diff --git a/paddle/fluid/operators/reduce_ops/reduce_mean_op.part.cu b/paddle/fluid/operators/reduce_ops/reduce_mean_op.part.cu
index 4cc2577f6b2ec2c53421257d207aba83b95965a3..a578c9f7d81083c533028b9c8912a24006ed0292 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_mean_op.part.cu
+++ b/paddle/fluid/operators/reduce_ops/reduce_mean_op.part.cu
@@ -17,15 +17,9 @@
 
 template <typename T>
 using CUDAReduceMeanGradKernel =
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, T,
-                          ops::MeanGradFunctor, true>;
-
-using FP16CUDAReduceMeanGradKernel =
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                          paddle::platform::float16, ops::FP16MeanGradFunctor,
-                          true>;
+    ops::ReduceCudaGradKernel<T, kps::DivideFunctor>;
 
 REGISTER_OP_CUDA_KERNEL(reduce_mean_grad, CUDAReduceMeanGradKernel<bool>,
-                        FP16CUDAReduceMeanGradKernel,
+                        CUDAReduceMeanGradKernel<paddle::platform::float16>,
                         CUDAReduceMeanGradKernel<float>,
                         CUDAReduceMeanGradKernel<double>);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.h b/paddle/fluid/operators/reduce_ops/reduce_op.h
index 661fb772f1c573f480def3e7162f62ba56a56d02..ff6a443638452d3ac01541c59245c08aedf5db81 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_op.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_op.h
@@ -623,11 +623,12 @@ class ReduceGradOp : public framework::OperatorWithKernel {
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    int in_dtype = ctx.Attr<int>("in_dtype");
+    int out_dtype = ctx.Attr<int>("out_dtype");
     auto input_data_type =
-        (in_dtype >= 0) ? static_cast<framework::proto::VarType::Type>(in_dtype)
-                        : OperatorWithKernel::IndicateVarDataType(
-                              ctx, framework::GradVarName("Out"));
+        (out_dtype >= 0)
+            ? static_cast<framework::proto::VarType::Type>(out_dtype)
+            : OperatorWithKernel::IndicateVarDataType(
+                  ctx, framework::GradVarName("Out"));
 #ifdef PADDLE_WITH_MKLDNN
     auto CanMKLDNNReduceGradBeUsed = [&]() {
       auto dx_dims = ctx.Input<Tensor>("X")->dims();
@@ -736,6 +737,55 @@ class ReduceCudaKernel : public framework::OpKernel<T> {
                                            pt_out.get());
   }
 };
+
+template <typename T, template <typename, typename> class TransformOp>
+class ReduceCudaGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    bool reduce_all = context.Attr<bool>("reduce_all");
+    std::vector<int> dims = context.Attr<std::vector<int>>("dim");
+    auto* in_x = context.Input<Tensor>("X");
+    auto* d_out =
+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* d_x = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto out_dtype = context.Attr<int>("in_dtype");
+    // get reduce_dim and reduce_num for reduce_mean_grad
+    int dim_size = in_x->dims().size();
+    std::vector<int> reduce_dims = GetReduceDim(dims, dim_size, reduce_all);
+    auto update_dims = vectorize(d_x->dims());
+    int reduce_num = 1;
+    for (auto i : reduce_dims) {
+      reduce_num *= (in_x->dims())[i];
+      update_dims[i] = 1;
+    }
+    // make new tensor
+    framework::Tensor new_d_out(d_out->type());
+    new_d_out.ShareDataWith(*d_out);
+    new_d_out.Resize(paddle::framework::make_ddim(update_dims));
+    auto& dev_ctx = context.cuda_device_context();
+    if (out_dtype > 0) {
+      d_x->mutable_data(
+          dev_ctx.GetPlace(),
+          static_cast<framework::proto::VarType::Type>(out_dtype));
+    } else {
+      d_x->mutable_data(
+          dev_ctx.GetPlace(),
+          static_cast<framework::proto::VarType::Type>(d_out->type()));
+    }
+    auto pt_d_out = paddle::experimental::MakePtenDenseTensor(new_d_out);
+    auto pt_d_x = paddle::experimental::MakePtenDenseTensor(*d_x);
+    auto pt_out_dtype = pten::TransToPtenDataType(
+        static_cast<framework::proto::VarType::Type>(out_dtype));
+    if (out_dtype <= 0) {
+      pt_out_dtype = pten::TransToPtenDataType(
+          static_cast<framework::proto::VarType::Type>(d_out->type()));
+    }
+    using MPType = typename kps::details::MPTypeTrait<T>::Type;
+    pten::ReduceGrad<T, TransformOp<T, MPType>>(
+        dev_ctx, pt_d_out.get(), pt_d_x.get(), pt_out_dtype,
+        TransformOp<T, MPType>(reduce_num));
+  }
+};
 #endif
 
 }  // namespace operators
diff --git a/paddle/fluid/operators/reduce_ops/reduce_sum_op.cc b/paddle/fluid/operators/reduce_ops/reduce_sum_op.cc
index 562a5719d74d90062fb92c5198bacf82499c9948..9a715eb98ef994f89e201656e8c371d819b11f19 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_sum_op.cc
+++ b/paddle/fluid/operators/reduce_ops/reduce_sum_op.cc
@@ -50,7 +50,7 @@ class ReduceSumOpGradMaker : public framework::SingleGradOpMaker<T> {
 
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const {
-    int in_dtype = ctx.Attr<int>("in_dtype");
+    int in_dtype = ctx.Attr<int>("out_dtype");
     if (in_dtype >= 0) {
       return framework::OpKernelType(
           static_cast<framework::proto::VarType::Type>(in_dtype),
diff --git a/paddle/fluid/operators/reduce_ops/reduce_sum_op.h b/paddle/fluid/operators/reduce_ops/reduce_sum_op.h
index 9782ce28da4af0b17f38219c039dd4fd62f46846..79b3480afbcd7750297e9e02575a43fdb6809126 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_sum_op.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_sum_op.h
@@ -74,7 +74,7 @@ class ReduceSumGradKernel : public framework::OpKernel<T> {
     auto dims = context.Attr<std::vector<int>>("dim");
     if (context.GetPlace().GetType() == platform::CPUPlace().GetType() &&
         dims.size() == 1) {
-      int in_dtype = context.Attr<int>("in_dtype");
+      int in_dtype = context.Attr<int>("out_dtype");
 
       if (in_dtype >= 0) {
         Tensor tmp_tensor;
diff --git a/paddle/fluid/operators/reduce_ops/reduce_sum_op.part.cu b/paddle/fluid/operators/reduce_ops/reduce_sum_op.part.cu
index c629663b19ebd7f42f3a16e69bd4b46784ff67dd..c3d3e0cf6ecd51f3bb2baa063878f80444db3563 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_sum_op.part.cu
+++ b/paddle/fluid/operators/reduce_ops/reduce_sum_op.part.cu
@@ -17,8 +17,7 @@
 
 template <typename T>
 using CUDAReduceSumGradKernel =
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, T,
-                          ops::SumGradFunctor, true>;
+    ops::ReduceCudaGradKernel<T, kps::IdentityFunctor>;
 
 REGISTER_OP_CUDA_KERNEL(
     reduce_sum_grad, CUDAReduceSumGradKernel<bool>,
diff --git a/paddle/pten/kernels/gpu/elementwise.h b/paddle/pten/kernels/gpu/elementwise.h
index 9a3ae7f12dfcd62a1a18154971fa99ab72c5561d..db85c596da945f77a7dd0fc3f88129c73bb0d7e5 100644
--- a/paddle/pten/kernels/gpu/elementwise.h
+++ b/paddle/pten/kernels/gpu/elementwise.h
@@ -134,12 +134,19 @@ struct DimensionsTransform {
   explicit DimensionsTransform(const std::vector<const DenseTensor *> &ins,
                                const pten::framework::DDim &dims,
                                int axis) {
-    const int N = ins.size();
+    const int N = max(static_cast<int>(ins.size()), 2);
     dim_size = dims.size();
     out_dims = pten::framework::vectorize<int64_t>(dims);
     in_dims.resize(N);
-    for (int j = 0; j < N; ++j) {
-      in_dims[j] = pten::framework::vectorize<int64_t>(ins[j]->dims());
+    if (ins.size() == 1) {
+      // when ins.size() = 1, broadcast input to output
+      in_dims[0] = pten::framework::vectorize<int64_t>(ins[0]->dims());
+      in_dims[1] = out_dims;
+      // Add out_dims to in_dims to avoid errors in dims merging
+    } else {
+      for (int j = 0; j < N; ++j) {
+        in_dims[j] = pten::framework::vectorize<int64_t>(ins[j]->dims());
+      }
     }
     InputDimensionsExtend(N, axis);
 
diff --git a/paddle/pten/kernels/gpu/reduce.h b/paddle/pten/kernels/gpu/reduce.h
index 10badf00a1e246a36b0e0b37525ac9ffba028e92..49a5e9e9b33a5e286d738e7198c8914c42c1136d 100644
--- a/paddle/pten/kernels/gpu/reduce.h
+++ b/paddle/pten/kernels/gpu/reduce.h
@@ -45,8 +45,7 @@ namespace cub = hipcub;
 #include "paddle/pten/api/ext/dispatch.h"
 #include "paddle/pten/backends/gpu/gpu_context.h"
 #include "paddle/pten/core/dense_tensor.h"
-#include "paddle/pten/kernels/funcs/elementwise_base.h"
-
+#include "paddle/pten/kernels/gpu/elementwise.h"
 // Reduce split or not, Whether to use ReduceHigherDim
 #define REDUCE_SPLIT_BOUNDARY 512
 #define REDUCE_VEC_SIZE 4
@@ -1254,6 +1253,24 @@ void Reduce(const GPUContext& dev_ctx,
             x, out, TransformOp<T, MPType>(reduce_num), reduce_dims, stream);
   }
 }
+
+template <typename InT, typename Functor>
+void ReduceGrad(const GPUContext& dev_ctx,
+                DenseTensor* d_out,
+                DenseTensor* d_x,
+                DataType out_dtype,
+                Functor functor) {
+  std::vector<const DenseTensor*> inputs = {d_out};
+  std::vector<DenseTensor*> outputs = {d_x};
+  PD_VISIT_ALL_TYPES(
+      out_dtype, "LaunchBroadcastElementwiseCudaKernel", ([&] {
+        LaunchBroadcastElementwiseCudaKernel<pten::ElementwiseType::kUnary,
+                                             InT,
+                                             data_t>(
+            dev_ctx, inputs, &outputs, 0, functor);
+      }));
+}
+
 }  // namespace pten
 
 #endif