diff --git a/src/tree/gpu_hist/evaluate_splits.cu b/src/tree/gpu_hist/evaluate_splits.cu
index a72357ec9..b49f389b2 100644
--- a/src/tree/gpu_hist/evaluate_splits.cu
+++ b/src/tree/gpu_hist/evaluate_splits.cu
@@ -332,6 +332,14 @@ __global__ __launch_bounds__(kBlockSize) void EvaluateSplitsKernel(
   }
 }
 
+template __global__ void EvaluateSplitsKernel<64>(
+    bst_feature_t max_active_features,
+    common::Span<const EvaluateSplitInputs> d_inputs,
+    const EvaluateSplitSharedInputs shared_inputs,
+    common::Span<bst_feature_t> sorted_idx,
+    const TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator,
+    common::Span<DeviceSplitCandidate> out_candidates);
+
 __device__ DeviceSplitCandidate operator+(const DeviceSplitCandidate &a,
                                           const DeviceSplitCandidate &b) {
   return b.loss_chg > a.loss_chg ? b : a;