diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h
index 6b82d93237b6baa20703c5b54b56f5381dd858df..292ffef1aef12732812b8c5b0020cad73b1d06fc 100644
--- a/paddle/fluid/platform/device_context.h
+++ b/paddle/fluid/platform/device_context.h
@@ -11,6 +11,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory>
+#include <mutex>  // NOLINT
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -100,6 +101,7 @@ class CUDADeviceContext : public DeviceContext {
 
   template <typename Callback>
   void RecordEvent(cudaEvent_t ev, Callback callback) {
+    std::lock_guard<std::mutex> guard(mtx_);
     callback();
     PADDLE_ENFORCE(cudaEventRecord(ev, stream_));
   }
@@ -116,6 +118,8 @@ class CUDADeviceContext : public DeviceContext {
   int compute_capability;
   int multi_process;
   int max_threads_per_mp;
+
+  std::mutex mtx_;
 };
 
 template <>