Change fetch op

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -33,11 +33,6 @@ void FetchOpHandle::Wait(platform::DeviceContext *waited_dev) {
 }
 
 void FetchOpHandle::WaitAndMergeCPUTensors() const {
-  // Wait fetch stream done.
-  for (auto &ctx : dev_ctx_) {
-    ctx.second->Wait();
-  }
-
   std::vector<const LoDTensor *> tensors_ptr;
   tensors_ptr.reserve(tensors_.size());
   for (auto &t : tensors_) {
@@ -72,6 +67,8 @@ void FetchOpHandle::RunImpl() {
       tensors_[i].ShareDataWith(t);
       tensors_[i].set_lod(t.lod());
     }
+
+    this->WaitAndMergeCPUTensors();
   }
 }
 

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -96,12 +96,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     for (auto *var : vars) {
       op->AddInput(var);
     }
-
-    dummy_vars.emplace_back();
-    auto *var = &dummy_vars.back();
-    var->generated_op_ = nullptr;
-    op->AddOutput(var);
-    InsertPendingVar(*var);
     InsertPendingOp(*op);
   }
 
@@ -176,8 +170,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   };
 
   // Wait FetchOps.
-  for (auto &fetch_op : fetch_ops) {
+  if (!fetch_ops.empty()) {
-    fetch_op.WaitAndMergeCPUTensors();
     sync_computation();
   }