_data_parallel_group())
    elif fleet_env._hcg.get_parallel_mode() == ParallelMode.TENSOR_PARALLEL:
        model = TensorParallel(model, fleet_env._hcg, strategy=strategy)
    elif fleet_env._hcg.get_parallel_mode() == ParallelMode.PIPELINE_PARALLEL: