diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index ee154207754f20791bdba55d0d66aea600e8dee6..e53a6a562ad1ed2ca02683b07cf6d4b56bc2cde7 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -53,7 +53,7 @@ if (WITH_GPU)
         op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
     endif()
     # conv_fusion_op needs cudnn 7 above
-    if (NOT ${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
+    if (NOT ${CUDNN_VERSION} VERSION_LESS 7100)
         op_library(conv_fusion_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_fusion);\n")
     endif()
diff --git a/paddle/fluid/operators/detection/density_prior_box_op.cu b/paddle/fluid/operators/detection/density_prior_box_op.cu
index acd5993154ed03f206f20082231feb5059ef32e1..6337a4837a64cef2ce0e7bae70d8ba5b8994958e 100644
--- a/paddle/fluid/operators/detection/density_prior_box_op.cu
+++ b/paddle/fluid/operators/detection/density_prior_box_op.cu
@@ -148,7 +148,7 @@ class DensityPriorBoxOpCUDAKernel : public framework::OpKernel<T> {
     // blockx is multiple of 32.
     int blockx = std::min(
         static_cast<int64_t>(((feature_width * num_priors + 31) >> 5) << 5),
-        512L);
+        static_cast<int64_t>(512L));
     int gridx = (feature_width * num_priors + blockx - 1) / blockx;
     dim3 threads(blockx, 1);
     dim3 grids(gridx, feature_height);
diff --git a/paddle/fluid/operators/huber_loss_op.h b/paddle/fluid/operators/huber_loss_op.h
index 9efda3dfc9871f197e0a66329772df2caedc4da4..fa21bd01cb052e5393385f0ef6c0203fc7c9e1a3 100644
--- a/paddle/fluid/operators/huber_loss_op.h
+++ b/paddle/fluid/operators/huber_loss_op.h
@@ -105,14 +105,16 @@ class HuberLossGradKernel : public framework::OpKernel<T> {
       out0->mutable_data<T>(context.GetPlace());
       auto x_grad = EigenVector<T>::Flatten(*out0);
       x_grad.device(place) =
-          out_grad * residual.unaryExpr(HuberLossBackward<T>(delta, -1.0));
+          residual.unaryExpr(HuberLossBackward<T>(delta, -1.0));
+      x_grad.device(place) = out_grad * x_grad;
     }
 
     if (out1) {
       out1->mutable_data<T>(context.GetPlace());
       auto y_grad = EigenVector<T>::Flatten(*out1);
       y_grad.device(place) =
-          out_grad * residual.unaryExpr(HuberLossBackward<T>(delta, 1.0));
+          residual.unaryExpr(HuberLossBackward<T>(delta, 1.0));
+      y_grad.device(place) = out_grad * y_grad;
     }
   }
 };
diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt
index 6d6fe245d8a0d9b3a29f11171e7d945e09a4133c..344130499506fbbecdda6551dc9abe3fca22d153 100644
--- a/python/paddle/fluid/tests/unittests/CMakeLists.txt
+++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt
@@ -25,7 +25,7 @@ endif(NOT WITH_DISTRIBUTE)
 
 if (NOT ${WITH_GPU})
     LIST(REMOVE_ITEM TEST_OPS test_conv2d_fusion_op)
-elseif(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
+elseif(${CUDNN_VERSION} VERSION_LESS 7100)
     LIST(REMOVE_ITEM TEST_OPS test_conv2d_fusion_op)
 endif()