diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index 585634fae9c85f77cc77d774ac166891014a025c..3917b9b65b5905be38ba8a236aa158f42586c825 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -561,6 +561,7 @@ AnalysisPredictor::~AnalysisPredictor() {
 }
 
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
+  std::lock_guard<std::mutex> lk(clone_mutex_);
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);
   return std::unique_ptr<PaddlePredictor>(x);
diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h
index a6e126c5d533f4299ccc3deed7d116cabc71f75b..6ca4b5e9bed7505fca3b833dfbb7026ff550d258 100644
--- a/paddle/fluid/inference/api/analysis_predictor.h
+++ b/paddle/fluid/inference/api/analysis_predictor.h
@@ -115,6 +115,8 @@ class AnalysisPredictor : public PaddlePredictor {
   // concurrency problems, wrong results and memory leak, so cache them.
   std::vector<framework::LoDTensor> feed_tensors_;
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
+  // A mutex help to make Clone thread safe.
+  std::mutex clone_mutex_;
 
  private:
   // Some status here that help to determine the status inside the predictor.
diff --git a/paddle/fluid/inference/api/analysis_predictor_tester.cc b/paddle/fluid/inference/api/analysis_predictor_tester.cc
index 6169e60541e4a14d560e719d56624b3219dbcefd..3df26cde3d5defac97074c9bc4086e81f9ec0c93 100644
--- a/paddle/fluid/inference/api/analysis_predictor_tester.cc
+++ b/paddle/fluid/inference/api/analysis_predictor_tester.cc
@@ -179,8 +179,9 @@ TEST(AnalysisPredictor, Clone) {
     threads.emplace_back([&predictors, &inputs, i] {
       LOG(INFO) << "thread #" << i << " running";
       std::vector<PaddleTensor> outputs;
+      auto predictor = predictors.front()->Clone();
       for (int j = 0; j < 10; j++) {
-        ASSERT_TRUE(predictors[i]->Run(inputs, &outputs));
+        ASSERT_TRUE(predictor->Run(inputs, &outputs));
       }
     });
   }
diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index e5189e958bdf6afe1f94dc9e2d721e3d2c8b904c..e18bc02d92eb517fa20dc83811694b8ac80ae316 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -161,6 +161,8 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
 }
 
 std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
+  std::lock_guard<std::mutex> lk(clone_mutex_);
+  VLOG(3) << "Predictor::clone";
   std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
   // Hot fix the bug that result diff in multi-thread.
   // TODO(Superjomn) re-implement a real clone here.
diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h
index d2133bd467376c723a80a98725ac7c70234c54b0..96b94777304382a9d4be115a84f80ead69249863 100644
--- a/paddle/fluid/inference/api/api_impl.h
+++ b/paddle/fluid/inference/api/api_impl.h
@@ -74,6 +74,8 @@ class NativePaddlePredictor : public PaddlePredictor {
   // Do not use unique_ptr, use parent scope to delete
   framework::Scope *sub_scope_{nullptr};
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
+  // A mutex to make Clone thread safe.
+  std::mutex clone_mutex_;
 };
 
 }  // namespace paddle