diff --git a/paddle/fluid/operators/multiplex_op.cu b/paddle/fluid/operators/multiplex_op.cu
index 45a2550793511f7cb8c20644ac79e9e88629ce7b..2f8a602f3c5c0a7c262235f99943ce336e20a7b4 100644
--- a/paddle/fluid/operators/multiplex_op.cu
+++ b/paddle/fluid/operators/multiplex_op.cu
@@ -33,7 +33,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
+    TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
     platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
@@ -69,7 +69,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
+    TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();