diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 5aaaf993323c2d4dbef688d0977ec6374fde6512..3e686b1c415e61a24e0f6729555e672721cf806f 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -1,5 +1,6 @@
 file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
 string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
+set(DEPS_OPS "")
 set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
 file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
 function(op_library TARGET)
@@ -48,6 +49,11 @@ function(op_library TARGET)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
     endif()
 
+    list(LENGTH op_library_DEPS op_library_DEPS_len)
+    if (${op_library_DEPS_len} GREATER 0)
+        set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
+    endif()
+    
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
@@ -181,55 +187,26 @@ endfunction()
 add_subdirectory(math)
 add_subdirectory(nccl)
 
-set(DEPS_OPS
-    cond_op
-    cross_entropy_op
-    recurrent_op
-    softmax_with_cross_entropy_op
-    softmax_op
-    sequence_softmax_op
-    sum_op
-    pool_op
-    maxout_op
-    unpool_op
-    pool_with_index_op
-    conv_op
-    conv_transpose_op
-    nccl_op
-    sequence_conv_op
-    sequence_pool_op
-    lod_rank_table_op
-    lod_tensor_to_array_op
-    array_to_lod_tensor_op
-    max_sequence_len_op
-    lstm_op
-    tensor_array_read_write_op
-    gru_op
-    adagrad_op
-    sgd_op
-    save_op
-    load_op
-    send_op
-    recv_op)
+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)
-op_library(send_op SRCS send_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
-set_source_files_properties(
-    send_op.cc
-    PROPERTIES
-    COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-
-op_library(recv_op SRCS recv_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
-set_source_files_properties(
-    recv_op.cc
-    PROPERTIES
-    COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-
-cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
+    add_subdirectory(detail)
+    set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
+    set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
+    set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    op_library(recv_op DEPS ${DISTRIBUTE_DEPS})
+    set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
+else()
+    set(DEPS_OPS ${DEPS_OPS} send_op recv_op)
 endif()
 
-op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
+op_library(cond_op DEPS framework_proto tensor net_op)
 op_library(cross_entropy_op DEPS cross_entropy)
 op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
 op_library(softmax_op DEPS softmax)
@@ -242,21 +219,16 @@ op_library(pool_op DEPS pooling)
 op_library(maxout_op DEPS maxouting)
 op_library(unpool_op DEPS unpooling)
 op_library(pool_with_index_op DEPS pooling)
-op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
-op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op)
-op_library(array_to_lod_tensor_op SRCS array_to_lod_tensor_op.cc DEPS lod_rank_table_op)
-op_library(max_sequence_len_op SRCS max_sequence_len_op.cc DEPS lod_rank_table)
-op_library(tensor_array_read_write_op SRCS tensor_array_read_write_op.cc)
-if(WITH_GPU)
-op_library(nccl_op DEPS nccl_common)
-endif()
+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)
 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 SRCS recurrent_op.cc DEPS executor)
-
+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)
@@ -269,13 +241,12 @@ endforeach()
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 
 
-
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
 cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
 cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
 if(WITH_GPU)
-  cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
+    cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
 endif()
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)