fix cpu test=develop

paddle/fluid/framework/details/parallel_ssa_graph_executor.cc

@@ -36,7 +36,6 @@ std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
   for (auto &op : graph->Get<GraphOps>(kGraphOps)) {
     auto &dev_ctx = op->DeviceContext();
     auto &p = dev_ctx.begin()->first;
-#ifdef PADDLE_WITH_CUDA
     int dev_id = boost::get<platform::CUDAPlace>(p).device;
     auto &dev_ops = graphs[dev_id]->Get<GraphOps>(kGraphOps);
     auto &dev_dummys = graphs[dev_id]->Get<GraphDepVars>(kGraphDepVars);
@@ -59,9 +58,6 @@ std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
           graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
       }
     }
-#else
-    PADDLE_THROW("Parallel Graph Execution only support CUDAPlace.");
-#endif
   }
 
   for (size_t dev_id = 0; dev_id < places.size(); ++dev_id) {

paddle/fluid/framework/parallel_executor.cc

@@ -304,6 +304,7 @@ ParallelExecutor::ParallelExecutor(
   }
 
   if (build_strategy.enable_parallel_graph_) {
+#ifdef PADDLE_WITH_CUDA
     auto parallel_graph =
         details::SeparateMultiDevicesGraph(member_->places_, std::move(graph));
     auto seq_allreduce_pass =
@@ -319,6 +320,10 @@ ParallelExecutor::ParallelExecutor(
     member_->executor_.reset(new details::ParallelSSAGraphExecutor(
         exec_strategy, member_->local_scopes_, member_->places_,
         std::move(parallel_graph)));
+#else
+    PADDLE_THROW(
+        "Paddle should be compiled with CUDA for ParallelGraph Execution.");
+#endif
   } else {
     if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
       member_->executor_.reset(new details::ThreadedSSAGraphExecutor(