Fix run program grad node mem

paddle/fluid/eager/to_static/run_program_op_func.h

@@ -88,7 +88,26 @@ inline void run_program_ad_func(
     // Set Attributes
     grad_node->SetAttrMap(attrs);
     // Set TensorWrappers
-    grad_node->SetFwdX(x);
+    auto* forward_global_block = PADDLE_GET_CONST(
+        paddle::framework::BlockDesc*, attrs.at("forward_global_block"));
+    auto* backward_global_block = PADDLE_GET_CONST(
+        paddle::framework::BlockDesc*, attrs.at("backward_global_block"));
+    // Clear unused x vars
+    auto temp_x = std::vector<paddle::Tensor>(x);
+    for (size_t i = 0; i < x.size(); i++) {
+      if (!backward_global_block->HasVar(x[i].name())) {
+        auto fake = paddle::Tensor(std::make_shared<phi::DenseTensor>());
+        fake.set_name("Fake_var");
+        temp_x[i] = fake;
+      }
+    }
+    grad_node->SetFwdX(temp_x);
+    // Clear unused out vars
+    for (size_t i = 0; i < out.size(); i++) {
+      if (!backward_global_block->HasVar(out[i]->name())) {
+        step_scope[0]->EraseVars({out[i]->name()});
+      }
+    }
 
     grad_node->SetFwdParams(params);
     grad_node->SetStepScope(step_scope);
@@ -97,10 +116,6 @@ inline void run_program_ad_func(
     // NOTE(@xiongkun): Not every tensor in x(list of tensor) is required
     // gradient. for example: x[1] is not used for output, the x[1] is ignored.
 
-    auto* forward_global_block = PADDLE_GET_CONST(
-        paddle::framework::BlockDesc*, attrs.at("forward_global_block"));
-    auto* backward_global_block = PADDLE_GET_CONST(
-        paddle::framework::BlockDesc*, attrs.at("backward_global_block"));
     std::vector<const paddle::Tensor*> x_require_grad;
     for (size_t i = 0; i < x.size(); ++i) {
       auto& name = x[i].name();

paddle/fluid/eager/to_static/run_program_op_node.h

@@ -480,8 +480,6 @@ inline void RunProgramAPI(
 }
 
 inline void RunProgramGradAPI(
-    const std::vector<paddle::Tensor> &x UNUSED,
-    const std::vector<paddle::Tensor> &params UNUSED,
     const std::vector<paddle::Tensor> &out_grad,
     const std::vector<paddle::framework::Scope *> &step_scope,  // NOLINT
     const paddle::framework::AttributeMap &attrs,
@@ -694,13 +692,8 @@ class GradNodeRunProgram : public egr::GradNodeBase {
     for (size_t i = 0; i < out_grad_names.size(); ++i) {
       hooked_grads[0][i].set_name(out_grad_names[i]);
     }
-    RunProgramGradAPI(x_,
-                      params_,
-                      hooked_grads[0],
-                      step_scope_,
-                      attrs_,
-                      x_grad_ptr,
-                      params_grad_ptr);
+    RunProgramGradAPI(
+        hooked_grads[0], step_scope_, attrs_, x_grad_ptr, params_grad_ptr);
     VLOG(3) << "End Eager Backward Node: GradNodeRunProgram";
 
     executed_ = true;