diff --git a/demo/ofa/bert/README.md b/demo/ofa/bert/README.md
index 83054a4c01bf78c5ff089d469823bf1920bc0d31..b4f39186ed656f294132fd85923fbf633cb31c18 100644
--- a/demo/ofa/bert/README.md
+++ b/demo/ofa/bert/README.md
@@ -1,6 +1,7 @@
 # OFA压缩PaddleNLP-BERT模型
 
-BERT-base模型是一个迁移能力很强的通用语义表示模型，但是模型中也有一些参数冗余。本教程将介绍如何使用PaddleSlim对[PaddleNLP](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/)中BERT-base模型进行压缩。
+BERT-base模型是一个迁移能力很强的通用语义表示模型，但是模型中也有一些参数冗余。本教程将介绍如何使用PaddleSlim对[PaddleNLP](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/)中BERT-base模型进行压缩。  
+本教程只会演示如何快速启动相应训练，详细教程请参考: [BERT](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0.0/docs/zh_cn/nlp/paddlenlp_slim_ofa_tutorial.md)
 
 ## 1. 压缩结果
 
@@ -82,7 +83,7 @@ BERT-base模型是一个迁移能力很强的通用语义表示模型，但是
                         <td rowspan=2 align=center> V100 </td>
                         <td rowspan=2 align=center> 16 </td>
                         <td style="text-align:center">
-                                <span style="font-size:18px;">BERT</span>
+                                <span style="font-size:18px;" align=center>BERT</span>
                         </td>
                         <td style="text-align:center">
                                 <span style="font-size:18px">N</span>
@@ -106,7 +107,7 @@ BERT-base模型是一个迁移能力很强的通用语义表示模型，但是
                         <td rowspan=2 align=center> Intel(R) Xeon(R) Gold 5117 CPU @ 2.00GHz </td>
                         <td rowspan=2 align=center> 16 </td>
                         <td style="text-align:center">
-                                <span style="font-size:18px;">BERT</span>
+                                <span style="font-size:18px;" align=center>BERT</span>
                         </td>
                         <td style="text-align:center">
                                 <span style="font-size:18px">N</span>
@@ -184,7 +185,7 @@ python -u ./run_glue_ofa.py --model_type bert \
 压缩训练之后在dev上的结果如表1-1中『Result with PaddleSlim』列所示，延时情况如表1-2所示。
 
 ## 3. OFA接口介绍
-OFA API介绍参考[API](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/ofa_api.rst)
+OFA API介绍参考[API](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0.0/docs/zh_cn/api_cn/dygraph/ofa/ofa_api.rst)
 
 # 基于本代码对TinyBERT(L=4, D=312)进行压缩
 下游任务模型是从TinyBERT官方repo转换得到。
diff --git a/demo/ofa/ernie/README.md b/demo/ofa/ernie/README.md
index 6143bfc6b921fa11d307eb6ed1bd0be5d8bfd1cb..32212b30c373c09b778d3743d467df243e5f09f9 100644
--- a/demo/ofa/ernie/README.md
+++ b/demo/ofa/ernie/README.md
@@ -2,6 +2,8 @@
 
 ERNIE是百度开创性提出的基于知识增强的持续学习语义理解框架，该框架将大数据预训练与多源丰富知识相结合，通过持续学习技术，不断吸收海量文本数据中词汇、结构、语义等方面的知识，实现模型效果不断进化。本教程讲介绍如何使用PaddleSlim对[ERNIE](https://github.com/PaddlePaddle/ERNIE)模型进行压缩。
 
+本教程只会演示如何快速启动相应训练，详细教程请参考：[ERNIE](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0.0/docs/zh_cn/nlp/ernie_slim_ofa_tutorial.md)
+
 使用本教程压缩算法可以在精度无损的情况下，对原始Tiny-ERNIE模型进行40%的加速。
 
 ## 1. 快速开始
@@ -41,4 +43,4 @@ python ./ofa_ernie.py \
 - `depth_mult_list` 表示压缩训练过程中，模型包含的Transformer Block数量的选择的范围。
 
 ## 2. OFA接口介绍
-TODO
+OFA API介绍参考[API](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0.0/docs/zh_cn/api_cn/dygraph/ofa/ofa_api.rst)
diff --git a/demo/ofa/ernie/ernie_supernet/importance.py b/demo/ofa/ernie/ernie_supernet/importance.py
index c27f89dd70cb681b30fb154925a4d60038b7241f..061f41b86df9be227745b97c7d0fe6558f76996e 100644
--- a/demo/ofa/ernie/ernie_supernet/importance.py
+++ b/demo/ofa/ernie/ernie_supernet/importance.py
@@ -52,12 +52,13 @@ def compute_neuron_head_importance(args, model, dev_ds, place, model_cfg):
     for eval_task in eval_task_names:
         for batch in dev_ds.start(place):
             ids, sids, label = batch
-            loss, _, _ = model(
+            out = model(
                 ids,
                 sids,
                 labels=label,
                 head_mask=head_mask,
                 num_layers=model_cfg['num_hidden_layers'])
+            loss = out[0]
             loss.backward()
             head_importance += L.abs(FD.to_variable(head_mask.gradient()))
 
diff --git a/paddleslim/nas/ofa/get_sub_model.py b/paddleslim/nas/ofa/get_sub_model.py
index 1b93e694703e4a7c87775303397f7ce6c7f6b5e4..a7c4a6c816a0c2276d9e496b8f202fce43db1da1 100644
--- a/paddleslim/nas/ofa/get_sub_model.py
+++ b/paddleslim/nas/ofa/get_sub_model.py
@@ -14,6 +14,7 @@
 
 import numpy as np
 import paddle
+from paddle.fluid import core
 
 __all__ = ['get_prune_params_config', 'prune_params']
 
@@ -96,11 +97,11 @@ def prune_params(model, param_config, super_model_sd=None):
 
         p = t_value._place()
         if p.is_cpu_place():
-            place = paddle.CPUPlace()
+            place = core.CPUPlace()
         elif p.is_cuda_pinned_place():
-            place = paddle.CUDAPinnedPlace()
+            place = core.CUDAPinnedPlace()
         else:
-            place = paddle.CUDAPlace(p.gpu_device_id())
+            place = core.CUDAPlace(p.gpu_device_id())
         t_value.set(prune_value, place)
         if param.trainable:
             param.clear_gradient()
diff --git a/paddleslim/nas/ofa/layers_old.py b/paddleslim/nas/ofa/layers_old.py
index 25dcd1525b8439d07fdcb70601b560a41b7de2fb..9b4058c182eec7a229772926f2e633ed11b70b26 100644
--- a/paddleslim/nas/ofa/layers_old.py
+++ b/paddleslim/nas/ofa/layers_old.py
@@ -931,10 +931,10 @@ class SuperBatchNorm(fluid.dygraph.BatchNorm):
                  "use_mkldnn", False, "fuse_with_relu", self._fuse_with_relu,
                  "use_global_stats", self._use_global_stats,
                  'trainable_statistics', self._trainable_statistics)
-        batch_norm_out, _, _, _, _, _ = core.ops.batch_norm(
+        batch_norm_out = core.ops.batch_norm(
             input, weight, bias, mean, variance, mean_out, variance_out, *attrs)
         return dygraph_utils._append_activation_in_dygraph(
-            batch_norm_out, act=self._act)
+            batch_norm_out[0], act=self._act)
 
 
 class SuperInstanceNorm(fluid.dygraph.InstanceNorm):