Add double grad in reduce sum (#27115)

* set default value to strategy in distributed_optimizer test=develop

paddle/fluid/operators/reduce_ops/reduce_sum_op.cc

@@ -51,6 +51,20 @@ class ReduceSumOpGradMaker : public framework::SingleGradOpMaker<T> {
   }
 };
 
+template <typename T>
+class ReduceSumDoubleOpGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetInput("X", this->OutputGrad(framework::GradVarName("X")));
+    op->SetOutput("Out", this->InputGrad(framework::GradVarName("Out")));
+    op->SetAttrMap(this->Attrs());
+    op->SetType("reduce_sum");
+  }
+};
+
 DECLARE_NO_NEED_BUFFER_VARS_INFERER(ReduceSumGradNoNeedBufferVarInferer, "X");
 class ReduceSumVarTypeInference : public paddle::framework::VarTypeInference {
  public:
@@ -77,6 +91,8 @@ REGISTER_OPERATOR(reduce_sum, ops::ReduceOp, ReduceSumOpMaker,
                   ops::ReduceSumOpGradMaker<paddle::framework::OpDesc>,
                   ops::ReduceSumOpGradMaker<paddle::imperative::OpBase>);
 REGISTER_OPERATOR(reduce_sum_grad, ops::ReduceGradOp,
+                  ops::ReduceSumDoubleOpGradMaker<paddle::framework::OpDesc>,
+                  ops::ReduceSumDoubleOpGradMaker<paddle::imperative::OpBase>,
                   ops::ReduceSumGradNoNeedBufferVarInferer);
 
 REGISTER_OP_CPU_KERNEL(

python/paddle/fluid/tests/unittests/test_nn_grad.py

@@ -101,6 +101,29 @@ class TestReduceMeanWithDimDoubleGradCheck(unittest.TestCase):
             self.func(p)
 
 
+class TestReduceSumWithDimDoubleGradCheck(unittest.TestCase):
+    @prog_scope()
+    def func(self, place):
+        shape = [7, 11]
+        eps = 0.05
+        dtype = np.float64
+
+        x = layers.data('x', shape, False, dtype)
+        x.persistable = True
+        y = layers.reduce_sum(x, dim=0)
+        x_arr = np.random.uniform(-1, 1, shape).astype(dtype)
+
+        gradient_checker.double_grad_check(
+            [x], y, x_init=x_arr, place=place, eps=eps)
+
+    def test_grad(self):
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
 class TestMulDoubleGradCheck(unittest.TestCase):
     @prog_scope()
     def func(self, place):