diff --git a/paddle/fluid/eager/pylayer/py_layer_node.cc b/paddle/fluid/eager/pylayer/py_layer_node.cc
index fad4fd50a5e3e1a57f191126f1c810dd31af5e9b..a00b292fe09152b470181d22ce959e7dc5ad76f1 100644
--- a/paddle/fluid/eager/pylayer/py_layer_node.cc
+++ b/paddle/fluid/eager/pylayer/py_layer_node.cc
@@ -106,8 +106,6 @@ GradNodePyLayer::operator()(
         pybind11::detail::error_string().c_str()));
   }
 
-  outputs_ = outputs;
-
   VLOG(6) << "PyLayer backward function finish...";
 
   PyObject* outputs_tuple = nullptr;
@@ -165,6 +163,9 @@ GradNodePyLayer::operator()(
   if (!PyTuple_Check(outputs)) {
     Py_XDECREF(outputs_tuple);
   }
+  Py_XDECREF(outputs);
+  Py_XDECREF(ctx_);
+  ctx_ = nullptr;
 
   return grad_out;
 }
diff --git a/paddle/fluid/eager/pylayer/py_layer_node.h b/paddle/fluid/eager/pylayer/py_layer_node.h
index b477d7a9ad9969c1221e300dfc17c1fcc638e858..c1a8c6e626b4fdb527fca05e2e86aa5dffacdb50 100644
--- a/paddle/fluid/eager/pylayer/py_layer_node.h
+++ b/paddle/fluid/eager/pylayer/py_layer_node.h
@@ -32,10 +32,7 @@ class GradNodePyLayer : public GradNodeBase {
     ctx_ = ctx;
   }
 
-  ~GradNodePyLayer() override {
-    Py_DECREF(ctx_);
-    Py_XDECREF(outputs_);
-  };
+  ~GradNodePyLayer() override { Py_XDECREF(ctx_); };
 
   virtual paddle::small_vector<std::vector<paddle::experimental::Tensor>,
                                kSlotSmallVectorSize>
@@ -50,9 +47,6 @@ class GradNodePyLayer : public GradNodeBase {
     return "GradNodePyLayer_" + std::string(Py_TYPE(ctx_)->tp_name);
   }
 
-  // for paddle.grad get result
-  PyObject* GetMutableOutputs() { return outputs_; }
-
   void SaveForwardOutputsMeta(
       const std::vector<std::vector<paddle::experimental::Tensor*>>&
           outputs_tensor) {
@@ -81,7 +75,6 @@ class GradNodePyLayer : public GradNodeBase {
 
  private:
   PyObject* ctx_{nullptr};
-  PyObject* outputs_{nullptr};
   std::vector<std::vector<phi::DenseTensorMeta>> forward_outputs_meta_;
   std::vector<std::vector<paddle::platform::Place>> forward_outputs_place_;
 };