Fuse gpu_hist all-reduce calls where possible (#7867)

src/tree/gpu_hist/evaluate_splits.cuh

@@ -103,7 +103,7 @@ class GPUHistEvaluator {
   }
 
   /**
-   * \brief Get sorted index storage based on the left node of inputs .
+   * \brief Get sorted index storage based on the left node of inputs.
    */
   auto SortedIdx(EvaluateSplitInputs<GradientSumT> left) {
     if (left.nidx == RegTree::kRoot && !cat_sorted_idx_.empty()) {

src/tree/gpu_hist/histogram.cu

@@ -247,15 +247,6 @@ void BuildGradientHistogram(EllpackDeviceAccessor const& matrix,
   dh::safe_cuda(cudaGetLastError());
 }
 
-template void BuildGradientHistogram<GradientPair>(
-    EllpackDeviceAccessor const& matrix,
-    FeatureGroupsAccessor const& feature_groups,
-    common::Span<GradientPair const> gpair,
-    common::Span<const uint32_t> ridx,
-    common::Span<GradientPair> histogram,
-    HistRounding<GradientPair> rounding,
-    bool force_global_memory);
-
 template void BuildGradientHistogram<GradientPairPrecise>(
     EllpackDeviceAccessor const& matrix,
     FeatureGroupsAccessor const& feature_groups,