test=develop

paddle/fluid/framework/details/reference_count_pass.cc

@@ -80,15 +80,15 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
       // This is weird but there is really some variables without var_desc
       // in computation_op
       if (var_desc == nullptr) {
-        if (compute_op->Node()->Op()->Block()->FindVar(var_name) == nullptr)
-          continue;
-      } else {
-        if (var_desc->Persistable()) continue;
-        auto var_type = var_desc->Proto()->type().type();
-        if (var_type != proto::VarType::LOD_TENSOR &&
-            var_type != proto::VarType::SELECTED_ROWS) {
-          continue;
-        }
+        var_desc = compute_op->Node()->Op()->Block()->FindVar(var_name);
+        if (var_desc == nullptr) continue;
+      }
+
+      if (var_desc->Persistable()) continue;
+      auto var_type = var_desc->Proto()->type().type();
+      if (var_type != proto::VarType::LOD_TENSOR &&
+          var_type != proto::VarType::SELECTED_ROWS) {
+        continue;
       }
 
       // compute op only runs in one device

paddle/fluid/framework/parallel_executor.cc

@@ -250,6 +250,13 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
 #ifdef PADDLE_WITH_CUDA
   if (!gcs_.empty()) {
     ResetReferenceCount();
+    for (auto &pair : cur_ref_cnts_) {
+      auto &name_map = *(pair.second);
+      for (auto &fetch_name : fetch_tensors) {
+        name_map.erase(fetch_name);
+      }
+      name_map.erase(fetched_var_name);
+    }
   }
 #endif
   auto fetch_data = member_->executor_->Run(fetch_tensors);