diff --git a/paddle/fluid/operators/scatter_op.h b/paddle/fluid/operators/scatter_op.h
index d29947b55e751a3e7993f765198364f4debe2472..181bb1af5cce7fad228e61e1a76ed66a9bd61b3e 100644
--- a/paddle/fluid/operators/scatter_op.h
+++ b/paddle/fluid/operators/scatter_op.h
@@ -35,7 +35,7 @@ class ScatterOpKernel : public framework::OpKernel<T> {
     auto *Out = ctx.Output<Tensor>("Out");
 
     // In place output: Out = X, Out[Ids] += Updates
-    Out->ShareDataWith(*X);
+    framework::TensorCopySync(*X, ctx.GetPlace(), Out);
     // Apply ScatterUpdate: Out[index] += Updates[:]
     ScatterAssign<T>(ctx.device_context(), *Updates, *Ids, Out);
   }
@@ -53,7 +53,7 @@ class ScatterGradientOpKernel : public framework::OpKernel<T> {
     auto *dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
     // In place gradient: dX = dO
-    dX->ShareDataWith(*dOut);
+    framework::TensorCopySync(*dOut, ctx.GetPlace(), dX);
     dUpdates->mutable_data<T>(ctx.GetPlace());
     // Gradient by Gather: dUpdates += dO[Ids]
     CPUGather<T>(ctx.device_context(), *dOut, *Ids, dUpdates);