diff --git a/paddle/operators/nccl_op_test.cu b/paddle/operators/nccl_op_test.cu
index 15d8bde933fcfa73902f725e75e0549321119939..a25e01baa4d539673f4149c72048e53a5613ed04 100644
--- a/paddle/operators/nccl_op_test.cu
+++ b/paddle/operators/nccl_op_test.cu
@@ -13,8 +13,11 @@
    limitations under the License. */
 #include "paddle/operators/nccl_op.h"
 
-#include "glog/logging.h"
-#include "gtest/gtest.h"
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+#include <thrust/device_vector.h>
+#include <memory>
+#include <vector>
 
 #include "paddle/framework/block_desc.h"
 #include "paddle/framework/op_desc.h"
@@ -24,10 +27,13 @@
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/gpu_info.h"
+#include "paddle/platform/place.h"
 
-#include <thrust/device_vector.h>
-#include <memory>
-#include <vector>
+USE_CPU_ONLY_OP(ncclInit);
+USE_GPU_ONLY_OP(ncclAllReduce);
+USE_GPU_ONLY_OP(ncclReduce);
+USE_GPU_ONLY_OP(ncclBcastSend);
+USE_GPU_ONLY_OP(ncclBcastRecv);
 
 static std::vector<int> gpu_list;
 
@@ -55,28 +61,28 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
   op->SetAttrMap(attrs);
 }
 
-TEST(NCCL, ncclInit) {
+// ncclInitOp with desc
+TEST(NCCL, ncclInitOp) {
   f::ProgramDescBind program;
   f::BlockDescBind *block = program.Block(0);
-  f::OpDescBind *op = block->AppendOp();
-
-  paddle::platform::Communicator comm;
-  op->SetType("ncclInit");
-  op->SetOutput("Communicator", )
-
-      AddOp("ncclInit", {}, {{"Communicator", {comm}}}, {{"gpus", {gpu_list}}},
-            block);
+  f::OpDescBind *op1 = block->AppendOp();
+
+  op1->SetType("ncclInit");
+  op1->SetOutput("Communicator", {"x1"});
+  op1->SetAttr("gpus", {gpu_list});
+  f::Scope g_scope;
+  paddle::platform::DeviceContext *ctx =
+      new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
+
+  auto *var = g_scope.Var("x1");
+  var->GetMutable<paddle::platform::Communicator>();
+
+  auto op = f::OpRegistry::CreateOp(*op1);
+  VLOG(1) << "invoke NCCLInitOp.";
+  op->Run(g_scope, *ctx);
+  VLOG(1) << "NCCLInitOp finished.";
 }
 
-// TEST(NCCL, ncclAllReduce) {
-//   f::ProgramDescBind program;
-//   f::BlockDescBind *block = program.Block(0);
-
-//   paddle::platform::Communicator comm;
-//   AddOp("ncclInit", {}, {{"Communicator", {comm}}, {"gpus", {gpu_list}}},
-//   block);
-// }
-
 int main(int argc, char **argv) {
   static int dev_count = paddle::platform::GetCUDADeviceCount();
   if (dev_count <= 1) {
diff --git a/python/paddle/v2/framework/tests/test_multigpu.py b/python/paddle/v2/framework/tests/test_multigpu.py
deleted file mode 100644
index b75d274d885e9014417faba258232e783fe2dda2..0000000000000000000000000000000000000000
--- a/python/paddle/v2/framework/tests/test_multigpu.py
+++ /dev/null
@@ -1,8 +0,0 @@
-import unittest, os
-import numpy as np
-import paddle.v2 as paddle
-from paddle.v2.framework.op import Operator
-import paddle.v2.framework.core as core
-from op_test import OpTest, create_op, set_input
-
-gpu_list = "0,1,2,3"
diff --git a/python/paddle/v2/framework/tests/test_nccl_ops.py b/python/paddle/v2/framework/tests/test_nccl_ops.py
deleted file mode 100644
index 6dd6231aa8dee863ade6e3eddf022d736c6a8a09..0000000000000000000000000000000000000000
--- a/python/paddle/v2/framework/tests/test_nccl_ops.py
+++ /dev/null
@@ -1,87 +0,0 @@
-import unittest, os
-import numpy as np
-import paddle.v2 as paddle
-from paddle.v2.framework.op import Operator
-import paddle.v2.framework.core as core
-from op_test import OpTest, create_op, set_input
-
-# gpu_list = os.environ["NV_LIST"]
-gpu_list = "0,1,2,3"
-
-if not core.is_compile_gpu() or not gpu_list:
-    exit(0)
-
-
-def allreduce(tensors, gpus):
-    num_device = len(gpus)
-    assert (len(tensors) == num_device), "not match of tensor and device"
-    Out = tensors
-    for i in range(1, len(tensors)):
-        Out[0] += Out[i]
-
-    for i in range(1, len(tensors)):
-        Out[i] = Out[0]
-
-    return Out
-
-
-class TestNCCLAllReduce(unittest.TestCase):
-    def setUp(self):
-
-        self.op_type = "ncclAllReduce"
-
-        self.gpus = [int(g) for g in gpu_list.split(",")]
-
-        self.g_scope = core.Scope()
-        self.g_ctx = core.DeviceContext.create(core.CPUPlace())
-        self.scopes = []
-        self.ops = []
-        self.places = []
-
-        self.input_data = []
-
-        for i in range(len(self.gpus)):
-            self.input_data.append(np.random.random((32, 32)))
-        self.output_data = allreduce(self.input_data, self.gpus)
-
-        nccl_init = Operator("ncclInit", Out="Communicator", gpus=self.gpus)
-        op.run(self.g_scope, self.g_ctx)
-
-        for i in range(len(self.gpus)):
-            # insert kid scope
-            scope = self.g_scope.new_scope()
-            place = core.GPUPlace(self.gpus[i])
-
-            inputs = {"X": self.input_data[i]}
-            outputs = {"Out": self.output_data[i]}
-            attrs = {"gpus": self.gpus}
-
-            op = create_op(scope, self.op_type, inputs, outputs, attrs)
-            set_input(scope, op, inputs, place)
-
-            self.scopes.append(scope)
-            self.ops.append(op)
-            self.places.append(place)
-
-    def test_output(self):
-        idx = 0
-        for scope, place, op in zip(self.scopes, self.places, self.ops):
-            ctx = core.DeviceContext.create(place)
-            op.run(scope, ctx)
-
-        for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
-            actual = np.array(scope.find_var(out_name).get_tensor())
-            expect = self.output_data[idx]
-
-            idx += 1
-            self.assertTrue(actual, expect), "has diff"
-
-
-# if __name__ == "__main__":
-#     unittest.main()
-# usage : export NV_LIST=0,1,2,3 python *.py
-
-# os.environ["NV_LIST"] = ["0,1,2,3"]
-
-if __name__ == "__main__":
-    unittest.main()