diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu
index f3ed692666cec2d5fda9c5af269d695432218ea4..f7533cdd644eea676044044e01fedfc68a9aea52 100644
--- a/paddle/operators/gather_op.cu
+++ b/paddle/operators/gather_op.cu
@@ -20,13 +20,8 @@
 namespace paddle {
 namespace operators {
 
-// template <typename T>
-__global__ void print_arr(const float *params, const int N) {
-  CUDA_1D_KERNEL_LOOP(i, N) { printf("device: %d, %f\n", i, params[i]); }
-}
-
 template <typename T>
-class GatherOpCUDAKernel : public framework::OpKernel {
+class GatherOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
@@ -42,7 +37,7 @@ class GatherOpCUDAKernel : public framework::OpKernel {
 };
 
 template <typename T>
-class GatherGradOpCUDAKernel : public framework::OpKernel {
+class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu
index e27a926c6ab74a1fd117838f7ad1eafdb6d1eb5a..89d23945e0c000a9a615ab3d9b3eca3ed7e937c2 100644
--- a/paddle/operators/scatter_op.cu
+++ b/paddle/operators/scatter_op.cu
@@ -20,7 +20,7 @@ namespace paddle {
 namespace operators {
 
 template <typename T>
-class ScatterOpCUDAKernel : public framework::OpKernel {
+class ScatterOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
@@ -37,7 +37,7 @@ class ScatterOpCUDAKernel : public framework::OpKernel {
 };
 
 template <typename T>
-class ScatterGradOpCUDAKernel : public framework::OpKernel {
+class ScatterGradOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),