diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc
index 98b4178177b0a8bafd6fe34a92be2a07a2fbc5a7..ab02a95f26b638606264c453a8165f05ff39a44a 100644
--- a/paddle/fluid/platform/device_context.cc
+++ b/paddle/fluid/platform/device_context.cc
@@ -159,6 +159,7 @@ CUDADeviceContext::~CUDADeviceContext() {
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
+  std::lock_guard<std::mutex> guard(mutex_);
   PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
   PADDLE_ENFORCE(cudaGetLastError());
 }
diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h
index 603b890af13b529c490c29112a73a09cc815d07a..df0a427b48b2e6d998d59a238985d190986cf20c 100644
--- a/paddle/fluid/platform/device_context.h
+++ b/paddle/fluid/platform/device_context.h
@@ -103,6 +103,7 @@ class CUDADeviceContext : public DeviceContext {
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
+  mutable std::mutex mutex_;
   cudaStream_t stream_;
   cudnnHandle_t cudnn_handle_;
   cublasHandle_t cublas_handle_;