Refactored GradNodeAccumulation data structure and behaviour (#39526)

* Refactored GradNodeAccumulation data structure and behaviour

* Fixed CI issues

* Fix compilation issues

* Fixed minor issues

* Reverted changes for intermediate and OverwriteOutput

* fixed minor issue

* Fixed code format issues

* Fixed CI-Coverage issue

* Fixed CI issues

paddle/fluid/eager/accumulation/accumulation_node.cc

@@ -25,6 +25,8 @@
 
 #include "glog/logging.h"
 
+namespace egr {
+
 static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
                             const paddle::experimental::Tensor& t) {
   if (!tensor->defined() || !tensor->initialized()) {
@@ -36,14 +38,6 @@ static void CopyOrAddTensor(paddle::experimental::Tensor* tensor,
   }
 }
 
-namespace egr {
-
-void GradNodeAccumulation::RetainGrad(
-    const std::function<paddle::experimental::Tensor(
-        const paddle::experimental::Tensor&)>& hook) {
-  retain_grad_hook_ = hook;
-}
-
 std::vector<std::vector<paddle::experimental::Tensor>> GradNodeAccumulation::
 operator()(
     const std::vector<std::vector<paddle::experimental::Tensor>>& grads) {
@@ -59,17 +53,18 @@ operator()(
                      "However received: %d in slot %d .",
                      grads[0].size(), 0));
   // Apply Gradient Hooks
+  paddle::experimental::Tensor grad_out;
   if (GradientHooksRegistered()) {
     std::vector<std::vector<paddle::experimental::Tensor>> hooked_grads =
         ApplyGradientHooks(grads);
-    // TODO(jiabin): It's little weird
-    CopyOrAddTensor(&accumulated_grad, hooked_grads[0][0]);
+    grad_out = hooked_grads[0][0];
   } else {
-    CopyOrAddTensor(&accumulated_grad, grads[0][0]);
+    grad_out = grads[0][0];
   }
 
-  if (retain_grad_hook_ != nullptr) {
-    retain_grad_hook_(accumulated_grad);
+  if (!weak_grad_.expired()) {
+    auto grad = weak_grad_.lock();
+    CopyOrAddTensor(grad.get(), grad_out);
   }
 
   // Apply Reduce Hooks
@@ -77,7 +72,7 @@ operator()(
     ApplyReduceHooks();
   }
 
-  return {{accumulated_grad}};
+  return {{grad_out}};
 }
 
 void GradNodeAccumulation::RegisterReduceHook(

paddle/fluid/eager/accumulation/accumulation_node.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include "paddle/fluid/eager/autograd_meta.h"
 #include "paddle/fluid/eager/grad_node_info.h"
 
 namespace egr {
@@ -21,7 +22,10 @@ namespace egr {
 class GradNodeAccumulation : public GradNodeBase {
  public:
   // Constructor: configure fwd input tensors to grad node
-  GradNodeAccumulation() : GradNodeBase(1, 1) { SetDefaultGradInOutMeta(); }
+  explicit GradNodeAccumulation(AutogradMeta* meta) : GradNodeBase(1, 1) {
+    weak_grad_ = meta->WeakGrad();
+    SetDefaultGradInOutMeta();
+  }
 
   ~GradNodeAccumulation() override = default;
 
@@ -30,11 +34,6 @@ class GradNodeAccumulation : public GradNodeBase {
       const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
       override;
 
-  void RetainGrad(const std::function<paddle::experimental::Tensor(
-                      const paddle::experimental::Tensor&)>& hook);
-
-  paddle::experimental::Tensor* Grad() { return &accumulated_grad; }
-
   std::string name() { return "GradNodeAccumulation"; }
 
   /**
@@ -49,7 +48,7 @@ class GradNodeAccumulation : public GradNodeBase {
   void ApplyReduceHooks();
 
  private:
-  paddle::experimental::Tensor accumulated_grad;
+  std::weak_ptr<paddle::experimental::Tensor> weak_grad_;
 
   std::function<paddle::experimental::Tensor(
       const paddle::experimental::Tensor&)>

paddle/fluid/eager/api/utils/hook_utils.cc

@@ -52,9 +52,15 @@ void RegisterReduceHookForTensor(const paddle::experimental::Tensor& tensor,
   }
 }
 
-void RetainGradForTensor(const paddle::experimental::Tensor& tensor) {
-  // TODO(jiabin): Support More Tensor type here
+static void RetainGradForRegularNode(
+    const paddle::experimental::Tensor& tensor) {
   AutogradMeta* meta = EagerUtils::unsafe_autograd_meta(tensor);
+  if (meta->RetainGrads()) {
+    return;
+  } else {
+    meta->SetRetainGrads(true);
+  }
+
   std::weak_ptr<paddle::experimental::Tensor> weak_grad_tensor =
       meta->WeakGrad();
 
@@ -79,21 +85,17 @@ void RetainGradForTensor(const paddle::experimental::Tensor& tensor) {
         }
       };
 
-  if (IsLeafTensor(tensor)) {
-    // Add RetainGrad as PostHook to AccumulationNode
-    std::shared_ptr<GradNodeBase> grad_node = EagerUtils::grad_node(tensor);
-    PADDLE_ENFORCE(
-        grad_node.get() != nullptr,
-        paddle::platform::errors::Fatal("Detected NULL grad_node"
-                                        "Leaf tensor should have had grad_node "
-                                        "with type: GradNodeAccumulation"));
-    auto accumulation_grad_node =
-        std::dynamic_pointer_cast<GradNodeAccumulation>(grad_node);
-    accumulation_grad_node->RetainGrad(hook);
+  // Append to GradientHooks
+  RegisterGradientHookForTensor(tensor, hook);
+}
 
+void RetainGradForTensor(const paddle::experimental::Tensor& tensor) {
+  if (IsLeafTensor(tensor)) {
+    // Leaf tensor's grad will always be retained
+    // Refer to implementation of AccumulationNode for more details
+    return;
   } else {
-    // Append to GradientHooks
-    RegisterGradientHookForTensor(tensor, hook);
+    RetainGradForRegularNode(tensor);
   }
 }
 

paddle/fluid/eager/api/utils/tensor_utils.cc

@@ -47,7 +47,7 @@ paddle::experimental::Tensor CreateTensorWithValue(
 
   auto meta = EagerUtils::autograd_meta(&out);
   if (is_leaf) {
-    auto accumulation_node = std::make_shared<GradNodeAccumulation>();
+    auto accumulation_node = std::make_shared<GradNodeAccumulation>(meta);
     meta->SetGradNode(accumulation_node);
     meta->SetStopGradient(false);
   }

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -1031,6 +1031,8 @@ static std::string GenerateGradNodeCreationContent(
     const std::string& output_name = output.name();
     const std::string& output_autograd_name = "p_autograd_" + output_name;
 
+    // Skip Intermediate Tensor
+
     if (output.duplicable()) {
       const char* GET_MULTI_AUTOGRAD_META_TEMPLATE =
           "  std::vector<egr::AutogradMeta*> %s = "
@@ -1145,6 +1147,8 @@ static std::string GenerateGradNodeCreationContent(
     const std::string& output_autograd_name = "p_autograd_" + output_name;
     size_t output_position = fwd_outputs_name_pos_map.at(output_name);
 
+    // Intermediate Tensor does not require SetHistory, nor RetainGrad
+
     if (output.duplicable()) {
       pass_stop_gradient_args += ", &" + output_autograd_name;
       const char* SET_OUT_RANK_TEMPLATE =

paddle/fluid/eager/autograd_meta.h

@@ -97,6 +97,7 @@ class AutogradMeta : public AbstractAutogradMeta {
             "Should Not set NULL as GradNode pointer, since "
             "our default Edge and autogradMeta has nullptr for "
             "grad node. Set Nullptr will lead error."));
+
     grad_node_ = grad_node;
   }
 

paddle/fluid/eager/grad_node_info.cc

@@ -53,7 +53,7 @@ void GradNodeBase::AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id) {
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       } else {
-        meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+        meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       }
@@ -76,7 +76,7 @@ void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     } else {
-      meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+      meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
       VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
               << this->name() << " to " << meta->GetMutableGradNode()->name();
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),

paddle/fluid/eager/tensor_wrapper.h

@@ -66,14 +66,13 @@ class TensorWrapper {
     }
 
     intermidiate_tensor_.set_name(tensor.name() + "@Saved");
-    PADDLE_ENFORCE_NOT_NULL(
-        EagerUtils::unsafe_autograd_meta(tensor),
-        paddle::platform::errors::Fatal(
-            "Full reserved Tensor should not have null autograd meta, since "
-            "tensor_wrapper is used to build backward info. There is no way "
-            "for us to build it with null autograd_meta."));
-    // copy output_rank
-    out_rank_info_ = EagerUtils::OutRankInfo(tensor);
+
+    // If an output is marked "intermedaite", we won't create
+    // autograd_meta for it.
+    // In that case, simply skip OutRankInfo Copy
+    if (EagerUtils::nullable_autograd_meta(tensor)) {
+      out_rank_info_ = EagerUtils::OutRankInfo(tensor);
+    }
   }
 
   paddle::experimental::Tensor recover(

paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc

@@ -17,11 +17,13 @@
 #include "gtest/gtest.h"
 
 #include "paddle/fluid/eager/accumulation/accumulation_node.h"
+#include "paddle/fluid/eager/api/utils/hook_utils.h"
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/fluid/eager/grad_node_info.h"
 #include "paddle/fluid/eager/grad_tensor_holder.h"
-#include "paddle/phi/api/lib/utils/allocator.h"
+#include "paddle/fluid/eager/utils.h"
 
+#include "paddle/phi/api/lib/utils/allocator.h"
 #include "paddle/phi/core/kernel_registry.h"
 
 // TODO(jiabin): remove nolint here!!!
@@ -37,7 +39,7 @@ TEST(AccumulationNode, Tensor) {
           .get(),
       meta);
   dt0->mutable_data<paddle::platform::float16>(
-      paddle::platform::CPUPlace())[0] = 10.0;
+      paddle::platform::CPUPlace())[0] = paddle::platform::float16(10.0f);
   paddle::experimental::Tensor et0 = paddle::experimental::Tensor(dt0);
 
   std::shared_ptr<phi::DenseTensor> dt1 = std::make_shared<phi::DenseTensor>(
@@ -47,84 +49,100 @@ TEST(AccumulationNode, Tensor) {
       meta);
 
   dt1->mutable_data<paddle::platform::float16>(
-      paddle::platform::CPUPlace())[0] = 20.0;
+      paddle::platform::CPUPlace())[0] = paddle::platform::float16(20.0f);
   paddle::experimental::Tensor et1 = paddle::experimental::Tensor(dt1);
 
+  std::shared_ptr<phi::DenseTensor> input_dt =
+      std::make_shared<phi::DenseTensor>(
+          std::make_unique<paddle::experimental::DefaultAllocator>(
+              paddle::platform::CPUPlace())
+              .get(),
+          meta);
+  paddle::experimental::Tensor input_et =
+      paddle::experimental::Tensor(input_dt);
+  auto grad_meta = EagerUtils::autograd_meta(&input_et);
+
+  // Initialize Grad Tensor
   std::shared_ptr<phi::DenseTensor> grad_dt =
       std::make_shared<phi::DenseTensor>(
           std::make_unique<paddle::experimental::DefaultAllocator>(
               paddle::platform::CPUPlace())
               .get(),
           meta);
-  paddle::experimental::Tensor grad_et = paddle::experimental::Tensor(grad_dt);
+  grad_dt->mutable_data<paddle::platform::float16>(
+      paddle::platform::CPUPlace())[0] = paddle::platform::float16(0.0f);
+  grad_meta->MutableGrad()->set_impl(grad_dt);
 
   // AccumulationNode
-  GradNodeAccumulation node = GradNodeAccumulation();
-
-  // Hook, RetainGrad
-  std::function<paddle::experimental::Tensor(
-      const paddle::experimental::Tensor&)>
-      hook = [&grad_et](const paddle::experimental::Tensor& t) {
-        grad_et.set_impl(t.impl());
-        return grad_et;
-      };
-  node.RetainGrad(hook);
+  auto node = std::make_shared<GradNodeAccumulation>(grad_meta);
+  grad_meta->SetGradNode(node);
+  grad_meta->SetStopGradient(false);
 
   // operator()
-  paddle::experimental::Tensor ret_et0 = node({{et0}})[0][0];
+  paddle::experimental::Tensor ret_et0 = node->operator()({{et0}})[0][0];
   auto* ret_et0_ptr =
       std::dynamic_pointer_cast<phi::DenseTensor>(ret_et0.impl())
           ->data<paddle::platform::float16>();
   CHECK_EQ(ret_et0_ptr[0], paddle::platform::float16(10.0f));
 
-  paddle::experimental::Tensor ret_et1 = node({{et1}})[0][0];
+  paddle::experimental::Tensor ret_et1 = node->operator()({{et1}})[0][0];
+
   auto* ret_et1_ptr =
       std::dynamic_pointer_cast<phi::DenseTensor>(ret_et1.impl())
           ->data<paddle::platform::float16>();
-  CHECK_EQ(ret_et1_ptr[0], paddle::platform::float16(30.0f));
+  CHECK_EQ(ret_et1_ptr[0], paddle::platform::float16(20.0f));
 
-  // Retain Grad
-  auto* ret_grad_et_ptr =
-      std::dynamic_pointer_cast<phi::DenseTensor>(grad_et.impl())
-          ->data<paddle::platform::float16>();
-  CHECK_EQ(ret_grad_et_ptr[0], paddle::platform::float16(30.0f));
+  // Check Retain Grad
+  CHECK_EQ(std::dynamic_pointer_cast<phi::DenseTensor>(et0.impl())
+               ->data<paddle::platform::float16>()[0],
+           paddle::platform::float16(10.0f));
+  paddle::experimental::Tensor* grad = EagerUtils::mutable_grad(input_et);
+  auto* grad_ptr = std::dynamic_pointer_cast<phi::DenseTensor>(grad->impl())
+                       ->data<paddle::platform::float16>();
+  CHECK_EQ(grad_ptr[0], paddle::platform::float16(30.0f));
 
   // Reduce Hook case 1: Call RegisterReduceHook and run operator()
   VLOG(6) << "Test Reduce Hook";
+  CHECK_EQ(std::dynamic_pointer_cast<phi::DenseTensor>(et0.impl())
+               ->data<paddle::platform::float16>()[0],
+           paddle::platform::float16(10.0f));
+
   auto reduce_hook_1 = [&](void) -> void {
-    auto* grad_et_ptr =
-        std::dynamic_pointer_cast<phi::DenseTensor>(grad_et.impl())
-            ->data<paddle::platform::float16>();
-    grad_et_ptr[0] = 36.0;
+    auto* input_et_ptr =
+        std::dynamic_pointer_cast<phi::DenseTensor>(input_et.impl())
+            ->mutable_data<paddle::platform::float16>(
+                paddle::platform::CPUPlace());
+    input_et_ptr[0] = 36.0;
     VLOG(6) << "Running Reduce Hook";
   };
 
-  node.RegisterReduceHook(reduce_hook_1);
+  node->RegisterReduceHook(reduce_hook_1);
 
   // operator()
-  paddle::experimental::Tensor _ret = node({{et0}})[0][0];
+  paddle::experimental::Tensor _ret = node->operator()({{et0}})[0][0];
 
   // Check operator() result, should be 36.0
   auto* _ret_ptr = std::dynamic_pointer_cast<phi::DenseTensor>(_ret.impl())
                        ->data<paddle::platform::float16>();
-  CHECK_EQ(_ret_ptr[0], paddle::platform::float16(36.0f));
+  CHECK_EQ(_ret_ptr[0], paddle::platform::float16(10.0f));
 
   // Check Retain Grad, should be 36.0
-  auto* _ret_grad_et_ptr =
-      std::dynamic_pointer_cast<phi::DenseTensor>(grad_et.impl())
+  auto* _ret_input_et_ptr =
+      std::dynamic_pointer_cast<phi::DenseTensor>(input_et.impl())
           ->data<paddle::platform::float16>();
-  CHECK_EQ(_ret_grad_et_ptr[0], paddle::platform::float16(36.0f));
+  CHECK_EQ(_ret_input_et_ptr[0], paddle::platform::float16(36.0f));
 
   // Reduce Hook case 2: Call RegisterReduceHook and ApplyReduceHooks directly
   VLOG(6) << "Test Reduce Hook";
   auto reduce_hook_2 = [&](void) -> void {
     auto* ret_et0_ptr = std::dynamic_pointer_cast<phi::DenseTensor>(et0.impl())
-                            ->data<paddle::platform::float16>();
+                            ->mutable_data<paddle::platform::float16>(
+                                paddle::platform::CPUPlace());
     ret_et0_ptr[0] = 100.0;  // set to 100.0
     VLOG(6) << "Running Reduce Hook";
   };
-  node.RegisterReduceHook(reduce_hook_2);
-  node.ApplyReduceHooks();
+  node->RegisterReduceHook(reduce_hook_2);
+  node->ApplyReduceHooks();
 
   // Check ApplyReduceHooks result
   CHECK_EQ(std::dynamic_pointer_cast<phi::DenseTensor>(et0.impl())

paddle/fluid/eager/tests/task_tests/backward_test.cc

@@ -59,22 +59,18 @@ TEST(Backward, SingleNodeEmptyGrad) {
     auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
     auto_grad_meta->SetStopGradient(false);
 
+    AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
+
     // Connect Tensor and AccumulationNode via AutoGradMeta
-    auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
+    auto acc_node_ptr =
+        std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
 
-    AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
     auto_grad_meta1->SetGradNode(
         std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
     auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
+    auto_grad_meta1->SetStopGradient(false);
 
-    egr_utils_api::RetainGradForTensor(leaf_tensor);
-
-    // Connect Node0 -> AccumulationNode via Edge
-    auto meta = egr::AutogradMeta();
-    meta.SetStopGradient(false);
-    meta.SetSingleOutRankWithSlot(0, 0);
-    meta.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res = {&meta};
+    std::vector<egr::AutogradMeta*> res = {auto_grad_meta1};
     node0_ptr->AddEdges(&res, 0);
   }
   std::vector<paddle::experimental::Tensor> outs = {target_tensor};
@@ -123,22 +119,17 @@ TEST(Backward, SingleNodeCustomGrad) {
         std::dynamic_pointer_cast<GradNodeBase>(node0_ptr));
     auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
     auto_grad_meta->SetStopGradient(false);
-    // Connect Tensor and AccumulationNode via AutoGradMeta
-    auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
 
     AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
+    // Connect Tensor and AccumulationNode via AutoGradMeta
+    auto acc_node_ptr =
+        std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
+
     auto_grad_meta1->SetGradNode(
         std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
     auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
-
-    egr_utils_api::RetainGradForTensor(leaf_tensor);
-
-    // Connect Node0 -> AccumulationNode via Edge
-    auto meta = egr::AutogradMeta();
-    meta.SetStopGradient(false);
-    meta.SetSingleOutRankWithSlot(0, 0);
-    meta.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res = {&meta};
+    auto_grad_meta1->SetStopGradient(false);
+    std::vector<egr::AutogradMeta*> res = {auto_grad_meta1};
     node0_ptr->AddEdges(&res, 0);
   }
 
@@ -201,22 +192,17 @@ TEST(Backward, LinearNodes) {
     std::vector<egr::AutogradMeta*> res0 = {&meta0};
     node0_ptr->AddEdges(&res0, 0);
 
+    AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
     // Connect Tensor and AccumulationNode via AutoGradMeta
-    auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
+    auto acc_node_ptr =
+        std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta1);
 
-    AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
     auto_grad_meta1->SetGradNode(
         std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
     auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
 
-    egr_utils_api::RetainGradForTensor(leaf_tensor);
-
-    // Connect Node1 -> AccumulationNode via Edge
-    auto meta1 = egr::AutogradMeta();
-    meta1.SetStopGradient(false);
-    meta1.SetSingleOutRankWithSlot(0, 0);
-    meta1.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res1 = {&meta1};
+    auto_grad_meta1->SetStopGradient(false);
+    std::vector<egr::AutogradMeta*> res1 = {auto_grad_meta1};
     node1_ptr->AddEdges(&res1, 0);
   }
 
@@ -311,22 +297,17 @@ TEST(Backward, WithAccumulation) {
     std::vector<egr::AutogradMeta*> res1 = {&meta1};
     node1_ptr->AddEdges(&res1, 0);
 
+    AutogradMeta* auto_grad_meta2 = EagerUtils::autograd_meta(&leaf_tensor);
     // Connect Tensor and AccumulationNode via AutoGradMeta
-    auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
+    auto acc_node_ptr =
+        std::make_shared<egr::GradNodeAccumulation>(auto_grad_meta2);
 
-    AutogradMeta* auto_grad_meta2 = EagerUtils::autograd_meta(&leaf_tensor);
     auto_grad_meta2->SetGradNode(
         std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
     auto_grad_meta2->SetSingleOutRankWithSlot(0, 0);
 
-    egr_utils_api::RetainGradForTensor(leaf_tensor);
-
-    // Connect Node2 -> AccumulationNode via Edge
-    auto meta2 = egr::AutogradMeta();
-    meta2.SetStopGradient(false);
-    meta2.SetSingleOutRankWithSlot(0, 0);
-    meta2.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res2 = {&meta2};
+    auto_grad_meta2->SetStopGradient(false);
+    std::vector<egr::AutogradMeta*> res2 = {auto_grad_meta2};
     node2_ptr->AddEdges(&res2, 0);
   }
 

paddle/fluid/eager/tests/task_tests/cross_batch_accumulation_test.cc

@@ -46,34 +46,26 @@ TEST(CrossBatchAccumulation, SingleScaleNode) {
   paddle::experimental::Tensor& target_tensor = target_tensors[0];
 
   paddle::experimental::Tensor leaf_tensor = paddle::experimental::Tensor();
-  {
-    auto scale_node_ptr = std::make_shared<GradNodeScale>(1, 1);
-    scale_node_ptr->SetAttributes_scale(5.0 /*scale*/);
-
-    scale_node_ptr->SetDefaultGradInOutMeta();
-
-    auto acc_node_ptr = std::make_shared<GradNodeAccumulation>();
-
-    AutogradMeta* auto_grad_meta = EagerUtils::autograd_meta(&target_tensor);
-    auto_grad_meta->SetGradNode(
-        std::dynamic_pointer_cast<GradNodeBase>(scale_node_ptr));
-    auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
-    auto_grad_meta->SetStopGradient(false);
-    egr_utils_api::RetainGradForTensor(target_tensor);  // result: 1.0
-
-    auto meta = AutogradMeta();
-    meta.SetSingleOutRankWithSlot(0, 0);
-    meta.SetStopGradient(false);
-    meta.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res = {&meta};
-    scale_node_ptr->AddEdges(&res, 0);
-
-    AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
-    auto_grad_meta1->SetGradNode(
-        std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
-    auto_grad_meta1->SetSingleOutRankWithSlot(0, 0);
-    egr_utils_api::RetainGradForTensor(leaf_tensor);
-  }
+
+  auto scale_node_ptr = std::make_shared<GradNodeScale>(1, 1);
+  scale_node_ptr->SetAttributes_scale(5.0 /*scale*/);
+
+  scale_node_ptr->SetDefaultGradInOutMeta();
+
+  AutogradMeta* auto_grad_meta = EagerUtils::autograd_meta(&target_tensor);
+  auto_grad_meta->SetGradNode(
+      std::dynamic_pointer_cast<GradNodeBase>(scale_node_ptr));
+  auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
+  auto_grad_meta->SetStopGradient(false);
+  egr_utils_api::RetainGradForTensor(target_tensor);  // result: 1.0
+
+  AutogradMeta* meta = EagerUtils::autograd_meta(&leaf_tensor);
+  auto acc_node_ptr = std::make_shared<GradNodeAccumulation>(meta);
+  meta->SetStopGradient(false);
+  meta->SetSingleOutRankWithSlot(0, 0);
+  meta->SetGradNode(acc_node_ptr);
+  std::vector<egr::AutogradMeta*> res = {meta};
+  scale_node_ptr->AddEdges(&res, 0);
 
   RunBackward(target_tensors, {});
 

paddle/fluid/eager/tests/task_tests/hook_test.cc

@@ -79,9 +79,6 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
   // Set grad in/out meta for node0
   scale_node_ptr->SetDefaultGradInOutMeta();
 
-  // Create AccumulationNode
-  auto acc_node_ptr = std::make_shared<GradNodeAccumulation>();
-
   // Connect Input Tensor and ScaleNode via AutoGradMeta
   // Apply RetainGrad
   {
@@ -102,16 +99,8 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
     egr_utils_api::RegisterGradientHookForTensor(target_tensor, hook);
     egr_utils_api::RetainGradForTensor(
         target_tensor);  // result: 1.0 + 3.0 = 4.0
-  }
-
-  // Connect ScaleNode -> AccumulationNode via Edge
-  {
-    auto meta = AutogradMeta();
-    meta.SetStopGradient(false);
-    meta.SetSingleOutRankWithSlot(0, 0);
-    meta.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res = {&meta};
-    scale_node_ptr->AddEdges(&res, 0);
+    egr_utils_api::RetainGradForTensor(
+        target_tensor);  // result: 1.0 + 3.0 = 4.0
   }
 
   // Retain Grad for leaf tensor1
@@ -123,9 +112,16 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
         hook = &hook_function;
 
     auto auto_grad_meta = std::make_shared<AutogradMeta>();
-    auto_grad_meta->SetGradNode(
-        std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
+
+    auto acc_node_ptr =
+        std::make_shared<GradNodeAccumulation>(auto_grad_meta.get());
+
+    auto_grad_meta->SetStopGradient(false);
+    auto_grad_meta->SetGradNode(acc_node_ptr);
     auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
+    std::vector<egr::AutogradMeta*> res = {auto_grad_meta.get()};
+    scale_node_ptr->AddEdges(&res, 0);
+
     leaf_tensor.set_autograd_meta(
         std::dynamic_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
             auto_grad_meta));
@@ -160,8 +156,6 @@ TEST(RetainGrad, HookAfterRetainGrad) {
   scale_node_ptr->SetAttributes_scale(5.0 /*scale*/);
   // Set grad in/out meta for node0
   scale_node_ptr->SetDefaultGradInOutMeta();
-  // Create AccumulationNode
-  auto acc_node_ptr = std::make_shared<GradNodeAccumulation>();
 
   // Connect Input Tensor and ScaleNode via AutoGradMeta
   // Apply RetainGrad
@@ -184,16 +178,6 @@ TEST(RetainGrad, HookAfterRetainGrad) {
     egr_utils_api::RegisterGradientHookForTensor(target_tensor, hook);
   }
 
-  // Connect ScaleNode -> AccumulationNode via Edge
-  {
-    auto meta = AutogradMeta();
-    meta.SetStopGradient(false);
-    meta.SetSingleOutRankWithSlot(0, 0);
-    meta.SetGradNode(acc_node_ptr);
-    std::vector<egr::AutogradMeta*> res = {&meta};
-    scale_node_ptr->AddEdges(&res, 0);
-  }
-
   // Retain Grad for leaf tensor1
   paddle::experimental::Tensor leaf_tensor = paddle::experimental::Tensor();
   {
@@ -203,17 +187,18 @@ TEST(RetainGrad, HookAfterRetainGrad) {
         hook = &hook_function;
 
     auto auto_grad_meta = std::make_shared<AutogradMeta>();
-    auto_grad_meta->SetGradNode(
-        std::dynamic_pointer_cast<GradNodeBase>(acc_node_ptr));
+    auto acc_node_ptr =
+        std::make_shared<GradNodeAccumulation>(auto_grad_meta.get());
+    auto_grad_meta->SetGradNode(acc_node_ptr);
+    auto_grad_meta->SetStopGradient(false);
+    std::vector<egr::AutogradMeta*> res = {auto_grad_meta.get()};
+    scale_node_ptr->AddEdges(&res, 0);
+
     auto_grad_meta->SetSingleOutRankWithSlot(0, 0);
     leaf_tensor.set_autograd_meta(
         std::dynamic_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
             auto_grad_meta));
 
-    egr_utils_api::RetainGradForTensor(
-        leaf_tensor);  // RetainGrad for leaf tensor gets
-                       // postponed, result: 4.0*5.0 + 3.0 =
-                       // 23.0
     egr_utils_api::RegisterGradientHookForTensor(leaf_tensor, hook);
   }
 

paddle/fluid/eager/utils.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/eager/utils.h"
+#include "paddle/fluid/eager/accumulation/accumulation_node.h"
 #include "paddle/fluid/eager/api/utils/global_utils.h"
 #include "paddle/fluid/eager/api/utils/hook_utils.h"
 #include "paddle/fluid/eager/tensor_wrapper.h"
@@ -21,7 +22,6 @@
 #include "paddle/phi/common/layout.h"
 #include "paddle/phi/core/tensor_meta.h"
 
-#include "paddle/fluid/eager/accumulation/accumulation_node.h"
 #include "paddle/fluid/framework/data_layout.h"
 #include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/framework/variable.h"
@@ -109,6 +109,16 @@ std::shared_ptr<GradNodeBase> EagerUtils::grad_node(
   }
 }
 
+paddle::experimental::Tensor* EagerUtils::mutable_grad(
+    const paddle::experimental::Tensor& target) {
+  auto* meta = nullable_autograd_meta(target);
+  if (meta) {
+    return meta->MutableGrad();
+  } else {
+    return nullptr;
+  }
+}
+
 void EagerUtils::SetHistory(std::vector<AutogradMeta*>* autograd_metas,
                             const std::shared_ptr<GradNodeBase>& grad_node) {
   for (const auto& autograd_meta : *autograd_metas) {
@@ -342,7 +352,8 @@ std::shared_ptr<egr::GradNodeBase> EagerUtils::GetGradAccumulationNode(
   } else {
     if (!autograd_ptr->StopGradient()) {
       VLOG(6) << "Add GradNodeAccumulation for tensor: " << tensor.name();
-      autograd_ptr->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+      autograd_ptr->SetGradNode(
+          std::make_shared<egr::GradNodeAccumulation>(autograd_ptr));
       return autograd_ptr->GetMutableGradNode();
     } else {
       return nullptr;

paddle/fluid/eager/utils.h

@@ -102,6 +102,8 @@ class EagerUtils {
 
   static std::shared_ptr<GradNodeBase> grad_node(
       const paddle::experimental::Tensor& target);
+  static paddle::experimental::Tensor* mutable_grad(
+      const paddle::experimental::Tensor& target);
 
   // Set history is used to set backward info during forward process, it will
   // set forward var's autograd meta's grad node as current backward node.

paddle/fluid/pybind/eager.cc

@@ -86,7 +86,8 @@ void EmptyTensorInitializer(TensorObject* self, const std::string& name,
   if (!autograd_meta->GetMutableGradNode()) {
     VLOG(3) << "Tensor(" << name
             << ") have not GradNode, add GradNodeAccumulation for it.";
-    autograd_meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+    autograd_meta->SetGradNode(
+        std::make_shared<egr::GradNodeAccumulation>(autograd_meta));
   }
 }
 

paddle/fluid/pybind/eager_method.cc

@@ -177,7 +177,7 @@ static PyObject* tensor_retain_grads(TensorObject* self, PyObject* args,
     if (!meta->GetMutableGradNode()) {
       VLOG(6) << "Make grad node of tensor: " << self->tensor.name()
               << "become accumulation node";
-      meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
+      meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
     }
     egr::egr_utils_api::RetainGradForTensor(self->tensor);
   }
@@ -199,17 +199,12 @@ static PyObject* tensor_clear_gradient(TensorObject* self, PyObject* args,
 
   paddle::experimental::Tensor* grad;
   if (egr::egr_utils_api::IsLeafTensor(self->tensor)) {
-    // Add RetainGrad as PostHook to AccumulationNode
-    std::shared_ptr<egr::GradNodeBase> grad_node =
-        egr::EagerUtils::grad_node(self->tensor);
-    PADDLE_ENFORCE(
-        grad_node.get() != nullptr,
-        paddle::platform::errors::Fatal("Detected NULL grad_node"
-                                        "Leaf tensor should have had grad_node "
-                                        "with type: GradNodeAccumulation"));
-    auto accumulation_grad_node =
-        std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
-    grad = accumulation_grad_node->Grad();
+    grad = egr::EagerUtils::mutable_grad(self->tensor);
+    PADDLE_ENFORCE(grad != nullptr,
+                   paddle::platform::errors::Fatal(
+                       "Detected NULL grad"
+                       "Please check if you have manually cleared"
+                       "the grad inside autograd_meta"));
   } else {
     auto meta = egr::EagerUtils::unsafe_autograd_meta(self->tensor);
     grad = meta->MutableGrad();
@@ -248,19 +243,15 @@ static PyObject* tensor__zero_grads(TensorObject* self, PyObject* args,
 
   if (egr::egr_utils_api::IsLeafTensor(self->tensor)) {
     // Add RetainGrad as PostHook to AccumulationNode
-    std::shared_ptr<egr::GradNodeBase> grad_node =
-        egr::EagerUtils::grad_node(self->tensor);
-    PADDLE_ENFORCE(
-        grad_node.get() != nullptr,
-        paddle::platform::errors::Fatal("Detected NULL grad_node"
-                                        "Leaf tensor should have had grad_node "
-                                        "with type: GradNodeAccumulation"));
-    auto accumulation_grad_node =
-        std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
-    if (accumulation_grad_node->Grad()->initialized()) {
-      accumulation_grad_node->Grad()->set_impl(
-          paddle::experimental::zeros_like(*(accumulation_grad_node->Grad()))
-              .impl());
+    paddle::experimental::Tensor* grad =
+        egr::EagerUtils::mutable_grad(self->tensor);
+    PADDLE_ENFORCE(grad != nullptr,
+                   paddle::platform::errors::Fatal(
+                       "Detected NULL grad"
+                       "Please check if you have manually cleared"
+                       "the grad inside autograd_meta"));
+    if (grad->initialized()) {
+      grad->set_impl(paddle::experimental::zeros_like(*(grad)).impl());
     }
   } else {
     auto meta = egr::EagerUtils::unsafe_autograd_meta(self->tensor);

paddle/fluid/pybind/eager_properties.cc

@@ -70,26 +70,13 @@ PyObject* tensor_properties_get_stop_gradient(TensorObject* self,
 
 PyObject* tensor_properties_get_grad(TensorObject* self, void* closure) {
   EAGER_TRY
-  if (egr::egr_utils_api::IsLeafTensor(self->tensor)) {
-    std::shared_ptr<egr::GradNodeBase> grad_node =
-        egr::EagerUtils::grad_node(self->tensor);
-    PADDLE_ENFORCE(
-        grad_node.get() != nullptr,
-        paddle::platform::errors::Fatal("Detected NULL grad_node"
-                                        "Leaf tensor should have had grad_node "
-                                        "with type: GradNodeAccumulation"));
-    auto accumulation_grad_node =
-        std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
-    return ToPyObject(*accumulation_grad_node->Grad());
+  VLOG(6) << "Get grad for tensor: " << self->tensor.name();
+  auto meta = egr::EagerUtils::nullable_autograd_meta(self->tensor);
+  if (meta) {
+    return ToPyObject(meta->Grad());
   } else {
-    VLOG(6) << "Get grad for tensor: " << self->tensor.name();
-    auto meta = egr::EagerUtils::nullable_autograd_meta(self->tensor);
-    if (meta) {
-      return ToPyObject(meta->Grad());
-    } else {
-      Py_INCREF(Py_None);
-      return Py_None;
-    }
+    Py_INCREF(Py_None);
+    return Py_None;
   }
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
@@ -101,16 +88,15 @@ int tensor_properties_set_grad(TensorObject* self, PyObject* value,
   PADDLE_ENFORCE(
       egr::egr_utils_api::IsLeafTensor(self->tensor),
       paddle::platform::errors::Fatal("Only leaf Tensor can be set grad."));
-  std::shared_ptr<egr::GradNodeBase> grad_node =
-      egr::EagerUtils::grad_node(self->tensor);
-  PADDLE_ENFORCE(
-      grad_node.get() != nullptr,
-      paddle::platform::errors::Fatal("Detected NULL grad_node"
-                                      "Leaf tensor should have had grad_node "
-                                      "with type: GradNodeAccumulation"));
-  auto accumulation_grad_node =
-      std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
-  accumulation_grad_node->Grad()->copy_(src, true);
+
+  paddle::experimental::Tensor* grad =
+      egr::EagerUtils::mutable_grad(self->tensor);
+  PADDLE_ENFORCE(grad != nullptr,
+                 paddle::platform::errors::Fatal(
+                     "Detected NULL grad"
+                     "Please check if you have manually cleared"
+                     "the grad inside autograd_meta"));
+  grad->copy_(src, true);
   return 0;
   EAGER_CATCH_AND_THROW_RETURN_ZERO
 }

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

@@ -606,8 +606,12 @@ class OpTest(unittest.TestCase):
 
             if is_input:
                 v = self._create_var_from_numpy(np_value_temp)
+
                 if if_return_inputs_grad_dict:
                     v.stop_gradient = False
+                    if _in_eager_mode():
+                        v.retain_grads()
+
                 if has_lod:
                     v.value().get_tensor().set_recursive_sequence_lengths(
                         lod_temp)
@@ -618,7 +622,6 @@ class OpTest(unittest.TestCase):
                     type=core.VarDesc.VarType.LOD_TENSOR,
                     persistable=False,
                     stop_gradient=False)
-
             return v
 
         # prepare variable for input or output
@@ -681,7 +684,6 @@ class OpTest(unittest.TestCase):
             # prepare input variable
             inputs = self.append_input_output_for_dygraph(op_proto, self.inputs,
                                                           True, False, block)
-
             # prepare output variable
             outputs = self.append_input_output_for_dygraph(
                 op_proto, self.outputs, False, False, block)
@@ -1741,6 +1743,7 @@ class OpTest(unittest.TestCase):
                 for attrs_name in self.attrs:
                     if self.attrs[attrs_name] is not None:
                         attrs_outputs[attrs_name] = self.attrs[attrs_name]
+
             block.append_op(
                 type=self.op_type,
                 inputs=inputs,
@@ -1817,7 +1820,9 @@ class OpTest(unittest.TestCase):
                         inputs={"X": loss_sum},
                         outputs={"Out": loss},
                         attrs={'scale': 1.0 / float(len(avg_sum))})
+
                 loss.backward()
+
                 fetch_list_grad = []
                 for inputs_to_check_name in inputs_to_check:
                     a = inputs_grad_dict[inputs_to_check_name].gradient()