backward.cc

// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "paddle/fluid/eager/backward.h"
#include <queue>

#include "paddle/fluid/eager/autograd_meta.h"
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/grad_tensor_holder.h"
#include "paddle/fluid/eager/utils.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/fluid/platform/profiler/event_tracing.h"

#include "glog/logging.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/errors.h"
#include "paddle/phi/kernels/autotune/switch_autotune.h"

namespace egr {

/*
* GeneralGrad is Helpper class to implement custom grad operation between
* outputs and inputs.
*
* **/
class GeneralGrad {
 public:
  static GeneralGrad& Instance() { return *general_grad_; }

  // Get inputs's / no_grad_vars's GradNodes and InputMeta Info
  void GetTargetNodesInfo(
      const std::vector<paddle::experimental::Tensor>& inputs,
      bool is_no_grad_vars) {
    std::string msg = is_no_grad_vars ? "no_grad_vars" : "inputs";
    VLOG(6) << "Running in GetTargetNodesInfo.";
    if (!inputs.empty()) {
      VLOG(6) << msg << " are not empty.";
      size_t num_inputs = inputs.size();
      for (size_t i = 0; i < num_inputs; i++) {
        AutogradMeta* auto_grad_meta =
            EagerUtils::unsafe_autograd_meta(inputs[i]);
        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].get();
        } else {
          VLOG(6) << "Unable to find target node in "
                     "orig_to_copied_node_mapping_, likely indicating an "
                     "unused input";
        }

        PADDLE_ENFORCE_NOT_NULL(target_node,
                                paddle::platform::errors::Fatal(
                                    "There is no grad op for %s:[%d] or it's"
                                    "stop_gradient=True.",
                                    msg, i));
        if (is_no_grad_vars) {
          (no_grad_var_nodes_inputmeta_map)[target_node] = auto_grad_meta;
        } else {  // normal input
          (input_target_nodes_inputmeta_map)[target_node] = auto_grad_meta;
        }
      }
    }
  }

  // Purify potential_startup_nodes, remove nodes those are the same as
  // input_target_nodes
  void PurifyPotentialStartUpNodes() {
    VLOG(6) << "Running in PurifyPotentialStartUpNodes";
    if (input_target_nodes_inputmeta_map.empty()) return;
    std::unordered_set<GradNodeBase*> potential_startup_nodes_to_be_erased;
    for (auto startup_op : potential_startup_nodes) {
      auto iter = input_target_nodes_inputmeta_map.find(startup_op);
      if (iter != input_target_nodes_inputmeta_map.end()) {
        potential_startup_nodes_to_be_erased.emplace(iter->first);
      }
    }
    if (!potential_startup_nodes_to_be_erased.empty()) {
      for (auto nodes : potential_startup_nodes_to_be_erased) {
        potential_startup_nodes.erase(nodes);
      }
    }
  }

  // Remove some nodes those doesn't need to be
  // stored in potential_stop_nodes、potential_startup_nodes
  void UpdateGraphInfo() {
    // Updated potential_sotp_nodes by depending_nodes,
    // make sure the path from root to target_node is ok
    std::unordered_set<GradNodeBase*> _startup_ops;
    VLOG(6) << "Running in UpdateGraphInfo";
    std::queue<GradNodeBase*> queue;
    for (auto& target_nodes_inputmeta_pair : input_target_nodes_inputmeta_map) {
      queue.emplace(target_nodes_inputmeta_pair.first);
    }

    while (!queue.empty()) {
      auto* target_node = queue.front();
      queue.pop();
      if (!(depending_nodes)[target_node].empty()) {
        auto precedding_nodes = (depending_nodes)[target_node];
        for (auto pre_nodes : precedding_nodes) {
          queue.emplace(pre_nodes);
          if (potential_stop_nodes.find(pre_nodes) !=
              potential_stop_nodes.end()) {
            potential_stop_nodes.erase(pre_nodes);
          }
        }
      } else {  // startup_ops have no precedding nodes
        VLOG(6) << "Emplace _startup_ops";
        _startup_ops.emplace(target_node);
      }
    }
    // Purify potential_startup_nodes again, remove some
    // potential startup_nodes that unreach to input target nodes
    if (!_startup_ops.empty()) {
      std::unordered_set<GradNodeBase*> potential_startup_nodes_to_be_erased;
      for (auto node : potential_startup_nodes) {
        if (_startup_ops.count(node) == 0) {
          VLOG(6) << "Set up potential_startup_nodes_to_be_erased";
          potential_startup_nodes_to_be_erased.emplace(node);
        }
      }
      if (!potential_startup_nodes_to_be_erased.empty()) {
        for (auto node : potential_startup_nodes_to_be_erased) {
          VLOG(6) << "Erase nodes in potential_startup_nodes_to_be_erased";
          potential_startup_nodes.erase(node);
        }
      }
    }
  }

  // Get Graph Info Betweent input target GradNode and outputs，
  // record depending_nodes、potential_stop_nodes、potential_startup_nodes
  void GetGraphInfoBetweenTargets(const std::queue<GradNodeBase*>& init_queue) {
    VLOG(6) << "Runing In GetGraphInfoBetweenTargets";

    // Calculate in_degree for each node
    std::unordered_map<GradNodeBase*, int> node_in_degree_map;

    // Copy nodes
    std::queue<GradNodeBase*> queue = init_queue;
    std::unordered_set<GradNodeBase*> visited;

    // Visit each node exactly once in any order
    while (!queue.empty()) {
      GradNodeBase* node = queue.front();
      queue.pop();

      if (visited.count(node)) {
        continue;
      }
      visited.insert(node);

      // Check node is target_nodes or not, if node is not target_node,
      // all the next_node will be marked in potential_stop_nodes
      bool is_potential_stop_nodes =
          input_target_nodes_inputmeta_map.count(node);

      // Find and append next nodes
      const paddle::small_vector<std::vector<GradSlotMeta>,
                                 kSlotSmallVectorSize>& metas =
          node->OutputMeta();
      for (const auto& meta_list : metas) {
        for (const GradSlotMeta& meta : meta_list) {
          const auto& edge = meta.GetEdge();
          GradNodeBase* next_node = edge.GetMutableGradNode().get();

          // Next node could be nullptr if it is leaf tensor with no
          // AccumulationNode attached
          // Or it could also originated from dispensable inputs
          if (!next_node) continue;

          // if node not in input_target_nodes,
          // all the next_nodes of current node will be inserted to
          // potential_stop_node
          if (is_potential_stop_nodes) {
            potential_stop_nodes.emplace(next_node);
          }

          // Update in_degree
          if (!node_in_degree_map.count(next_node))
            node_in_degree_map[next_node] = 0;
          node_in_degree_map[next_node]++;

          // Record depending relationship
          (depending_nodes)[next_node].emplace(node);
          queue.push(next_node);
        }
      }
    }
    // Update Graph Info, remove some nodes in
    // potential_stop_nodes、potential_startup_nodes、
    UpdateGraphInfo();
  }

  void ModifyReadyQueue(std::queue<GradNodeBase*>* queue) {
    std::queue<GradNodeBase*> tmp_queue;
    for (auto nodes : potential_startup_nodes) {
      tmp_queue.emplace(nodes);
    }
    tmp_queue.swap(*queue);
  }

  // Set result for input target grad_var when potential_startup_nodes is empty
  void SetResultForInputTargetVar(
      const std::unordered_map<GradNodeBase*,
                               std::unique_ptr<GradTensorHolder>>&
          node_input_buffers_dict) {
    if (potential_startup_nodes.size() == 0) {
      for (auto input_target_node : *GetInPutTargetNodesInputMetaMap()) {
        // out rank_info of forward op
        auto rank_info = input_target_node.second->OutRankInfo();
        auto iter = node_input_buffers_dict.find(input_target_node.first);
        if (iter != node_input_buffers_dict.end()) {
          auto& target_result =
              (iter->second)->Buffers()[rank_info.first][rank_info.second];
          // save the target result
          results_map[input_target_node.first] = target_result;
        }
      }
    }
  }

  // Set input target grad_var from node_input_buffer by inputmeta
  void SetResultForInputTargetVar(GradTensorHolder input_buffers,
                                  GradNodeBase* node) {
    auto iter = GetInPutTargetNodesInputMetaMap()->find(node);
    if (iter != GetInPutTargetNodesInputMetaMap()->end()) {
      VLOG(6) << "Get target result by by inputmeta";
      // out rank_info of forward op
      auto rank_info = (iter->second)->OutRankInfo();
      // rank_info is a pair, first means slot_id, second means rank.
      auto& target_result =
          input_buffers.Buffers()[rank_info.first][rank_info.second];
      // save the target result
      results_map[node] = target_result;
    }
  }

  std::vector<paddle::experimental::Tensor> GetResults(
      const std::vector<paddle::experimental::Tensor>& inputs,
      bool allow_unused, bool create_graph) {
    VLOG(6) << "Running in GetResults";
    if (inputs.empty()) return {};

    std::vector<paddle::experimental::Tensor> results;
    results.reserve(inputs.size());

    for (size_t i = 0; i < inputs.size(); ++i) {
      auto& input = inputs[i];
      AutogradMeta* auto_grad_meta = EagerUtils::unsafe_autograd_meta(input);

      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].get();
      } else {
        VLOG(6) << "Unable to find target node in "
                   "orig_to_copied_node_mapping_, likely indicating an unused "
                   "input";
      }

      auto iter = results_map.find(target_node);
      if (iter != results_map.end()) {
        // set StopGradient = !create_graph
        AutogradMeta* tensor_auto_grad_meta =
            EagerUtils::autograd_meta(&(iter->second));
        tensor_auto_grad_meta->SetStopGradient(!create_graph);
        results.emplace_back(iter->second);
      } else {
        PADDLE_ENFORCE_EQ(allow_unused, true,
                          paddle::platform::errors::InvalidArgument(
                              "The %d-th input does not appear in the backward "
                              "graph. Please check the input tensor or set "
                              "allow_unused=True to get None result.",
                              i));
        results.emplace_back();
      }
    }
    Clear();
    return results;
  }

  void PreparedForGeneralGrad(
      const std::vector<paddle::experimental::Tensor>& inputs,
      const std::vector<paddle::experimental::Tensor>& no_grad_vars,
      std::queue<GradNodeBase*>* queue,
      const std::unordered_map<GradNodeBase*,
                               std::unique_ptr<GradTensorHolder>>&
          node_input_buffers_dict) {
    // Get no_grad_vars's GradNodes and InputMeta Info
    GetTargetNodesInfo(no_grad_vars, true /* is_no_grad_vars */);
    // Get inputs's GradNodes and InputMeta Info
    GetTargetNodesInfo(inputs, false /* is_no_grad_vars */);
    // Purify potential_startup_ops, remove those nodes that are the same as
    // input_target_nodes
    PurifyPotentialStartUpNodes();
    // Get Graph Info Betweent input target gradnode and outputs
    // Record the depending_nodes and
    // potential_stop_nodes、potential_startup_nodes
    GetGraphInfoBetweenTargets(*queue);
    // Reset queue. Queue is empty only when
    // 1.input equals to output. 2.input can not reach to output.
    ModifyReadyQueue(queue);
    // Set result for input target grad_var when queue is empty
    if (queue->empty()) SetResultForInputTargetVar(node_input_buffers_dict);
  }

  bool IsPotentialStopNodes(GradNodeBase* node) {
    return potential_stop_nodes.count(node);
  }

  std::unordered_map<GradNodeBase*, AutogradMeta*>*
  GetNoGradVarNodesInputMetaMap() {
    return &no_grad_var_nodes_inputmeta_map;
  }

  std::unordered_map<GradNodeBase*, AutogradMeta*>*
  GetInPutTargetNodesInputMetaMap() {
    return &input_target_nodes_inputmeta_map;
  }

  std::unordered_set<GradNodeBase*>* GetPotentialStopNodes() {
    return &potential_stop_nodes;
  }

  std::unordered_set<GradNodeBase*>* GetPotentialStartupNodes() {
    return &potential_startup_nodes;
  }

  void Clear() {
    no_grad_var_nodes_inputmeta_map.clear();
    input_target_nodes_inputmeta_map.clear();
    potential_startup_nodes.clear();
    potential_stop_nodes.clear();
    depending_nodes.clear();
    results_map.clear();
    copied_grad_nodes_.clear();
    orig_to_copied_node_mapping_.clear();
  }

  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()].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;
    copied_grad_nodes_.push_back(copied_node);

    return copied_node.get();
  }

  void ReconstructBackwardGraph(
      const std::queue<GradNodeBase*>& orig_init_queue) {
    std::queue<GradNodeBase*> queue = orig_init_queue;
    std::unordered_set<GradNodeBase*> visited;

    // BFS and recursively copy the grad nodes
    while (!queue.empty()) {
      GradNodeBase* orig_node = queue.front();
      queue.pop();
      if (visited.count(orig_node)) {
        continue;
      }
      visited.insert(orig_node);

      PADDLE_ENFORCE(
          orig_to_copied_node_mapping_.count(orig_node),
          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].get();

      const paddle::small_vector<std::vector<GradSlotMeta>,
                                 kSlotSmallVectorSize>& orig_meta =
          orig_node->OutputMeta();
      paddle::small_vector<std::vector<GradSlotMeta>, kSlotSmallVectorSize>&
          copied_edges = copied_node->MutableOutputMeta();
      for (size_t i = 0; i < orig_meta.size(); i++) {
        for (size_t j = 0; j < orig_meta[i].size(); j++) {
          const Edge& orig_edge = orig_meta[i][j].GetEdge();
          Edge& copied_edge = copied_edges[i][j].GetMutableEdge();

          std::shared_ptr<GradNodeBase> orig_next_node =
              orig_edge.GetMutableGradNode();
          if (!orig_next_node) continue;

          // Copy Next Node
          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()];

          } else {
            copied_next_node = orig_next_node->Copy();
            orig_to_copied_node_mapping_[orig_next_node.get()] =
                copied_next_node;
            copied_grad_nodes_.push_back(copied_next_node);
          }

          // Update Edge's Grad Node
          copied_edge.SetGradNode(copied_next_node);

          // Update BFS queue
          queue.push(orig_next_node.get());
        }
      }
    }
  }

 private:
  GeneralGrad() = default;
  static GeneralGrad* general_grad_;
  // no_grad_vars's GradNode and GradNode's InputMeta.
  std::unordered_map<GradNodeBase*, AutogradMeta* /* InputMeta */>
      no_grad_var_nodes_inputmeta_map;
  // inputs's GradNode and GradNode's InputMeta.
  std::unordered_map<GradNodeBase*, AutogradMeta* /* InputMeta */>
      input_target_nodes_inputmeta_map;
  // Record all the potential startup_nodes, will be changed.
  std::unordered_set<GradNodeBase*> potential_startup_nodes;
  // Record all the potential stop nodes, will be changed.
  std::unordered_set<GradNodeBase*> potential_stop_nodes;
  std::unordered_map<GradNodeBase* /* next node */,
                     std::unordered_set<GradNodeBase*> /* pre nodes */>
      depending_nodes;
  std::unordered_map<GradNodeBase*, paddle::experimental::Tensor> results_map;

  std::vector<std::shared_ptr<GradNodeBase>> copied_grad_nodes_;
  std::unordered_map<GradNodeBase*, std::shared_ptr<GradNodeBase>>
      orig_to_copied_node_mapping_;

  DISABLE_COPY_AND_ASSIGN(GeneralGrad);
};

std::unordered_map<GradNodeBase*, int> getInDegreeMap(
    const std::queue<GradNodeBase*>& init_queue) {
  // Calculate in_degree for each node
  // We can completely remove this pass, if in_degree were set during forward
  // pass
  std::unordered_map<GradNodeBase*, int> node_in_degree_map;

  // Copy nodes
  std::queue<GradNodeBase*> queue = init_queue;
  std::unordered_set<GradNodeBase*> visited;

  // Visit each node exactly once in any order
  while (!queue.empty()) {
    GradNodeBase* node = queue.front();
    queue.pop();

    if (visited.count(node)) {
      continue;
    }
    visited.insert(node);

    PADDLE_ENFORCE_NOT_NULL(
        node,
        paddle::platform::errors::Fatal(
            "We got null node when we traverse the backward graph, and this "
            "should not happened please check your code and contact us."));
    // Find and append next nodes
    const paddle::small_vector<std::vector<GradSlotMeta>, kSlotSmallVectorSize>&
        metas = node->OutputMeta();
    for (const auto& meta_list : metas) {
      for (const GradSlotMeta& meta : meta_list) {
        const auto& edge = meta.GetEdge();
        GradNodeBase* next_node = edge.GetMutableGradNode().get();
        // Next node could be nullptr if it is leaf tensor with no
        // AccumulationNode attached
        // Or it could also originated from dispensable inputs
        if (!next_node) continue;

        // Update in_degree
        if (!node_in_degree_map.count(next_node))
          node_in_degree_map[next_node] = 0;
        node_in_degree_map[next_node]++;
        queue.push(next_node);
      }
    }
  }

  return node_in_degree_map;
}

// Enforce GradNode has TensorWrappers as Input
void EnforceGradNodeHasInput(GradNodeBase* node) {
  VLOG(6) << "Running in EnforceGradNodeHasInput";
  PADDLE_ENFORCE_NE(
      node->IsTensorWrappersCleared(), true,
      paddle::platform::errors::Fatal(
          "The TensorWrappers of %s do not exist. This may be because:\n"
          "You calculate backward twice for the same subgraph without "
          "setting retain_graph=True. Please set retain_graph=True in the "
          "first backward/grad call.\n",
          node->name()));
}

void DuplicateCheck(const std::vector<paddle::experimental::Tensor>& inputs,
                    bool is_input) {
  std::unordered_set<AutogradMeta*> visisted_ins;
  std::string msg = is_input ? "inputs" : "outputs";
  for (auto in : inputs) {
    AutogradMeta* auto_grad_meta = EagerUtils::unsafe_autograd_meta(in);
    PADDLE_ENFORCE_EQ(
        visisted_ins.count(auto_grad_meta), 0,
        paddle::platform::errors::AlreadyExists(
            "%s contain duplicate tensor %s, please check %s carefully.", msg,
            in.name(), msg));
    visisted_ins.insert(auto_grad_meta);
  }
}

GeneralGrad* GeneralGrad::general_grad_ = new GeneralGrad();

std::vector<paddle::experimental::Tensor> RunBackward(
    const std::vector<paddle::experimental::Tensor>& tensors,  // output
    const std::vector<paddle::experimental::Tensor>& grad_tensors,
    bool retain_graph, bool create_graph = false,
    const std::vector<paddle::experimental::Tensor>& inputs = {},
    bool allow_unused = false,
    const std::vector<paddle::experimental::Tensor>& no_grad_vars = {}) {
  VLOG(6) << "Start Backward";

  // *Gradient Hook should happen at node-level
  // *Inplace version check should perform at node-level
  // *Cross-batch accumulation happens at forward pass

  // GeneralGrad
  bool is_general_grad = !inputs.empty();
  if (is_general_grad) GeneralGrad::Instance().Clear();

  /* --- Initialization --- */
  // 1. Init queue with starting nodes
  // 2. Prepare initial input buffers
  std::queue<GradNodeBase*> queue;
  std::queue<GradNodeBase*> orig_queue;
  std::unordered_map<GradNodeBase*, std::unique_ptr<GradTensorHolder>>
      node_input_buffers_dict;
  for (size_t i = 0; i < tensors.size(); i++) {
    const paddle::experimental::Tensor& tensor = tensors[i];

    AutogradMeta* auto_grad_meta = EagerUtils::nullable_autograd_meta(tensor);
    if (auto_grad_meta == nullptr) {
      VLOG(3) << "Skip auto grad since there is no grad op for var or loss is "
                 "stop_gradient=True: "
              << tensor.name();
      continue;
    }
    // Get grad input info from target tensors
    auto input_info = auto_grad_meta->OutRankInfo();

    VLOG(2) << "Out Rank of Tensor is slot: " << input_info.first
            << ", rank: " << input_info.second;
    // Get target GradNodeBase from target tensors
    auto shared_grad_node = auto_grad_meta->GetMutableGradNode();

    if (shared_grad_node == nullptr || shared_grad_node.get() == nullptr ||
        auto_grad_meta->StopGradient()) {
      VLOG(3) << "Skip auto grad since there is no grad op for var or loss is "
                 "stop_gradient=True: "
              << tensor.name();
      continue;
    }

    // TODO(zhanlve): Copy and Modify GradNode if is_general_grad
    GradNodeBase* grad_node = shared_grad_node.get();
    if (is_general_grad) {
      // Save orig grad node
      orig_queue.push(grad_node);

      // Replace grad_node with copied grad_node
      grad_node = GeneralGrad::Instance().CopyGradNode(shared_grad_node);

      // Record potential startup grad node
      GeneralGrad::Instance().GetPotentialStartupNodes()->insert(grad_node);
    }

    // Prepare GradTensorHolder
    if (!node_input_buffers_dict.count(grad_node)) {
      VLOG(6) << "Create Value for grad input tensor " << i
              << " of grad node: " << grad_node->name();
      node_input_buffers_dict[grad_node] =
          std::make_unique<GradTensorHolder>(grad_node->InputMeta());
    }
    bool copy_from_grad_t =
        grad_tensors.size() > 0 && grad_tensors[i].initialized();
    if (copy_from_grad_t) {
      PADDLE_ENFORCE(
          grad_tensors.size() == tensors.size(),
          paddle::platform::errors::Fatal(
              "Detected size mismatch between tensors and grad_tensors"
              "grad_tensors should either have "
              "size = 0 or same size as tensors."));
      // Feed given tensor if it's provided
      VLOG(6) << "Fill grad input tensor " << i << "with give grad tensor";

      // Deep copy
      node_input_buffers_dict[grad_node]->CopyValueFromTensor(
          input_info.first, input_info.second, grad_tensors[i]);
    } else {
      VLOG(6) << "Fill grad input tensor " << i << " with 1.0";
      // Initialize tensor with 1.0
      // Forward Tensor "tensor" is passed to indicate tensortype, datatype and
      // dims
      // GradTensorHolder will initialize another tensor with same tensortype,
      // datatype and dims but filled with 1.0
      node_input_buffers_dict[grad_node]->CopyValueFromTensor(
          input_info.first, input_info.second, tensor, true /*fill_one=true*/);
    }

    // Prepare queue, potential startup_nodes
    queue.push(grad_node);
  }

  if (is_general_grad) {
    // Copy Backward Graph
    GeneralGrad::Instance().ReconstructBackwardGraph(orig_queue);
  }

  VLOG(6) << "Update In degree Map for backward";
  // 3. Compute in_degree for each node
  std::unordered_map<GradNodeBase*, int> node_in_degree_map =
      getInDegreeMap(queue);

  if (is_general_grad) {
    // Prepare several vital preprocess for GeneralGrad
    GeneralGrad::Instance().PreparedForGeneralGrad(inputs, no_grad_vars, &queue,
                                                   node_input_buffers_dict);
  }

  VLOG(6) << " startup_ops' size is :" << queue.size();

  /* --- Topological Visit --- */
  // 1. Pop queue
  // 2. Run node
  //    |- Check and capture target result
  //    |- node(grads)
  //    |- Prepare for next node
  // 3. Update queue
  VLOG(6) << "Run Backward";
  while (!queue.empty()) {
    GradNodeBase* node = queue.front();
    VLOG(6) << "Running GradNode:" << node->name();

    paddle::platform::RecordEvent node_record_event(
        std::string((*node).name()) + " grad_node",
        paddle::platform::TracerEventType::Operator, 1);

    if (queue.size() > 1 && node_in_degree_map[node] != 0) {
      queue.pop();
      continue;
    }
    queue.pop();

    // Run node: This is where Hook happens
    PADDLE_ENFORCE(
        node_input_buffers_dict.count(node),
        paddle::platform::errors::Fatal(
            "Unable to find next node in the GradTensorHolder \n"
            "Trying to run Node without configuring its GradTensorHolder."));

    std::unique_ptr<GradTensorHolder> node_input_buffer =
        std::move(node_input_buffers_dict[node]);

    // Set input target grad_var from node_input_buffer by inputmeta
    if (!inputs.empty() && is_general_grad) {
      GeneralGrad::Instance().SetResultForInputTargetVar(*node_input_buffer,
                                                         node);
    }

    // no_grad_vars
    if (!no_grad_vars.empty() && is_general_grad) {
      auto iter =
          GeneralGrad::Instance().GetNoGradVarNodesInputMetaMap()->find(node);
      if (iter !=
          GeneralGrad::Instance().GetNoGradVarNodesInputMetaMap()->end()) {
        VLOG(6) << "Change the input buffer[slot][rank] by Zeros";
        auto rank_info = (iter->second)->OutRankInfo();
        node_input_buffer->SetBufferSlotRankZeros(rank_info.first,
                                                  rank_info.second);
      }
    }

    VLOG(6) << "Running GradNode:" << node->name();

    // Check input
    EnforceGradNodeHasInput(node);

    VLOG(6) << "Run Backward Kernel with GradTensorHolder.";
    // Run Pre Backward Node and get outputs
    paddle::small_vector<std::vector<paddle::experimental::Tensor>,
                         kSlotSmallVectorSize>
        grad_output_tensors = (*node)(node_input_buffer->Buffers(),
                                      create_graph, is_general_grad);

    // retain_grad or not
    if (!retain_graph) {
      VLOG(6)
          << "retain_graph is false, need to clear the TensorWrapper of nodes.";
      node->ClearTensorWrappers();
    }

    // TODO(jiabin): Should we erase it or find a more efficient way.

    node_input_buffers_dict.erase(node);

    // Prepare GradTensorHolder for next node
    const paddle::small_vector<std::vector<GradSlotMeta>, kSlotSmallVectorSize>&
        metas = node->OutputMeta();
    PADDLE_ENFORCE(metas.size() == grad_output_tensors.size() || metas.empty(),
                   paddle::platform::errors::Fatal(
                       "Number of edges should be either empty ( for leaf node "
                       ") or the same as number of output grad tensors, but we "
                       "got edges size is: %d, grad_output size is: %d",
                       metas.size(), grad_output_tensors.size()));

    for (size_t i = 0; i < metas.size(); i++) {
      for (size_t j = 0; j < metas[i].size(); j++) {
        const Edge& edge = metas[i][j].GetEdge();
        if (!edge.IsInitialized()) {
          continue;
        }
        auto edge_rank = edge.GetEdgeRankInfo();
        // Since we make edge has as same rank as bwd outputs, we indexing them
        // with
        // the same rank(i, j)
        auto next_node_shared = edge.GetMutableGradNode();

        // Next node could be nullptr if it is leaf tensor with no
        // AccumulationNode attached
        // Or it could also originated from dispensable inputs
        if (!next_node_shared || !next_node_shared.get() ||
            grad_output_tensors[i].empty()) {
          continue;
        }

        PADDLE_ENFORCE_LT(
            j, grad_output_tensors[i].size(),
            paddle::platform::errors::Fatal(
                "Rank of grad_output_tensors should be less than "
                "grad_output_tensors[i].size(), which is: %d. This error may "
                "indicate autoprune or autograd api error. ",
                grad_output_tensors.size()));
        paddle::experimental::Tensor& grad_output_tensor =
            grad_output_tensors[i][j];

        if ((!grad_output_tensor.defined() ||
             !grad_output_tensor.initialized())) {
          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();

        auto* next_node = next_node_shared.get();
        if (!node_input_buffers_dict.count(next_node)) {
          const auto& input_meta = next_node->InputMeta();
          auto grad_tensor_holder =
              std::make_unique<GradTensorHolder>(input_meta);
          VLOG(6) << "Construct GradTensorHolder for grad node: "
                  << next_node->name();
          node_input_buffers_dict[next_node] = std::move(grad_tensor_holder);
        }

        VLOG(6) << "Sum grad inputs for edge slot: " << edge_rank.first
                << ", rank: " << edge_rank.second;

        node_input_buffers_dict[next_node]->add(
            edge_rank.first, edge_rank.second, grad_output_tensor,
            create_graph);

        // Update queue
        node_in_degree_map[next_node]--;

        PADDLE_ENFORCE(
            node_in_degree_map[next_node] >= 0,
            paddle::platform::errors::Fatal(
                "Detected in-degree value smaller than zero. For Node: %s"
                "Node's in-degree cannot be negative.",
                next_node->name()));

        if (is_general_grad) {
          bool is_potential_stop_node =
              GeneralGrad::Instance().GetPotentialStopNodes()->count(next_node);
          if (node_in_degree_map[next_node] == 0 && !is_potential_stop_node) {
            queue.emplace(std::move(next_node));
          }
        } else {
          if (node_in_degree_map[next_node] == 0) {
            queue.emplace(std::move(next_node));
          }
        }
      }
    }
  }

  if (!is_general_grad) return {};
  return GeneralGrad::Instance().GetResults(inputs, allow_unused, create_graph);
}

void Backward(
    const std::vector<paddle::experimental::Tensor>& tensors,  // outputs
    const std::vector<paddle::experimental::Tensor>& grad_tensors,
    bool retain_graph) {
  VLOG(6) << "Run in Backward";
  paddle::platform::RecordEvent backward_record_event(
      "backward", paddle::platform::TracerEventType::Operator, 1);
  RunBackward(tensors, grad_tensors, retain_graph);
  phi::autotune::AutoTuneStatus::Instance().Update();
}

std::vector<paddle::experimental::Tensor> Grad(
    const std::vector<paddle::experimental::Tensor>& tensors,  // outputs
    const std::vector<paddle::experimental::Tensor>& inputs,
    const std::vector<paddle::experimental::Tensor>& grad_tensors,
    bool retain_graph, bool create_graph, bool only_inputs, bool allow_unused,
    const std::vector<paddle::experimental::Tensor>& no_grad_vars) {
  VLOG(6) << "Run in Grad";

  DuplicateCheck(inputs, true /* is_input */);
  DuplicateCheck(tensors, false /* is_input */);

  return RunBackward(tensors, grad_tensors, retain_graph, create_graph, inputs,
                     allow_unused, no_grad_vars);
}
}  // namespace egr