update reduce config (#18335)

test=release/1.5

paddle/fluid/framework/parallel_executor.cc

@@ -314,25 +314,23 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
   member_->use_all_reduce_ =
       build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;
   member_->nranks_ = build_strategy.num_trainers_ * places.size();
-#if defined(PADDLE_WITH_CUDA) && defined(_WIN32)
-  if (member_->use_cuda_) {
-    PADDLE_ENFORCE(places.size() == 1, "Windows can support Single GPU only.");
-  }
-#endif
-  if (!member_->use_all_reduce_) {
-    if (places.size() == 1) {
+  if (!member_->use_all_reduce_ && member_->nranks_ == 1) {
     LOG(INFO) << "If you set build_strategy.reduce with 'Reduce',"
                  "the number of places should be greater than 1.";
+    member_->build_strategy_.reduce_ =
+        BuildStrategy::ReduceStrategy::kAllReduce;
     member_->use_all_reduce_ = true;
   }
+#if defined(PADDLE_WITH_CUDA) && defined(_WIN32)
+  if (member_->use_cuda_) {
+    PADDLE_ENFORCE(places.size() == 1, "Windows can support Single GPU only.");
   }
-
+#endif
   LOG(INFO) << string::Sprintf(
       "The number of %s, which is used in ParallelExecutor, is %lu. And "
       "the Program will be copied %lu copies",
       (member_->use_cuda_ ? "CUDAPlace" : "CPUPlace"), places.size(),
       places.size());
-
   // Step 1. Bcast the bcast_vars to devs.
   // Create local scopes
   if (local_scopes.empty()) {