[DoubleGrad] Enabled test_autograd_functional_dynamic.py under eager mode (#41668)

* [DoubleGrad] Enabled double grad test cases in eager_mode for test_imperative_double_grad

* Fixed elementwise issue

* Addressed CI failures

* [DoubleGrad] Enabled test_imperative_triple_grad test cases under eager_mode

* [DoubleGrad] Enabled test_autograd_functional_dynamic.py under eager mode

* Enabled more test cases

* Fixed performance issues

* Fixed minor issue

paddle/fluid/eager/auto_code_generator/final_state_generator/codegen_utils.py

@@ -22,9 +22,16 @@ import os
 ### Global Variables ###
 ########################
 ops_to_fill_zero_for_empty_grads = set([
-    "split_grad", "rnn_grad", "matmul_double_grad", "matmul_triple_grad",
-    "sigmoid_double_grad", "sigmoid_triple_grad", "add_double_grad",
-    "add_triple_grad"
+    "split_grad",
+    "rnn_grad",
+    "matmul_double_grad",
+    "matmul_triple_grad",
+    "sigmoid_double_grad",
+    "sigmoid_triple_grad",
+    "add_double_grad",
+    "add_triple_grad",
+    "multiply_double_grad",
+    "multiply_triple_grad",
 ])
 
 # For API dispatch used at python-level

paddle/fluid/eager/autograd_meta.h

@@ -107,6 +107,8 @@ class AutogradMeta : public AbstractAutogradMeta {
 
   GradNodeBase* GradNode() const { return grad_node_.get(); }
 
+  void ResetGradNode() { grad_node_.reset(); }
+
   void SetSingleOutRankWithSlot(size_t slot_id, size_t rank) {
     out_slot_id_ = slot_id;
     out_rank_ = rank;

paddle/fluid/eager/backward.cc

@@ -53,7 +53,7 @@ class GeneralGrad {
         auto* target_node = auto_grad_meta->GetMutableGradNode().get();
 
         if (orig_to_copied_node_mapping_.count(target_node)) {
-          target_node = orig_to_copied_node_mapping_[target_node];
+          target_node = orig_to_copied_node_mapping_[target_node].get();
         } else {
           VLOG(6) << "Unable to find target node in "
                      "orig_to_copied_node_mapping_, likely indicating an "
@@ -261,7 +261,7 @@ class GeneralGrad {
 
       auto* target_node = auto_grad_meta->GetMutableGradNode().get();
       if (orig_to_copied_node_mapping_.count(target_node)) {
-        target_node = orig_to_copied_node_mapping_[target_node];
+        target_node = orig_to_copied_node_mapping_[target_node].get();
       } else {
         VLOG(6) << "Unable to find target node in "
                    "orig_to_copied_node_mapping_, likely indicating an unused "
@@ -349,12 +349,12 @@ class GeneralGrad {
 
   GradNodeBase* CopyGradNode(const std::shared_ptr<GradNodeBase>& orig_node) {
     if (orig_to_copied_node_mapping_.count(orig_node.get())) {
-      return orig_to_copied_node_mapping_[orig_node.get()];
+      return orig_to_copied_node_mapping_[orig_node.get()].get();
     }
     std::shared_ptr<GradNodeBase> copied_node = orig_node->Copy();
 
     // Save node and update mapping
-    orig_to_copied_node_mapping_[orig_node.get()] = copied_node.get();
+    orig_to_copied_node_mapping_[orig_node.get()] = copied_node;
     copied_grad_nodes_.push_back(copied_node);
 
     return copied_node.get();
@@ -379,7 +379,7 @@ class GeneralGrad {
           paddle::platform::errors::Fatal(
               "Cannot reconstruct backward graph,"
               "unable to find copied target for certain grad node."));
-      GradNodeBase* copied_node = orig_to_copied_node_mapping_[orig_node];
+      GradNodeBase* copied_node = orig_to_copied_node_mapping_[orig_node].get();
 
       const std::vector<std::vector<Edge>>& orig_edges = orig_node->GetEdges();
       std::vector<std::vector<Edge>>& copied_edges =
@@ -397,13 +397,12 @@ class GeneralGrad {
           std::shared_ptr<GradNodeBase> copied_next_node;
           if (orig_to_copied_node_mapping_.count(orig_next_node.get())) {
             copied_next_node =
-                orig_to_copied_node_mapping_[orig_next_node.get()]
-                    ->shared_from_this();
+                orig_to_copied_node_mapping_[orig_next_node.get()];
 
           } else {
             copied_next_node = orig_next_node->Copy();
             orig_to_copied_node_mapping_[orig_next_node.get()] =
-                copied_next_node.get();
+                copied_next_node;
             copied_grad_nodes_.push_back(copied_next_node);
           }
 
@@ -436,7 +435,8 @@ class GeneralGrad {
   std::unordered_map<GradNodeBase*, paddle::experimental::Tensor> results_map;
 
   std::vector<std::shared_ptr<GradNodeBase>> copied_grad_nodes_;
-  std::unordered_map<GradNodeBase*, GradNodeBase*> orig_to_copied_node_mapping_;
+  std::unordered_map<GradNodeBase*, std::shared_ptr<GradNodeBase>>
+      orig_to_copied_node_mapping_;
 
   DISABLE_COPY_AND_ASSIGN(GeneralGrad);
 };
@@ -534,6 +534,7 @@ std::vector<paddle::experimental::Tensor> RunBackward(
 
   // GeneralGrad
   bool is_general_grad = !inputs.empty();
+  if (is_general_grad) GeneralGrad::Instance().Clear();
 
   /* --- Initialization --- */
   // 1. Init queue with starting nodes
@@ -746,6 +747,7 @@ std::vector<paddle::experimental::Tensor> RunBackward(
           VLOG(6) << "We get grad_output_tensor with slot: " << i
                   << ", rank: " << j << " as uninitialized or undefined tensor";
         }
+
         VLOG(6) << "Get Edge and grad_output_tensor with slot: " << i
                 << ", rank: " << j
                 << " 's name is: " << grad_output_tensor.name();

paddle/fluid/eager/grad_node_info.h

@@ -87,7 +87,7 @@ class GradSlotMeta {
   std::shared_ptr<phi::DenseTensorMeta> meta_ = nullptr;
 };
 
-class GradNodeBase : public std::enable_shared_from_this<GradNodeBase> {
+class GradNodeBase {
  public:
   GradNodeBase() { VLOG(6) << "Construct GradNodeBase"; }
   GradNodeBase(size_t bwd_in_slot_num, size_t bwd_out_slot_num);

paddle/fluid/eager/tensor_wrapper.h

@@ -79,9 +79,9 @@ class TensorWrapper {
 
     auto* tensor_autograd_meta = EagerUtils::nullable_autograd_meta(tensor);
     if (tensor_autograd_meta) {
-      auto autograd_meta = std::make_shared<AutogradMeta>(
-          Edge(nullptr, EagerUtils::OutRankInfo(tensor)));
-      autograd_meta->SetStopGradient(tensor_autograd_meta->StopGradient());
+      auto autograd_meta =
+          std::make_shared<AutogradMeta>(*tensor_autograd_meta);
+      autograd_meta->ResetGradNode();
       intermidiate_tensor_.set_autograd_meta(autograd_meta);
       weak_grad_node_ = tensor_autograd_meta->GetMutableGradNode();
     }
@@ -98,8 +98,11 @@ class TensorWrapper {
     check_inplace_version();
 
     // if it's full_reserved just return the full copy of tensor
+    if (full_reserved_) {
+      return intermidiate_tensor_;
+    } else {
       paddle::experimental::Tensor recovered_tensor = intermidiate_tensor_;
-    if (!full_reserved_) {
+
       std::shared_ptr<GradNodeBase> new_grad_node = weak_grad_node_.lock();
       if (new_grad_node) {
         VLOG(3) << "Recovered TensorWrapper with GradNode "
@@ -109,18 +112,16 @@ class TensorWrapper {
       }
       auto* intermediate_autograd_meta =
           EagerUtils::unsafe_autograd_meta(intermidiate_tensor_);
-      auto p_ab_autograd_meta = std::make_shared<AutogradMeta>(
-          Edge(new_grad_node, intermediate_autograd_meta->OutRankInfo()));
-      p_ab_autograd_meta->SetStopGradient(
-          intermediate_autograd_meta->StopGradient());
-
-      recovered_tensor.set_autograd_meta(
-          std::static_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
-              p_ab_autograd_meta));
+      auto p_ab_autograd_meta =
+          std::make_shared<AutogradMeta>(*intermediate_autograd_meta);
+      if (new_grad_node) {
+        p_ab_autograd_meta->SetGradNode(new_grad_node);
       }
+      recovered_tensor.set_autograd_meta(p_ab_autograd_meta);
 
       return recovered_tensor;
     }
+  }
 
   void check_inplace_version() {
     if (no_need_buffer_) {

paddle/phi/infermeta/backward.h

@@ -100,6 +100,8 @@ void GatherNdGradInferMeta(const MetaTensor& x,
                            const MetaTensor& out_grad,
                            MetaTensor* x_grad);
 
+void GeneralUnaryGradInferMeta(const MetaTensor& x, MetaTensor* dx);
+
 void GeneralBinaryGradInferMeta(const MetaTensor& x,
                                 const MetaTensor& y,
                                 MetaTensor* dx,
@@ -132,8 +134,6 @@ void GeneralQuinaryGradInferMeta(const MetaTensor& x,
                                  MetaTensor* dk,
                                  MetaTensor* dl);
 
-void GeneralUnaryGradInferMeta(const MetaTensor& x, MetaTensor* dx);
-
 void GumbelSoftmaxGradInferMeta(const MetaTensor& out,
                                 const MetaTensor& dout,
                                 int axis,

python/paddle/autograd/functional.py

@@ -943,8 +943,10 @@ def batch_jacobian(func, inputs, create_graph=False, allow_unused=False):
             #        [0., 1., 0., 1., 0., 1., 0., 1.]]))
 
     '''
+
     inputs = _as_tensors(inputs)
     outputs = _as_tensors(func(*inputs))
+
     batch_size = inputs[0].shape[0]
     for input in inputs:
         assert input.shape[
@@ -961,12 +963,14 @@ def batch_jacobian(func, inputs, create_graph=False, allow_unused=False):
     for i, flat_output in enumerate(flat_outputs):
         jac_i = list([] for _ in range(fin_size))
         for k in range(flat_output.shape[1]):
+
             row_k = paddle.grad(
                 flat_output[:, k],
                 inputs,
                 create_graph=create_graph,
                 retain_graph=True,
                 allow_unused=allow_unused)
+
             for j in range(fin_size):
                 jac_i[j].append(
                     paddle.reshape(

python/paddle/fluid/tests/unittests/autograd/test_autograd_functional_dynamic.py

@@ -205,7 +205,14 @@ class TestVJP(TestAutogradFunctional):
         self.check_results(ref_result, aliased_result)
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.func_vjp_i1o1()
+            self.func_vjp_i2o1()
+            self.func_vjp_i2o2()
+            self.func_vjp_i2o2_omitting_v()
+            self.func_vjp_nested()
+            self.func_vjp_aliased_input()
+
         self.func_vjp_i1o1()
         self.func_vjp_i2o1()
         self.func_vjp_i2o2()
@@ -227,7 +234,8 @@ class TestVJPException(unittest.TestCase):
                                 paddle.to_tensor(self.v))
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.func_vjp()
         self.func_vjp()
 
 
@@ -303,7 +311,11 @@ class TestJVP(TestAutogradFunctional):
             self.check_results(results_omitting_v, results_with_v)
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.func_jvp_i1o1()
+            self.func_jvp_i2o1()
+            self.func_jvp_i2o2()
+            self.func_jvp_i2o2_omitting_v()
         self.func_jvp_i1o1()
         self.func_jvp_i2o1()
         self.func_jvp_i2o2()
@@ -328,12 +340,12 @@ class TestJacobianClassNoBatch(unittest.TestCase):
         self._atol = config.TOLERANCE.get(str(self._dtype)).get(
             "first_order_grad").get("atol")
 
-        self.xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
+    def func_jacobian(self):
+        xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
             self.xs, typing.Sequence) else paddle.to_tensor(self.xs)
-        self._actual = paddle.autograd.Jacobian(self.func, self.xs, False)
-        self._expected = self._expected()
+        self._actual = paddle.autograd.Jacobian(self.func, xs, False)
+        self._expected = self._get_expected()
 
-    def func_jacobian(self):
         Index = collections.namedtuple('Index', ('type', 'value'))
         indexes = (Index('all', (slice(0, None, None), slice(0, None, None))),
                    Index('row', (0, slice(0, None, None))),
@@ -349,13 +361,16 @@ class TestJacobianClassNoBatch(unittest.TestCase):
                 err_msg=f'Testcase {index.type} index not passed, value is {index.value}'
             )
 
-    def _expected(self):
-        jac = utils._compute_numerical_jacobian(self.func, self.xs, self._eps,
+    def _get_expected(self):
+        xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
+            self.xs, typing.Sequence) else paddle.to_tensor(self.xs)
+        jac = utils._compute_numerical_jacobian(self.func, xs, self._eps,
                                                 self._dtype)
         return utils._np_concat_matrix_sequence(jac, utils.MatrixFormat.NM)
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.func_jacobian()
         self.func_jacobian()
 
 
@@ -375,12 +390,12 @@ class TestJacobianClassBatchFirst(unittest.TestCase):
         self._atol = config.TOLERANCE.get(str(self._dtype)).get(
             "first_order_grad").get("atol")
 
-        self.xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
+    def func_jacobian(self):
+        xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
             self.xs, typing.Sequence) else paddle.to_tensor(self.xs)
-        self._actual = paddle.autograd.Jacobian(self.func, self.xs, True)
-        self._expected = self._expected()
+        self._actual = paddle.autograd.Jacobian(self.func, xs, True)
+        self._expected = self._get_expected()
 
-    def func_jacobian(self):
         Index = collections.namedtuple('Index', ('type', 'value'))
         indexes = (
             Index('all', (slice(0, None, None), slice(0, None, None),
@@ -402,15 +417,18 @@ class TestJacobianClassBatchFirst(unittest.TestCase):
                 err_msg=f'Testcase {index.type} index not passed, value is {index.value}'
             )
 
-    def _expected(self):
-        jac = utils._compute_numerical_batch_jacobian(
-            self.func, self.xs, self._eps, self._dtype, False)
+    def _get_expected(self):
+        xs = [paddle.to_tensor(x) for x in self.xs] if isinstance(
+            self.xs, typing.Sequence) else paddle.to_tensor(self.xs)
+        jac = utils._compute_numerical_batch_jacobian(self.func, xs, self._eps,
+                                                      self._dtype, False)
         jac = utils._np_concat_matrix_sequence(jac, utils.MatrixFormat.NBM)
         return utils._np_transpose_matrix_format(jac, utils.MatrixFormat.NBM,
                                                  utils.MatrixFormat.BNM)
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.func_jacobian()
         self.func_jacobian()
 
 
@@ -492,7 +510,14 @@ class TestHessianClassNoBatch(unittest.TestCase):
             paddle.autograd.Hessian(func, paddle.ones([3]))
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_single_input()
+            self.func_multi_input()
+            self.func_allow_unused_true()
+            self.func_create_graph_true()
+            self.func_out_not_single()
+        self.setUpClass()
         self.func_single_input()
         self.func_multi_input()
         self.func_allow_unused_true()
@@ -599,7 +624,14 @@ class TestHessianClassBatchFirst(unittest.TestCase):
             paddle.autograd.Hessian(func, paddle.ones((3, 3)), is_batched=True)
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_single_input()
+            self.func_multi_input()
+            self.func_allow_unused()
+            self.func_stop_gradient()
+            self.func_out_not_single()
+        self.setUpClass()
         self.func_single_input()
         self.func_multi_input()
         self.func_allow_unused()
@@ -619,6 +651,7 @@ class TestHessian(unittest.TestCase):
             "second_order_grad").get("rtol")
         self.atol = config.TOLERANCE.get(self.dtype).get(
             "second_order_grad").get("atol")
+
         self.x = paddle.rand(shape=self.shape, dtype=self.dtype)
         self.y = paddle.rand(shape=self.shape, dtype=self.dtype)
 
@@ -694,9 +727,10 @@ class TestHessian(unittest.TestCase):
                                    self.rtol, self.atol)
         try:
             paddle.grad(hessian, self.x)
-        except RuntimeError as e:
+        except Exception as e:
             error_msg = cpt.get_exception_message(e)
-            assert error_msg.find("has no gradient") > 0
+            assert error_msg.find("has no gradient") > 0 or error_msg.find(
+                "does not appear") > 0
 
     def func_create_graph_true(self):
         def func(x):
@@ -713,7 +747,15 @@ class TestHessian(unittest.TestCase):
         assert triple_grad is not None
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_single_input()
+            self.func_multi_input()
+            self.func_allow_unused_false()
+            self.func_allow_unused_true()
+            self.func_create_graph_false()
+            self.func_create_graph_true()
+        self.setUpClass()
         self.func_single_input()
         self.func_multi_input()
         self.func_allow_unused_false()
@@ -830,9 +872,10 @@ class TestBatchHessian(unittest.TestCase):
                                    self.rtol, self.atol)
         try:
             paddle.grad(hessian, self.x)
-        except RuntimeError as e:
+        except Exception as e:
             error_msg = cpt.get_exception_message(e)
-            assert error_msg.find("has no gradient") > 0
+            assert error_msg.find("has no gradient") > 0 or error_msg.find(
+                "does not appear") > 0
 
     def func_create_graph_true(self):
         def func(x):
@@ -849,7 +892,15 @@ class TestBatchHessian(unittest.TestCase):
         assert triple_grad is not None
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_single_input()
+            self.func_multi_input()
+            self.func_allow_unused_false()
+            self.func_allow_unused_true()
+            self.func_create_graph_false()
+            self.func_create_graph_true()
+        self.setUpClass()
         self.func_single_input()
         self.func_multi_input()
         self.func_allow_unused_false()
@@ -985,7 +1036,14 @@ class TestVHP(unittest.TestCase):
         assert triple_grad is not None
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_v_default()
+            self.func_multi_input()
+            self.func_single_input()
+            self.func_allow_unused_true()
+            self.func_create_graph_true()
+        self.setUpClass()
         self.func_v_default()
         self.func_multi_input()
         self.func_single_input()
@@ -1100,9 +1158,10 @@ class TestJacobian(unittest.TestCase):
                                        self.atol)
         try:
             paddle.grad(jacobian[0], [self.x, self.y])
-        except RuntimeError as e:
+        except Exception as e:
             error_msg = cpt.get_exception_message(e)
-            assert error_msg.find("has no gradient") > 0
+            assert error_msg.find("has no gradient") > 0 or error_msg.find(
+                "does not appear") > 0
 
     def func_create_graph_true(self):
         def func(x, y):
@@ -1123,7 +1182,17 @@ class TestJacobian(unittest.TestCase):
         assert double_grad is not None
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_multi_input_and_multi_output()
+            self.func_multi_input_and_single_output()
+            self.func_single_input_and_multi_output()
+            self.func_single_input_and_single_output()
+            self.func_allow_unused_false()
+            self.func_allow_unused_true()
+            self.func_create_graph_false()
+            self.func_create_graph_true()
+        self.setUpClass()
         self.func_multi_input_and_multi_output()
         self.func_multi_input_and_single_output()
         self.func_single_input_and_multi_output()
@@ -1269,9 +1338,10 @@ class TestJacobianBatch(unittest.TestCase):
                                        self.atol)
         try:
             paddle.grad(jacobian[0], [self.x, self.y])
-        except RuntimeError as e:
+        except Exception as e:
             error_msg = cpt.get_exception_message(e)
-            assert error_msg.find("has no gradient") > 0
+            assert error_msg.find("has no gradient") > 0 or error_msg.find(
+                "does not appear") > 0
 
     def func_create_graph_true(self):
         def func(x, y):
@@ -1292,7 +1362,17 @@ class TestJacobianBatch(unittest.TestCase):
         assert double_grad is not None
 
     def test_all_cases(self):
-        if _in_legacy_dygraph():
+        with _test_eager_guard():
+            self.setUpClass()
+            self.func_batch_single_input_and_batch_single_output()
+            self.func_batch_single_input_and_batch_multi_output()
+            self.func_batch_multi_input_and_batch_single_output()
+            self.func_batch_multi_input_and_batch_multi_output()
+            self.func_allow_unused_false()
+            self.func_allow_unused_true()
+            self.func_create_graph_false()
+            self.func_create_graph_true()
+        self.setUpClass()
         self.func_batch_single_input_and_batch_single_output()
         self.func_batch_single_input_and_batch_multi_output()
         self.func_batch_multi_input_and_batch_single_output()

python/paddle/tensor/linalg.py

@@ -1195,7 +1195,15 @@ def t(input, name=None):
             "Input(input) only support N-D (N<=2) tensor, but received "
             "length of Input(input) is %s. Perhaps you can use paddle."
             "tensor.transpose() instead." % len(input.shape))
-    if paddle.in_dynamic_mode():
+    if in_dygraph_mode():
+        if len(input.shape) == 1:
+            return input
+        # 2-D tensor
+        perm = [1, 0]
+        out = _C_ops.final_state_transpose(input, perm)
+        return out
+
+    if _in_legacy_dygraph():
         if len(input.shape) == 1:
             return input
         # 2-D tensor

python/paddle/utils/code_gen/backward.yaml

@@ -1097,6 +1097,7 @@
   kernel :
     func : multiply_double_grad
   optional : grad_x_grad, grad_y_grad
+  backward : multiply_triple_grad
 
 - backward_api : multiply_grad
   forward : multiply (Tensor x, Tensor y) -> Tensor(out)
@@ -1109,6 +1110,17 @@
     func : multiply_grad
   backward : multiply_double_grad
 
+- backward_api : multiply_triple_grad
+  forward : multiply_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, int aixs = -1) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
+  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, int axis = -1)
+  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
+  infer_meta :
+    func : GeneralQuinaryGradInferMeta
+    param : [x, y, fwd_grad_out, x, y]
+  kernel :
+    func : multiply_triple_grad
+  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_grad_out_grad
+
 - backward_api : mv_grad
   forward : mv (Tensor x, Tensor vec) -> Tensor(out)
   args : (Tensor x, Tensor vec, Tensor out_grad)
@@ -1286,6 +1298,16 @@
     func : relu_grad
   backward: relu_double_grad
 
+- backward_api : reshape_double_grad
+  forward : reshape_grad (Tensor xshape, Tensor grad_out) -> Tensor(grad_x)
+  args : (Tensor grad_out, Tensor grad_x_grad)
+  output : Tensor(grad_out_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [grad_out]
+  kernel :
+    func : reshape_double_grad
+
 - backward_api : reshape_grad
   forward : reshape_with_xshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
   args : (Tensor xshape, Tensor out_grad)
@@ -1299,6 +1321,7 @@
     data_type: out_grad
     backend: out_grad
     layout: out_grad
+  backward : reshape_double_grad
 
 - backward_api : roi_align_grad
   forward : roi_align (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned) -> Tensor(out)
@@ -1592,6 +1615,13 @@
     func : subtract_grad
   no_need_buffer : x, y
 
+- backward_api : sum_double_grad
+  forward : sum_grad (Tensor x, Tensor grad_out, int64_t[] dims, bool keep_dim, bool reduce_all=false) -> Tensor(grad_x)
+  args : (Tensor grad_x_grad, int64_t[] dims={}, bool keep_dim=false)
+  output : Tensor(grad_out_grad)
+  invoke : sum(grad_x_grad, dims, grad_x_grad.dtype(), keep_dim)
+  backward : sum_triple_grad
+
 - backward_api : sum_grad
   forward : sum (Tensor x, int64_t[] dims={}, DataType out_dtype=paddle::experimental::DataType::UNDEFINED, bool keep_dim=false) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int64_t[] dims, bool keep_dim, bool reduce_all=false)
@@ -1601,6 +1631,13 @@
     param : [x]
   kernel :
     func : sum_grad
+  backward : sum_double_grad
+
+- backward_api : sum_triple_grad
+  forward : sum_double_grad (Tensor grad_grad_x, int64_t[] dims={}, bool keep_dim=false) -> Tensor(grad_grad_out)
+  args : (Tensor grad_grad_x, Tensor grad_grad_out_grad, int64_t[] dims={}, bool keep_dim=false, bool reduce_all=false)
+  output : Tensor(grad_grad_x_grad)
+  invoke : sum_grad(grad_grad_x, grad_grad_out_grad, dims, keep_dim, reduce_all)
   no_need_buffer : x
 
 - backward_api : swish_grad
@@ -1695,6 +1732,12 @@
     func : trace_grad
   no_need_buffer : x
 
+- backward_api : transpose_double_grad
+  forward : transpose_grad (Tensor grad_out, int[] axis) -> Tensor(grad_x)
+  args : (Tensor grad_x_grad, int[] axis)
+  output : Tensor(grad_out_grad)
+  invoke : transpose(grad_x_grad, axis)
+
 - backward_api : transpose_grad
   forward : transpose (Tensor x, int[] axis) -> Tensor(out)
   args : (Tensor out_grad, int[] axis)
@@ -1704,6 +1747,7 @@
     param : [out_grad, axis]
   kernel :
     func : transpose_grad
+  backward : transpose_double_grad
 
 - backward_api : triangular_solve_grad
   forward : triangular_solve (Tensor x, Tensor y, bool upper, bool tranpose, bool unitriangular) -> Tensor(out)