diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index 652845e1aa661cb7d17bda88f50ebf5877d90d5c..6257983b7666e08252d6b77388ba80d07ce285c4 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -57,8 +57,8 @@ if(WITH_COVERAGE OR NOT WITH_AVX OR WIN32)
     SET(OP_MKL_DEPS ${OP_MKL_DEPS} pyramid_hash_op)
 endif()
 
-register_operators(EXCLUDES py_func_op warpctc_op dgc_op conv_fusion_op
-	sync_batch_norm_op multihead_matmul_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS})
+register_operators(EXCLUDES py_func_op warpctc_op dgc_op
+	sync_batch_norm_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS})
 
 if (WITH_GPU)
     # warpctc_op needs cudnn 7 above
@@ -67,17 +67,10 @@ if (WITH_GPU)
     else()
         op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
     endif()
-    # conv_fusion_op needs cudnn 7 above
-    if (NOT ${CUDNN_VERSION} VERSION_LESS 7100)
-        op_library(conv_fusion_op)
-        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_fusion);\n")
-    endif()
     if (NOT WIN32)
         op_library(sync_batch_norm_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n")
     endif()
-    op_library(multihead_matmul_op)
-    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(multihead_matmul);\n")
 else()
     op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
 endif()
diff --git a/paddle/fluid/operators/fused/CMakeLists.txt b/paddle/fluid/operators/fused/CMakeLists.txt
index a31531c599a71e7da0697825a12ab86f5d809a51..42529f02920f91f55ef846108710292af93e1e01 100644
--- a/paddle/fluid/operators/fused/CMakeLists.txt
+++ b/paddle/fluid/operators/fused/CMakeLists.txt
@@ -1,12 +1,29 @@
 include(operators)
-register_operators(EXCLUDES fusion_transpose_flatten_concat_op fusion_conv_inception_op fused_fc_elementwise_layernorm_op)
+register_operators(EXCLUDES
+    conv_fusion_op
+    fusion_transpose_flatten_concat_op
+    fusion_conv_inception_op
+    fused_fc_elementwise_layernorm_op
+    multihead_matmul_op)
+
 if (WITH_GPU)
-  op_library(fusion_transpose_flatten_concat_op)
-  file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fusion_transpose_flatten_concat);\n")
-  if (NOT ${CUDNN_VERSION} VERSION_LESS 7100)
-      op_library(fusion_conv_inception_op)
-      file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_inception_fusion);\n")
-  endif()
-  op_library(fused_fc_elementwise_layernorm_op)
-  file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fused_fc_elementwise_layernorm);\n")
+    # conv_fusion_op needs cudnn 7 above
+    if (NOT ${CUDNN_VERSION} VERSION_LESS 7100)
+        op_library(conv_fusion_op)
+        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_fusion);\n")
+    endif()
+    # fusion_transpose_flatten_concat_op
+    op_library(fusion_transpose_flatten_concat_op)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fusion_transpose_flatten_concat);\n")
+    # fusion_conv_inception_op needs cudnn 7 above
+    if (NOT ${CUDNN_VERSION} VERSION_LESS 7100)
+        op_library(fusion_conv_inception_op)
+        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_inception_fusion);\n")
+    endif()
+    # fused_fc_elementwise_layernorm_op
+    op_library(fused_fc_elementwise_layernorm_op)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fused_fc_elementwise_layernorm);\n")
+    # multihead_matmul_op
+    op_library(multihead_matmul_op)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(multihead_matmul);\n")
 endif()
diff --git a/paddle/fluid/operators/conv_fusion_op.cc b/paddle/fluid/operators/fused/conv_fusion_op.cc
similarity index 88%
rename from paddle/fluid/operators/conv_fusion_op.cc
rename to paddle/fluid/operators/fused/conv_fusion_op.cc
index 23b8087e781da30ed7b66ba651f8071ecb7aaf50..dd74d278111002dfee7a77b06c7126cceaa2afc1 100644
--- a/paddle/fluid/operators/conv_fusion_op.cc
+++ b/paddle/fluid/operators/fused/conv_fusion_op.cc
@@ -62,10 +62,10 @@ class Conv2DFusionOpMaker : public Conv2DOpMaker {
 class Conv2DFusionOpInferShape : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("Input"),
-                   "Input(Input) of ConvOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasInput("Filter"),
-                   "Input(Filter) of ConvOp should not be null.");
+    PADDLE_ENFORCE_EQ(ctx->HasInput("Input"), true,
+                      "Input(Input) of ConvOp should not be null.");
+    PADDLE_ENFORCE_EQ(ctx->HasInput("Filter"), true,
+                      "Input(Filter) of ConvOp should not be null.");
     auto in_dims = ctx->GetInputDim("Input");
     auto filter_dims = ctx->GetInputDim("Filter");
 
@@ -79,14 +79,14 @@ class Conv2DFusionOpInferShape : public framework::InferShapeBase {
       oshape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
                                       dilations[i], paddings[i], strides[i]));
     }
-    PADDLE_ENFORCE(ctx->HasOutput("Output"),
-                   "Output(Output) of ConvOp should not be null.");
+    PADDLE_ENFORCE_EQ(ctx->HasOutput("Output"), true,
+                      "Output(Output) of ConvOp should not be null.");
     ctx->SetOutputDim("Output", framework::make_ddim(oshape));
     std::vector<int> channels =
         ctx->Attrs().Get<std::vector<int>>("split_channels");
     if (channels.size()) {
-      PADDLE_ENFORCE(ctx->HasOutputs("Outputs"),
-                     "Output(Outputs) of ConvOp should not be null.");
+      PADDLE_ENFORCE_EQ(ctx->HasOutputs("Outputs"), true,
+                        "Output(Outputs) of ConvOp should not be null.");
       std::vector<framework::DDim> oshapes;
       oshapes.reserve(channels.size());
       for (size_t i = 0; i < channels.size(); ++i) {
diff --git a/paddle/fluid/operators/conv_fusion_op.cu.cc b/paddle/fluid/operators/fused/conv_fusion_op.cu.cc
similarity index 99%
rename from paddle/fluid/operators/conv_fusion_op.cu.cc
rename to paddle/fluid/operators/fused/conv_fusion_op.cu.cc
index 566daa6608282d89a92418e2c168bdf2c09c65c1..0e2f3e1d8869f46f36c2bd90130a21df81f054e7 100644
--- a/paddle/fluid/operators/conv_fusion_op.cu.cc
+++ b/paddle/fluid/operators/fused/conv_fusion_op.cu.cc
@@ -41,7 +41,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
     auto* input = ctx.Input<Tensor>("Input");
     auto* filter = ctx.Input<Tensor>("Filter");
     auto* bias = ctx.Input<Tensor>("Bias");
-    PADDLE_ENFORCE(bias, "The bias should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(bias, "The bias should not be null.");
     auto* residual = ctx.Input<Tensor>("ResidualData");
     auto* output = ctx.Output<Tensor>("Output");
 
diff --git a/paddle/fluid/operators/multihead_matmul_op.cc b/paddle/fluid/operators/fused/multihead_matmul_op.cc
similarity index 100%
rename from paddle/fluid/operators/multihead_matmul_op.cc
rename to paddle/fluid/operators/fused/multihead_matmul_op.cc
diff --git a/paddle/fluid/operators/multihead_matmul_op.cu b/paddle/fluid/operators/fused/multihead_matmul_op.cu
similarity index 100%
rename from paddle/fluid/operators/multihead_matmul_op.cu
rename to paddle/fluid/operators/fused/multihead_matmul_op.cu