diff --git a/src/operator/nn/layer_norm.cu b/src/operator/nn/layer_norm.cu
index b2de5726fc18..0d7bad530311 100644
--- a/src/operator/nn/layer_norm.cu
+++ b/src/operator/nn/layer_norm.cu
@@ -312,7 +312,7 @@ void LayerNormGPUContig(const LayerNormParam param,
   cudaStream_t stream = Stream<gpu>::GetStream(ctx.get_stream<gpu>());
   const dim3 dimBlock(32, nthread_y);
   MXNET_ACC_TYPE_SWITCH(in_data.type_flag_, DType, AccType, {
-    typedef std::conditional<safe_acc, AccType, DType>::type AType;
+    typedef typename std::conditional<safe_acc, AccType, DType>::type AType;
     int nshared = nthread_y > 1 ? nthread_y * 32 * sizeof(AType)
                                   + (nthread_y / 2) * 32 * sizeof(int) : 0;
     CheckLaunchParam(dimGrid, dimBlock);
@@ -637,7 +637,7 @@ void LayerNormGradGPUContig(const LayerNormParam param,
                                &gb_block_dim, &gb_grid_dim, &npart);
   if (gamma_grad_req != kNullOp || beta_grad_req != kNullOp) {
     MXNET_ACC_TYPE_SWITCH(in_data.type_flag_, DType, AccType, {
-      typedef std::conditional<safe_acc, AccType, DType>::type AType;
+      typedef typename std::conditional<safe_acc, AccType, DType>::type AType;
       Tensor<gpu, 1, AType> workspace =
         ctx.requested[0].get_space_typed<gpu, 1, AType>(Shape1(2 * npart * nchannel), s);
       AType* part_gamma_grad_ptr = workspace.dptr_;
@@ -696,7 +696,7 @@ void LayerNormGradGPUContig(const LayerNormParam param,
   const int LOAD_UNROLL = 4;
   if (data_grad_req != kNullOp) {
     MXNET_ACC_TYPE_SWITCH(in_data.type_flag_, DType, AccType, {
-      typedef std::conditional<safe_acc, AccType, DType>::type AType;
+      typedef typename std::conditional<safe_acc, AccType, DType>::type AType;
       int nshared = data_block_dim.y > 1 ? data_block_dim.y * data_block_dim.x * sizeof(AType) : 0;
       CheckLaunchParam(data_grid_dim, data_block_dim);
       if (data_grad_req == kAddTo) {