diff --git a/paddle/fluid/operators/optimizers/rmsprop_op.cc b/paddle/fluid/operators/optimizers/rmsprop_op.cc
index 99d1156ee6d5fc88161e25bfa581a265707e6f92..eeee008cdc53c457146074060d526d8d0e8b43aa 100644
--- a/paddle/fluid/operators/optimizers/rmsprop_op.cc
+++ b/paddle/fluid/operators/optimizers/rmsprop_op.cc
@@ -143,4 +143,5 @@ http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    rmsprop, ops::RmspropOpKernel<paddle::platform::CPUDeviceContext, float>);
+    rmsprop, ops::RmspropOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::RmspropOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/fluid/operators/optimizers/rmsprop_op.cu b/paddle/fluid/operators/optimizers/rmsprop_op.cu
index 8b17d6a0204045a9b20adb79dbad72dff5ba267e..bf11ee686757c6c5e54e05f055eaa19f6553f915 100644
--- a/paddle/fluid/operators/optimizers/rmsprop_op.cu
+++ b/paddle/fluid/operators/optimizers/rmsprop_op.cu
@@ -15,4 +15,5 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
-    rmsprop, ops::RmspropOpKernel<paddle::platform::CUDADeviceContext, float>);
+    rmsprop, ops::RmspropOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::RmspropOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/python/paddle/optimizer/adam.py b/python/paddle/optimizer/adam.py
index 708aaa788f60d56a2adb41c8a571079354b3c192..24cebf8e6e6388a2d1e9711e3f862090918876a3 100644
--- a/python/paddle/optimizer/adam.py
+++ b/python/paddle/optimizer/adam.py
@@ -282,14 +282,13 @@ class Adam(Optimizer):
         for param in self._parameter_list:
             if not param.trainable:
                 continue
-            if hasattr(
-                    param, "_is_sparse"
-            ) and param._is_sparse and self.regularization is not None:
-                raise RuntimeError(
-                    "Adam don't support weight_decay with sparse parameters, please set it to None."
-                )
             if param._grad_ivar() is not None:
                 grad_var = param._grad_ivar()
+                if hasattr(grad_var, "_is_sparse") and grad_var._is_sparse(
+                ) and self.regularization is not None:
+                    raise RuntimeError(
+                        "Adam don't support weight_decay with sparse parameters, please set it to None."
+                    )
                 params_grads.append((param, grad_var))
 
         optimize_ops = self._apply_optimize(