diff --git a/src/tree/gpu_hist/histogram.cu b/src/tree/gpu_hist/histogram.cu
index 426343901..c6f6b79b2 100644
--- a/src/tree/gpu_hist/histogram.cu
+++ b/src/tree/gpu_hist/histogram.cu
@@ -306,12 +306,14 @@ void BuildGradientHistogram(CUDAContext const* ctx, EllpackDeviceAccessor const&
     dh::safe_cuda(cudaDeviceGetAttribute(&n_mps, cudaDevAttrMultiProcessorCount, device));
     int n_blocks_per_mp = 0;
     dh::safe_cuda(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&n_blocks_per_mp, kernel,
+                                                                kBlockThreads, smem_size));
 #elif defined(XGBOOST_USE_HIP)
     dh::safe_cuda(hipDeviceGetAttribute(&n_mps, hipDeviceAttributeMultiprocessorCount, device));
     int n_blocks_per_mp = 0;
     dh::safe_cuda(hipOccupancyMaxActiveBlocksPerMultiprocessor(&n_blocks_per_mp, kernel,
-#endif
                                                                 kBlockThreads, smem_size));
+#endif
+
     // This gives the number of blocks to keep the device occupied
     // Use this as the maximum number of blocks
     unsigned grid_size = n_blocks_per_mp * n_mps;