diff --git a/plsc/core/grad_clip.py b/plsc/core/grad_clip.py
index bb2356bd35fa1..1668f307a514c 100644
--- a/plsc/core/grad_clip.py
+++ b/plsc/core/grad_clip.py
@@ -14,6 +14,7 @@
 
 import warnings
 import paddle
+from paddle import _legacy_C_ops as _C_ops
 
 
 def _squared_l2_norm(x):
@@ -22,7 +23,7 @@ def _squared_l2_norm(x):
         sum_square = paddle.sum(square)
         return sum_square
 
-    return paddle._C_ops.squared_l2_norm(x)
+    return _C_ops.squared_l2_norm(x)
 
 
 class ClipGradByGlobalNorm(object):
diff --git a/plsc/core/grad_scaler.py b/plsc/core/grad_scaler.py
index a56e8712d0378..4ba1527665a5f 100644
--- a/plsc/core/grad_scaler.py
+++ b/plsc/core/grad_scaler.py
@@ -15,7 +15,7 @@
 from collections import defaultdict
 from paddle.amp import GradScaler as FrameworkGradScaler
 from paddle.fluid.dygraph.amp import OptimizerState
-from paddle import _C_ops
+from paddle import _legacy_C_ops as _C_ops
 import paddle
 
 
diff --git a/plsc/optimizer/adamw.py b/plsc/optimizer/adamw.py
index 03a26ca813e0b..7ff4614dfbc28 100644
--- a/plsc/optimizer/adamw.py
+++ b/plsc/optimizer/adamw.py
@@ -18,6 +18,7 @@
 
 import math
 import paddle
+from paddle import _legacy_C_ops as _C_ops
 from .optimizer import Optimizer
 from plsc.utils import logger
 
@@ -100,7 +101,7 @@ def step(self):
                         paddle.float16, paddle.bfloat16
                 }:
                     master_param = state['master_param']
-                _, _, _, _, _, _ = paddle._C_ops.adamw(
+                _, _, _, _, _, _ = _C_ops.adamw(
                     p, grad,
                     paddle.to_tensor(lr), exp_avg, exp_avg_sq, beta1_pow,
                     beta2_pow, master_param, p, exp_avg, exp_avg_sq, beta1_pow,
diff --git a/plsc/optimizer/momentum.py b/plsc/optimizer/momentum.py
index 3849e04dcca3e..84121e3502b4a 100644
--- a/plsc/optimizer/momentum.py
+++ b/plsc/optimizer/momentum.py
@@ -18,6 +18,7 @@
 
 import math
 import paddle
+from paddle import _legacy_C_ops as _C_ops
 from .optimizer import Optimizer
 from plsc.utils import logger
 
@@ -101,7 +102,7 @@ def step(self):
                     axis = getattr(p, 'axis', None)
                     assert index is not None
                     assert axis is not None
-                    _, _, _ = paddle._C_ops.sparse_momentum(
+                    _, _, _ = _C_ops.sparse_momentum(
                         p,
                         grad,
                         exp_avg,
@@ -125,7 +126,7 @@ def step(self):
                         'multi_precision',
                         master_param is not None)
                 else:
-                    _, _, _ = paddle._C_ops.momentum(
+                    _, _, _ = _C_ops.momentum(
                         p,
                         grad,
                         exp_avg,