diff --git a/PaddleNLP/paddlenlp/seq2vec/encoder.py b/PaddleNLP/paddlenlp/seq2vec/encoder.py
index 728d4e97760fdf8bb189f8a8fb1160f7e79ac955..3b8dc67e331932d31c27bfdb67b67d3adf416b7d 100644
--- a/PaddleNLP/paddlenlp/seq2vec/encoder.py
+++ b/PaddleNLP/paddlenlp/seq2vec/encoder.py
@@ -129,13 +129,13 @@ class CNNEncoder(nn.Layer):
         self._activation = conv_layer_activation
         self._output_dim = output_dim
 
-        self.convs = [
+        self.convs = paddle.nn.LayerList([
             nn.Conv2D(
                 in_channels=1,
                 out_channels=self._num_filter,
                 kernel_size=(i, self._emb_dim),
                 **kwargs) for i in self._ngram_filter_sizes
-        ]
+        ])
 
         maxpool_output_dim = self._num_filter * len(self._ngram_filter_sizes)
         if self._output_dim:
diff --git a/PaddleNLP/paddlenlp/transformers/roberta/modeling.py b/PaddleNLP/paddlenlp/transformers/roberta/modeling.py
index fc797cfee5ff2ddca44257b503906c769cdf9791..d14296568887ab50cb5ad9e81b1334554491fa88 100644
--- a/PaddleNLP/paddlenlp/transformers/roberta/modeling.py
+++ b/PaddleNLP/paddlenlp/transformers/roberta/modeling.py
@@ -50,8 +50,10 @@ class RobertaEmbeddings(nn.Layer):
     def forward(self, input_ids, token_type_ids=None, position_ids=None):
         if position_ids is None:
             # maybe need use shape op to unify static graph and dynamic graph
-            seq_length = input_ids.shape[1]
-            position_ids = paddle.arange(0, seq_length, dtype="int64")
+            ones = paddle.ones_like(input_ids, dtype="int64")
+            seq_length = paddle.cumsum(ones, axis=1)
+            position_ids = seq_length - ones
+            position_ids.stop_gradient = True
         if token_type_ids is None:
             token_type_ids = paddle.zeros_like(input_ids, dtype="int64")