Fix test rnn memory helper op (#37474)

* clear LoDTensorArray

* fix  bugs

* fix

* fix gpu

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -343,7 +343,6 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
   Scope* local_scope = create_local_scope_
                            ? global_scope_->GetMutableLocalScope()
                            : global_scope_->GetMutableScope();
-
   auto op_with_kernel = dynamic_cast<const framework::OperatorWithKernel*>(op);
   {
     platform::RecordEvent infershape_event("InferShape");
@@ -354,8 +353,7 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
 
   if (op_with_kernel != nullptr &&
       FLAGS_new_executor_use_inplace) {  // TODO(xiongkun03) Does operator
-                                         // base support
-                                         // inplace ?
+                                         // base support inplace ?
     for (auto& pair : instr_node.InplaceInfo()) {
       const auto& in = paddle::framework::details::GetTensorFromVar(pair.first);
       auto* out =

paddle/fluid/framework/new_executor/interpretercore_garbage_collector.cc

@@ -79,6 +79,7 @@ void InterpreterCoreGarbageCollector::Add(paddle::framework::Variable* var,
     for (auto& t : *tensor_arr) {
       Add(t.MoveMemoryHolder(), event, ctx);
     }
+    tensor_arr->clear();
   } else if (var->IsType<std::vector<Scope*>>()) {
     // NOTE(@xiongkun03) conditional_op / while_op will create a STEP_SCOPE
     // refer to executor.cc to see what old garbage collector does.

paddle/fluid/framework/new_executor/interpretercore_util.cc

@@ -411,6 +411,7 @@ void build_op_func_list(const platform::Place& place,
         for (auto& t : *lod_tensor_arr) {
           garbages->emplace_back(t.MoveMemoryHolder());
         }
+        lod_tensor_arr->clear();
       } else {
         PADDLE_THROW(platform::errors::Unimplemented(
             "Type %s of variable %s is not supported eager deletion.",

paddle/fluid/operators/rnn_memory_helper_op.cc

@@ -109,7 +109,11 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(dev_place);
 
-    if (out_grad_var == nullptr) {
+    // NOTE(xiongkun03): In standalone executor, after each run, the
+    // var.tensor.holder will be delete instead of variable. So we need exam the
+    // IsInitialized().
+    if (out_grad_var == nullptr ||
+        !out_grad_var->Get<framework::LoDTensor>().IsInitialized()) {
       VLOG(5) << "Using fill constant 0 as starting gradient";
       auto in_var_name = Input("X");
       auto *in_var = scope.FindVar(in_var_name);

paddle/fluid/operators/sum_op.h

@@ -129,6 +129,7 @@ template <typename DeviceContext, typename T>
 class SumKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
+    VLOG(10) << "start sum kernel";
     auto in_vars = context.MultiInputVar("X");
     size_t in_num = in_vars.size();
     auto out_var = context.OutputVar("Out");
@@ -138,7 +139,8 @@ class SumKernel : public framework::OpKernel<T> {
     if (out_var->IsType<framework::LoDTensor>()) {
       auto *out = out_var->GetMutable<framework::LoDTensor>();
       auto *out_ptr = out->mutable_data<T>(context.GetPlace());
-      if (in_num >= 1 && in_vars[0]->IsType<framework::LoDTensor>()) {
+      if (in_num >= 1 && in_vars[0]->IsType<framework::LoDTensor>() &&
+          in_vars[0]->Get<framework::LoDTensor>().IsInitialized()) {
         auto &in_0_tensor = in_vars[0]->Get<framework::LoDTensor>();
         if (in_0_tensor.numel() > 0) {
           in_place = (in_0_tensor.data<T>() == out_ptr);
@@ -151,7 +153,9 @@ class SumKernel : public framework::OpKernel<T> {
       int start = in_place ? 1 : 0;
       if (!in_place) {
         if ((in_num >= 2) && in_vars[0]->IsType<framework::LoDTensor>() &&
-            in_vars[1]->IsType<framework::LoDTensor>()) {
+            in_vars[1]->IsType<framework::LoDTensor>() &&
+            in_vars[0]->Get<framework::LoDTensor>().IsInitialized() &&
+            in_vars[1]->Get<framework::LoDTensor>().IsInitialized()) {
           auto &in_0 = in_vars[0]->Get<framework::LoDTensor>();
           auto &in_1 = in_vars[1]->Get<framework::LoDTensor>();
           if (in_0.numel() && in_1.numel()) {
@@ -162,6 +166,7 @@ class SumKernel : public framework::OpKernel<T> {
           }
         }
         if (start != 2) {
+          VLOG(10) << "Fill with constant = 0 in sum kernel.";
           math::SetConstant<DeviceContext, T> constant_functor;
           constant_functor(context.template device_context<DeviceContext>(),
                            out, static_cast<T>(0));
@@ -173,7 +178,7 @@ class SumKernel : public framework::OpKernel<T> {
       for (size_t i = start; i < in_num; i++) {
         if (in_vars[i]->IsType<framework::LoDTensor>()) {
           auto &in_t = in_vars[i]->Get<framework::LoDTensor>();
-          if (in_t.numel() == 0) {
+          if (!in_t.IsInitialized() || in_t.numel() == 0) {
             continue;
           }
           auto in = EigenVector<T>::Flatten(in_t);
@@ -200,6 +205,7 @@ class SumKernel : public framework::OpKernel<T> {
           "unsupport type: %s.",
           framework::ToTypeName(out_var->Type())));
     }
+    VLOG(10) << "end sum kernel";
   }
 };
 }  // namespace operators