[Dy2static] Auto Remove Step Scope while GradRunProgramNode GCed. (#51411)

* merge

* fix bugs while backward multi-times.

* code format by ci

paddle/fluid/eager/to_static/run_program_op_func.h

@@ -60,11 +60,6 @@ inline void run_program_ad_func(
     std::vector<paddle::framework::Scope*>& step_scope,  // NOLINT
     std::vector<paddle::Tensor*>& dout,                  // NOLINT
     const paddle::framework::AttributeMap& attrs) {
-  VLOG(2) << "start run run_program";
-  // Call forward function
-  RunProgramAPI(x, params, out, step_scope, dout, attrs);
-  VLOG(2) << "start run run_program grad";
-
   // Prepare Autograd Meta
   auto deref_out = details::DereferenceTensors(out);
   std::vector<egr::AutogradMeta*> p_autograd_x =
@@ -78,6 +73,13 @@ inline void run_program_ad_func(
   bool require_any_grad = egr::EagerUtils::ComputeRequireGrad(
       trace_backward, &p_autograd_x, &p_autograd_params);
 
+  VLOG(2) << "start run run_program with require_any_grad = "
+          << require_any_grad;
+  // Call forward function
+  // if require_any_grad is False, don't save any middle vars.
+  RunProgramAPI(x, params, out, step_scope, dout, require_any_grad, attrs);
+  VLOG(2) << "start run run_program grad";
+
   if (require_any_grad) {
     egr::EagerUtils::PassStopGradient(false, &p_autograd_outs);
     // Create GradOpNode (1 means [out_grad], 2 means [x_grad, paramx_grad])

paddle/fluid/eager/to_static/run_program_op_node.h

@@ -283,6 +283,7 @@ inline void RunProgramAPI(
     std::vector<paddle::Tensor *> &out,                   // NOLINT
     std::vector<paddle::framework::Scope *> &step_scope,  // NOLINT
     std::vector<paddle::Tensor *> &dout,                  // NOLINT
+    bool require_any_grad,
     const paddle::framework::AttributeMap &attrs) {
   VLOG(2) << "RunProgramOpKernel Compute";
   // In the original run_program OP, the default value of the is_test
@@ -436,8 +437,10 @@ inline void RunProgramAPI(
 
     VLOG(3) << paddle::framework::GenScopeTreeDebugInfo(out_scope_vec->front());
 
-    if (is_test || !egr::Controller::Instance().HasGrad()) {
-      VLOG(4) << "is test, set this scope can reused";
+    if (is_test || !require_any_grad) {
+      VLOG(4) << "don't require any grad, set this scope can reused";
+      VLOG(4) << "is_test: " << is_test
+              << ", require_any_grad: " << require_any_grad;
       global_inner_scope->SetCanReuesd(true);
       details::GcScope(global_inner_scope);
     } else {
@@ -576,7 +579,8 @@ inline void RunProgramGradAPI(
                                                    *backward_global_block,
                                                    global_inner_scope);
     VLOG(4) << "after backward gc all vars";
-    global_inner_scope->SetCanReuesd(true);
+    global_inner_scope->SetCanReuesd(
+        false);  // can't reuse util call `~GradNodeRunProgram`
     details::GcScope(global_inner_scope);
   }
 }
@@ -586,7 +590,15 @@ class GradNodeRunProgram : public egr::GradNodeBase {
   GradNodeRunProgram(size_t bwd_in_slot_num, size_t bwd_out_slot_num)
       : egr::GradNodeBase(bwd_in_slot_num, bwd_out_slot_num) {}
 
-  ~GradNodeRunProgram() override = default;
+  ~GradNodeRunProgram() {
+    auto *out_scope_vec = &step_scope_;
+    // Normally out_scope_vec.size() == 1. for safty, we add for-loop here.
+    for (size_t i = 0; i < out_scope_vec->size(); ++i) {
+      paddle::framework::Scope *global_inner_scope = out_scope_vec->at(i);
+      global_inner_scope->SetCanReuesd(true);  // set this to reuse scope.
+      details::GcScope(global_inner_scope);
+    }
+  }
   // Functor: perform backward computations
   virtual paddle::small_vector<std::vector<paddle::Tensor>,
                                egr::kSlotSmallVectorSize>