diff --git a/PaddleNLP/examples/dialogue/dgu/main.py b/PaddleNLP/examples/dialogue/dgu/main.py
index b7a12664550b60990974993bb0477ad805394d78..352f0ef0f6931a6584b1f280667a9ef993ee12da 100644
--- a/PaddleNLP/examples/dialogue/dgu/main.py
+++ b/PaddleNLP/examples/dialogue/dgu/main.py
@@ -128,22 +128,13 @@ def train(args, model, train_data_loader, dev_data_loader, metric, rank):
         max_train_steps=max_train_steps)
     lr_scheduler = LambdaDecay(args.learning_rate, factor_fn)
     optimizer = AdamW(
-        learning_rate=lr_scheduler,
-        parameters=model.parameters(),
-        weight_decay=args.weight_decay,
-        apply_decay_param_fun=lambda x: x in [
-            params.name for params in model.parameters()
-            if not any(nd in params.name for nd in ['bias', 'norm'])],
-        grad_clip=nn.ClipGradByGlobalNorm(args.max_grad_norm)
-    )
-    optimizer = paddle.optimizer.AdamW(
         learning_rate=lr_scheduler,
         parameters=model.parameters(),
         weight_decay=args.weight_decay,
         apply_decay_param_fun=lambda x: x in [
             p.name for n, p in model.named_parameters()
-            if not any(nd in n for nd in ["bias", "norm"])
-        ])
+            if not any(nd in n for nd in ["bias", "norm"])],
+        grad_clip=nn.ClipGradByGlobalNorm(args.max_grad_norm))
     loss_fn = DGULossFunction(args.task_name)
 
     load_ckpt(args, model, optimizer)