【Hackathon 6th Fundable Projects 4 No.5】remove scale_loss_grad_op_han…

…dle -part (#66293)

* remove scale_loss_grad_op_handle

* fix ci

paddle/fluid/framework/compiled_program.cc

@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/convert_utils.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/op_handle_base.h"
-#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/framework/ir/memory_optimize_pass/memory_optimization_var_info.h"

paddle/fluid/framework/details/CMakeLists.txt

@@ -2,7 +2,6 @@ set(op_handle_srcs
     nan_inf_utils_detail.cc
     var_handle.cc
     op_handle_base.cc
-    scale_loss_grad_op_handle.cc
     share_tensor_buffer_functor.cc
     computation_op_handle.cc
     share_tensor_buffer_op_handle.cc

paddle/fluid/framework/details/build_strategy.h

@@ -74,16 +74,7 @@ struct BuildStrategy {
   // `FLAGS_cpu_deterministic=true` to env.
   enum class ReduceStrategy { kAllReduce = 0, kReduce = 1, kNoReduce = 2 };
 
-  enum class GradientScaleStrategy {
-    kCoeffNumDevice = 0,
-    kOne = 1,
-    // user can customize gradient scale to use, and just feed
-    // it into exe.run().
-    kCustomized = 2,
-  };
-
   ReduceStrategy reduce_{ReduceStrategy::kAllReduce};
-  GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice};
 
   std::string debug_graphviz_path_{""};
 
@@ -238,8 +229,6 @@ inline std::ostream &operator<<(std::ostream &os,
                                 const BuildStrategy &strategy) {
   os << "BuildStrategy: " << &strategy << std::endl;
   os << "reduce_: " << static_cast<int>(strategy.reduce_) << std::endl;
-  os << "gradient_scale_: " << static_cast<int>(strategy.gradient_scale_)
-     << std::endl;
   os << "debug_graphviz_path_: " << strategy.debug_graphviz_path_ << std::endl;
   os << "enable_backward_optimizer_op_deps_: "
      << strategy.enable_backward_optimizer_op_deps_ << std::endl;

paddle/fluid/framework/ir/graph_helper.cc

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <stack>
 
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
-#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
 #include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_utils.h"
@@ -474,37 +473,6 @@ void RemoveControlDepInputAndOuput(OpDesc *op_desc) {
   op_desc->Flush();
 }
 
-static OpDesc *ReplaceScaleLossGradOp(const Node &node, OpDesc *desc) {
-  desc->SetType("fill_constant");
-  desc->SetAttr("shape", std::vector<int64_t>({1}));
-  desc->SetAttr("value", 1.0f);
-
-  if (node.IsWrappedBy<details::OpHandleBase>()) {
-    details::OpHandleBase &op_hander =
-        const_cast<Node *>(&node)->Wrapper<details::OpHandleBase>();
-    desc->SetAttr(
-        "dtype",
-        dynamic_cast<details::ScaleLossGradOpHandle *>(&op_hander)->DType());
-    desc->SetAttr(
-        "value",
-        dynamic_cast<details::ScaleLossGradOpHandle *>(&op_hander)->Coeff());
-  }
-
-  desc->SetAttr("force_cpu", false);
-  desc->SetAttr(
-      OpProtoAndCheckerMaker::OpRoleAttrName(),
-      (static_cast<int>(OpRole::kBackward) | static_cast<int>(OpRole::kLoss)));
-  // TODO(Ruibiao) : Set OpDeviceAttrName when needed
-
-  std::vector<std::string> output_names;
-  output_names.reserve(node.outputs.size());
-  for (auto out : node.outputs) {
-    output_names.emplace_back(out->Name());
-  }
-  desc->SetOutput("Out", output_names);
-  return desc;
-}
-
 void ReplaceAllReduceOp(const Node &node,
                         proto::BlockDesc *block,
                         std::vector<OpDesc> *ops) {
@@ -662,21 +630,17 @@ static void GetGraphOpDesc(const std::vector<Node *> &nodes,
   for (Node *n : nodes) {
     // if node is not Op, skip
     if (!n->IsOp()) continue;
-    // create fill_constant op
-    if (n->Name() == "scale_loss_grad") {
-      VLOG(4) << "convert op node scale_loss_grad to desc fill_constant";
-      ops->emplace_back();
-      auto &desc = ops->back();
-      ReplaceScaleLossGradOp(*n, &desc);
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-    } else if ((n->Name() == "allreduce" || n->Name() == "fused_all_reduce") &&
-               dynamic_cast<details::NCCLOpHandleBase *>(
-                   &(n->Wrapper<details::OpHandleBase>())) != nullptr) {
+    if ((n->Name() == "allreduce" || n->Name() == "fused_all_reduce") &&
+        dynamic_cast<details::NCCLOpHandleBase *>(
+            &(n->Wrapper<details::OpHandleBase>())) != nullptr) {
       VLOG(4) << "convert op node " << n->Name() << " to desc c_allreduce_sum";
       ReplaceAllReduceOp(*n, block, ops);
       VLOG(4) << n->ToString();
+      continue;
+    }
 #endif
-    } else if (n->Op()) {
+    if (n->Op()) {
       VLOG(4) << "convert op node to desc " << n->Op()->Type();
       if (is_fused_opt(n)) {
         OpDesc depend_desc(n->Op()->Block());

paddle/fluid/framework/ir/graph_to_program_pass_test.cc

@@ -410,81 +410,6 @@ TEST(GraphToProgramPass, MultiBlock) {
   FLAGS_convert_all_blocks = flag_temp;
 }
 
-void BuildProgramWithScaleLossGrad(Graph* g) {
-  OpDesc op1;
-  op1.SetType("op1");
-  OpDesc op2;
-  op2.SetType("op2");
-  OpDesc op3;
-  op3.SetType("op3");
-  OpDesc op4;
-  op4.SetType("op4");
-  VarDesc var1("var1");
-  VarDesc var2("var2");
-
-  ir::Node* o1 = g->CreateOpNode(&op1);
-  ir::Node* o2 = g->CreateOpNode(&op2);
-  ir::Node* o3 =
-      g->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation);
-  ir::Node* o4 =
-      g->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation);
-  ir::Node* v1 = g->CreateVarNode(&var1);
-  ir::Node* v2 = g->CreateVarNode(&var2);
-
-  // o1->v1->o2
-  o1->outputs.push_back(v1);
-  o2->inputs.push_back(v1);
-  v1->inputs.push_back(o1);
-  v1->outputs.push_back(o2);
-  // o3->v1
-  o3->outputs.push_back(v1);
-  v1->inputs.push_back(o1);
-  v1->inputs.push_back(o3);
-  // o4->v2
-  o4->outputs.push_back(v2);
-  v2->inputs.push_back(o4);
-}
-
-TEST(GraphToProgramPass, ReplaceScaleLossGrad) {
-  // Step1: Build a program with multi block
-  ProgramDesc before_prog;
-  Graph before_graph(before_prog);
-  BuildProgramWithScaleLossGrad(&before_graph);
-
-  // Step2 : Convert graph back to program
-  auto pass = paddle::framework::ir::PassRegistry::Instance().Get(
-      "graph_to_program_pass");
-
-  ProgramDesc after_prog;
-  pass->SetNotOwned<paddle::framework::ProgramDesc>("program", &after_prog);
-  pass->Apply(&before_graph);
-
-  // Step3 : statistics scale_loss_grad and fill_constant number
-  int scale_node_num = 0, fill_node_num = 0;
-  const auto& before_nodes_set = before_graph.Nodes();
-  for (const auto& n : before_nodes_set) {
-    if (n->Name() == "scale_loss_grad") {
-      ++scale_node_num;
-    } else if (n->Name() == "fill_constant") {
-      ++fill_node_num;
-    }
-  }
-
-  int scale_op_num = 0, fill_op_num = 0;
-  const auto& block = after_prog.Block(0);
-  for (const auto& op : block.AllOps()) {
-    if (op->Type() == "fill_constant") {
-      ++fill_op_num;
-    } else if (op->Type() == "scale_loss_grad") {
-      ++scale_op_num;
-    }
-  }
-
-  // Check pass OK
-  EXPECT_EQ(scale_op_num, 0);
-  EXPECT_EQ(scale_node_num + fill_node_num, fill_op_num);
-}
-
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle