diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc
index 2e79d495d5ff00000000029ac0f6eb486aaea94a..ef4142f334e503380dc7ccd74c348404ffe52ee6 100644
--- a/paddle/fluid/inference/analysis/analyzer.cc
+++ b/paddle/fluid/inference/analysis/analyzer.cc
@@ -107,6 +107,9 @@ void Analyzer::Run(Argument* argument) {
     passes.push_back("mkldnn_placement_pass");
   }
 #endif
+  // infer_clean_graph_pass should be the first default pass
+  // after mkldnn_placement_pass.
+  passes.push_back("infer_clean_graph_pass");
   for (auto& pass : ir_passes_) {
     if (!disabled_ir_passes_.count(pass)) {
       passes.push_back(pass);
diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h
index c51a4fdb2f6b27e54637481c23bf6f1f6ec97718..7114f5222c5904bb4422b9e67ad035b85bbb770c 100644
--- a/paddle/fluid/inference/analysis/analyzer.h
+++ b/paddle/fluid/inference/analysis/analyzer.h
@@ -67,7 +67,6 @@ class Analyzer : public OrderedRegistry<PassManager> {
   // larger fusion.
   const std::vector<std::string> all_ir_passes_{{
       // Manual update the passes here.
-      "infer_clean_graph_pass",         //
       "attention_lstm_fuse_pass",       //
       "seqconv_eltadd_relu_fuse_pass",  //
       "embedding_fc_lstm_fuse_pass",    //