diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 038ad859db5494a5b0e43ace7e0db2ccd5d7d0d4..3e686b1c415e61a24e0f6729555e672721cf806f 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -187,6 +187,12 @@ endfunction()
 add_subdirectory(math)
 add_subdirectory(nccl)
 
+if(WITH_GPU)
+    op_library(nccl_op DEPS nccl_common)
+else()
+    set(DEPS_OPS ${DEPS_OPS} nccl_op)
+endif()
+
 if(WITH_DISTRIBUTE)
     add_subdirectory(detail)
     set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
@@ -217,16 +223,12 @@ op_library(lod_rank_table_op DEPS lod_rank_table)
 op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
 op_library(array_to_lod_tensor_op DEPS lod_rank_table_op)
 op_library(max_sequence_len_op DEPS lod_rank_table)
-if(WITH_GPU)
-    op_library(nccl_op DEPS nccl_common)
-endif()
 op_library(sequence_conv_op DEPS context_project)
 op_library(sequence_pool_op DEPS sequence_pooling)
 op_library(lstm_op DEPS sequence2batch lstm_compute)
 op_library(conv_transpose_op DEPS vol2col)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op DEPS executor)
-
 # FIXME(typhoonzero): save/load depends lodtensor serialization functions
 op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)