diff --git a/PaddleNLP/pretrain_language_models/BERT/utils/init.py b/PaddleNLP/pretrain_language_models/BERT/utils/init.py
index b6d15f9b04db9c8fb10137737b617445b3c78230..091e5c34e156922d94844daa215b83917b059d49 100644
--- a/PaddleNLP/pretrain_language_models/BERT/utils/init.py
+++ b/PaddleNLP/pretrain_language_models/BERT/utils/init.py
@@ -30,7 +30,7 @@ def cast_fp32_to_fp16(exe, main_program):
             param_t = fluid.global_scope().find_var(param.name).get_tensor()
             data = np.array(param_t)
             if param.name.find("layer_norm") == -1:
-                param_t.set(np.float16(data).view(np.uint16), exe.place)
+                param_t.set(np.float16(data), exe.place)
             master_param_var = fluid.global_scope().find_var(param.name +
                                                              ".master")
             if master_param_var is not None:
diff --git a/PaddleNLP/pretrain_language_models/XLNet/utils/init.py b/PaddleNLP/pretrain_language_models/XLNet/utils/init.py
index d495a42edffe644389d393b657d2c2e64bb5df0d..61e86a5d9f9552972ca1c83c0ed80cc8d32d37a1 100644
--- a/PaddleNLP/pretrain_language_models/XLNet/utils/init.py
+++ b/PaddleNLP/pretrain_language_models/XLNet/utils/init.py
@@ -32,7 +32,7 @@ def cast_fp32_to_fp16(exe, main_program):
             if param.name.find("layer_norm") == -1 and param.name.find(
                     "embedding") == -1:
                 print("shkip params", param.name)
-                param_t.set(np.float16(data).view(np.uint16), exe.place)
+                param_t.set(np.float16(data), exe.place)
             master_param_var = fluid.global_scope().find_var(param.name +
                                                              ".master")
             if master_param_var is not None:
diff --git a/dygraph/bert/utils/init.py b/dygraph/bert/utils/init.py
index 6b69d87e7d3a877cccedd7fd5f42788279ef1470..e0963eef8c815c30ea741a67fc61eda076fa7870 100644
--- a/dygraph/bert/utils/init.py
+++ b/dygraph/bert/utils/init.py
@@ -30,7 +30,7 @@ def cast_fp32_to_fp16(exe, main_program):
             param_t = fluid.global_scope().find_var(param.name).get_tensor()
             data = np.array(param_t)
             if param.name.find("layer_norm") == -1:
-                param_t.set(np.float16(data).view(np.uint16), exe.place)
+                param_t.set(np.float16(data), exe.place)
             master_param_var = fluid.global_scope().find_var(param.name +
                                                              ".master")
             if master_param_var is not None: