diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index 90d4444accd21a3804fde1bc0a4c0c4078d40892..753c87253237b1f2b46d657605efed98d7302fe9 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -500,6 +500,7 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
       std::string flag = "--fraction_of_gpu_memory_to_use=" +
                          std::to_string(fraction_of_gpu_memory);
       flags.push_back(flag);
+      flags.push_back("--cudnn_deterministic=True");
       VLOG(3) << "set flag: " << flag;
       framework::InitGflags(flags);
     }
diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu b/paddle/fluid/operators/conv_transpose_cudnn_op.cu
index d6c7f964c68e8f80d79064dc7f6dab7767a601be..6294b16a5458d4a87084df21a3b1835d0f95af53 100644
--- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu
+++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu
@@ -243,6 +243,10 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
         cudnn_output_desc, CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
         workspace_size_limit, &algo));
 
+    if (algo == 0 && FLAGS_cudnn_deterministic) {
+      algo = static_cast<cudnnConvolutionBwdDataAlgo_t>(1);
+    }
+
     // get workspace size able to allocate
     CUDNN_ENFORCE(
         platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(