gpu (#1310)

Former-commit-id: 82681d523fa9e521e2c04b5fd32e6f435f9ba722

oneflow/core/graph/logical_graph.cpp

@@ -522,7 +522,7 @@ void LogicalGraph::AddAllReduce(LogicalNode* src, LogicalNode* dst) {
   std::shared_ptr<const ParallelDesc> dst_pd = dst->parallel_desc();
   CHECK_EQ(src_pd->parallel_num(), dst_pd->parallel_num());
   CHECK_EQ(src_pd->device_type(), dst_pd->device_type());
-  if (Global<JobDesc>::Get()->enable_nccl()) {
+  if (Global<JobDesc>::Get()->enable_nccl() && src_pd->device_type() == DeviceType::kGPU) {
     if (src_pd->sorted_machine_ids().size() == 1
         || Global<JobDesc>::Get()->use_nccl_inter_node_communication()) {
       AddNcclAllReduce(src, dst);