diff --git a/cpp/bench/common/ml_benchmark.hpp b/cpp/bench/common/ml_benchmark.hpp
index 71feb7608b..fb90d64aca 100644
--- a/cpp/bench/common/ml_benchmark.hpp
+++ b/cpp/bench/common/ml_benchmark.hpp
@@ -21,7 +21,7 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/common/utils.hpp>
 #include <memory>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <sstream>
 #include <string>
 #include <vector>
diff --git a/cpp/bench/sg/arima_loglikelihood.cu b/cpp/bench/sg/arima_loglikelihood.cu
index 34ae8f8b35..9c9088fb65 100644
--- a/cpp/bench/sg/arima_loglikelihood.cu
+++ b/cpp/bench/sg/arima_loglikelihood.cu
@@ -25,7 +25,7 @@
 #include <rmm/device_uvector.hpp>
 
 #include "benchmark.cuh"
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/bench/sg/benchmark.cuh b/cpp/bench/sg/benchmark.cuh
index 3dbc50e087..5299d9c5fe 100644
--- a/cpp/bench/sg/benchmark.cuh
+++ b/cpp/bench/sg/benchmark.cuh
@@ -22,8 +22,8 @@
 #include <benchmark/benchmark.h>
 #include <cuda_runtime.h>
 #include <cuml/common/logger.hpp>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/bench/sg/dataset.cuh b/cpp/bench/sg/dataset.cuh
index 6d965af53f..5b15535fbc 100644
--- a/cpp/bench/sg/dataset.cuh
+++ b/cpp/bench/sg/dataset.cuh
@@ -19,12 +19,12 @@
 #include <cuml/datasets/make_blobs.hpp>
 #include <fstream>
 #include <iostream>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/transpose.cuh>
 #include <raft/linalg/unary_op.cuh>
 #include <raft/random/make_regression.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <sstream>
 #include <string>
 #include <vector>
diff --git a/cpp/bench/sg/dataset_ts.cuh b/cpp/bench/sg/dataset_ts.cuh
index 481a6b899f..385c5f62ea 100644
--- a/cpp/bench/sg/dataset_ts.cuh
+++ b/cpp/bench/sg/dataset_ts.cuh
@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/random/rng.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/bench/sg/umap.cu b/cpp/bench/sg/umap.cu
index c75dc0e51e..73366639e8 100644
--- a/cpp/bench/sg/umap.cu
+++ b/cpp/bench/sg/umap.cu
@@ -17,7 +17,7 @@
 #include "benchmark.cuh"
 #include <cuml/manifold/umap.hpp>
 #include <cuml/manifold/umapparams.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <utility>
 
 namespace ML {
diff --git a/cpp/examples/symreg/symreg_example.cpp b/cpp/examples/symreg/symreg_example.cpp
index 1ed530b5c4..a7043f278c 100644
--- a/cpp/examples/symreg/symreg_example.cpp
+++ b/cpp/examples/symreg/symreg_example.cpp
@@ -28,7 +28,7 @@
 #include <cuml/genetic/genetic.h>
 #include <cuml/genetic/program.h>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/include/cuml/common/utils.hpp b/cpp/include/cuml/common/utils.hpp
index 2a61d04e4f..fcabb32fcd 100644
--- a/cpp/include/cuml/common/utils.hpp
+++ b/cpp/include/cuml/common/utils.hpp
@@ -18,7 +18,7 @@
 
 #include <cuda_runtime.h>
 #include <execinfo.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cstdio>
 
diff --git a/cpp/include/cuml/fil/multi_sum.cuh b/cpp/include/cuml/fil/multi_sum.cuh
index 2b61cbdf6e..180134a494 100644
--- a/cpp/include/cuml/fil/multi_sum.cuh
+++ b/cpp/include/cuml/fil/multi_sum.cuh
@@ -15,7 +15,7 @@
  */
 /** @file multi_sum.cuh */
 #pragma once
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 /**
  template parameters: data [T]ype, reduction [R]adix
  function parameters:
diff --git a/cpp/include/cuml/neighbors/knn.hpp b/cpp/include/cuml/neighbors/knn.hpp
index a0730933c5..86d1bb2b30 100644
--- a/cpp/include/cuml/neighbors/knn.hpp
+++ b/cpp/include/cuml/neighbors/knn.hpp
@@ -18,7 +18,7 @@
 
 #include <raft/distance/distance_types.hpp>
 #include <raft/spatial/knn/ann_common.h>
-#include <raft/spatial/knn/ball_cover_common.h>
+#include <raft/spatial/knn/ball_cover_types.hpp>
 
 namespace raft {
 class handle_t;
diff --git a/cpp/include/cuml/tsa/arima_common.h b/cpp/include/cuml/tsa/arima_common.h
index feacae5b3e..ff60bc0cd5 100644
--- a/cpp/include/cuml/tsa/arima_common.h
+++ b/cpp/include/cuml/tsa/arima_common.h
@@ -20,7 +20,7 @@
 
 #include <algorithm>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <thrust/execution_policy.h>
 #include <thrust/for_each.h>
 #include <thrust/iterator/counting_iterator.h>
diff --git a/cpp/src/arima/batched_arima.cu b/cpp/src/arima/batched_arima.cu
index 4d0840c1d8..a989bc5812 100644
--- a/cpp/src/arima/batched_arima.cu
+++ b/cpp/src/arima/batched_arima.cu
@@ -34,8 +34,8 @@
 #include <raft/core/handle.hpp>
 #include <raft/core/nvtx.hpp>
 #include <raft/linalg/matrix_vector_op.cuh>
-#include <raft/stats/common.hpp>
 #include <raft/stats/information_criterion.cuh>
+#include <raft/stats/stats_types.hpp>
 #include <raft/util/cuda_utils.cuh>
 #include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
diff --git a/cpp/src/common/cumlHandle.cpp b/cpp/src/common/cumlHandle.cpp
index 187d8ede4f..4db5b7e342 100644
--- a/cpp/src/common/cumlHandle.cpp
+++ b/cpp/src/common/cumlHandle.cpp
@@ -17,7 +17,7 @@
 #include "cumlHandle.hpp"
 
 #include <cuml/common/logger.hpp>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
 // #TODO: Replace with public header when ready
diff --git a/cpp/src/common/cuml_api.cpp b/cpp/src/common/cuml_api.cpp
index 7ce6728395..b8b0e39b0e 100644
--- a/cpp/src/common/cuml_api.cpp
+++ b/cpp/src/common/cuml_api.cpp
@@ -19,7 +19,7 @@
 #include <cuml/common/utils.hpp>
 #include <cuml/cuml_api.h>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cstddef>
 #include <functional>
diff --git a/cpp/src/dbscan/dbscan.cu b/cpp/src/dbscan/dbscan.cu
index ec65ebe725..4121321546 100644
--- a/cpp/src/dbscan/dbscan.cu
+++ b/cpp/src/dbscan/dbscan.cu
@@ -17,7 +17,7 @@
 #include <cuml/cluster/dbscan.hpp>
 
 #include "dbscan.cuh"
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace ML {
 namespace Dbscan {
diff --git a/cpp/src/dbscan/runner.cuh b/cpp/src/dbscan/runner.cuh
index 284009650d..c433fa59e8 100644
--- a/cpp/src/dbscan/runner.cuh
+++ b/cpp/src/dbscan/runner.cuh
@@ -23,10 +23,10 @@
 #include "mergelabels/tree_reduction.cuh"
 #include "vertexdeg/runner.cuh"
 #include <common/nvtx.hpp>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/nvtx.hpp>
 #include <raft/label/classlabels.cuh>
 #include <raft/sparse/csr.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cuml/common/logger.hpp>
 
diff --git a/cpp/src/dbscan/vertexdeg/precomputed.cuh b/cpp/src/dbscan/vertexdeg/precomputed.cuh
index df70f81676..7725d4b5d2 100644
--- a/cpp/src/dbscan/vertexdeg/precomputed.cuh
+++ b/cpp/src/dbscan/vertexdeg/precomputed.cuh
@@ -19,11 +19,11 @@
 #include <cub/cub.cuh>
 #include <cuda_runtime.h>
 #include <math.h>
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
-#include <raft/device_atomics.cuh>
 #include <raft/linalg/coalesced_reduction.cuh>
 #include <raft/linalg/reduce.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
+#include <raft/util/device_atomics.cuh>
 
 #include "pack.h"
 
diff --git a/cpp/src/decisiontree/batched-levelalgo/bins.cuh b/cpp/src/decisiontree/batched-levelalgo/bins.cuh
index 6accfe3c28..312c4f2b51 100644
--- a/cpp/src/decisiontree/batched-levelalgo/bins.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/bins.cuh
@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 #pragma once
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace DT {
diff --git a/cpp/src/decisiontree/batched-levelalgo/builder.cuh b/cpp/src/decisiontree/batched-levelalgo/builder.cuh
index c27d21f12f..b7045fbcf6 100644
--- a/cpp/src/decisiontree/batched-levelalgo/builder.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/builder.cuh
@@ -25,7 +25,7 @@
 #include <common/Timer.h>
 #include <cuml/common/pinned_host_vector.hpp>
 #include <cuml/tree/flatnode.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <deque>
 #include <raft/core/nvtx.hpp>
diff --git a/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels_impl.cuh b/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels_impl.cuh
index f93e25e58c..a42e53c8d2 100644
--- a/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels_impl.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels_impl.cuh
@@ -19,7 +19,7 @@
 
 #include <common/grid_sync.cuh>
 #include <cub/cub.cuh>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <thrust/binary_search.h>
 
 #include "builder_kernels.cuh"
diff --git a/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh b/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
index 438d03b953..174605a22d 100644
--- a/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
@@ -20,7 +20,7 @@
 #include <iostream>
 #include <memory>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
 #include <thrust/fill.h>
diff --git a/cpp/src/decisiontree/batched-levelalgo/split.cuh b/cpp/src/decisiontree/batched-levelalgo/split.cuh
index ea80412cfc..bb4bd5408a 100644
--- a/cpp/src/decisiontree/batched-levelalgo/split.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/split.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/unary_op.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace DT {
diff --git a/cpp/src/decisiontree/decisiontree.cuh b/cpp/src/decisiontree/decisiontree.cuh
index a6b8b0c229..cd425c6760 100644
--- a/cpp/src/decisiontree/decisiontree.cuh
+++ b/cpp/src/decisiontree/decisiontree.cuh
@@ -19,8 +19,8 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/tree/flatnode.h>
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include "treelite_util.h"
 #include <treelite/c_api.h>
diff --git a/cpp/src/explainer/kernel_shap.cu b/cpp/src/explainer/kernel_shap.cu
index 13dc46e15c..b90eec4991 100644
--- a/cpp/src/explainer/kernel_shap.cu
+++ b/cpp/src/explainer/kernel_shap.cu
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cuml/explainer/kernel_shap.hpp>
 
diff --git a/cpp/src/explainer/permutation_shap.cu b/cpp/src/explainer/permutation_shap.cu
index 54e560709d..8cb69b06a0 100644
--- a/cpp/src/explainer/permutation_shap.cu
+++ b/cpp/src/explainer/permutation_shap.cu
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cuml/explainer/permutation_shap.hpp>
 
diff --git a/cpp/src/fil/common.cuh b/cpp/src/fil/common.cuh
index c4877216a8..34eace065e 100644
--- a/cpp/src/fil/common.cuh
+++ b/cpp/src/fil/common.cuh
@@ -24,7 +24,7 @@
 #include <string>
 
 #include <cuml/fil/fil.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include "internal.cuh"
 
diff --git a/cpp/src/fil/fil.cu b/cpp/src/fil/fil.cu
index 8c3965ded6..8721019529 100644
--- a/cpp/src/fil/fil.cu
+++ b/cpp/src/fil/fil.cu
@@ -22,9 +22,9 @@ creation and prediction (the main inference kernel is defined in infer.cu). */
 
 #include <cuml/fil/fil.h>  // for algo_t,
 
-#include <raft/core/cudart_utils.hpp>  // for RAFT_CUDA_TRY, cudaStream_t,
 #include <raft/core/error.hpp>         // for ASSERT
 #include <raft/core/handle.hpp>        // for handle_t
+#include <raft/util/cudart_utils.hpp>  // for RAFT_CUDA_TRY, cudaStream_t,
 #include <rmm/device_uvector.hpp>      // for device_uvector
 #include <thrust/host_vector.h>        // for host_vector
 
diff --git a/cpp/src/fil/infer.cu b/cpp/src/fil/infer.cu
index 54ff1590ad..04acd71003 100644
--- a/cpp/src/fil/infer.cu
+++ b/cpp/src/fil/infer.cu
@@ -20,8 +20,8 @@
 
 #include <cuml/fil/multi_sum.cuh>
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <thrust/functional.h>
 
diff --git a/cpp/src/fil/internal.cuh b/cpp/src/fil/internal.cuh
index 487f2c702a..eb0566e5a3 100644
--- a/cpp/src/fil/internal.cuh
+++ b/cpp/src/fil/internal.cuh
@@ -23,7 +23,7 @@
 #include <iostream>
 #include <numeric>
 #include <raft/core/error.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 #include <treelite/c_api.h>
 #include <treelite/tree.h>
diff --git a/cpp/src/fil/treelite_import.cu b/cpp/src/fil/treelite_import.cu
index edb69b8ac1..f01b9cf2f7 100644
--- a/cpp/src/fil/treelite_import.cu
+++ b/cpp/src/fil/treelite_import.cu
@@ -24,9 +24,9 @@
 #include <cuml/fil/fil.h>  // for algo_t, from_treelite, storage_type_repr, storage_type_t, treelite_params_t
 #include <cuml/fil/fnv_hash.h>  // for fowler_noll_vo_fingerprint64_32
 
-#include <raft/core/cudart_utils.hpp>  // for RAFT_CUDA_TRY
 #include <raft/core/error.hpp>         // for ASSERT
 #include <raft/core/handle.hpp>        // for handle_t
+#include <raft/util/cudart_utils.hpp>  // for RAFT_CUDA_TRY
 
 #include <treelite/base.h>   // for Operator, SplitFeatureType, kGE, kGT, kLE, kLT, kNumerical
 #include <treelite/c_api.h>  // for ModelHandle
diff --git a/cpp/src/genetic/fitness.cuh b/cpp/src/genetic/fitness.cuh
index a440489339..82e00c2e24 100644
--- a/cpp/src/genetic/fitness.cuh
+++ b/cpp/src/genetic/fitness.cuh
@@ -14,7 +14,6 @@
  * limitations under the License.
  */
 
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/eltwise.cuh>
 #include <raft/linalg/matrix_vector_op.cuh>
 #include <raft/linalg/strided_reduction.cuh>
@@ -24,6 +23,7 @@
 #include <raft/stats/mean_center.cuh>
 #include <raft/stats/stddev.cuh>
 #include <raft/stats/sum.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
@@ -40,7 +40,7 @@
 #include <thrust/sort.h>
 #include <thrust/transform.h>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace cuml {
 namespace genetic {
diff --git a/cpp/src/genetic/node.cuh b/cpp/src/genetic/node.cuh
index b999940b41..a0d5ca95cf 100644
--- a/cpp/src/genetic/node.cuh
+++ b/cpp/src/genetic/node.cuh
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <cuml/genetic/node.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace cuml {
 namespace genetic {
diff --git a/cpp/src/genetic/reg_stack.cuh b/cpp/src/genetic/reg_stack.cuh
index 4696f6e975..92c852eb4f 100644
--- a/cpp/src/genetic/reg_stack.cuh
+++ b/cpp/src/genetic/reg_stack.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #ifndef CUDA_PRAGMA_UNROLL
 #ifdef __CUDA_ARCH__
diff --git a/cpp/src/glm/qn/glm_linear.cuh b/cpp/src/glm/qn/glm_linear.cuh
index eae1b8b8f7..1fb024abed 100644
--- a/cpp/src/glm/qn/glm_linear.cuh
+++ b/cpp/src/glm/qn/glm_linear.cuh
@@ -18,8 +18,8 @@
 
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_logistic.cuh b/cpp/src/glm/qn/glm_logistic.cuh
index 07d654fce3..267589507b 100644
--- a/cpp/src/glm/qn/glm_logistic.cuh
+++ b/cpp/src/glm/qn/glm_logistic.cuh
@@ -18,8 +18,8 @@
 
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_softmax.cuh b/cpp/src/glm/qn/glm_softmax.cuh
index 84555e4ad5..066e330a64 100644
--- a/cpp/src/glm/qn/glm_softmax.cuh
+++ b/cpp/src/glm/qn/glm_softmax.cuh
@@ -18,8 +18,8 @@
 
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_svm.cuh b/cpp/src/glm/qn/glm_svm.cuh
index 846c378d99..1d2bc3623e 100644
--- a/cpp/src/glm/qn/glm_svm.cuh
+++ b/cpp/src/glm/qn/glm_svm.cuh
@@ -18,8 +18,8 @@
 
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/qn_solvers.cuh b/cpp/src/glm/qn/qn_solvers.cuh
index 57322d692a..d52bdffeeb 100644
--- a/cpp/src/glm/qn/qn_solvers.cuh
+++ b/cpp/src/glm/qn/qn_solvers.cuh
@@ -44,7 +44,7 @@
 #include "qn_util.cuh"
 #include "simple_mat.cuh"
 #include <cuml/common/logger.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/glm/qn/qn_util.cuh b/cpp/src/glm/qn/qn_util.cuh
index a5c28172c6..43174f91ef 100644
--- a/cpp/src/glm/qn/qn_util.cuh
+++ b/cpp/src/glm/qn/qn_util.cuh
@@ -20,7 +20,7 @@
 
 #include <cuml/common/logger.hpp>
 #include <limits>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/simple_mat/base.hpp b/cpp/src/glm/qn/simple_mat/base.hpp
index 5d0cfefd2a..5b60094274 100644
--- a/cpp/src/glm/qn/simple_mat/base.hpp
+++ b/cpp/src/glm/qn/simple_mat/base.hpp
@@ -17,7 +17,7 @@
 
 #include <raft/core/handle.hpp>
 #include <raft/core/interruptible.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/glm/qn/simple_mat/dense.hpp b/cpp/src/glm/qn/simple_mat/dense.hpp
index 11c646b46a..f9b15809d2 100644
--- a/cpp/src/glm/qn/simple_mat/dense.hpp
+++ b/cpp/src/glm/qn/simple_mat/dense.hpp
@@ -20,9 +20,9 @@
 
 #include "base.hpp"
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/ternary_op.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <raft/util/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
diff --git a/cpp/src/hdbscan/condensed_hierarchy.cu b/cpp/src/hdbscan/condensed_hierarchy.cu
index 1d45df8f1f..4f3cdcc73b 100644
--- a/cpp/src/hdbscan/condensed_hierarchy.cu
+++ b/cpp/src/hdbscan/condensed_hierarchy.cu
@@ -19,7 +19,7 @@
 #include <cub/cub.cuh>
 
 #include <cuml/common/logger.hpp>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/hdbscan/detail/condense.cuh b/cpp/src/hdbscan/detail/condense.cuh
index b4b9685e6e..d19a22c69c 100644
--- a/cpp/src/hdbscan/detail/condense.cuh
+++ b/cpp/src/hdbscan/detail/condense.cuh
@@ -20,7 +20,7 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/hdbscan/detail/membership.cuh b/cpp/src/hdbscan/detail/membership.cuh
index c9fd087d3f..6a67ae2d51 100644
--- a/cpp/src/hdbscan/detail/membership.cuh
+++ b/cpp/src/hdbscan/detail/membership.cuh
@@ -21,7 +21,7 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/op/sort.cuh>
diff --git a/cpp/src/hdbscan/detail/predict.cuh b/cpp/src/hdbscan/detail/predict.cuh
index 74852f516d..34af443a5e 100644
--- a/cpp/src/hdbscan/detail/predict.cuh
+++ b/cpp/src/hdbscan/detail/predict.cuh
@@ -19,8 +19,8 @@
 #include "kernels/predict.cuh"
 #include "reachability.cuh"
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/hdbscan/detail/reachability_faiss.cuh b/cpp/src/hdbscan/detail/reachability_faiss.cuh
index 625c26539e..c2d970727b 100644
--- a/cpp/src/hdbscan/detail/reachability_faiss.cuh
+++ b/cpp/src/hdbscan/detail/reachability_faiss.cuh
@@ -39,8 +39,8 @@
 #include <faiss/gpu/utils/MatrixMult.cuh>
 
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/spatial/knn/faiss_mr.hpp>
+#include <raft/util/cuda_utils.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/hdbscan/detail/select.cuh b/cpp/src/hdbscan/detail/select.cuh
index c0a68a540a..9253fb5961 100644
--- a/cpp/src/hdbscan/detail/select.cuh
+++ b/cpp/src/hdbscan/detail/select.cuh
@@ -21,7 +21,7 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/op/sort.cuh>
diff --git a/cpp/src/hdbscan/detail/soft_clustering.cuh b/cpp/src/hdbscan/detail/soft_clustering.cuh
index 082a26255f..2139ba67ad 100644
--- a/cpp/src/hdbscan/detail/soft_clustering.cuh
+++ b/cpp/src/hdbscan/detail/soft_clustering.cuh
@@ -22,8 +22,8 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/op/sort.cuh>
diff --git a/cpp/src/hdbscan/detail/stabilities.cuh b/cpp/src/hdbscan/detail/stabilities.cuh
index b766e61fb3..7be40bfbcb 100644
--- a/cpp/src/hdbscan/detail/stabilities.cuh
+++ b/cpp/src/hdbscan/detail/stabilities.cuh
@@ -21,7 +21,7 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/op/sort.cuh>
diff --git a/cpp/src/hdbscan/detail/utils.h b/cpp/src/hdbscan/detail/utils.h
index dcf03ef25c..d97c71e468 100644
--- a/cpp/src/hdbscan/detail/utils.h
+++ b/cpp/src/hdbscan/detail/utils.h
@@ -18,7 +18,7 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/op/sort.cuh>
diff --git a/cpp/src/hdbscan/hdbscan.cu b/cpp/src/hdbscan/hdbscan.cu
index f9f8b6c971..9e939723f3 100644
--- a/cpp/src/hdbscan/hdbscan.cu
+++ b/cpp/src/hdbscan/hdbscan.cu
@@ -17,10 +17,10 @@
 #include "detail/condense.cuh"
 #include "detail/predict.cuh"
 #include <cuml/cluster/hdbscan.hpp>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include "runner.h"
 
diff --git a/cpp/src/hdbscan/prediction_data.cu b/cpp/src/hdbscan/prediction_data.cu
index 4113329268..fd1b94c79d 100644
--- a/cpp/src/hdbscan/prediction_data.cu
+++ b/cpp/src/hdbscan/prediction_data.cu
@@ -16,8 +16,8 @@
 
 #include "detail/utils.h"
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cuml/cluster/hdbscan.hpp>
 
diff --git a/cpp/src/hdbscan/runner.h b/cpp/src/hdbscan/runner.h
index e212bceeb4..5e9dc69e06 100644
--- a/cpp/src/hdbscan/runner.h
+++ b/cpp/src/hdbscan/runner.h
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/core/handle.hpp>
 #include <raft/core/kvp.hpp>
diff --git a/cpp/src/holtwinters/internal/hw_decompose.cuh b/cpp/src/holtwinters/internal/hw_decompose.cuh
index fd6f3934f9..2b289c0339 100644
--- a/cpp/src/holtwinters/internal/hw_decompose.cuh
+++ b/cpp/src/holtwinters/internal/hw_decompose.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
 // #TODO: Replace with public header when ready
diff --git a/cpp/src/holtwinters/internal/hw_eval.cuh b/cpp/src/holtwinters/internal/hw_eval.cuh
index acd7d0802b..0115be99fc 100644
--- a/cpp/src/holtwinters/internal/hw_eval.cuh
+++ b/cpp/src/holtwinters/internal/hw_eval.cuh
@@ -18,8 +18,8 @@
 
 #include "hw_utils.cuh"
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 template <typename Dtype>
 __device__ Dtype holtwinters_eval_device(int tid,
diff --git a/cpp/src/holtwinters/internal/hw_optim.cuh b/cpp/src/holtwinters/internal/hw_optim.cuh
index e267e4a938..164fdd7c50 100644
--- a/cpp/src/holtwinters/internal/hw_optim.cuh
+++ b/cpp/src/holtwinters/internal/hw_optim.cuh
@@ -19,7 +19,7 @@
 #include "hw_eval.cuh"
 #include "hw_utils.cuh"
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 template <typename Dtype>
 __device__ Dtype golden_step(Dtype a, Dtype b, Dtype c)
diff --git a/cpp/src/holtwinters/runner.cuh b/cpp/src/holtwinters/runner.cuh
index f8f093ebdc..83a7557be7 100644
--- a/cpp/src/holtwinters/runner.cuh
+++ b/cpp/src/holtwinters/runner.cuh
@@ -21,7 +21,7 @@
 #include "internal/hw_forecast.cuh"
 #include "internal/hw_optim.cuh"
 #include <cuml/tsa/holtwinters_params.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
 #include <raft/linalg/transpose.cuh>
diff --git a/cpp/src/knn/knn.cu b/cpp/src/knn/knn.cu
index 6e790a0827..b6d6e181fd 100644
--- a/cpp/src/knn/knn.cu
+++ b/cpp/src/knn/knn.cu
@@ -16,10 +16,10 @@
 
 #include <cuda_runtime.h>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/label/classlabels.cuh>
 #include <raft/spatial/knn/ann.cuh>
 #include <raft/spatial/knn/ball_cover.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <raft/spatial/knn/knn.cuh>
 #include <raft/spatial/knn/specializations.cuh>
diff --git a/cpp/src/knn/knn_sparse.cu b/cpp/src/knn/knn_sparse.cu
index ba308c6bb4..0f7b95c681 100644
--- a/cpp/src/knn/knn_sparse.cu
+++ b/cpp/src/knn/knn_sparse.cu
@@ -18,7 +18,7 @@
 #include <raft/core/handle.hpp>
 
 #include <raft/sparse/selection/knn.cuh>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 
 namespace ML {
 namespace Sparse {
diff --git a/cpp/src/metrics/accuracy_score.cu b/cpp/src/metrics/accuracy_score.cu
index 2fb58714cc..f99feacd3f 100644
--- a/cpp/src/metrics/accuracy_score.cu
+++ b/cpp/src/metrics/accuracy_score.cu
@@ -17,7 +17,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/accuracy.hpp>
+#include <raft/stats/accuracy.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/adjusted_rand_index.cu b/cpp/src/metrics/adjusted_rand_index.cu
index 542fc7b889..a061d8a883 100644
--- a/cpp/src/metrics/adjusted_rand_index.cu
+++ b/cpp/src/metrics/adjusted_rand_index.cu
@@ -17,7 +17,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/adjusted_rand_index.hpp>
+#include <raft/stats/adjusted_rand_index.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/entropy.cu b/cpp/src/metrics/entropy.cu
index 739b7c1e6c..3153ee7119 100644
--- a/cpp/src/metrics/entropy.cu
+++ b/cpp/src/metrics/entropy.cu
@@ -17,7 +17,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/entropy.hpp>
+#include <raft/stats/entropy.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/homogeneity_score.cu b/cpp/src/metrics/homogeneity_score.cu
index c41ef061f9..050f0ffff3 100644
--- a/cpp/src/metrics/homogeneity_score.cu
+++ b/cpp/src/metrics/homogeneity_score.cu
@@ -17,7 +17,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/homogeneity_score.hpp>
+#include <raft/stats/homogeneity_score.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/kl_divergence.cu b/cpp/src/metrics/kl_divergence.cu
index 9a86af7331..8aa6e7fd7e 100644
--- a/cpp/src/metrics/kl_divergence.cu
+++ b/cpp/src/metrics/kl_divergence.cu
@@ -17,7 +17,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/kl_divergence.hpp>
+#include <raft/stats/kl_divergence.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/mutual_info_score.cu b/cpp/src/metrics/mutual_info_score.cu
index 5468c41cc0..a359a14b19 100644
--- a/cpp/src/metrics/mutual_info_score.cu
+++ b/cpp/src/metrics/mutual_info_score.cu
@@ -19,7 +19,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/mutual_info_score.hpp>
+#include <raft/stats/mutual_info_score.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/r2_score.cu b/cpp/src/metrics/r2_score.cu
index e0553ce47f..6dff9e5986 100644
--- a/cpp/src/metrics/r2_score.cu
+++ b/cpp/src/metrics/r2_score.cu
@@ -16,7 +16,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/r2_score.hpp>
+#include <raft/stats/r2_score.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/rand_index.cu b/cpp/src/metrics/rand_index.cu
index f337904f00..1c15c02ebc 100644
--- a/cpp/src/metrics/rand_index.cu
+++ b/cpp/src/metrics/rand_index.cu
@@ -19,7 +19,7 @@
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/stats/rand_index.hpp>
+#include <raft/stats/rand_index.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/trustworthiness.cu b/cpp/src/metrics/trustworthiness.cu
index 20169d17f4..f1c0ed8ce0 100644
--- a/cpp/src/metrics/trustworthiness.cu
+++ b/cpp/src/metrics/trustworthiness.cu
@@ -14,23 +14,21 @@
  * limitations under the License.
  */
 
-#include <raft/stats/trustworthiness_score.hpp>
+#include <raft/stats/trustworthiness_score.cuh>
 
 #include <cuml/metrics/metrics.hpp>
 #include <raft/core/handle.hpp>
 
 #if defined RAFT_DISTANCE_COMPILED
-#include <raft/distance/specializations.hpp>
+#include <raft/distance/specializations.cuh>
 #endif
 
 #if defined RAFT_NN_COMPILED
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 #endif
 
 #include <raft/distance/distance.cuh>
 
-#include <raft/core/handle.hpp>
-
 namespace ML {
 namespace Metrics {
 
diff --git a/cpp/src/ml_cuda_utils.h b/cpp/src/ml_cuda_utils.h
index 5cd0dde268..b8cc4330f7 100644
--- a/cpp/src/ml_cuda_utils.h
+++ b/cpp/src/ml_cuda_utils.h
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <cuda_runtime.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace ML {
 
diff --git a/cpp/src/ml_mg_utils.cuh b/cpp/src/ml_mg_utils.cuh
index 2e696fac0f..c43627be41 100644
--- a/cpp/src/ml_mg_utils.cuh
+++ b/cpp/src/ml_mg_utils.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <cuda_runtime.h>
 
diff --git a/cpp/src/random_projection/rproj.cuh b/cpp/src/random_projection/rproj.cuh
index 81a6f77b3f..4e2fbd56ff 100644
--- a/cpp/src/random_projection/rproj.cuh
+++ b/cpp/src/random_projection/rproj.cuh
@@ -20,8 +20,8 @@
 
 #include <cuml/random_projection/rproj_c.h>
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 // TODO: This needs to be removed.
 #include <raft/sparse/detail/cusparse_wrappers.h>
diff --git a/cpp/src/randomforest/randomforest.cuh b/cpp/src/randomforest/randomforest.cuh
index fcddddb9b4..9b1d14bb00 100644
--- a/cpp/src/randomforest/randomforest.cuh
+++ b/cpp/src/randomforest/randomforest.cuh
@@ -20,7 +20,7 @@
 #include <decisiontree/decisiontree.cuh>
 #include <decisiontree/treelite_util.h>
 
-#include <raft/random/permute.hpp>
+#include <raft/random/permute.cuh>
 
 #include <raft/core/handle.hpp>
 #include <raft/core/nvtx.hpp>
diff --git a/cpp/src/svm/ws_util.cu b/cpp/src/svm/ws_util.cu
index cd5740b1e3..326288747e 100644
--- a/cpp/src/svm/ws_util.cu
+++ b/cpp/src/svm/ws_util.cu
@@ -16,7 +16,7 @@
 
 #include <cub/cub.cuh>
 #include <limits.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace SVM {
diff --git a/cpp/src/tsa/auto_arima.cuh b/cpp/src/tsa/auto_arima.cuh
index a5d9bbcda8..7bfb56c88b 100644
--- a/cpp/src/tsa/auto_arima.cuh
+++ b/cpp/src/tsa/auto_arima.cuh
@@ -28,7 +28,7 @@
 #include <thrust/transform.h>
 
 #include <common/fast_int_div.cuh>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/tsne/barnes_hut_kernels.cuh b/cpp/src/tsne/barnes_hut_kernels.cuh
index f3831e3297..3742a36892 100644
--- a/cpp/src/tsne/barnes_hut_kernels.cuh
+++ b/cpp/src/tsne/barnes_hut_kernels.cuh
@@ -34,8 +34,8 @@
 #define FACTOR7 1
 
 #include <float.h>
-#include <raft/core/cudart_utils.hpp>
-#include <raft/device_atomics.cuh>
+#include <raft/util/cudart_utils.hpp>
+#include <raft/util/device_atomics.cuh>
 
 namespace ML {
 namespace TSNE {
diff --git a/cpp/src/tsne/distances.cuh b/cpp/src/tsne/distances.cuh
index d7928a0748..c0535dedcd 100644
--- a/cpp/src/tsne/distances.cuh
+++ b/cpp/src/tsne/distances.cuh
@@ -17,13 +17,13 @@
 #pragma once
 
 #include <cuml/neighbors/knn_sparse.hpp>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
 #include <raft/distance/distance_types.hpp>
 #include <raft/linalg/eltwise.cuh>
 #include <raft/sparse/coo.hpp>
 #include <raft/sparse/linalg/symmetrize.cuh>
 #include <raft/sparse/selection/knn.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <selection/knn.cuh>
 
 #include <cuml/manifold/common.hpp>
diff --git a/cpp/src/tsne/exact_tsne.cuh b/cpp/src/tsne/exact_tsne.cuh
index f219b2e242..f67758e951 100644
--- a/cpp/src/tsne/exact_tsne.cuh
+++ b/cpp/src/tsne/exact_tsne.cuh
@@ -18,7 +18,7 @@
 #include "exact_kernels.cuh"
 #include "utils.cuh"
 #include <cuml/common/logger.hpp>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.h>
diff --git a/cpp/src/tsne/tsne.cu b/cpp/src/tsne/tsne.cu
index 66ea5c94bc..83ea8a08db 100644
--- a/cpp/src/tsne/tsne.cu
+++ b/cpp/src/tsne/tsne.cu
@@ -19,7 +19,7 @@
 #include <raft/core/handle.hpp>
 
 #if defined RAFT_DISTANCE_COMPILED
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 #endif
 
 #include <raft/distance/distance_types.hpp>
diff --git a/cpp/src/tsne/tsne_runner.cuh b/cpp/src/tsne/tsne_runner.cuh
index 09cc88f6f5..82c02bf52d 100644
--- a/cpp/src/tsne/tsne_runner.cuh
+++ b/cpp/src/tsne/tsne_runner.cuh
@@ -20,9 +20,9 @@
 #include "utils.cuh"
 #include <cuml/common/logger.hpp>
 #include <cuml/manifold/common.hpp>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
 #include <raft/distance/distance_types.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/transform.h>
diff --git a/cpp/src/tsne/utils.cuh b/cpp/src/tsne/utils.cuh
index 43368bb5fd..4126871512 100644
--- a/cpp/src/tsne/utils.cuh
+++ b/cpp/src/tsne/utils.cuh
@@ -39,7 +39,7 @@
 #include <iostream>
 #include <unistd.h>
 
-#include <raft/device_atomics.cuh>
+#include <raft/util/device_atomics.cuh>
 
 /**
  * @brief Performs P + P.T.
diff --git a/cpp/src/tsvd/tsvd.cuh b/cpp/src/tsvd/tsvd.cuh
index 7d6929bd62..1c04acb9c1 100644
--- a/cpp/src/tsvd/tsvd.cuh
+++ b/cpp/src/tsvd/tsvd.cuh
@@ -17,7 +17,6 @@
 #pragma once
 
 #include <cuml/decomposition/params.hpp>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/eig.cuh>
@@ -30,6 +29,7 @@
 #include <raft/stats/mean.cuh>
 #include <raft/stats/stddev.cuh>
 #include <raft/stats/sum.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/umap/fuzzy_simpl_set/naive.cuh b/cpp/src/umap/fuzzy_simpl_set/naive.cuh
index f4c24e5567..bd14876e37 100644
--- a/cpp/src/umap/fuzzy_simpl_set/naive.cuh
+++ b/cpp/src/umap/fuzzy_simpl_set/naive.cuh
@@ -20,8 +20,8 @@
 #include <cuml/manifold/umapparams.h>
 #include <cuml/neighbors/knn.hpp>
 
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <raft/sparse/coo.hpp>
 #include <raft/sparse/linalg/symmetrize.cuh>
diff --git a/cpp/src/umap/knn_graph/algo.cuh b/cpp/src/umap/knn_graph/algo.cuh
index 2c63985c3a..b13ffc82a2 100644
--- a/cpp/src/umap/knn_graph/algo.cuh
+++ b/cpp/src/umap/knn_graph/algo.cuh
@@ -25,7 +25,7 @@
 #include <raft/sparse/selection/knn.cuh>
 
 #if defined RAFT_DISTANCE_COMPILED
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 #endif
 
 #include <raft/spatial/knn/knn.cuh>
diff --git a/cpp/src/umap/runner.cuh b/cpp/src/umap/runner.cuh
index aba437673a..eeaed18cbe 100644
--- a/cpp/src/umap/runner.cuh
+++ b/cpp/src/umap/runner.cuh
@@ -44,7 +44,7 @@
 #include <raft/sparse/op/filter.cuh>
 #include <raft/sparse/op/sort.cuh>
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <common/nvtx.hpp>
 #include <cuda_runtime.h>
diff --git a/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh b/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh
index 3cdb056fd1..d19cc4c54c 100644
--- a/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh
+++ b/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh
@@ -20,8 +20,8 @@
 
 #include <cuml/manifold/umapparams.h>
 
-#include <raft/cuda_utils.cuh>
 #include <raft/random/rng.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <raft/util/cudart_utils.hpp>
 
 #include <cstddef>
diff --git a/cpp/src/umap/supervised.cuh b/cpp/src/umap/supervised.cuh
index 071cbc7c10..f9590bcb47 100644
--- a/cpp/src/umap/supervised.cuh
+++ b/cpp/src/umap/supervised.cuh
@@ -26,7 +26,7 @@
 #include "init_embed/runner.cuh"
 #include "knn_graph/runner.cuh"
 #include "simpl_set_embed/runner.cuh"
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <thrust/count.h>
 #include <thrust/device_ptr.h>
@@ -43,7 +43,7 @@
 #include <raft/sparse/linalg/symmetrize.cuh>
 #include <raft/sparse/op/filter.cuh>
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <cuda_runtime.h>
 
diff --git a/cpp/src/umap/umap.cu b/cpp/src/umap/umap.cu
index 4de1169948..9a653772c3 100644
--- a/cpp/src/umap/umap.cu
+++ b/cpp/src/umap/umap.cu
@@ -19,9 +19,8 @@
 #include <cuml/manifold/umap.hpp>
 #include <cuml/manifold/umapparams.h>
 #include <raft/core/handle.hpp>
-#include <raft/spatial/knn/specializations.hpp>
-
-#include <raft/cuda_utils.cuh>
+#include <raft/spatial/knn/specializations.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 #include <iostream>
 
diff --git a/cpp/src_prims/common/device_loads_stores.cuh b/cpp/src_prims/common/device_loads_stores.cuh
index c802293710..c3f2ba0604 100644
--- a/cpp/src_prims/common/device_loads_stores.cuh
+++ b/cpp/src_prims/common/device_loads_stores.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 
diff --git a/cpp/src_prims/common/device_utils.cuh b/cpp/src_prims/common/device_utils.cuh
index e90abf4353..797fc6e011 100644
--- a/cpp/src_prims/common/device_utils.cuh
+++ b/cpp/src_prims/common/device_utils.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <utility>  // pair
 
 namespace MLCommon {
diff --git a/cpp/src_prims/common/fast_int_div.cuh b/cpp/src_prims/common/fast_int_div.cuh
index 394aae2c74..d20b2126d2 100644
--- a/cpp/src_prims/common/fast_int_div.cuh
+++ b/cpp/src_prims/common/fast_int_div.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <stdint.h>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/common/grid_sync.cuh b/cpp/src_prims/common/grid_sync.cuh
index 4575928293..323c89e61d 100644
--- a/cpp/src_prims/common/grid_sync.cuh
+++ b/cpp/src_prims/common/grid_sync.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 
diff --git a/cpp/src_prims/common/iota.cuh b/cpp/src_prims/common/iota.cuh
index 5139fda739..99608da878 100644
--- a/cpp/src_prims/common/iota.cuh
+++ b/cpp/src_prims/common/iota.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 
diff --git a/cpp/src_prims/decoupled_lookback.cuh b/cpp/src_prims/decoupled_lookback.cuh
index 6861adb4ec..6203c67fc2 100644
--- a/cpp/src_prims/decoupled_lookback.cuh
+++ b/cpp/src_prims/decoupled_lookback.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 
diff --git a/cpp/src_prims/functions/hinge.cuh b/cpp/src_prims/functions/hinge.cuh
index 29cb3c194f..9961a03a1f 100644
--- a/cpp/src_prims/functions/hinge.cuh
+++ b/cpp/src_prims/functions/hinge.cuh
@@ -18,7 +18,6 @@
 
 #include "penalty.cuh"
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/eltwise.cuh>
 #include <raft/linalg/gemm.cuh>
@@ -30,6 +29,7 @@
 #include <raft/matrix/matrix.cuh>
 #include <raft/stats/mean.cuh>
 #include <raft/stats/sum.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/functions/linearReg.cuh b/cpp/src_prims/functions/linearReg.cuh
index cf86059c01..c7396826ce 100644
--- a/cpp/src_prims/functions/linearReg.cuh
+++ b/cpp/src_prims/functions/linearReg.cuh
@@ -18,7 +18,6 @@
 
 #include "penalty.cuh"
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/eltwise.cuh>
 #include <raft/linalg/gemm.cuh>
@@ -28,6 +27,7 @@
 #include <raft/matrix/matrix.cuh>
 #include <raft/stats/mean.cuh>
 #include <raft/stats/sum.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/functions/logisticReg.cuh b/cpp/src_prims/functions/logisticReg.cuh
index 17dcca5d8c..fa7d84041b 100644
--- a/cpp/src_prims/functions/logisticReg.cuh
+++ b/cpp/src_prims/functions/logisticReg.cuh
@@ -19,7 +19,6 @@
 #include "penalty.cuh"
 #include "sigmoid.cuh"
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/eltwise.cuh>
 #include <raft/linalg/gemm.cuh>
@@ -29,6 +28,7 @@
 #include <raft/matrix/matrix.cuh>
 #include <raft/stats/mean.cuh>
 #include <raft/stats/sum.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/functions/sigmoid.cuh b/cpp/src_prims/functions/sigmoid.cuh
index 9d0177660a..d4d2b75c6c 100644
--- a/cpp/src_prims/functions/sigmoid.cuh
+++ b/cpp/src_prims/functions/sigmoid.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
 #include <raft/linalg/unary_op.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/src_prims/linalg/batched/gemv.cuh b/cpp/src_prims/linalg/batched/gemv.cuh
index d3916a814c..57985e499a 100644
--- a/cpp/src_prims/linalg/batched/gemv.cuh
+++ b/cpp/src_prims/linalg/batched/gemv.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
-#include <raft/vectorized.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/vectorized.cuh>
 
 namespace MLCommon {
 namespace LinAlg {
diff --git a/cpp/src_prims/linalg/batched/make_symm.cuh b/cpp/src_prims/linalg/batched/make_symm.cuh
index e7b4d9c7a4..47c8f2ee59 100644
--- a/cpp/src_prims/linalg/batched/make_symm.cuh
+++ b/cpp/src_prims/linalg/batched/make_symm.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 namespace LinAlg {
diff --git a/cpp/src_prims/linalg/block.cuh b/cpp/src_prims/linalg/block.cuh
index d401b989e5..2585f423f6 100644
--- a/cpp/src_prims/linalg/block.cuh
+++ b/cpp/src_prims/linalg/block.cuh
@@ -17,8 +17,8 @@
 #include <cub/cub.cuh>
 
 #include <raft/common/device_loads_stores.cuh>
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 // Anonymous namespace for internal auxiliary functions
 namespace {
diff --git a/cpp/src_prims/linalg/init.h b/cpp/src_prims/linalg/init.h
index c3dadd2895..bdb7b08bae 100644
--- a/cpp/src_prims/linalg/init.h
+++ b/cpp/src_prims/linalg/init.h
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
diff --git a/cpp/src_prims/matrix/reverse.cuh b/cpp/src_prims/matrix/reverse.cuh
index dd1b0ec975..7cae75d780 100644
--- a/cpp/src_prims/matrix/reverse.cuh
+++ b/cpp/src_prims/matrix/reverse.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include <raft/cuda_utils.cuh>
-#include <raft/vectorized.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/vectorized.cuh>
 
 namespace MLCommon {
 namespace Matrix {
diff --git a/cpp/src_prims/selection/kselection.cuh b/cpp/src_prims/selection/kselection.cuh
index a10beddb65..8eb4a89477 100644
--- a/cpp/src_prims/selection/kselection.cuh
+++ b/cpp/src_prims/selection/kselection.cuh
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <limits>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 #include <stdlib.h>
 
 namespace MLCommon {
diff --git a/cpp/test/mg/knn.cu b/cpp/test/mg/knn.cu
index 30b947e7e9..dd6f903758 100644
--- a/cpp/test/mg/knn.cu
+++ b/cpp/test/mg/knn.cu
@@ -23,7 +23,7 @@
 
 #include <raft/comms/mpi_comms.hpp>
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace KNN {
diff --git a/cpp/test/mg/knn_test_helper.cuh b/cpp/test/mg/knn_test_helper.cuh
index 26fddf4d80..c9b18757a3 100644
--- a/cpp/test/mg/knn_test_helper.cuh
+++ b/cpp/test/mg/knn_test_helper.cuh
@@ -26,7 +26,7 @@
 #include <raft/linalg/reduce_rows_by_key.cuh>
 #include <selection/knn.cuh>
 
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace ML {
 namespace KNN {
diff --git a/cpp/test/mg/test_opg_utils.h b/cpp/test/mg/test_opg_utils.h
index 438162d024..7990532b47 100644
--- a/cpp/test/mg/test_opg_utils.h
+++ b/cpp/test/mg/test_opg_utils.h
@@ -18,7 +18,7 @@
 
 #include <gtest/gtest.h>
 #include <mpi.h>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace MLCommon {
 namespace Test {
diff --git a/cpp/test/prims/batched/csr.cu b/cpp/test/prims/batched/csr.cu
index 9ad7825d49..7fa547df49 100644
--- a/cpp/test/prims/batched/csr.cu
+++ b/cpp/test/prims/batched/csr.cu
@@ -20,7 +20,7 @@
 #include <linalg/batched/matrix.cuh>
 #include <sparse/batched/csr.cuh>
 
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 #include <gtest/gtest.h>
 
diff --git a/cpp/test/prims/decoupled_lookback.cu b/cpp/test/prims/decoupled_lookback.cu
index 79b3482394..bcc99bda5e 100644
--- a/cpp/test/prims/decoupled_lookback.cu
+++ b/cpp/test/prims/decoupled_lookback.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <decoupled_lookback.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/interruptible.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/device_utils.cu b/cpp/test/prims/device_utils.cu
index 181cdfcb60..e4e683827d 100644
--- a/cpp/test/prims/device_utils.cu
+++ b/cpp/test/prims/device_utils.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <common/device_utils.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/interruptible.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/fast_int_div.cu b/cpp/test/prims/fast_int_div.cu
index 607e632a6d..b3a2cba489 100644
--- a/cpp/test/prims/fast_int_div.cu
+++ b/cpp/test/prims/fast_int_div.cu
@@ -17,7 +17,7 @@
 #include "test_utils.h"
 #include <common/fast_int_div.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/fillna.cu b/cpp/test/prims/fillna.cu
index 2745eb63c7..f35d4d0a36 100644
--- a/cpp/test/prims/fillna.cu
+++ b/cpp/test/prims/fillna.cu
@@ -19,9 +19,9 @@
 #include <random>
 #include <vector>
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include "test_utils.h"
diff --git a/cpp/test/prims/grid_sync.cu b/cpp/test/prims/grid_sync.cu
index 8515c0301a..c460cd8aa3 100644
--- a/cpp/test/prims/grid_sync.cu
+++ b/cpp/test/prims/grid_sync.cu
@@ -17,7 +17,7 @@
 #include "test_utils.h"
 #include <common/grid_sync.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/jones_transform.cu b/cpp/test/prims/jones_transform.cu
index 2ae8b57f58..5a9194d6be 100644
--- a/cpp/test/prims/jones_transform.cu
+++ b/cpp/test/prims/jones_transform.cu
@@ -16,8 +16,8 @@
 #include <algorithm>
 #include <gtest/gtest.h>
 #include <iostream>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <random>
 #include <rmm/device_uvector.hpp>
 #include <timeSeries/jones_transform.cuh>
diff --git a/cpp/test/prims/log.cu b/cpp/test/prims/log.cu
index 140904a475..f7b4d04e27 100644
--- a/cpp/test/prims/log.cu
+++ b/cpp/test/prims/log.cu
@@ -17,7 +17,7 @@
 #include "test_utils.h"
 #include <functions/log.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/make_arima.cu b/cpp/test/prims/make_arima.cu
index 1bb566d799..118e880c23 100644
--- a/cpp/test/prims/make_arima.cu
+++ b/cpp/test/prims/make_arima.cu
@@ -19,9 +19,9 @@
 #include <thrust/device_vector.h>
 
 #include "test_utils.h"
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/interruptible.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <random/make_arima.cuh>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/sigmoid.cu b/cpp/test/prims/sigmoid.cu
index bc8126779d..195a4f5973 100644
--- a/cpp/test/prims/sigmoid.cu
+++ b/cpp/test/prims/sigmoid.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <functions/sigmoid.cuh>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/test_utils.h b/cpp/test/prims/test_utils.h
index 1e7e00b246..8a50f0a065 100644
--- a/cpp/test/prims/test_utils.h
+++ b/cpp/test/prims/test_utils.h
@@ -18,9 +18,9 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <memory>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/interruptible.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 namespace MLCommon {
 
diff --git a/cpp/test/sg/dbscan_test.cu b/cpp/test/sg/dbscan_test.cu
index c2ad7575a1..6ac408dbe4 100644
--- a/cpp/test/sg/dbscan_test.cu
+++ b/cpp/test/sg/dbscan_test.cu
@@ -15,19 +15,17 @@
  */
 
 #include <gtest/gtest.h>
-#include <raft/core/handle.hpp>
-#include <raft/util/cuda_utils.cuh>
-#include <raft/util/cudart_utils.hpp>
 #include <vector>
 
 #include <cuml/cluster/dbscan.hpp>
 #include <cuml/datasets/make_blobs.hpp>
 #include <cuml/metrics/metrics.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/distance/distance.cuh>
 #include <raft/distance/distance_types.hpp>
-
-#include <raft/core/handle.hpp>
 #include <raft/linalg/transpose.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 
 #include <test_utils.h>
 
diff --git a/cpp/test/sg/genetic/evolution_test.cu b/cpp/test/sg/genetic/evolution_test.cu
index 410f84bde5..41f1122c58 100644
--- a/cpp/test/sg/genetic/evolution_test.cu
+++ b/cpp/test/sg/genetic/evolution_test.cu
@@ -23,8 +23,8 @@
 #include <cuml/genetic/program.h>
 #include <gtest/gtest.h>
 #include <iostream>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/genetic/node_test.cpp b/cpp/test/sg/genetic/node_test.cpp
index 5ea426a65d..f68858ecbb 100644
--- a/cpp/test/sg/genetic/node_test.cpp
+++ b/cpp/test/sg/genetic/node_test.cpp
@@ -16,7 +16,7 @@
 
 #include <cuml/genetic/node.h>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
+#include <raft/util/cudart_utils.hpp>
 
 namespace cuml {
 namespace genetic {
diff --git a/cpp/test/sg/genetic/program_test.cu b/cpp/test/sg/genetic/program_test.cu
index d11a86f813..9296d7be0a 100644
--- a/cpp/test/sg/genetic/program_test.cu
+++ b/cpp/test/sg/genetic/program_test.cu
@@ -21,8 +21,8 @@
 #include <cuml/genetic/program.h>
 #include <gtest/gtest.h>
 #include <iostream>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/hdbscan_test.cu b/cpp/test/sg/hdbscan_test.cu
index 13eb3b229c..d7040c67b8 100644
--- a/cpp/test/sg/hdbscan_test.cu
+++ b/cpp/test/sg/hdbscan_test.cu
@@ -30,8 +30,8 @@
 #include <hdbscan/detail/soft_clustering.cuh>
 #include <hdbscan/detail/utils.h>
 
-#include <raft/spatial/knn/specializations.hpp>
-#include <raft/stats/adjusted_rand_index.hpp>
+#include <raft/spatial/knn/specializations.cuh>
+#include <raft/stats/adjusted_rand_index.cuh>
 
 #include <raft/cluster/detail/agglomerative.cuh>
 
diff --git a/cpp/test/sg/holtwinters_test.cu b/cpp/test/sg/holtwinters_test.cu
index 417e657023..3a1d71dc7e 100644
--- a/cpp/test/sg/holtwinters_test.cu
+++ b/cpp/test/sg/holtwinters_test.cu
@@ -21,10 +21,10 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/tsa/holtwinters.h>
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
 #include <raft/core/math.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 
diff --git a/cpp/test/sg/knn_test.cu b/cpp/test/sg/knn_test.cu
index 0a1a2bcbdb..967b985657 100644
--- a/cpp/test/sg/knn_test.cu
+++ b/cpp/test/sg/knn_test.cu
@@ -28,7 +28,7 @@
 #include <cuml/datasets/make_blobs.hpp>
 
 #include <cuml/neighbors/knn.hpp>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/test/sg/lars_test.cu b/cpp/test/sg/lars_test.cu
index 3a559ba5fc..aec1b8d215 100644
--- a/cpp/test/sg/lars_test.cu
+++ b/cpp/test/sg/lars_test.cu
@@ -16,8 +16,8 @@
 
 #include <gtest/gtest.h>
 #include <iomanip>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
+#include <raft/util/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cusolver_wrappers.hpp>
 #include <raft/random/rng.cuh>
diff --git a/cpp/test/sg/linkage_test.cu b/cpp/test/sg/linkage_test.cu
index c58048d4b7..ccb8161020 100644
--- a/cpp/test/sg/linkage_test.cu
+++ b/cpp/test/sg/linkage_test.cu
@@ -15,9 +15,9 @@
  */
 
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <vector>
 
 #include <cuml/cluster/linkage.hpp>
diff --git a/cpp/test/sg/shap_kernel.cu b/cpp/test/sg/shap_kernel.cu
index 40d328d7a6..f71960731b 100644
--- a/cpp/test/sg/shap_kernel.cu
+++ b/cpp/test/sg/shap_kernel.cu
@@ -18,9 +18,9 @@
 
 #include <test_utils.h>
 
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/count.h>
diff --git a/cpp/test/sg/trustworthiness_test.cu b/cpp/test/sg/trustworthiness_test.cu
index b015772597..72282a4be3 100644
--- a/cpp/test/sg/trustworthiness_test.cu
+++ b/cpp/test/sg/trustworthiness_test.cu
@@ -17,9 +17,9 @@
 #include <cuml/metrics/metrics.hpp>
 
 #include <gtest/gtest.h>
-#include <raft/core/cudart_utils.hpp>
 #include <raft/core/handle.hpp>
-#include <raft/cuda_utils.cuh>
+#include <raft/util/cuda_utils.cuh>
+#include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 #include <vector>
 
diff --git a/cpp/test/sg/tsne_test.cu b/cpp/test/sg/tsne_test.cu
index 97d530692b..14964a5955 100644
--- a/cpp/test/sg/tsne_test.cu
+++ b/cpp/test/sg/tsne_test.cu
@@ -29,7 +29,7 @@
 #include <raft/core/handle.hpp>
 
 #if defined RAFT_DISTANCE_COMPILED
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 #endif
 
 #include <raft/util/cudart_utils.hpp>
diff --git a/cpp/test/sg/umap_parametrizable_test.cu b/cpp/test/sg/umap_parametrizable_test.cu
index 9313738255..bd73273141 100644
--- a/cpp/test/sg/umap_parametrizable_test.cu
+++ b/cpp/test/sg/umap_parametrizable_test.cu
@@ -27,7 +27,7 @@
 #include <datasets/digits.h>
 
 #if defined RAFT_NN_COMPILED
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/specializations.cuh>
 #endif
 
 #include <test_utils.h>