speed up gc scope (#73485)

huangjiyi · web-flow · commit 023c04f091e9 · 2025-06-21T13:31:22.000+08:00
diff --git a/paddle/fluid/eager/to_static/run_program_op_node.h b/paddle/fluid/eager/to_static/run_program_op_node.h
@@ -99,6 +99,7 @@ static bool IsVariableRefArray(const Tensor &tensor) {
 
 static auto GetNameFromValue(const std::vector<::pir::Value> &values) {
   std::vector<std::string> names;
+  names.reserve(values.size());
   std::transform(
       values.begin(),
       values.end(),
@@ -170,14 +171,13 @@ static void ShareTensorsIntoScopeWithName(
     const std::vector<std::string> &tensor_names,
     paddle::framework::Scope *scope) {
   for (size_t i = 0; i < tensors.size(); ++i) {
-    auto name = tensor_names[i];
+    const auto &name = tensor_names[i];
     VLOG(4) << "Share Tensor Into Scope: " << name;
     if (name == paddle::framework::kFakeVarName ||
         name == paddle::framework::kEmptyVarName) {
       continue;
     }
     auto *var = scope->Var(name);
-    CheckInputVarStatus(tensors[i]);
     // share tensor
     auto tensor_base = tensors[i].impl();
     if (phi::DenseTensor::classof(tensor_base.get())) {
@@ -352,30 +352,25 @@ static void BuildScopeByBlock(
 }
 
 static void GcScope(paddle::framework::Scope *scope) {
-  std::deque<std::shared_ptr<paddle::memory::Allocation>> *garbages =
-      new std::deque<std::shared_ptr<paddle::memory::Allocation>>();
-
-  for (auto &var : scope->LocalVars()) {
+  for (auto &[_, var] : scope->LocalVarsMap()) {
     if (var != nullptr) {
       if (var->IsType<phi::DenseTensor>()) {
-        garbages->emplace_back(
-            var->GetMutable<phi::DenseTensor>()->MoveMemoryHolder());
+        var->GetMutable<phi::DenseTensor>()->MoveMemoryHolder();
       }
       if (var->IsType<phi::SelectedRows>()) {
-        garbages->emplace_back(var->GetMutable<phi::SelectedRows>()
-                                   ->mutable_value()
-                                   ->MoveMemoryHolder());
+        var->GetMutable<phi::SelectedRows>()
+            ->mutable_value()
+            ->MoveMemoryHolder();
       }
       if (var->IsType<phi::TensorArray>()) {
         auto *lod_tensor_arr = var->GetMutable<phi::TensorArray>();
         for (auto &t : *lod_tensor_arr) {
-          garbages->emplace_back(t.MoveMemoryHolder());
+          t.MoveMemoryHolder();
         }
         lod_tensor_arr->clear();
       }
     }
   }
-  delete garbages;  // free mem
 }
 
 template <class T>
diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h
@@ -48,6 +48,15 @@ namespace framework {
  * scope.
  */
 class TEST_API Scope {
+ protected:
+  struct KeyHasher {
+    std::size_t operator()(const std::string& key) const {
+      return XXH32(key.c_str(), key.size(), 1);
+    }
+  };
+  using KeyVarMap =
+      std::unordered_map<std::string, std::unique_ptr<Variable>, KeyHasher>;
+
  public:
   Scope();
   ~Scope();
@@ -114,6 +123,9 @@ class TEST_API Scope {
   // enumerate all the variables which current scope contains.
   std::vector<Variable*> LocalVars();
 
+  // Get a reference to the vars map
+  const KeyVarMap& LocalVarsMap() { return vars_; }
+
   // Rename variable to a new name
   void Rename(const std::string& origin_name,
               const std::string& new_name) const;
@@ -130,14 +142,7 @@ class TEST_API Scope {
   void SetCanReused(bool can_reused) { can_reused_ = can_reused; }
 
  protected:
-  struct KeyHasher {
-    std::size_t operator()(const std::string& key) const {
-      return XXH32(key.c_str(), key.size(), 1);
-    }
-  };
-
-  mutable std::unordered_map<std::string, std::unique_ptr<Variable>, KeyHasher>
-      vars_;
+  mutable KeyVarMap vars_;
 
  private:
   // Call Scope::NewScope for a sub-scope.
