diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc
index aee94e12340597e981ac385a01335d2ffa069191..85910c10e7409ac629c8fb5265550a5cfa6bc2c2 100644
--- a/paddle/fluid/inference/api/analysis_config.cc
+++ b/paddle/fluid/inference/api/analysis_config.cc
@@ -141,7 +141,6 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
 
 void AnalysisConfig::EnableMKLDNN() {
 #ifdef PADDLE_WITH_MKLDNN
-  pass_builder()->EnableMKLDNN();
   use_mkldnn_ = true;
 #else
   LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
@@ -234,16 +233,13 @@ void AnalysisConfig::Update() {
   }
 
   if (use_mkldnn_) {
+#ifdef PADDLE_WITH_MKLDNN
     if (!enable_ir_optim_) {
       LOG(ERROR)
           << "EnableMKLDNN() only works when IR optimization is enabled.";
+    } else {
+      pass_builder()->EnableMKLDNN();
     }
-#ifdef PADDLE_WITH_MKLDNN
-    pass_builder()->EnableMKLDNN();
-    use_mkldnn_ = true;
-#else
-    LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
-    use_mkldnn_ = false;
 #endif
   }
 
@@ -255,9 +251,6 @@ void AnalysisConfig::Update() {
     }
 #ifdef PADDLE_WITH_MKLDNN
     pass_builder()->EnableMkldnnQuantizer();
-#else
-    LOG(ERROR) << "Please compile with MKLDNN first to use MkldnnQuantizer";
-    use_mkldnn_quantizer_ = false;
 #endif
   }
 
diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc
index 1d1d39e44096b9f50e5bc9603fa12aba92b0e8e2..87e02a02caebd93d701dfd9e51c35fb974c770ed 100644
--- a/paddle/fluid/inference/api/paddle_pass_builder.cc
+++ b/paddle/fluid/inference/api/paddle_pass_builder.cc
@@ -64,10 +64,12 @@ void PaddlePassBuilder::DeletePass(size_t idx) {
   passes_.erase(std::begin(passes_) + idx);
 }
 
-void GpuPassStrategy::EnableMKLDNN() {
-  LOG(ERROR) << "GPU not support MKLDNN yet";
+void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
+  analysis_passes_.push_back(pass);
 }
 
+void PaddlePassBuilder::ClearPasses() { passes_.clear(); }
+
 // The following passes works for Anakin sub-graph engine.
 const std::vector<std::string> kAnakinSubgraphPasses({
     "infer_clean_graph_pass",                       //
@@ -102,12 +104,12 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   use_gpu_ = true;
 }
 
-void GpuPassStrategy::EnableMkldnnQuantizer() {
-  LOG(ERROR) << "GPU not support MKL-DNN quantization";
+void GpuPassStrategy::EnableMKLDNN() {
+  LOG(ERROR) << "GPU not support MKLDNN yet";
 }
 
-void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
-  analysis_passes_.push_back(pass);
+void GpuPassStrategy::EnableMkldnnQuantizer() {
+  LOG(ERROR) << "GPU not support MKL-DNN quantization";
 }
 
 CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
@@ -135,5 +137,39 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   });
   use_gpu_ = false;
 }
-void PaddlePassBuilder::ClearPasses() { passes_.clear(); }
+
+void CpuPassStrategy::EnableMKLDNN() {
+// TODO(Superjomn) Consider the way to mix CPU with GPU.
+#ifdef PADDLE_WITH_MKLDNN
+  if (!use_mkldnn_) {
+    passes_.insert(passes_.begin(), "mkldnn_placement_pass");
+
+    for (auto &pass : std::vector<std::string>(
+             {"depthwise_conv_mkldnn_pass",    //
+              "conv_bn_fuse_pass",             // Execute BN passes again to
+              "conv_eltwiseadd_bn_fuse_pass",  // preserve correct pass order
+              "conv_bias_mkldnn_fuse_pass",    //
+              "conv3d_bias_mkldnn_fuse_pass",  //
+              "conv_elementwise_add_mkldnn_fuse_pass",
+              "conv_relu_mkldnn_fuse_pass"})) {
+      passes_.push_back(pass);
+    }
+  }
+  use_mkldnn_ = true;
+#else
+  use_mkldnn_ = false;
+#endif
+}
+
+void CpuPassStrategy::EnableMkldnnQuantizer() {
+#ifdef PADDLE_WITH_MKLDNN
+  if (!use_mkldnn_quantizer_) {
+    passes_.push_back("cpu_quantize_placement_pass");
+  }
+  use_mkldnn_quantizer_ = true;
+#else
+  use_mkldnn_quantizer_ = false;
+#endif
+}
+
 }  // namespace paddle
diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h
index 48da8c156f426477011bcc060260c812ad94df23..09ef195d5e66aff0cef17f1594de34c656187a35 100644
--- a/paddle/fluid/inference/api/paddle_pass_builder.h
+++ b/paddle/fluid/inference/api/paddle_pass_builder.h
@@ -109,43 +109,16 @@ class CpuPassStrategy : public PassStrategy {
   CpuPassStrategy();
 
   explicit CpuPassStrategy(const CpuPassStrategy &other)
-      : PassStrategy(other.AllPasses()) {}
+      : PassStrategy(other.AllPasses()) {
+    use_gpu_ = other.use_gpu_;
+    use_mkldnn_ = other.use_mkldnn_;
+    use_mkldnn_quantizer_ = other.use_mkldnn_quantizer_;
+  }
 
   virtual ~CpuPassStrategy() = default;
 
-  void EnableMKLDNN() override {
-// TODO(Superjomn) Consider the way to mix CPU with GPU.
-#ifdef PADDLE_WITH_MKLDNN
-    if (!use_mkldnn_) {
-      passes_.insert(passes_.begin(), "mkldnn_placement_pass");
-
-      for (auto &pass : std::vector<std::string>(
-               {"depthwise_conv_mkldnn_pass",    //
-                "conv_bn_fuse_pass",             // Execute BN passes again to
-                "conv_eltwiseadd_bn_fuse_pass",  // preserve correct pass order
-                "conv_bias_mkldnn_fuse_pass",    //
-                "conv3d_bias_mkldnn_fuse_pass",  //
-                "conv_relu_mkldnn_fuse_pass",    //
-                "conv_elementwise_add_mkldnn_fuse_pass"})) {
-        passes_.push_back(pass);
-      }
-    }
-    use_mkldnn_ = true;
-#else
-    use_mkldnn_ = false;
-#endif
-  }
-
-  void EnableMkldnnQuantizer() override {
-#ifdef PADDLE_WITH_MKLDNN
-    if (!use_mkldnn_quantizer_) {
-      passes_.push_back("cpu_quantize_placement_pass");
-    }
-    use_mkldnn_quantizer_ = true;
-#else
-    use_mkldnn_quantizer_ = false;
-#endif
-  }
+  void EnableMKLDNN() override;
+  void EnableMkldnnQuantizer() override;
 
  protected:
   bool use_mkldnn_quantizer_{false};