From 6f42ced4d6f84cc04ea27f8940449d57fa08ab51 Mon Sep 17 00:00:00 2001
From: Micka <9810050+lowener@users.noreply.github.com>
Date: Fri, 9 Sep 2022 00:14:24 +0200
Subject: [PATCH] Remove Raft deprecated headers (#4858)

This should be merged before the Raft PR: https://github.com/rapidsai/raft/pull/753.

Authors:
  - Micka (https://github.com/lowener)

Approvers:
  - Corey J. Nolet (https://github.com/cjnolet)
  - William Hicks (https://github.com/wphicks)

URL: https://github.com/rapidsai/cuml/pull/4858
---
 cpp/bench/common/ml_benchmark.hpp             |  2 +-
 cpp/bench/prims/add.cu                        |  2 +-
 cpp/bench/prims/distance_common.cuh           |  6 ++---
 cpp/bench/prims/fused_l2_nn.cu                | 12 +++++-----
 cpp/bench/prims/gram_matrix.cu                |  2 +-
 cpp/bench/prims/make_blobs.cu                 |  2 +-
 cpp/bench/prims/map_then_reduce.cu            |  2 +-
 cpp/bench/prims/matrix_vector_op.cu           |  2 +-
 cpp/bench/prims/permute.cu                    |  6 ++---
 cpp/bench/prims/reduce.cu                     |  2 +-
 cpp/bench/prims/rng.cu                        |  4 ++--
 cpp/bench/sg/arima_loglikelihood.cu           |  6 ++---
 cpp/bench/sg/benchmark.cuh                    |  4 ++--
 cpp/bench/sg/dataset.cuh                      | 10 ++++----
 cpp/bench/sg/dataset_ts.cuh                   |  6 ++---
 cpp/examples/dbscan/dbscan_example.cpp        |  2 +-
 cpp/examples/kmeans/kmeans_example.cpp        |  4 ++--
 cpp/examples/symreg/symreg_example.cpp        |  2 +-
 cpp/include/cuml/cluster/hdbscan.hpp          |  2 +-
 cpp/include/cuml/cluster/linkage.hpp          |  2 +-
 cpp/include/cuml/common/utils.hpp             |  4 ++--
 .../cuml/decomposition/sign_flip_mg.hpp       |  2 +-
 cpp/include/cuml/genetic/genetic.h            |  2 +-
 cpp/include/cuml/genetic/program.h            |  2 +-
 cpp/include/cuml/linear_model/glm.hpp         |  2 +-
 .../cuml/linear_model/preprocess_mg.hpp       |  4 ++--
 cpp/include/cuml/neighbors/knn_mg.hpp         |  2 +-
 cpp/include/cuml/random_projection/rproj_c.h  |  4 ++--
 cpp/include/cuml/solvers/lars.hpp             |  2 +-
 cpp/include/cuml/svm/linear.hpp               |  4 ++--
 cpp/include/cuml/svm/svc.hpp                  |  2 +-
 cpp/include/cuml/tsa/arima_common.h           |  4 ++--
 cpp/src/arima/batched_arima.cu                |  8 +++----
 cpp/src/arima/batched_kalman.cu               |  8 +++----
 cpp/src/common/cumlHandle.cpp                 |  2 +-
 cpp/src/common/cumlHandle.hpp                 |  2 +-
 cpp/src/common/cuml_api.cpp                   |  2 +-
 cpp/src/common/nvtx.hpp                       | 12 +++++-----
 cpp/src/common/tensor.hpp                     |  4 ++--
 cpp/src/datasets/make_blobs.cu                |  2 +-
 cpp/src/datasets/make_regression.cu           |  2 +-
 cpp/src/dbscan/adjgraph/algo.cuh              |  2 +-
 cpp/src/dbscan/corepoints/compute.cuh         |  2 +-
 cpp/src/dbscan/corepoints/exchange.cuh        |  4 ++--
 cpp/src/dbscan/dbscan.cu                      |  2 +-
 cpp/src/dbscan/dbscan.cuh                     |  2 +-
 cpp/src/dbscan/mergelabels/runner.cuh         |  4 ++--
 cpp/src/dbscan/mergelabels/tree_reduction.cuh |  4 ++--
 cpp/src/dbscan/runner.cuh                     |  6 ++---
 cpp/src/dbscan/vertexdeg/algo.cuh             |  6 ++---
 cpp/src/dbscan/vertexdeg/precomputed.cuh      |  6 ++---
 .../batched-levelalgo/builder.cuh             |  4 ++--
 .../kernels/builder_kernels.cuh               |  2 +-
 .../batched-levelalgo/quantiles.cuh           |  4 ++--
 .../decisiontree/batched-levelalgo/split.cuh  |  2 +-
 cpp/src/decisiontree/decisiontree.cu          |  2 +-
 cpp/src/decisiontree/decisiontree.cuh         |  6 ++---
 cpp/src/explainer/kernel_shap.cu              |  4 ++--
 cpp/src/explainer/permutation_shap.cu         |  4 ++--
 cpp/src/explainer/tree_shap.cu                |  4 ++--
 cpp/src/fil/fil.cu                            | 10 ++++----
 cpp/src/fil/infer.cu                          |  2 +-
 cpp/src/fil/internal.cuh                      |  2 +-
 cpp/src/fil/treelite_import.cu                |  6 ++---
 cpp/src/genetic/fitness.cuh                   | 20 ++++++++--------
 cpp/src/genetic/genetic.cu                    |  8 +++----
 cpp/src/genetic/program.cu                    |  4 ++--
 cpp/src/glm/ols_mg.cu                         | 10 ++++----
 cpp/src/glm/preprocess.cuh                    | 18 +++++++-------
 cpp/src/glm/preprocess_mg.cu                  | 10 ++++----
 cpp/src/glm/qn/glm_base.cuh                   | 12 +++++-----
 cpp/src/glm/qn/glm_linear.cuh                 |  2 +-
 cpp/src/glm/qn/glm_logistic.cuh               |  2 +-
 cpp/src/glm/qn/glm_regularizer.cuh            |  8 +++----
 cpp/src/glm/qn/glm_softmax.cuh                |  2 +-
 cpp/src/glm/qn/glm_svm.cuh                    |  2 +-
 cpp/src/glm/qn/qn.cuh                         |  2 +-
 cpp/src/glm/qn/simple_mat/base.hpp            |  4 ++--
 cpp/src/glm/qn/simple_mat/dense.hpp           | 12 +++++-----
 cpp/src/glm/qn/simple_mat/sparse.hpp          | 12 +++++-----
 cpp/src/glm/ridge.cuh                         | 24 +++++++++----------
 cpp/src/glm/ridge_mg.cu                       | 12 +++++-----
 cpp/src/hdbscan/condensed_hierarchy.cu        |  8 +++----
 cpp/src/hdbscan/detail/condense.cuh           |  6 ++---
 cpp/src/hdbscan/detail/extract.cuh            |  8 +++----
 cpp/src/hdbscan/detail/membership.cuh         |  8 +++----
 cpp/src/hdbscan/detail/predict.cuh            |  2 +-
 cpp/src/hdbscan/detail/reachability.cuh       | 12 +++++-----
 cpp/src/hdbscan/detail/reachability_faiss.cuh |  2 +-
 cpp/src/hdbscan/detail/select.cuh             |  8 +++----
 cpp/src/hdbscan/detail/soft_clustering.cuh    | 16 ++++++-------
 cpp/src/hdbscan/detail/stabilities.cuh        |  8 +++----
 cpp/src/hdbscan/detail/utils.h                | 12 +++++-----
 cpp/src/hdbscan/hdbscan.cu                    |  5 +---
 cpp/src/hdbscan/prediction_data.cu            |  8 +++----
 cpp/src/hdbscan/runner.h                      |  4 ++--
 cpp/src/hierarchy/linkage.cu                  |  2 +-
 cpp/src/hierarchy/pw_dist_graph.cuh           |  2 +-
 cpp/src/holtwinters/internal/hw_decompose.cuh |  4 ++--
 cpp/src/holtwinters/internal/hw_eval.cuh      |  4 ++--
 cpp/src/holtwinters/internal/hw_optim.cuh     |  2 +-
 cpp/src/holtwinters/internal/hw_utils.cuh     |  4 ++--
 cpp/src/holtwinters/runner.cuh                |  4 ++--
 cpp/src/kmeans/common.cuh                     | 20 ++++++++--------
 cpp/src/kmeans/kmeans_mg_impl.cuh             |  2 +-
 cpp/src/kmeans/sg_impl.cuh                    |  2 +-
 cpp/src/knn/knn.cu                            | 10 ++++----
 cpp/src/knn/knn_opg_common.cuh                |  6 ++---
 cpp/src/knn/knn_sparse.cu                     |  4 ++--
 cpp/src/metrics/mutual_info_score.cu          |  4 ++--
 cpp/src/metrics/pairwise_distance.cu          |  8 +++----
 cpp/src/metrics/pairwise_distance_canberra.cu |  4 ++--
 .../metrics/pairwise_distance_canberra.cuh    |  8 +++----
 .../metrics/pairwise_distance_chebyshev.cu    |  4 ++--
 .../metrics/pairwise_distance_chebyshev.cuh   |  8 +++----
 .../metrics/pairwise_distance_correlation.cu  |  4 ++--
 .../metrics/pairwise_distance_correlation.cuh |  8 +++----
 cpp/src/metrics/pairwise_distance_cosine.cu   |  4 ++--
 cpp/src/metrics/pairwise_distance_cosine.cuh  |  8 +++----
 .../metrics/pairwise_distance_euclidean.cu    |  4 ++--
 .../metrics/pairwise_distance_euclidean.cuh   |  8 +++----
 cpp/src/metrics/pairwise_distance_hamming.cu  |  4 ++--
 cpp/src/metrics/pairwise_distance_hamming.cuh |  8 +++----
 .../metrics/pairwise_distance_hellinger.cu    |  4 ++--
 .../metrics/pairwise_distance_hellinger.cuh   |  8 +++----
 .../pairwise_distance_jensen_shannon.cu       |  4 ++--
 .../pairwise_distance_jensen_shannon.cuh      |  8 +++----
 .../pairwise_distance_kl_divergence.cu        |  4 ++--
 .../pairwise_distance_kl_divergence.cuh       |  8 +++----
 cpp/src/metrics/pairwise_distance_l1.cu       |  4 ++--
 cpp/src/metrics/pairwise_distance_l1.cuh      |  8 +++----
 .../metrics/pairwise_distance_minkowski.cu    |  4 ++--
 .../metrics/pairwise_distance_minkowski.cuh   |  8 +++----
 .../metrics/pairwise_distance_russell_rao.cu  |  4 ++--
 .../metrics/pairwise_distance_russell_rao.cuh |  8 +++----
 cpp/src/metrics/rand_index.cu                 |  4 ++--
 cpp/src/metrics/trustworthiness.cu            |  8 +++----
 cpp/src/ml_cuda_utils.h                       |  4 ++--
 cpp/src/ml_mg_utils.cuh                       |  2 +-
 cpp/src/pca/pca.cuh                           | 18 +++++++-------
 cpp/src/pca/pca_mg.cu                         | 10 ++++----
 cpp/src/pca/sign_flip_mg.cu                   |  8 +++----
 cpp/src/random_projection/rproj.cuh           |  2 +-
 cpp/src/random_projection/rproj_utils.cuh     |  4 ++--
 cpp/src/randomforest/randomforest.cu          |  2 +-
 cpp/src/randomforest/randomforest.cuh         |  6 ++---
 cpp/src/solver/cd.cuh                         |  5 ++--
 cpp/src/solver/cd_mg.cu                       | 18 +++++++-------
 cpp/src/solver/lars_impl.cuh                  | 12 +++++-----
 cpp/src/solver/sgd.cuh                        | 22 ++++++++---------
 cpp/src/spectral/spectral.cu                  |  4 ++--
 cpp/src/svm/kernelcache.cuh                   |  6 ++---
 cpp/src/svm/linear.cu                         | 16 ++++++-------
 cpp/src/svm/results.cuh                       | 10 ++++----
 cpp/src/svm/smosolver.cuh                     | 10 ++++----
 cpp/src/svm/svc.cu                            |  2 +-
 cpp/src/svm/svc_impl.cuh                      |  4 ++--
 cpp/src/svm/svr.cu                            |  2 +-
 cpp/src/svm/svr_impl.cuh                      |  4 ++--
 cpp/src/svm/workingset.cuh                    |  6 ++---
 cpp/src/tsa/auto_arima.cu                     |  4 ++--
 cpp/src/tsa/auto_arima.cuh                    |  2 +-
 cpp/src/tsne/barnes_hut_kernels.cuh           |  4 ++--
 cpp/src/tsne/barnes_hut_tsne.cuh              |  4 ++--
 cpp/src/tsne/distances.cuh                    | 10 ++++----
 cpp/src/tsne/exact_kernels.cuh                |  4 ++--
 cpp/src/tsne/exact_tsne.cuh                   |  2 +-
 cpp/src/tsne/fft_tsne.cuh                     |  4 ++--
 cpp/src/tsne/tsne_runner.cuh                  |  2 +-
 cpp/src/tsne/utils.cuh                        |  8 +++----
 cpp/src/tsvd/tsvd.cuh                         | 24 +++++++++----------
 cpp/src/tsvd/tsvd_mg.cu                       | 10 ++++----
 cpp/src/umap/fuzzy_simpl_set/naive.cuh        |  8 +++----
 cpp/src/umap/init_embed/random_algo.cuh       |  4 ++--
 cpp/src/umap/init_embed/spectral_algo.cuh     |  6 ++---
 cpp/src/umap/knn_graph/algo.cuh               |  8 +++----
 cpp/src/umap/optimize.cuh                     | 14 +++++------
 cpp/src/umap/runner.cuh                       | 10 ++++----
 cpp/src/umap/simpl_set_embed/algo.cuh         |  6 ++---
 .../simpl_set_embed/optimize_batch_kernel.cuh |  4 ++--
 cpp/src/umap/supervised.cuh                   | 12 +++++-----
 cpp/src_prims/cache/cache.cuh                 |  4 ++--
 cpp/src_prims/cufft_utils.h                   |  4 ++--
 cpp/src_prims/functions/hinge.cuh             | 22 ++++++++---------
 cpp/src_prims/functions/linearReg.cuh         | 18 +++++++-------
 cpp/src_prims/functions/log.cuh               |  2 +-
 cpp/src_prims/functions/logisticReg.cuh       | 18 +++++++-------
 cpp/src_prims/functions/penalty.cuh           | 12 +++++-----
 cpp/src_prims/functions/sigmoid.cuh           |  2 +-
 cpp/src_prims/functions/sign.cuh              |  2 +-
 cpp/src_prims/functions/softThres.cuh         |  2 +-
 cpp/src_prims/label/classlabels.cuh           |  8 +++----
 cpp/src_prims/label/merge_labels.cuh          |  4 ++--
 cpp/src_prims/linalg/batched/matrix.cuh       |  6 ++---
 cpp/src_prims/linalg/block.cuh                |  2 +-
 cpp/src_prims/linalg/init.h                   |  2 +-
 cpp/src_prims/matrix/grammatrix.cuh           |  4 ++--
 cpp/src_prims/matrix/kernelfactory.cuh        |  2 +-
 cpp/src_prims/matrix/kernelmatrices.cuh       |  4 ++--
 cpp/src_prims/metrics/adjusted_rand_index.cuh |  8 +++----
 .../metrics/batched/information_criterion.cuh |  2 +-
 cpp/src_prims/metrics/contingencyMatrix.cuh   |  2 +-
 cpp/src_prims/metrics/dispersion.cuh          |  6 ++---
 cpp/src_prims/metrics/entropy.cuh             |  6 ++---
 cpp/src_prims/metrics/kl_divergence.cuh       |  4 ++--
 cpp/src_prims/metrics/mutual_info_score.cuh   |  4 ++--
 cpp/src_prims/metrics/rand_index.cuh          |  2 +-
 cpp/src_prims/metrics/scores.cuh              | 12 +++++-----
 cpp/src_prims/metrics/silhouette_score.cuh    | 12 +++++-----
 .../metrics/trustworthiness_score.cuh         |  6 ++---
 cpp/src_prims/random/make_arima.cuh           |  2 +-
 cpp/src_prims/selection/knn.cuh               |  4 ++--
 cpp/src_prims/selection/processing.cuh        | 10 ++++----
 cpp/src_prims/sparse/batched/csr.cuh          |  4 ++--
 cpp/src_prims/timeSeries/arima_helpers.cuh    |  6 ++---
 cpp/src_prims/timeSeries/fillna.cuh           |  6 ++---
 cpp/src_prims/timeSeries/jones_transform.cuh  |  4 ++--
 cpp/src_prims/timeSeries/stationarity.cuh     |  8 +++----
 cpp/test/mg/knn.cu                            |  2 +-
 cpp/test/mg/knn_test_helper.cuh               |  2 +-
 cpp/test/mg/pca.cu                            |  4 ++--
 cpp/test/prims/add_sub_dev_scalar.cu          | 10 ++++----
 cpp/test/prims/adjusted_rand_index.cu         |  2 +-
 cpp/test/prims/batched/csr.cu                 |  2 +-
 cpp/test/prims/batched/gemv.cu                |  4 ++--
 .../prims/batched/information_criterion.cu    |  2 +-
 cpp/test/prims/batched/make_symm.cu           |  4 ++--
 cpp/test/prims/batched/matrix.cu              |  4 ++--
 cpp/test/prims/cache.cu                       |  4 ++--
 cpp/test/prims/columnSort.cu                  |  2 +-
 cpp/test/prims/completeness_score.cu          |  2 +-
 cpp/test/prims/contingencyMatrix.cu           |  2 +-
 cpp/test/prims/decoupled_lookback.cu          |  4 ++--
 cpp/test/prims/device_utils.cu                |  4 ++--
 cpp/test/prims/dispersion.cu                  |  4 ++--
 cpp/test/prims/dist_adj.cu                    |  4 ++--
 cpp/test/prims/distance_base.cuh              |  4 ++--
 cpp/test/prims/eltwise2d.cu                   |  4 ++--
 cpp/test/prims/entropy.cu                     |  4 ++--
 cpp/test/prims/fast_int_div.cu                |  2 +-
 cpp/test/prims/fillna.cu                      |  4 ++--
 cpp/test/prims/gather.cu                      |  6 ++---
 cpp/test/prims/gram.cu                        |  4 ++--
 cpp/test/prims/grid_sync.cu                   |  2 +-
 cpp/test/prims/hinge.cu                       |  4 ++--
 cpp/test/prims/homogeneity_score.cu           |  2 +-
 cpp/test/prims/jones_transform.cu             |  4 ++--
 cpp/test/prims/kl_divergence.cu               |  2 +-
 cpp/test/prims/knn_classify.cu                |  8 +++----
 cpp/test/prims/knn_regression.cu              |  8 +++----
 cpp/test/prims/kselection.cu                  |  4 ++--
 cpp/test/prims/label.cu                       |  4 ++--
 cpp/test/prims/linalg_block.cu                |  6 ++---
 cpp/test/prims/linearReg.cu                   |  4 ++--
 cpp/test/prims/log.cu                         |  2 +-
 cpp/test/prims/logisticReg.cu                 |  4 ++--
 cpp/test/prims/make_arima.cu                  |  4 ++--
 cpp/test/prims/merge_labels.cu                |  4 ++--
 cpp/test/prims/mutual_info_score.cu           |  2 +-
 cpp/test/prims/penalty.cu                     |  4 ++--
 cpp/test/prims/rand_index.cu                  |  2 +-
 cpp/test/prims/reverse.cu                     |  4 ++--
 cpp/test/prims/score.cu                       |  6 ++---
 cpp/test/prims/sigmoid.cu                     |  2 +-
 cpp/test/prims/silhouette_score.cu            |  2 +-
 cpp/test/prims/test_utils.h                   |  4 ++--
 cpp/test/prims/trustworthiness.cu             |  4 ++--
 cpp/test/prims/v_measure.cu                   |  2 +-
 cpp/test/sg/dbscan_test.cu                    |  8 +++----
 cpp/test/sg/fil_test.cu                       |  4 ++--
 cpp/test/sg/fnv_hash_test.cpp                 |  4 ++--
 cpp/test/sg/genetic/evolution_test.cu         |  4 ++--
 cpp/test/sg/genetic/node_test.cpp             |  4 ++--
 cpp/test/sg/genetic/program_test.cu           |  4 ++--
 cpp/test/sg/hdbscan_test.cu                   |  6 ++---
 cpp/test/sg/holtwinters_test.cu               |  4 ++--
 cpp/test/sg/kmeans_test.cu                    |  4 ++--
 cpp/test/sg/knn_test.cu                       |  4 ++--
 cpp/test/sg/lars_test.cu                      |  6 ++---
 cpp/test/sg/linear_svm_test.cu                | 10 ++++----
 cpp/test/sg/linkage_test.cu                   |  4 ++--
 cpp/test/sg/multi_sum_test.cu                 |  4 ++--
 cpp/test/sg/ols.cu                            |  2 +-
 cpp/test/sg/pca_test.cu                       |  4 ++--
 cpp/test/sg/quasi_newton.cu                   |  6 ++---
 cpp/test/sg/rf_test.cu                        |  8 +++----
 cpp/test/sg/ridge.cu                          |  2 +-
 cpp/test/sg/rproj_test.cu                     |  6 ++---
 cpp/test/sg/sgd.cu                            |  4 ++--
 cpp/test/sg/shap_kernel.cu                    |  4 ++--
 cpp/test/sg/svc_test.cu                       | 10 ++++----
 cpp/test/sg/trustworthiness_test.cu           |  4 ++--
 cpp/test/sg/tsne_test.cu                      |  4 ++--
 cpp/test/sg/tsvd_test.cu                      |  4 ++--
 cpp/test/sg/umap_parametrizable_test.cu       |  8 +++----
 wiki/cpp/DEVELOPER_GUIDE.md                   |  6 ++---
 296 files changed, 814 insertions(+), 816 deletions(-)

diff --git a/cpp/bench/common/ml_benchmark.hpp b/cpp/bench/common/ml_benchmark.hpp
index 23c7ef7e5a..71feb7608b 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <sstream>
 #include <string>
 #include <vector>
diff --git a/cpp/bench/prims/add.cu b/cpp/bench/prims/add.cu
index 5a9340cd2f..d3e89921e0 100644
--- a/cpp/bench/prims/add.cu
+++ b/cpp/bench/prims/add.cu
@@ -15,7 +15,7 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/distance_common.cuh b/cpp/bench/prims/distance_common.cuh
index dcc8292f82..9ce14ff897 100644
--- a/cpp/bench/prims/distance_common.cuh
+++ b/cpp/bench/prims/distance_common.cuh
@@ -15,9 +15,9 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/fused_l2_nn.cu b/cpp/bench/prims/fused_l2_nn.cu
index c949e119d3..475f384081 100644
--- a/cpp/bench/prims/fused_l2_nn.cu
+++ b/cpp/bench/prims/fused_l2_nn.cu
@@ -16,12 +16,12 @@
 
 #include <common/ml_benchmark.hpp>
 #include <limits>
-#include <raft/cudart_utils.h>
-#include <raft/distance/fused_l2_nn.hpp>
-#include <raft/handle.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/random/rng.hpp>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/fused_l2_nn.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/random/rng.cuh>
+#include <raft/spatial/knn/specializations.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/gram_matrix.cu b/cpp/bench/prims/gram_matrix.cu
index be1e180bc8..fec6cd1b8e 100644
--- a/cpp/bench/prims/gram_matrix.cu
+++ b/cpp/bench/prims/gram_matrix.cu
@@ -21,7 +21,7 @@
 #include <memory>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <sstream>
 #include <string>
 #include <vector>
diff --git a/cpp/bench/prims/make_blobs.cu b/cpp/bench/prims/make_blobs.cu
index 963679d77c..6b3beac598 100644
--- a/cpp/bench/prims/make_blobs.cu
+++ b/cpp/bench/prims/make_blobs.cu
@@ -15,7 +15,7 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/random/make_blobs.hpp>
+#include <raft/random/make_blobs.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/map_then_reduce.cu b/cpp/bench/prims/map_then_reduce.cu
index 0520562f7b..cb0c173750 100644
--- a/cpp/bench/prims/map_then_reduce.cu
+++ b/cpp/bench/prims/map_then_reduce.cu
@@ -15,7 +15,7 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
+#include <raft/linalg/map_then_reduce.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/matrix_vector_op.cu b/cpp/bench/prims/matrix_vector_op.cu
index e117d96bb2..168cd1a45d 100644
--- a/cpp/bench/prims/matrix_vector_op.cu
+++ b/cpp/bench/prims/matrix_vector_op.cu
@@ -15,7 +15,7 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/permute.cu b/cpp/bench/prims/permute.cu
index 9f6c6d40a7..50888a5688 100644
--- a/cpp/bench/prims/permute.cu
+++ b/cpp/bench/prims/permute.cu
@@ -15,9 +15,9 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/random/permute.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/permute.cuh>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/reduce.cu b/cpp/bench/prims/reduce.cu
index bdfe17c62d..f6df432dcf 100644
--- a/cpp/bench/prims/reduce.cu
+++ b/cpp/bench/prims/reduce.cu
@@ -15,7 +15,7 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/linalg/reduce.hpp>
+#include <raft/linalg/reduce.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/rng.cu b/cpp/bench/prims/rng.cu
index aca727eb1b..b8135198cf 100644
--- a/cpp/bench/prims/rng.cu
+++ b/cpp/bench/prims/rng.cu
@@ -15,8 +15,8 @@
  */
 
 #include <common/ml_benchmark.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/sg/arima_loglikelihood.cu b/cpp/bench/sg/arima_loglikelihood.cu
index 119eb5027a..34ae8f8b35 100644
--- a/cpp/bench/sg/arima_loglikelihood.cu
+++ b/cpp/bench/sg/arima_loglikelihood.cu
@@ -20,12 +20,12 @@
 
 #include <cuml/tsa/arima_common.h>
 #include <cuml/tsa/batched_arima.hpp>
-#include <raft/handle.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include "benchmark.cuh"
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/bench/sg/benchmark.cuh b/cpp/bench/sg/benchmark.cuh
index 279c85c6e4..3dbc50e087 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/bench/sg/dataset.cuh b/cpp/bench/sg/dataset.cuh
index af5975e1f0..6d965af53f 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/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/random/make_regression.hpp>
+#include <raft/linalg/transpose.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/random/make_regression.cuh>
 #include <sstream>
 #include <string>
 #include <vector>
diff --git a/cpp/bench/sg/dataset_ts.cuh b/cpp/bench/sg/dataset_ts.cuh
index 37392a8273..481a6b899f 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/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 namespace ML {
 namespace Bench {
diff --git a/cpp/examples/dbscan/dbscan_example.cpp b/cpp/examples/dbscan/dbscan_example.cpp
index 203e7b9afb..385aa66ae2 100644
--- a/cpp/examples/dbscan/dbscan_example.cpp
+++ b/cpp/examples/dbscan/dbscan_example.cpp
@@ -23,7 +23,7 @@
 #include <sstream>
 #include <vector>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/cluster/dbscan.hpp>
 
diff --git a/cpp/examples/kmeans/kmeans_example.cpp b/cpp/examples/kmeans/kmeans_example.cpp
index 81eab915cb..5149ffcd2d 100644
--- a/cpp/examples/kmeans/kmeans_example.cpp
+++ b/cpp/examples/kmeans/kmeans_example.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -23,7 +23,7 @@
 
 #include <cuda_runtime.h>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/cluster/kmeans.hpp>
 
diff --git a/cpp/examples/symreg/symreg_example.cpp b/cpp/examples/symreg/symreg_example.cpp
index 498c994d82..1ed530b5c4 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/include/cuml/cluster/hdbscan.hpp b/cpp/include/cuml/cluster/hdbscan.hpp
index 0e72ae030a..a061d00525 100644
--- a/cpp/include/cuml/cluster/hdbscan.hpp
+++ b/cpp/include/cuml/cluster/hdbscan.hpp
@@ -18,7 +18,7 @@
 
 #include <raft/distance/distance_type.hpp>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/include/cuml/cluster/linkage.hpp b/cpp/include/cuml/cluster/linkage.hpp
index eb6e88ff81..1da6d849be 100644
--- a/cpp/include/cuml/cluster/linkage.hpp
+++ b/cpp/include/cuml/cluster/linkage.hpp
@@ -19,7 +19,7 @@
 #include <raft/distance/distance_type.hpp>
 #include <raft/sparse/hierarchy/common.h>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace raft {
 class handle_t;
diff --git a/cpp/include/cuml/common/utils.hpp b/cpp/include/cuml/common/utils.hpp
index 120b81b061..2a61d04e4f 100644
--- a/cpp/include/cuml/common/utils.hpp
+++ b/cpp/include/cuml/common/utils.hpp
@@ -18,11 +18,11 @@
 
 #include <cuda_runtime.h>
 #include <execinfo.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <cstdio>
 
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 #include <sstream>
 #include <stdexcept>
 #include <string>
diff --git a/cpp/include/cuml/decomposition/sign_flip_mg.hpp b/cpp/include/cuml/decomposition/sign_flip_mg.hpp
index f64c75ce35..58f07e0331 100644
--- a/cpp/include/cuml/decomposition/sign_flip_mg.hpp
+++ b/cpp/include/cuml/decomposition/sign_flip_mg.hpp
@@ -18,7 +18,7 @@
 
 #include <cumlprims/opg/matrix/data.hpp>
 #include <cumlprims/opg/matrix/part_descriptor.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace PCA {
diff --git a/cpp/include/cuml/genetic/genetic.h b/cpp/include/cuml/genetic/genetic.h
index df50808646..207502bc77 100644
--- a/cpp/include/cuml/genetic/genetic.h
+++ b/cpp/include/cuml/genetic/genetic.h
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include "common.h"
 #include "program.h"
diff --git a/cpp/include/cuml/genetic/program.h b/cpp/include/cuml/genetic/program.h
index 10b949ed3c..42fcdb4eac 100644
--- a/cpp/include/cuml/genetic/program.h
+++ b/cpp/include/cuml/genetic/program.h
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <random>
 
 #include "common.h"
diff --git a/cpp/include/cuml/linear_model/glm.hpp b/cpp/include/cuml/linear_model/glm.hpp
index c4875c50fb..1085586f94 100644
--- a/cpp/include/cuml/linear_model/glm.hpp
+++ b/cpp/include/cuml/linear_model/glm.hpp
@@ -17,7 +17,7 @@
 
 #include <cuml/linear_model/qn.h>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/include/cuml/linear_model/preprocess_mg.hpp b/cpp/include/cuml/linear_model/preprocess_mg.hpp
index 5c30c0dca0..9373f25733 100644
--- a/cpp/include/cuml/linear_model/preprocess_mg.hpp
+++ b/cpp/include/cuml/linear_model/preprocess_mg.hpp
@@ -18,8 +18,8 @@
 
 #include <cumlprims/opg/matrix/data.hpp>
 #include <cumlprims/opg/matrix/part_descriptor.hpp>
-#include <raft/comms/comms.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/include/cuml/neighbors/knn_mg.hpp b/cpp/include/cuml/neighbors/knn_mg.hpp
index eca0ec0ac3..7ee9771fc0 100644
--- a/cpp/include/cuml/neighbors/knn_mg.hpp
+++ b/cpp/include/cuml/neighbors/knn_mg.hpp
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <vector>
 
 #include <cumlprims/opg/matrix/data.hpp>
diff --git a/cpp/include/cuml/random_projection/rproj_c.h b/cpp/include/cuml/random_projection/rproj_c.h
index 7e14e14e0d..5e6b148556 100644
--- a/cpp/include/cuml/random_projection/rproj_c.h
+++ b/cpp/include/cuml/random_projection/rproj_c.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2018-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/include/cuml/solvers/lars.hpp b/cpp/include/cuml/solvers/lars.hpp
index 796be49ca7..c740b64672 100644
--- a/cpp/include/cuml/solvers/lars.hpp
+++ b/cpp/include/cuml/solvers/lars.hpp
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace Solver {
diff --git a/cpp/include/cuml/svm/linear.hpp b/cpp/include/cuml/svm/linear.hpp
index 8df5c6e606..2717cec307 100644
--- a/cpp/include/cuml/svm/linear.hpp
+++ b/cpp/include/cuml/svm/linear.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace SVM {
diff --git a/cpp/include/cuml/svm/svc.hpp b/cpp/include/cuml/svm/svc.hpp
index 5ed13e60ac..5413929ab1 100644
--- a/cpp/include/cuml/svm/svc.hpp
+++ b/cpp/include/cuml/svm/svc.hpp
@@ -20,7 +20,7 @@
 #include "svm_parameter.h"
 #include <cuml/common/logger.hpp>
 #include <cuml/matrix/kernelparams.h>
-#include <raft/handle.hpp>
+#include <raft/core/handle.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 1f4f3554a3..feacae5b3e 100644
--- a/cpp/include/cuml/tsa/arima_common.h
+++ b/cpp/include/cuml/tsa/arima_common.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -20,7 +20,7 @@
 
 #include <algorithm>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/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 3bf5384342..1004df53c0 100644
--- a/cpp/src/arima/batched_arima.cu
+++ b/cpp/src/arima/batched_arima.cu
@@ -32,11 +32,11 @@
 #include <common/nvtx.hpp>
 #include <linalg/batched/matrix.cuh>
 #include <metrics/batched/information_criterion.cuh>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/core/nvtx.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
 #include <rmm/device_uvector.hpp>
 #include <timeSeries/arima_helpers.cuh>
 #include <timeSeries/fillna.cuh>
diff --git a/cpp/src/arima/batched_kalman.cu b/cpp/src/arima/batched_kalman.cu
index c05cce367f..f4cc3b6adf 100644
--- a/cpp/src/arima/batched_kalman.cu
+++ b/cpp/src/arima/batched_kalman.cu
@@ -24,17 +24,17 @@
 #include <thrust/for_each.h>
 #include <thrust/iterator/counting_iterator.h>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <linalg/batched/matrix.cuh>
 #include <linalg/block.cuh>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 #include <timeSeries/arima_helpers.cuh>
 
 namespace ML {
diff --git a/cpp/src/common/cumlHandle.cpp b/cpp/src/common/cumlHandle.cpp
index 42ec14f6df..187d8ede4f 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/cudart_utils.h>
+#include <raft/core/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/cumlHandle.hpp b/cpp/src/common/cumlHandle.hpp
index 3e0ca98dd5..3d7fa3b031 100644
--- a/cpp/src/common/cumlHandle.hpp
+++ b/cpp/src/common/cumlHandle.hpp
@@ -18,7 +18,7 @@
 
 #include <cuml/cuml_api.h>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 
diff --git a/cpp/src/common/cuml_api.cpp b/cpp/src/common/cuml_api.cpp
index fea33b4826..7ce6728395 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <cstddef>
 #include <functional>
diff --git a/cpp/src/common/nvtx.hpp b/cpp/src/common/nvtx.hpp
index 1d9a2a3336..21c61c48f4 100644
--- a/cpp/src/common/nvtx.hpp
+++ b/cpp/src/common/nvtx.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 namespace ML {
 
@@ -25,7 +25,7 @@ namespace ML {
  * @param name range name
  * @param stream stream to synchronize
  */
-[[deprecated("Use new raft::common::nvtx::push_range from <raft/common/nvtx.hpp>")]] inline void
+[[deprecated("Use new raft::common::nvtx::push_range from <raft/core/nvtx.hpp>")]] inline void
 PUSH_RANGE(const char* name, cudaStream_t stream)
 {
   raft::common::nvtx::push_range(name);
@@ -35,7 +35,7 @@ PUSH_RANGE(const char* name, cudaStream_t stream)
  * @brief Synchronize CUDA stream and pop the latest nvtx range
  * @param stream stream to synchronize
  */
-[[deprecated("Use new raft::common::nvtx::pop_range from <raft/common/nvtx.hpp>")]] inline void
+[[deprecated("Use new raft::common::nvtx::pop_range from <raft/core/nvtx.hpp>")]] inline void
 POP_RANGE(cudaStream_t stream)
 {
   raft::common::nvtx::pop_range();
@@ -45,14 +45,14 @@ POP_RANGE(cudaStream_t stream)
  * @brief Push a named nvtx range
  * @param name range name
  */
-[[deprecated("Use new raft::common::nvtx::push_range from <raft/common/nvtx.hpp>")]] inline void
+[[deprecated("Use new raft::common::nvtx::push_range from <raft/core/nvtx.hpp>")]] inline void
 PUSH_RANGE(const char* name)
 {
   raft::common::nvtx::push_range(name);
 }
 
 /** Pop the latest range */
-[[deprecated("Use new raft::common::nvtx::pop_range from <raft/common/nvtx.hpp>")]] inline void
+[[deprecated("Use new raft::common::nvtx::pop_range from <raft/core/nvtx.hpp>")]] inline void
 POP_RANGE()
 {
   raft::common::nvtx::pop_range();
diff --git a/cpp/src/common/tensor.hpp b/cpp/src/common/tensor.hpp
index b76428b6c3..61bb8b84e9 100644
--- a/cpp/src/common/tensor.hpp
+++ b/cpp/src/common/tensor.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 
 #include <vector>
diff --git a/cpp/src/datasets/make_blobs.cu b/cpp/src/datasets/make_blobs.cu
index 400a366588..bfe8f9b12c 100644
--- a/cpp/src/datasets/make_blobs.cu
+++ b/cpp/src/datasets/make_blobs.cu
@@ -15,7 +15,7 @@
  */
 
 #include <cuml/datasets/make_blobs.hpp>
-#include <raft/random/make_blobs.hpp>
+#include <raft/random/make_blobs.cuh>
 
 namespace ML {
 namespace Datasets {
diff --git a/cpp/src/datasets/make_regression.cu b/cpp/src/datasets/make_regression.cu
index 55a54368a7..463908bd0e 100644
--- a/cpp/src/datasets/make_regression.cu
+++ b/cpp/src/datasets/make_regression.cu
@@ -15,7 +15,7 @@
  */
 
 #include <cuml/datasets/make_regression.hpp>
-#include <raft/random/make_regression.hpp>
+#include <raft/random/make_regression.cuh>
 
 namespace ML {
 namespace Datasets {
diff --git a/cpp/src/dbscan/adjgraph/algo.cuh b/cpp/src/dbscan/adjgraph/algo.cuh
index 0e49a8fab1..543f820daa 100644
--- a/cpp/src/dbscan/adjgraph/algo.cuh
+++ b/cpp/src/dbscan/adjgraph/algo.cuh
@@ -22,8 +22,8 @@
 
 #include "pack.h"
 
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
 #include <raft/sparse/convert/csr.cuh>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/dbscan/corepoints/compute.cuh b/cpp/src/dbscan/corepoints/compute.cuh
index 547604b508..0cf8b14457 100644
--- a/cpp/src/dbscan/corepoints/compute.cuh
+++ b/cpp/src/dbscan/corepoints/compute.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <thrust/for_each.h>
 #include <thrust/iterator/counting_iterator.h>
diff --git a/cpp/src/dbscan/corepoints/exchange.cuh b/cpp/src/dbscan/corepoints/exchange.cuh
index 61897cd1a7..449147fd64 100644
--- a/cpp/src/dbscan/corepoints/exchange.cuh
+++ b/cpp/src/dbscan/corepoints/exchange.cuh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 namespace ML {
 namespace Dbscan {
diff --git a/cpp/src/dbscan/dbscan.cu b/cpp/src/dbscan/dbscan.cu
index df0a357809..ec65ebe725 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 namespace ML {
 namespace Dbscan {
diff --git a/cpp/src/dbscan/dbscan.cuh b/cpp/src/dbscan/dbscan.cuh
index 6b8c38435f..ef4be6f66f 100644
--- a/cpp/src/dbscan/dbscan.cuh
+++ b/cpp/src/dbscan/dbscan.cuh
@@ -18,7 +18,7 @@
 
 #include "runner.cuh"
 
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 #include <cuml/cluster/dbscan.hpp>
 #include <cuml/common/logger.hpp>
diff --git a/cpp/src/dbscan/mergelabels/runner.cuh b/cpp/src/dbscan/mergelabels/runner.cuh
index 506239e04a..4dc40e2d09 100644
--- a/cpp/src/dbscan/mergelabels/runner.cuh
+++ b/cpp/src/dbscan/mergelabels/runner.cuh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,7 +18,7 @@
 
 #include <label/merge_labels.cuh>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 namespace ML {
 namespace Dbscan {
 namespace MergeLabels {
diff --git a/cpp/src/dbscan/mergelabels/tree_reduction.cuh b/cpp/src/dbscan/mergelabels/tree_reduction.cuh
index 6128f9d7fd..3bb9fdd6cd 100644
--- a/cpp/src/dbscan/mergelabels/tree_reduction.cuh
+++ b/cpp/src/dbscan/mergelabels/tree_reduction.cuh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,7 +18,7 @@
 
 #include "runner.cuh"
 
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 #include <cuml/common/logger.hpp>
 
diff --git a/cpp/src/dbscan/runner.cuh b/cpp/src/dbscan/runner.cuh
index ba14717b26..7a5fb34822 100644
--- a/cpp/src/dbscan/runner.cuh
+++ b/cpp/src/dbscan/runner.cuh
@@ -24,13 +24,13 @@
 #include "vertexdeg/runner.cuh"
 #include <common/nvtx.hpp>
 #include <label/classlabels.cuh>
-#include <raft/common/nvtx.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/nvtx.hpp>
 #include <raft/sparse/csr.hpp>
 
 #include <cuml/common/logger.hpp>
 
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 #include <label/classlabels.cuh>
 
diff --git a/cpp/src/dbscan/vertexdeg/algo.cuh b/cpp/src/dbscan/vertexdeg/algo.cuh
index 216e009a3e..6743596dc5 100644
--- a/cpp/src/dbscan/vertexdeg/algo.cuh
+++ b/cpp/src/dbscan/vertexdeg/algo.cuh
@@ -20,10 +20,10 @@
 #include <cuda_runtime.h>
 #include <math.h>
 #include <raft/device_atomics.cuh>
-#include <raft/linalg/coalesced_reduction.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
+#include <raft/linalg/coalesced_reduction.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
 #include <raft/linalg/norm.cuh>
-#include <raft/spatial/knn/epsilon_neighborhood.hpp>
+#include <raft/spatial/knn/epsilon_neighborhood.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/dbscan/vertexdeg/precomputed.cuh b/cpp/src/dbscan/vertexdeg/precomputed.cuh
index e9fd3cf345..df70f81676 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/cudart_utils.h>
 #include <raft/device_atomics.cuh>
-#include <raft/linalg/coalesced_reduction.hpp>
-#include <raft/linalg/reduce.hpp>
+#include <raft/linalg/coalesced_reduction.cuh>
+#include <raft/linalg/reduce.cuh>
 
 #include "pack.h"
 
diff --git a/cpp/src/decisiontree/batched-levelalgo/builder.cuh b/cpp/src/decisiontree/batched-levelalgo/builder.cuh
index 0a7b9097af..c27d21f12f 100644
--- a/cpp/src/decisiontree/batched-levelalgo/builder.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/builder.cuh
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <memory>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include "kernels/builder_kernels.cuh"
@@ -28,7 +28,7 @@
 #include <raft/cuda_utils.cuh>
 
 #include <deque>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 #include <utility>
 
 namespace ML {
diff --git a/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels.cuh b/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels.cuh
index 7440e7618b..6909355b9f 100644
--- a/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/kernels/builder_kernels.cuh
@@ -20,7 +20,7 @@
 #include "../objectives.cuh"
 #include "../quantiles.h"
 
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 #include <cub/cub.cuh>
 
diff --git a/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh b/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
index fd90f37822..438d03b953 100644
--- a/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/quantiles.cuh
@@ -19,13 +19,13 @@
 #include <cub/cub.cuh>
 #include <iostream>
 #include <memory>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
 #include <thrust/fill.h>
 
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 #include "quantiles.h"
 
diff --git a/cpp/src/decisiontree/batched-levelalgo/split.cuh b/cpp/src/decisiontree/batched-levelalgo/split.cuh
index e69fb81489..ea80412cfc 100644
--- a/cpp/src/decisiontree/batched-levelalgo/split.cuh
+++ b/cpp/src/decisiontree/batched-levelalgo/split.cuh
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace ML {
 namespace DT {
diff --git a/cpp/src/decisiontree/decisiontree.cu b/cpp/src/decisiontree/decisiontree.cu
index 0b0e5df15d..46516aaa4a 100644
--- a/cpp/src/decisiontree/decisiontree.cu
+++ b/cpp/src/decisiontree/decisiontree.cu
@@ -16,7 +16,7 @@
 
 #include <cuml/tree/decisiontree.hpp>
 #include <cuml/tree/flatnode.h>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include "decisiontree.cuh"
 
diff --git a/cpp/src/decisiontree/decisiontree.cuh b/cpp/src/decisiontree/decisiontree.cuh
index e4f6f56f2e..03fc8af413 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 
 #include "treelite_util.h"
 #include <treelite/c_api.h>
@@ -32,7 +32,7 @@
 #include <locale>
 #include <map>
 #include <numeric>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 #include <random>
 #include <vector>
 
diff --git a/cpp/src/explainer/kernel_shap.cu b/cpp/src/explainer/kernel_shap.cu
index adae146012..13dc46e15c 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/explainer/kernel_shap.hpp>
 
diff --git a/cpp/src/explainer/permutation_shap.cu b/cpp/src/explainer/permutation_shap.cu
index 9b19b9b90b..54e560709d 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/explainer/permutation_shap.hpp>
 
diff --git a/cpp/src/explainer/tree_shap.cu b/cpp/src/explainer/tree_shap.cu
index cd57800ba2..df49232633 100644
--- a/cpp/src/explainer/tree_shap.cu
+++ b/cpp/src/explainer/tree_shap.cu
@@ -24,8 +24,8 @@
 #include <limits>
 #include <memory>
 #include <numeric>
-#include <raft/error.hpp>
-#include <raft/span.hpp>
+#include <raft/core/error.hpp>
+#include <raft/core/span.hpp>
 #include <rmm/device_uvector.hpp>
 #include <thrust/device_ptr.h>
 #include <thrust/device_vector.h>
diff --git a/cpp/src/fil/fil.cu b/cpp/src/fil/fil.cu
index 204b2b61f1..8c3965ded6 100644
--- a/cpp/src/fil/fil.cu
+++ b/cpp/src/fil/fil.cu
@@ -22,11 +22,11 @@ creation and prediction (the main inference kernel is defined in infer.cu). */
 
 #include <cuml/fil/fil.h>  // for algo_t,
 
-#include <raft/cudart_utils.h>     // for RAFT_CUDA_TRY, cudaStream_t,
-#include <raft/error.hpp>          // for ASSERT
-#include <raft/handle.hpp>         // for handle_t
-#include <rmm/device_uvector.hpp>  // for device_uvector
-#include <thrust/host_vector.h>    // for host_vector
+#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 <rmm/device_uvector.hpp>      // for device_uvector
+#include <thrust/host_vector.h>        // for host_vector
 
 #include <cmath>    // for expf
 #include <cstddef>  // for size_t
diff --git a/cpp/src/fil/infer.cu b/cpp/src/fil/infer.cu
index 5af31aa76e..54ff1590ad 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/cudart_utils.h>
 
 #include <thrust/functional.h>
 
diff --git a/cpp/src/fil/internal.cuh b/cpp/src/fil/internal.cuh
index 633eefde04..487f2c702a 100644
--- a/cpp/src/fil/internal.cuh
+++ b/cpp/src/fil/internal.cuh
@@ -22,8 +22,8 @@
 #include <cuml/fil/fil.h>
 #include <iostream>
 #include <numeric>
+#include <raft/core/error.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/error.hpp>
 #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 aeb113dd3b..edb69b8ac1 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/cudart_utils.h>  // for RAFT_CUDA_TRY
-#include <raft/error.hpp>       // for ASSERT
-#include <raft/handle.hpp>      // for handle_t
+#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 <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 a149ed23b9..a440489339 100644
--- a/cpp/src/genetic/fitness.cuh
+++ b/cpp/src/genetic/fitness.cuh
@@ -15,15 +15,15 @@
  */
 
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/strided_reduction.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
-#include <raft/stats/stddev.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/strided_reduction.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
+#include <raft/stats/stddev.cuh>
+#include <raft/stats/sum.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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 namespace cuml {
 namespace genetic {
diff --git a/cpp/src/genetic/genetic.cu b/cpp/src/genetic/genetic.cu
index fb46efa745..172147fc61 100644
--- a/cpp/src/genetic/genetic.cu
+++ b/cpp/src/genetic/genetic.cu
@@ -21,11 +21,11 @@
 #include <cuml/genetic/genetic.h>
 #include <cuml/genetic/program.h>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/random/rng.cuh>
 
 #include <algorithm>
 #include <numeric>
diff --git a/cpp/src/genetic/program.cu b/cpp/src/genetic/program.cu
index cd69056ca9..9b5707cc88 100644
--- a/cpp/src/genetic/program.cu
+++ b/cpp/src/genetic/program.cu
@@ -17,8 +17,8 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/genetic/node.h>
 #include <cuml/genetic/program.h>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <algorithm>
diff --git a/cpp/src/glm/ols_mg.cu b/cpp/src/glm/ols_mg.cu
index 3906c1fc6b..a2b45dd000 100644
--- a/cpp/src/glm/ols_mg.cu
+++ b/cpp/src/glm/ols_mg.cu
@@ -20,12 +20,12 @@
 #include <cumlprims/opg/linalg/lstsq.hpp>
 #include <cumlprims/opg/stats/mean.hpp>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/glm/preprocess.cuh b/cpp/src/glm/preprocess.cuh
index 88f20b76fc..8d6c069d53 100644
--- a/cpp/src/glm/preprocess.cuh
+++ b/cpp/src/glm/preprocess.cuh
@@ -16,15 +16,15 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
-#include <raft/stats/meanvar.hpp>
-#include <raft/stats/stddev.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
+#include <raft/stats/meanvar.cuh>
+#include <raft/stats/stddev.cuh>
 #include <raft/stats/weighted_mean.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
diff --git a/cpp/src/glm/preprocess_mg.cu b/cpp/src/glm/preprocess_mg.cu
index a9874aff79..b9cdcc446e 100644
--- a/cpp/src/glm/preprocess_mg.cu
+++ b/cpp/src/glm/preprocess_mg.cu
@@ -20,12 +20,12 @@
 #include <cumlprims/opg/matrix/math.hpp>
 #include <cumlprims/opg/stats/mean.hpp>
 #include <cumlprims/opg/stats/mean_center.hpp>
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/matrix/math.hpp>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/matrix/math.cuh>
 #include <rmm/device_uvector.hpp>
 
 using namespace MLCommon;
diff --git a/cpp/src/glm/qn/glm_base.cuh b/cpp/src/glm/qn/glm_base.cuh
index 73faac80ed..364e8eac70 100644
--- a/cpp/src/glm/qn/glm_base.cuh
+++ b/cpp/src/glm/qn/glm_base.cuh
@@ -17,13 +17,13 @@
 #pragma once
 
 #include "simple_mat.cuh"
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/map.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/map.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/stats/mean.cuh>
 
 #include <thrust/execution_policy.h>
 #include <thrust/functional.h>
diff --git a/cpp/src/glm/qn/glm_linear.cuh b/cpp/src/glm/qn/glm_linear.cuh
index d5f7f51aed..eae1b8b8f7 100644
--- a/cpp/src/glm/qn/glm_linear.cuh
+++ b/cpp/src/glm/qn/glm_linear.cuh
@@ -19,7 +19,7 @@
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_logistic.cuh b/cpp/src/glm/qn/glm_logistic.cuh
index 5129006ec9..07d654fce3 100644
--- a/cpp/src/glm/qn/glm_logistic.cuh
+++ b/cpp/src/glm/qn/glm_logistic.cuh
@@ -19,7 +19,7 @@
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_regularizer.cuh b/cpp/src/glm/qn/glm_regularizer.cuh
index 0f99df716e..4d75f22046 100644
--- a/cpp/src/glm/qn/glm_regularizer.cuh
+++ b/cpp/src/glm/qn/glm_regularizer.cuh
@@ -17,11 +17,11 @@
 #pragma once
 
 #include "simple_mat.cuh"
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/stats/mean.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_softmax.cuh b/cpp/src/glm/qn/glm_softmax.cuh
index abf7708e03..84555e4ad5 100644
--- a/cpp/src/glm/qn/glm_softmax.cuh
+++ b/cpp/src/glm/qn/glm_softmax.cuh
@@ -19,7 +19,7 @@
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/glm_svm.cuh b/cpp/src/glm/qn/glm_svm.cuh
index 944bfefe7a..846c378d99 100644
--- a/cpp/src/glm/qn/glm_svm.cuh
+++ b/cpp/src/glm/qn/glm_svm.cuh
@@ -19,7 +19,7 @@
 #include "glm_base.cuh"
 #include "simple_mat.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 
 namespace ML {
 namespace GLM {
diff --git a/cpp/src/glm/qn/qn.cuh b/cpp/src/glm/qn/qn.cuh
index 6c55be2b95..6fa67b653f 100644
--- a/cpp/src/glm/qn/qn.cuh
+++ b/cpp/src/glm/qn/qn.cuh
@@ -27,7 +27,7 @@
 
 #include <cuml/linear_model/qn.h>
 
-#include <raft/matrix/math.hpp>
+#include <raft/matrix/math.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/glm/qn/simple_mat/base.hpp b/cpp/src/glm/qn/simple_mat/base.hpp
index e72119a310..5d0cfefd2a 100644
--- a/cpp/src/glm/qn/simple_mat/base.hpp
+++ b/cpp/src/glm/qn/simple_mat/base.hpp
@@ -15,9 +15,9 @@
  */
 #pragma once
 
+#include <raft/core/handle.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
-#include <raft/interruptible.hpp>
 
 namespace ML {
 
diff --git a/cpp/src/glm/qn/simple_mat/dense.hpp b/cpp/src/glm/qn/simple_mat/dense.hpp
index 67f62cacfd..b655d6500e 100644
--- a/cpp/src/glm/qn/simple_mat/dense.hpp
+++ b/cpp/src/glm/qn/simple_mat/dense.hpp
@@ -19,16 +19,16 @@
 #include <vector>
 
 #include "base.hpp"
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 #include <raft/linalg/ternary_op.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/glm/qn/simple_mat/sparse.hpp b/cpp/src/glm/qn/simple_mat/sparse.hpp
index e71691cd3e..2b53023c4f 100644
--- a/cpp/src/glm/qn/simple_mat/sparse.hpp
+++ b/cpp/src/glm/qn/simple_mat/sparse.hpp
@@ -19,15 +19,15 @@
 #include <vector>
 
 #include "base.hpp"
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
 #include <raft/linalg/ternary_op.cuh>
 
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <raft/sparse/detail/cusparse_wrappers.h>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/glm/ridge.cuh b/cpp/src/glm/ridge.cuh
index 2f13cb74e2..93956ff4af 100644
--- a/cpp/src/glm/ridge.cuh
+++ b/cpp/src/glm/ridge.cuh
@@ -16,18 +16,18 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/svd.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
-#include <raft/stats/stddev.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/svd.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
+#include <raft/stats/stddev.cuh>
+#include <raft/stats/sum.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include "preprocess.cuh"
diff --git a/cpp/src/glm/ridge_mg.cu b/cpp/src/glm/ridge_mg.cu
index 30a571cce3..663944a562 100644
--- a/cpp/src/glm/ridge_mg.cu
+++ b/cpp/src/glm/ridge_mg.cu
@@ -21,13 +21,13 @@
 #include <cumlprims/opg/linalg/svd.hpp>
 #include <cumlprims/opg/stats/mean.hpp>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/hdbscan/condensed_hierarchy.cu b/cpp/src/hdbscan/condensed_hierarchy.cu
index 15afb14366..1d45df8f1f 100644
--- a/cpp/src/hdbscan/condensed_hierarchy.cu
+++ b/cpp/src/hdbscan/condensed_hierarchy.cu
@@ -14,18 +14,18 @@
  * limitations under the License.
  */
 
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 
 #include <cub/cub.cuh>
 
 #include <cuml/common/logger.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
diff --git a/cpp/src/hdbscan/detail/condense.cuh b/cpp/src/hdbscan/detail/condense.cuh
index 50202b6130..b4b9685e6e 100644
--- a/cpp/src/hdbscan/detail/condense.cuh
+++ b/cpp/src/hdbscan/detail/condense.cuh
@@ -20,13 +20,13 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
diff --git a/cpp/src/hdbscan/detail/extract.cuh b/cpp/src/hdbscan/detail/extract.cuh
index e778124216..14fc377742 100644
--- a/cpp/src/hdbscan/detail/extract.cuh
+++ b/cpp/src/hdbscan/detail/extract.cuh
@@ -21,13 +21,13 @@
 #include "stabilities.cuh"
 #include "utils.h"
 
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/cudart_utils.h>
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #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 87d26235ed..c9fd087d3f 100644
--- a/cpp/src/hdbscan/detail/membership.cuh
+++ b/cpp/src/hdbscan/detail/membership.cuh
@@ -21,14 +21,14 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 
 #include <algorithm>
 
diff --git a/cpp/src/hdbscan/detail/predict.cuh b/cpp/src/hdbscan/detail/predict.cuh
index c4ec1020ba..74852f516d 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/cudart_utils.h>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/hdbscan/detail/reachability.cuh b/cpp/src/hdbscan/detail/reachability.cuh
index b3e533e35a..a818697ee8 100644
--- a/cpp/src/hdbscan/detail/reachability.cuh
+++ b/cpp/src/hdbscan/detail/reachability.cuh
@@ -18,21 +18,21 @@
 
 #include "reachability_faiss.cuh"
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
-#include <raft/sparse/convert/csr.hpp>
+#include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/hierarchy/detail/connectivities.cuh>
-#include <raft/sparse/linalg/symmetrize.hpp>
-#include <raft/sparse/selection/knn_graph.hpp>
+#include <raft/sparse/linalg/symmetrize.cuh>
+#include <raft/sparse/selection/knn_graph.cuh>
 
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
 
 #include <cuml/neighbors/knn.hpp>
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/iterator/zip_iterator.h>
diff --git a/cpp/src/hdbscan/detail/reachability_faiss.cuh b/cpp/src/hdbscan/detail/reachability_faiss.cuh
index 804c8b500d..625c26539e 100644
--- a/cpp/src/hdbscan/detail/reachability_faiss.cuh
+++ b/cpp/src/hdbscan/detail/reachability_faiss.cuh
@@ -38,8 +38,8 @@
 #include <faiss/gpu/utils/Limits.cuh>
 #include <faiss/gpu/utils/MatrixMult.cuh>
 
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
 #include <raft/spatial/knn/faiss_mr.hpp>
 
 #include <rmm/device_uvector.hpp>
diff --git a/cpp/src/hdbscan/detail/select.cuh b/cpp/src/hdbscan/detail/select.cuh
index 8236e22a7c..c0a68a540a 100644
--- a/cpp/src/hdbscan/detail/select.cuh
+++ b/cpp/src/hdbscan/detail/select.cuh
@@ -21,14 +21,14 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 
 #include <algorithm>
 
diff --git a/cpp/src/hdbscan/detail/soft_clustering.cuh b/cpp/src/hdbscan/detail/soft_clustering.cuh
index 214d376563..bca750c659 100644
--- a/cpp/src/hdbscan/detail/soft_clustering.cuh
+++ b/cpp/src/hdbscan/detail/soft_clustering.cuh
@@ -22,20 +22,20 @@
 
 #include <cub/cub.cuh>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 #include <raft/distance/distance_type.hpp>
-#include <raft/label/classlabels.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/matrix/math.hpp>
+#include <raft/label/classlabels.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/matrix/math.cuh>
 
 #include <algorithm>
 #include <cmath>
diff --git a/cpp/src/hdbscan/detail/stabilities.cuh b/cpp/src/hdbscan/detail/stabilities.cuh
index 3a145b200b..de5876b6ad 100644
--- a/cpp/src/hdbscan/detail/stabilities.cuh
+++ b/cpp/src/hdbscan/detail/stabilities.cuh
@@ -21,14 +21,14 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 
 #include <algorithm>
 
diff --git a/cpp/src/hdbscan/detail/utils.h b/cpp/src/hdbscan/detail/utils.h
index 0a080cf243..dcf03ef25c 100644
--- a/cpp/src/hdbscan/detail/utils.h
+++ b/cpp/src/hdbscan/detail/utils.h
@@ -18,16 +18,16 @@
 
 #include <cub/cub.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/label/classlabels.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/norm.hpp>
+#include <raft/label/classlabels.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/norm.cuh>
 
 #include <algorithm>
 
diff --git a/cpp/src/hdbscan/hdbscan.cu b/cpp/src/hdbscan/hdbscan.cu
index 853e39aa80..45a72f7f17 100644
--- a/cpp/src/hdbscan/hdbscan.cu
+++ b/cpp/src/hdbscan/hdbscan.cu
@@ -18,11 +18,8 @@
 #include "detail/predict.cuh"
 #include <cuml/cluster/hdbscan.hpp>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-
-#include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
 #include "runner.h"
 
diff --git a/cpp/src/hdbscan/prediction_data.cu b/cpp/src/hdbscan/prediction_data.cu
index ab7d0ce3b3..b8e1ece633 100644
--- a/cpp/src/hdbscan/prediction_data.cu
+++ b/cpp/src/hdbscan/prediction_data.cu
@@ -16,14 +16,14 @@
 
 #include "detail/utils.h"
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
 #include <cuml/cluster/hdbscan.hpp>
 
-#include <raft/matrix/math.hpp>
-#include <raft/sparse/convert/csr.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/matrix/math.cuh>
+#include <raft/sparse/convert/csr.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <algorithm>
 #include <cmath>
diff --git a/cpp/src/hdbscan/runner.h b/cpp/src/hdbscan/runner.h
index a26dd61a82..fccda626c3 100644
--- a/cpp/src/hdbscan/runner.h
+++ b/cpp/src/hdbscan/runner.h
@@ -16,9 +16,9 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <cuml/common/logger.hpp>
diff --git a/cpp/src/hierarchy/linkage.cu b/cpp/src/hierarchy/linkage.cu
index 0874cfd38a..dae2b43199 100644
--- a/cpp/src/hierarchy/linkage.cu
+++ b/cpp/src/hierarchy/linkage.cu
@@ -16,7 +16,7 @@
 
 #include "pw_dist_graph.cuh"
 #include <cuml/cluster/linkage.hpp>
-#include <raft/sparse/hierarchy/single_linkage.hpp>
+#include <raft/sparse/hierarchy/single_linkage.cuh>
 
 namespace raft {
 class handle_t;
diff --git a/cpp/src/hierarchy/pw_dist_graph.cuh b/cpp/src/hierarchy/pw_dist_graph.cuh
index 716a6a3e00..302aa2780a 100644
--- a/cpp/src/hierarchy/pw_dist_graph.cuh
+++ b/cpp/src/hierarchy/pw_dist_graph.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
 #include <cuml/metrics/metrics.hpp>
 
diff --git a/cpp/src/holtwinters/internal/hw_decompose.cuh b/cpp/src/holtwinters/internal/hw_decompose.cuh
index d94cf30ec3..fd6f3934f9 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.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 93592811e0..acd7d0802b 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.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 0036f76fe2..e267e4a938 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 template <typename Dtype>
 __device__ Dtype golden_step(Dtype a, Dtype b, Dtype c)
diff --git a/cpp/src/holtwinters/internal/hw_utils.cuh b/cpp/src/holtwinters/internal/hw_utils.cuh
index 7de8c5e577..4631170084 100644
--- a/cpp/src/holtwinters/internal/hw_utils.cuh
+++ b/cpp/src/holtwinters/internal/hw_utils.cuh
@@ -18,8 +18,8 @@
 #include <cuda_runtime.h>
 #include <cuml/tsa/holtwinters_params.h>
 #include <iostream>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/eltwise.cuh>
 #include <vector>
 
 #define IDX(n, m, N) (n + (m) * (N))
diff --git a/cpp/src/holtwinters/runner.cuh b/cpp/src/holtwinters/runner.cuh
index e06bd50543..f8f093ebdc 100644
--- a/cpp/src/holtwinters/runner.cuh
+++ b/cpp/src/holtwinters/runner.cuh
@@ -21,10 +21,10 @@
 #include "internal/hw_forecast.cuh"
 #include "internal/hw_optim.cuh"
 #include <cuml/tsa/holtwinters_params.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/linalg/transpose.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/kmeans/common.cuh b/cpp/src/kmeans/common.cuh
index 7b46b29742..2ca5493363 100644
--- a/cpp/src/kmeans/common.cuh
+++ b/cpp/src/kmeans/common.cuh
@@ -27,16 +27,16 @@
 #include <matrix/gather.cuh>
 #include <raft/linalg/reduce_cols_by_key.cuh>
 #include <raft/linalg/reduce_rows_by_key.cuh>
-#include <raft/random/permute.hpp>
-
-#include <raft/comms/comms.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/distance/fused_l2_nn.hpp>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/mean_squared_error.hpp>
-#include <raft/linalg/reduce.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/permute.cuh>
+
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/distance/fused_l2_nn.cuh>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/mean_squared_error.cuh>
+#include <raft/linalg/reduce.cuh>
+#include <raft/random/rng.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/kmeans/kmeans_mg_impl.cuh b/cpp/src/kmeans/kmeans_mg_impl.cuh
index eb85141738..af5427e765 100644
--- a/cpp/src/kmeans/kmeans_mg_impl.cuh
+++ b/cpp/src/kmeans/kmeans_mg_impl.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 #include <cuml/cluster/kmeans.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/kmeans/sg_impl.cuh b/cpp/src/kmeans/sg_impl.cuh
index 65b7b66ec7..ff308db2bd 100644
--- a/cpp/src/kmeans/sg_impl.cuh
+++ b/cpp/src/kmeans/sg_impl.cuh
@@ -18,7 +18,7 @@
 
 #include "common.cuh"
 #include <cuml/cluster/kmeans.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/knn/knn.cu b/cpp/src/knn/knn.cu
index e9586c53dd..da5bd634b0 100644
--- a/cpp/src/knn/knn.cu
+++ b/cpp/src/knn/knn.cu
@@ -16,12 +16,12 @@
 
 #include <cuda_runtime.h>
 #include <raft/cuda_utils.cuh>
-#include <raft/label/classlabels.hpp>
-#include <raft/spatial/knn/ann.hpp>
-#include <raft/spatial/knn/ball_cover.hpp>
+#include <raft/label/classlabels.cuh>
+#include <raft/spatial/knn/ann.cuh>
+#include <raft/spatial/knn/ball_cover.cuh>
 
-#include <raft/spatial/knn/knn.hpp>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/spatial/knn/knn.cuh>
+#include <raft/spatial/knn/specializations.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <cuml/common/logger.hpp>
diff --git a/cpp/src/knn/knn_opg_common.cuh b/cpp/src/knn/knn_opg_common.cuh
index 2bd0e52344..5e83e59154 100644
--- a/cpp/src/knn/knn_opg_common.cuh
+++ b/cpp/src/knn/knn_opg_common.cuh
@@ -24,10 +24,10 @@
 #include <cumlprims/opg/matrix/data.hpp>
 #include <cumlprims/opg/matrix/part_descriptor.hpp>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/spatial/knn/knn.hpp>
+#include <raft/spatial/knn/knn.cuh>
 
 #include <cstddef>
 #include <memory>
diff --git a/cpp/src/knn/knn_sparse.cu b/cpp/src/knn/knn_sparse.cu
index 22fe9a186e..c5ddfec160 100644
--- a/cpp/src/knn/knn_sparse.cu
+++ b/cpp/src/knn/knn_sparse.cu
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,7 +17,7 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/neighbors/knn_sparse.hpp>
 
-#include <raft/sparse/selection/knn.hpp>
+#include <raft/sparse/selection/knn.cuh>
 
 #include <cusparse_v2.h>
 
diff --git a/cpp/src/metrics/mutual_info_score.cu b/cpp/src/metrics/mutual_info_score.cu
index 1c2cf4c2a3..e0eeb0b8c1 100644
--- a/cpp/src/metrics/mutual_info_score.cu
+++ b/cpp/src/metrics/mutual_info_score.cu
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -15,7 +15,7 @@
  * limitations under the License.
  */
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/metrics/metrics.hpp>
 #include <metrics/mutual_info_score.cuh>
diff --git a/cpp/src/metrics/pairwise_distance.cu b/cpp/src/metrics/pairwise_distance.cu
index bdba9032f6..1928091c33 100644
--- a/cpp/src/metrics/pairwise_distance.cu
+++ b/cpp/src/metrics/pairwise_distance.cu
@@ -28,11 +28,11 @@
 #include "pairwise_distance_minkowski.cuh"
 #include "pairwise_distance_russell_rao.cuh"
 #include <cuml/metrics/metrics.hpp>
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 #include <raft/sparse/distance/common.h>
-#include <raft/sparse/distance/distance.hpp>
+#include <raft/sparse/distance/distance.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_canberra.cu b/cpp/src/metrics/pairwise_distance_canberra.cu
index 6ec30e0371..757b2f8dab 100644
--- a/cpp/src/metrics/pairwise_distance_canberra.cu
+++ b/cpp/src/metrics/pairwise_distance_canberra.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_canberra.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_canberra.cuh b/cpp/src/metrics/pairwise_distance_canberra.cuh
index 21dea2b73b..231d00fcaf 100644
--- a/cpp/src/metrics/pairwise_distance_canberra.cuh
+++ b/cpp/src/metrics/pairwise_distance_canberra.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_chebyshev.cu b/cpp/src/metrics/pairwise_distance_chebyshev.cu
index 28b0d5ea2a..0dbf77914d 100644
--- a/cpp/src/metrics/pairwise_distance_chebyshev.cu
+++ b/cpp/src/metrics/pairwise_distance_chebyshev.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_chebyshev.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_chebyshev.cuh b/cpp/src/metrics/pairwise_distance_chebyshev.cuh
index 22303f4367..217bc9ce2c 100644
--- a/cpp/src/metrics/pairwise_distance_chebyshev.cuh
+++ b/cpp/src/metrics/pairwise_distance_chebyshev.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,9 +16,9 @@
  */
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_correlation.cu b/cpp/src/metrics/pairwise_distance_correlation.cu
index 5e3bc89877..630282d510 100644
--- a/cpp/src/metrics/pairwise_distance_correlation.cu
+++ b/cpp/src/metrics/pairwise_distance_correlation.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_correlation.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_correlation.cuh b/cpp/src/metrics/pairwise_distance_correlation.cuh
index 4fc877032a..587bb52e61 100644
--- a/cpp/src/metrics/pairwise_distance_correlation.cuh
+++ b/cpp/src/metrics/pairwise_distance_correlation.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_cosine.cu b/cpp/src/metrics/pairwise_distance_cosine.cu
index dc9180683f..3e4f9c6e81 100644
--- a/cpp/src/metrics/pairwise_distance_cosine.cu
+++ b/cpp/src/metrics/pairwise_distance_cosine.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_cosine.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_cosine.cuh b/cpp/src/metrics/pairwise_distance_cosine.cuh
index 309e4382cd..9587636378 100644
--- a/cpp/src/metrics/pairwise_distance_cosine.cuh
+++ b/cpp/src/metrics/pairwise_distance_cosine.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,9 +16,9 @@
  */
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_euclidean.cu b/cpp/src/metrics/pairwise_distance_euclidean.cu
index 63ea9af1c4..0a73883ba9 100644
--- a/cpp/src/metrics/pairwise_distance_euclidean.cu
+++ b/cpp/src/metrics/pairwise_distance_euclidean.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_euclidean.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_euclidean.cuh b/cpp/src/metrics/pairwise_distance_euclidean.cuh
index 02e07727b0..8e1291399f 100644
--- a/cpp/src/metrics/pairwise_distance_euclidean.cuh
+++ b/cpp/src/metrics/pairwise_distance_euclidean.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -15,9 +15,9 @@
  * limitations under the License.
  */
 #pragma once
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_hamming.cu b/cpp/src/metrics/pairwise_distance_hamming.cu
index d3f5c33446..26ac60c368 100644
--- a/cpp/src/metrics/pairwise_distance_hamming.cu
+++ b/cpp/src/metrics/pairwise_distance_hamming.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_hamming.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_hamming.cuh b/cpp/src/metrics/pairwise_distance_hamming.cuh
index 3c77107e84..aac0a5f631 100644
--- a/cpp/src/metrics/pairwise_distance_hamming.cuh
+++ b/cpp/src/metrics/pairwise_distance_hamming.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_hellinger.cu b/cpp/src/metrics/pairwise_distance_hellinger.cu
index 4824a7d4d3..e6afeedcc7 100644
--- a/cpp/src/metrics/pairwise_distance_hellinger.cu
+++ b/cpp/src/metrics/pairwise_distance_hellinger.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_hellinger.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_hellinger.cuh b/cpp/src/metrics/pairwise_distance_hellinger.cuh
index 130bd24286..ade8458dcc 100644
--- a/cpp/src/metrics/pairwise_distance_hellinger.cuh
+++ b/cpp/src/metrics/pairwise_distance_hellinger.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,9 +16,9 @@
  */
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_jensen_shannon.cu b/cpp/src/metrics/pairwise_distance_jensen_shannon.cu
index b2e2f76ad5..3de240da98 100644
--- a/cpp/src/metrics/pairwise_distance_jensen_shannon.cu
+++ b/cpp/src/metrics/pairwise_distance_jensen_shannon.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_jensen_shannon.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_jensen_shannon.cuh b/cpp/src/metrics/pairwise_distance_jensen_shannon.cuh
index 48bf518f99..0317f5d2f0 100644
--- a/cpp/src/metrics/pairwise_distance_jensen_shannon.cuh
+++ b/cpp/src/metrics/pairwise_distance_jensen_shannon.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_kl_divergence.cu b/cpp/src/metrics/pairwise_distance_kl_divergence.cu
index 61ca9b9039..fd5146a295 100644
--- a/cpp/src/metrics/pairwise_distance_kl_divergence.cu
+++ b/cpp/src/metrics/pairwise_distance_kl_divergence.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_kl_divergence.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_kl_divergence.cuh b/cpp/src/metrics/pairwise_distance_kl_divergence.cuh
index 3ee4e0ce11..b213f7dd3c 100644
--- a/cpp/src/metrics/pairwise_distance_kl_divergence.cuh
+++ b/cpp/src/metrics/pairwise_distance_kl_divergence.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_l1.cu b/cpp/src/metrics/pairwise_distance_l1.cu
index b96723d1ed..5c2edd6c66 100644
--- a/cpp/src/metrics/pairwise_distance_l1.cu
+++ b/cpp/src/metrics/pairwise_distance_l1.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_l1.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_l1.cuh b/cpp/src/metrics/pairwise_distance_l1.cuh
index c0fb8bdea6..791ee6189d 100644
--- a/cpp/src/metrics/pairwise_distance_l1.cuh
+++ b/cpp/src/metrics/pairwise_distance_l1.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,9 +16,9 @@
  */
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_minkowski.cu b/cpp/src/metrics/pairwise_distance_minkowski.cu
index 924fe322b8..d7dfcc96d0 100644
--- a/cpp/src/metrics/pairwise_distance_minkowski.cu
+++ b/cpp/src/metrics/pairwise_distance_minkowski.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_minkowski.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_minkowski.cuh b/cpp/src/metrics/pairwise_distance_minkowski.cuh
index 9d493a80f1..fded8672ec 100644
--- a/cpp/src/metrics/pairwise_distance_minkowski.cuh
+++ b/cpp/src/metrics/pairwise_distance_minkowski.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/pairwise_distance_russell_rao.cu b/cpp/src/metrics/pairwise_distance_russell_rao.cu
index c005eaea56..ab11531219 100644
--- a/cpp/src/metrics/pairwise_distance_russell_rao.cu
+++ b/cpp/src/metrics/pairwise_distance_russell_rao.cu
@@ -16,8 +16,8 @@
  */
 
 #include "pairwise_distance_russell_rao.cuh"
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/metrics/pairwise_distance_russell_rao.cuh b/cpp/src/metrics/pairwise_distance_russell_rao.cuh
index 86c7464155..0e06af702f 100644
--- a/cpp/src/metrics/pairwise_distance_russell_rao.cuh
+++ b/cpp/src/metrics/pairwise_distance_russell_rao.cuh
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
 namespace ML {
 
diff --git a/cpp/src/metrics/rand_index.cu b/cpp/src/metrics/rand_index.cu
index 021b0e1b28..6d25490e20 100644
--- a/cpp/src/metrics/rand_index.cu
+++ b/cpp/src/metrics/rand_index.cu
@@ -1,6 +1,6 @@
 
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -15,7 +15,7 @@
  * limitations under the License.
  */
 
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include <cuml/metrics/metrics.hpp>
 #include <metrics/rand_index.cuh>
diff --git a/cpp/src/metrics/trustworthiness.cu b/cpp/src/metrics/trustworthiness.cu
index e97776df47..813d248594 100644
--- a/cpp/src/metrics/trustworthiness.cu
+++ b/cpp/src/metrics/trustworthiness.cu
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2018-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,10 +18,10 @@
 
 #include <cuml/metrics/metrics.hpp>
 
-#include <raft/distance/distance.hpp>
-#include <raft/distance/specializations.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/distance/specializations.cuh>
 
-#include <raft/handle.hpp>
+#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 fc8df559d7..5cd0dde268 100644
--- a/cpp/src/ml_cuda_utils.h
+++ b/cpp/src/ml_cuda_utils.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <cuda_runtime.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 namespace ML {
 
diff --git a/cpp/src/ml_mg_utils.cuh b/cpp/src/ml_mg_utils.cuh
index da1b3a934d..2e696fac0f 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/cudart_utils.h>
 
 #include <cuda_runtime.h>
 
diff --git a/cpp/src/pca/pca.cuh b/cpp/src/pca/pca.cuh
index 52d6b40665..459c2e96d8 100644
--- a/cpp/src/pca/pca.cuh
+++ b/cpp/src/pca/pca.cuh
@@ -17,16 +17,16 @@
 #pragma once
 
 #include <cuml/decomposition/params.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
-#include <raft/linalg/eig.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/cov.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
+#include <raft/linalg/eig.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/cov.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
 #include <rmm/device_uvector.hpp>
 #include <tsvd/tsvd.cuh>
 
diff --git a/cpp/src/pca/pca_mg.cu b/cpp/src/pca/pca_mg.cu
index 483f76b470..2b11608d15 100644
--- a/cpp/src/pca/pca_mg.cu
+++ b/cpp/src/pca/pca_mg.cu
@@ -26,12 +26,12 @@
 #include <cumlprims/opg/stats/mean.hpp>
 #include <cumlprims/opg/stats/mean_center.hpp>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/transpose.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/stats/mean_center.hpp>
+#include <raft/linalg/transpose.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/stats/mean_center.cuh>
 
 #include <cstddef>
 
diff --git a/cpp/src/pca/sign_flip_mg.cu b/cpp/src/pca/sign_flip_mg.cu
index 4ef824c901..65ebafa9ab 100644
--- a/cpp/src/pca/sign_flip_mg.cu
+++ b/cpp/src/pca/sign_flip_mg.cu
@@ -18,11 +18,11 @@
 #include <thrust/device_vector.h>
 #include <thrust/execution_policy.h>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/exec_policy.hpp>
 
 #include <thrust/device_vector.h>
diff --git a/cpp/src/random_projection/rproj.cuh b/cpp/src/random_projection/rproj.cuh
index 5bee38fda4..81a6f77b3f 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/cudart_utils.h>
 
 // TODO: This needs to be removed.
 #include <raft/sparse/detail/cusparse_wrappers.h>
diff --git a/cpp/src/random_projection/rproj_utils.cuh b/cpp/src/random_projection/rproj_utils.cuh
index c9db13d020..b6a4d8da78 100644
--- a/cpp/src/random_projection/rproj_utils.cuh
+++ b/cpp/src/random_projection/rproj_utils.cuh
@@ -18,9 +18,9 @@
 
 #include <cuml/random_projection/rproj_c.h>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/randomforest/randomforest.cu b/cpp/src/randomforest/randomforest.cu
index b04927fbd6..c17bc6a666 100644
--- a/cpp/src/randomforest/randomforest.cu
+++ b/cpp/src/randomforest/randomforest.cu
@@ -21,7 +21,7 @@
 #include <treelite/c_api.h>
 #include <treelite/tree.h>
 
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 
 #include <cstddef>
 #include <cstdio>
diff --git a/cpp/src/randomforest/randomforest.cuh b/cpp/src/randomforest/randomforest.cuh
index d631a3c366..4e1119f191 100644
--- a/cpp/src/randomforest/randomforest.cuh
+++ b/cpp/src/randomforest/randomforest.cuh
@@ -22,9 +22,9 @@
 
 #include <metrics/scores.cuh>
 
-#include <raft/common/nvtx.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/random/permute.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/nvtx.hpp>
+#include <raft/random/permute.cuh>
 #include <raft/random/rng.cuh>
 
 #include <thrust/execution_policy.h>
diff --git a/cpp/src/solver/cd.cuh b/cpp/src/solver/cd.cuh
index 73226ef7eb..e24eba3a2b 100644
--- a/cpp/src/solver/cd.cuh
+++ b/cpp/src/solver/cd.cuh
@@ -22,9 +22,9 @@
 #include <functions/penalty.cuh>
 #include <functions/softThres.cuh>
 #include <glm/preprocess.cuh>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/nvtx.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <raft/linalg/add.cuh>
 #include <raft/linalg/axpy.cuh>
 #include <raft/linalg/eltwise.cuh>
@@ -171,6 +171,7 @@ void cdFit(const raft::handle_t& handle,
     raft::linalg::multiplyScalar(
       sample_weight, sample_weight, (math_t)n_rows / h_sum_sw, n_rows, stream);
   }
+
   if (fit_intercept) {
     mu_input.resize(n_cols, stream);
     mu_labels.resize(1, stream);
diff --git a/cpp/src/solver/cd_mg.cu b/cpp/src/solver/cd_mg.cu
index 7300ac26fc..18516e57f5 100644
--- a/cpp/src/solver/cd_mg.cu
+++ b/cpp/src/solver/cd_mg.cu
@@ -25,16 +25,16 @@
 #include <cumlprims/opg/linalg/norm.hpp>
 
 #include "shuffle.h"
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/multiply.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/multiply.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
 
 #include <cstddef>
 
diff --git a/cpp/src/solver/lars_impl.cuh b/cpp/src/solver/lars_impl.cuh
index 7810eccc21..42767aa6c7 100644
--- a/cpp/src/solver/lars_impl.cuh
+++ b/cpp/src/solver/lars_impl.cuh
@@ -24,15 +24,15 @@
 #include <cache/cache_util.cuh>
 #include <cub/cub.cuh>
 #include <cuml/common/logger.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/cholesky_r1_update.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/cholesky_r1_update.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/gemv.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/gemv.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <thrust/copy.h>
diff --git a/cpp/src/solver/sgd.cuh b/cpp/src/solver/sgd.cuh
index d7236d2122..256c8d490d 100644
--- a/cpp/src/solver/sgd.cuh
+++ b/cpp/src/solver/sgd.cuh
@@ -23,18 +23,18 @@
 #include <functions/linearReg.cuh>
 #include <functions/logisticReg.cuh>
 #include <glm/preprocess.cuh>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemv.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemv.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace ML {
diff --git a/cpp/src/spectral/spectral.cu b/cpp/src/spectral/spectral.cu
index 7c05d01a94..8244f470a1 100644
--- a/cpp/src/spectral/spectral.cu
+++ b/cpp/src/spectral/spectral.cu
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #include <raft/sparse/coo.hpp>
 
-#include <raft/sparse/linalg/spectral.hpp>
+#include <raft/sparse/linalg/spectral.cuh>
 
 namespace raft {
 class handle_t;
diff --git a/cpp/src/svm/kernelcache.cuh b/cpp/src/svm/kernelcache.cuh
index b2900f4674..68f08d27a2 100644
--- a/cpp/src/svm/kernelcache.cuh
+++ b/cpp/src/svm/kernelcache.cuh
@@ -23,10 +23,10 @@
 #include <linalg/init.h>
 #include <matrix/grammatrix.cuh>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/gemm.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/gemm.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/svm/linear.cu b/cpp/src/svm/linear.cu
index 65a1fc6758..9a8c751b5b 100644
--- a/cpp/src/svm/linear.cu
+++ b/cpp/src/svm/linear.cu
@@ -24,15 +24,15 @@
 #include <label/classlabels.cuh>
 #include <matrix/kernelfactory.cuh>
 #include <omp.h>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/gemv.hpp>
-#include <raft/linalg/map.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/gemv.cuh>
+#include <raft/linalg/map.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_uvector.hpp>
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
diff --git a/cpp/src/svm/results.cuh b/cpp/src/svm/results.cuh
index 67a6f87f19..0d638d3f2a 100644
--- a/cpp/src/svm/results.cuh
+++ b/cpp/src/svm/results.cuh
@@ -25,11 +25,11 @@
 #include "ws_util.cuh"
 #include <cub/device/device_select.cuh>
 #include <linalg/init.h>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_uvector.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 
diff --git a/cpp/src/svm/smosolver.cuh b/cpp/src/svm/smosolver.cuh
index 0ee698a948..af8674ec08 100644
--- a/cpp/src/svm/smosolver.cuh
+++ b/cpp/src/svm/smosolver.cuh
@@ -24,13 +24,13 @@
 
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/gemv.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/gemv.cuh>
+#include <raft/linalg/unary_op.cuh>
 
 #include <iostream>
 #include <limits>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <string>
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
@@ -48,8 +48,8 @@
 #include <cuml/matrix/kernelparams.h>
 #include <matrix/grammatrix.cuh>
 #include <matrix/kernelfactory.cuh>
-#include <raft/linalg/gemv.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/gemv.cuh>
+#include <raft/linalg/unary_op.cuh>
 
 #include "results.cuh"
 
diff --git a/cpp/src/svm/svc.cu b/cpp/src/svm/svc.cu
index ea3d7032f6..95072d7c61 100644
--- a/cpp/src/svm/svc.cu
+++ b/cpp/src/svm/svc.cu
@@ -22,7 +22,7 @@
 #include <cuml/svm/svc.hpp>
 #include <label/classlabels.cuh>
 #include <matrix/kernelfactory.cuh>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace ML {
 namespace SVM {
diff --git a/cpp/src/svm/svc_impl.cuh b/cpp/src/svm/svc_impl.cuh
index 3aeeb60882..77751d9063 100644
--- a/cpp/src/svm/svc_impl.cuh
+++ b/cpp/src/svm/svc_impl.cuh
@@ -30,10 +30,10 @@
 #include <cuml/svm/svm_parameter.h>
 #include <label/classlabels.cuh>
 #include <matrix/kernelfactory.cuh>
-#include <raft/label/classlabels.hpp>
+#include <raft/label/classlabels.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_uvector.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 #include <thrust/copy.h>
diff --git a/cpp/src/svm/svr.cu b/cpp/src/svm/svr.cu
index 3aabe68b1f..a8ce1cc477 100644
--- a/cpp/src/svm/svr.cu
+++ b/cpp/src/svm/svr.cu
@@ -22,7 +22,7 @@
 #include <cuml/svm/svc.hpp>
 #include <label/classlabels.cuh>
 #include <matrix/kernelfactory.cuh>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace ML {
 namespace SVM {
diff --git a/cpp/src/svm/svr_impl.cuh b/cpp/src/svm/svr_impl.cuh
index f8ae2406c4..ed3f9d9815 100644
--- a/cpp/src/svm/svr_impl.cuh
+++ b/cpp/src/svm/svr_impl.cuh
@@ -30,8 +30,8 @@
 #include <cuml/svm/svm_parameter.h>
 #include <label/classlabels.cuh>
 #include <matrix/kernelfactory.cuh>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
 #include <thrust/iterator/counting_iterator.h>
diff --git a/cpp/src/svm/workingset.cuh b/cpp/src/svm/workingset.cuh
index 6240ea7594..f2a3569152 100644
--- a/cpp/src/svm/workingset.cuh
+++ b/cpp/src/svm/workingset.cuh
@@ -27,10 +27,10 @@
 #include "smo_sets.cuh"
 #include "ws_util.cuh"
 #include <cub/cub.cuh>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/handle.hpp>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <thrust/device_ptr.h>
diff --git a/cpp/src/tsa/auto_arima.cu b/cpp/src/tsa/auto_arima.cu
index 17cc2f1bd4..921f019bc9 100644
--- a/cpp/src/tsa/auto_arima.cu
+++ b/cpp/src/tsa/auto_arima.cu
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -13,7 +13,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 
 #include "auto_arima.cuh"
 
diff --git a/cpp/src/tsa/auto_arima.cuh b/cpp/src/tsa/auto_arima.cuh
index 6608a0678d..428c9b29e5 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/cudart_utils.h>
+#include <raft/core/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 a0bfccf2a6..f3831e3297 100644
--- a/cpp/src/tsne/barnes_hut_kernels.cuh
+++ b/cpp/src/tsne/barnes_hut_kernels.cuh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -34,7 +34,7 @@
 #define FACTOR7 1
 
 #include <float.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/device_atomics.cuh>
 
 namespace ML {
diff --git a/cpp/src/tsne/barnes_hut_tsne.cuh b/cpp/src/tsne/barnes_hut_tsne.cuh
index 1ca21d35ba..a666547eb3 100644
--- a/cpp/src/tsne/barnes_hut_tsne.cuh
+++ b/cpp/src/tsne/barnes_hut_tsne.cuh
@@ -19,8 +19,8 @@
 #include "utils.cuh"
 #include <cuml/common/logger.hpp>
 #include <cuml/manifold/tsne.h>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/eltwise.cuh>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.h>
diff --git a/cpp/src/tsne/distances.cuh b/cpp/src/tsne/distances.cuh
index 37f405ab61..1f68f50c25 100644
--- a/cpp/src/tsne/distances.cuh
+++ b/cpp/src/tsne/distances.cuh
@@ -17,17 +17,17 @@
 #pragma once
 
 #include <cuml/neighbors/knn_sparse.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/linalg/eltwise.cuh>
 #include <raft/sparse/coo.hpp>
-#include <raft/sparse/linalg/symmetrize.hpp>
-#include <raft/sparse/selection/knn.hpp>
+#include <raft/sparse/linalg/symmetrize.cuh>
+#include <raft/sparse/selection/knn.cuh>
 #include <selection/knn.cuh>
 
 #include <cuml/manifold/common.hpp>
 
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 
 #include "utils.cuh"
 
diff --git a/cpp/src/tsne/exact_kernels.cuh b/cpp/src/tsne/exact_kernels.cuh
index 81270a44d0..e0588247de 100644
--- a/cpp/src/tsne/exact_kernels.cuh
+++ b/cpp/src/tsne/exact_kernels.cuh
@@ -18,8 +18,8 @@
 
 #include <float.h>
 #include <math.h>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/eltwise.cuh>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.h>
diff --git a/cpp/src/tsne/exact_tsne.cuh b/cpp/src/tsne/exact_tsne.cuh
index 8f034d27d4..f219b2e242 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.h>
diff --git a/cpp/src/tsne/fft_tsne.cuh b/cpp/src/tsne/fft_tsne.cuh
index 0eacec6cf0..b47f060e58 100644
--- a/cpp/src/tsne/fft_tsne.cuh
+++ b/cpp/src/tsne/fft_tsne.cuh
@@ -29,8 +29,8 @@
 #include <common/device_utils.cuh>
 #include <cufft_utils.h>
 #include <linalg/init.h>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/stats/sum.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/tsne/tsne_runner.cuh b/cpp/src/tsne/tsne_runner.cuh
index 3135824f2b..678652ef46 100644
--- a/cpp/src/tsne/tsne_runner.cuh
+++ b/cpp/src/tsne/tsne_runner.cuh
@@ -20,7 +20,7 @@
 #include "utils.cuh"
 #include <cuml/common/logger.hpp>
 #include <cuml/manifold/common.hpp>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/distance/distance_type.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src/tsne/utils.cuh b/cpp/src/tsne/utils.cuh
index bcf1f1bcea..43368bb5fd 100644
--- a/cpp/src/tsne/utils.cuh
+++ b/cpp/src/tsne/utils.cuh
@@ -22,8 +22,8 @@
 #include <stdio.h>
 
 #include <cuml/common/logger.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/norm.hpp>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/norm.cuh>
 
 #include <cuda_runtime.h>
 
@@ -31,8 +31,8 @@
 #include <thrust/reduce.h>
 #include <thrust/transform.h>
 
-#include <raft/random/rng.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/random/rng.cuh>
+#include <raft/stats/sum.cuh>
 #include <sys/time.h>
 
 #include <chrono>
diff --git a/cpp/src/tsvd/tsvd.cuh b/cpp/src/tsvd/tsvd.cuh
index ea77c74391..7d6929bd62 100644
--- a/cpp/src/tsvd/tsvd.cuh
+++ b/cpp/src/tsvd/tsvd.cuh
@@ -17,19 +17,19 @@
 #pragma once
 
 #include <cuml/decomposition/params.hpp>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eig.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemm.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eig.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemm.cuh>
 #include <raft/linalg/rsvd.cuh>
-#include <raft/linalg/transpose.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/stddev.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/transpose.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/stddev.cuh>
+#include <raft/stats/sum.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <rmm/exec_policy.hpp>
diff --git a/cpp/src/tsvd/tsvd_mg.cu b/cpp/src/tsvd/tsvd_mg.cu
index 09b3999f81..c4cd9c6825 100644
--- a/cpp/src/tsvd/tsvd_mg.cu
+++ b/cpp/src/tsvd/tsvd_mg.cu
@@ -25,12 +25,12 @@
 #include <cumlprims/opg/stats/mean_center.hpp>
 #include <cumlprims/opg/stats/stddev.hpp>
 
-#include <raft/comms/comms.hpp>
+#include <raft/core/comms.hpp>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/stats/mean_center.hpp>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/stats/mean_center.cuh>
 
 #include <cstddef>
 
diff --git a/cpp/src/umap/fuzzy_simpl_set/naive.cuh b/cpp/src/umap/fuzzy_simpl_set/naive.cuh
index 3b7d7080ed..f4c24e5567 100644
--- a/cpp/src/umap/fuzzy_simpl_set/naive.cuh
+++ b/cpp/src/umap/fuzzy_simpl_set/naive.cuh
@@ -20,13 +20,13 @@
 #include <cuml/manifold/umapparams.h>
 #include <cuml/neighbors/knn.hpp>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
 #include <raft/sparse/coo.hpp>
-#include <raft/sparse/linalg/symmetrize.hpp>
-#include <raft/sparse/op/sort.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/sparse/linalg/symmetrize.cuh>
+#include <raft/sparse/op/sort.cuh>
+#include <raft/stats/mean.cuh>
 
 #include <cuda_runtime.h>
 
diff --git a/cpp/src/umap/init_embed/random_algo.cuh b/cpp/src/umap/init_embed/random_algo.cuh
index 20c5382fe7..340280bd80 100644
--- a/cpp/src/umap/init_embed/random_algo.cuh
+++ b/cpp/src/umap/init_embed/random_algo.cuh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <cuml/manifold/umapparams.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 namespace UMAPAlgo {
 namespace InitEmbed {
diff --git a/cpp/src/umap/init_embed/spectral_algo.cuh b/cpp/src/umap/init_embed/spectral_algo.cuh
index 2b191aa9a2..d8e54e80c1 100644
--- a/cpp/src/umap/init_embed/spectral_algo.cuh
+++ b/cpp/src/umap/init_embed/spectral_algo.cuh
@@ -19,9 +19,9 @@
 #include <cuml/cluster/spectral.hpp>
 #include <cuml/manifold/umapparams.h>
 
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/random/rng.cuh>
 #include <raft/sparse/coo.hpp>
 
 #include <thrust/device_ptr.h>
diff --git a/cpp/src/umap/knn_graph/algo.cuh b/cpp/src/umap/knn_graph/algo.cuh
index 8b2743a148..c027684bf4 100644
--- a/cpp/src/umap/knn_graph/algo.cuh
+++ b/cpp/src/umap/knn_graph/algo.cuh
@@ -21,13 +21,13 @@
 #include <cuml/neighbors/knn_sparse.hpp>
 #include <iostream>
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/sparse/selection/knn.hpp>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/sparse/selection/knn.cuh>
 #include <selection/knn.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 
 namespace UMAPAlgo {
 namespace kNNGraph {
diff --git a/cpp/src/umap/optimize.cuh b/cpp/src/umap/optimize.cuh
index 438d58d591..5ccb12369c 100644
--- a/cpp/src/umap/optimize.cuh
+++ b/cpp/src/umap/optimize.cuh
@@ -19,14 +19,14 @@
 #include <cuml/common/logger.hpp>
 #include <cuml/manifold/umapparams.h>
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/multiply.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/multiply.cuh>
 #include <raft/linalg/power.cuh>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/stats/mean.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <cuda_runtime.h>
diff --git a/cpp/src/umap/runner.cuh b/cpp/src/umap/runner.cuh
index 3a89f58613..16e5fb0081 100644
--- a/cpp/src/umap/runner.cuh
+++ b/cpp/src/umap/runner.cuh
@@ -37,17 +37,17 @@
 #include <thrust/scan.h>
 #include <thrust/system/cuda/execution_policy.h>
 
-#include <raft/sparse/convert/csr.hpp>
+#include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/coo.hpp>
-#include <raft/sparse/linalg/norm.hpp>
-#include <raft/sparse/op/filter.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/linalg/norm.cuh>
+#include <raft/sparse/op/filter.cuh>
+#include <raft/sparse/op/sort.cuh>
 
 #include <raft/cuda_utils.cuh>
 
 #include <common/nvtx.hpp>
 #include <cuda_runtime.h>
-#include <raft/common/nvtx.hpp>
+#include <raft/core/nvtx.hpp>
 
 namespace UMAPAlgo {
 
diff --git a/cpp/src/umap/simpl_set_embed/algo.cuh b/cpp/src/umap/simpl_set_embed/algo.cuh
index 2241ce26f5..15b6bff14d 100644
--- a/cpp/src/umap/simpl_set_embed/algo.cuh
+++ b/cpp/src/umap/simpl_set_embed/algo.cuh
@@ -28,15 +28,15 @@
 #include <common/fast_int_div.cuh>
 #include <cstdlib>
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/unary_op.cuh>
 #include <raft/sparse/coo.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include "optimize_batch_kernel.cuh"
 #include <string>
 
-#include <raft/sparse/op/filter.hpp>
+#include <raft/sparse/op/filter.cuh>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.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 a7b8076299..ff38858b52 100644
--- a/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh
+++ b/cpp/src/umap/simpl_set_embed/optimize_batch_kernel.cuh
@@ -20,9 +20,9 @@
 
 #include <cuml/manifold/umapparams.h>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 #include <cstddef>
 
diff --git a/cpp/src/umap/supervised.cuh b/cpp/src/umap/supervised.cuh
index fb1fc32e9c..3513a737b0 100644
--- a/cpp/src/umap/supervised.cuh
+++ b/cpp/src/umap/supervised.cuh
@@ -25,7 +25,7 @@
 #include "init_embed/runner.cuh"
 #include "knn_graph/runner.cuh"
 #include "simpl_set_embed/runner.cuh"
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <thrust/count.h>
 #include <thrust/device_ptr.h>
@@ -35,12 +35,12 @@
 #include <thrust/scan.h>
 #include <thrust/system/cuda/execution_policy.h>
 
-#include <raft/sparse/convert/csr.hpp>
+#include <raft/sparse/convert/csr.cuh>
 #include <raft/sparse/coo.hpp>
-#include <raft/sparse/linalg/add.hpp>
-#include <raft/sparse/linalg/norm.hpp>
-#include <raft/sparse/linalg/symmetrize.hpp>
-#include <raft/sparse/op/filter.hpp>
+#include <raft/sparse/linalg/add.cuh>
+#include <raft/sparse/linalg/norm.cuh>
+#include <raft/sparse/linalg/symmetrize.cuh>
+#include <raft/sparse/op/filter.cuh>
 
 #include <raft/cuda_utils.cuh>
 
diff --git a/cpp/src_prims/cache/cache.cuh b/cpp/src_prims/cache/cache.cuh
index 788f08431f..958835be8d 100644
--- a/cpp/src_prims/cache/cache.cuh
+++ b/cpp/src_prims/cache/cache.cuh
@@ -21,9 +21,9 @@
 
 #include <cuml/common/logger.hpp>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/cufft_utils.h b/cpp/src_prims/cufft_utils.h
index c87e8ac65b..d46ab4fa72 100644
--- a/cpp/src_prims/cufft_utils.h
+++ b/cpp/src_prims/cufft_utils.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <cufft.h>
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 
 // TODO move to raft https://github.com/rapidsai/raft/issues/91
 namespace raft {
diff --git a/cpp/src_prims/functions/hinge.cuh b/cpp/src_prims/functions/hinge.cuh
index f6ced59818..1c194e6f65 100644
--- a/cpp/src_prims/functions/hinge.cuh
+++ b/cpp/src_prims/functions/hinge.cuh
@@ -18,17 +18,17 @@
 
 #include "penalty.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/sum.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 5d0fa4210c..cbc9a6e980 100644
--- a/cpp/src_prims/functions/linearReg.cuh
+++ b/cpp/src_prims/functions/linearReg.cuh
@@ -18,15 +18,15 @@
 
 #include "penalty.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/sum.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/functions/log.cuh b/cpp/src_prims/functions/log.cuh
index 328276bc32..4dbaf7b218 100644
--- a/cpp/src_prims/functions/log.cuh
+++ b/cpp/src_prims/functions/log.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/src_prims/functions/logisticReg.cuh b/cpp/src_prims/functions/logisticReg.cuh
index af5489d017..6cb8c86bd3 100644
--- a/cpp/src_prims/functions/logisticReg.cuh
+++ b/cpp/src_prims/functions/logisticReg.cuh
@@ -19,15 +19,15 @@
 #include "penalty.cuh"
 #include "sigmoid.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/gemm.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
-#include <raft/stats/mean.hpp>
-#include <raft/stats/sum.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/gemm.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/sum.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/functions/penalty.cuh b/cpp/src_prims/functions/penalty.cuh
index 247bae10ba..22c3d66c42 100644
--- a/cpp/src_prims/functions/penalty.cuh
+++ b/cpp/src_prims/functions/penalty.cuh
@@ -17,13 +17,13 @@
 #pragma once
 
 #include "sign.cuh"
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/matrix/math.hpp>
-#include <raft/matrix/matrix.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/matrix/math.cuh>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/functions/sigmoid.cuh b/cpp/src_prims/functions/sigmoid.cuh
index 537d171b18..9d0177660a 100644
--- a/cpp/src_prims/functions/sigmoid.cuh
+++ b/cpp/src_prims/functions/sigmoid.cuh
@@ -17,7 +17,7 @@
 #pragma once
 
 #include <raft/cuda_utils.cuh>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/src_prims/functions/sign.cuh b/cpp/src_prims/functions/sign.cuh
index 85d18787da..0b1d990fb9 100644
--- a/cpp/src_prims/functions/sign.cuh
+++ b/cpp/src_prims/functions/sign.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/src_prims/functions/softThres.cuh b/cpp/src_prims/functions/softThres.cuh
index 05bf651a26..8088c2f042 100644
--- a/cpp/src_prims/functions/softThres.cuh
+++ b/cpp/src_prims/functions/softThres.cuh
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/src_prims/label/classlabels.cuh b/cpp/src_prims/label/classlabels.cuh
index 278c3685b6..ecb4e55871 100644
--- a/cpp/src_prims/label/classlabels.cuh
+++ b/cpp/src_prims/label/classlabels.cuh
@@ -18,11 +18,11 @@
 
 #include <cub/cub.cuh>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/label/classlabels.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/label/classlabels.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/label/merge_labels.cuh b/cpp/src_prims/label/merge_labels.cuh
index 4cff001a2c..d9efc3729e 100644
--- a/cpp/src_prims/label/merge_labels.cuh
+++ b/cpp/src_prims/label/merge_labels.cuh
@@ -20,9 +20,9 @@
 #include <math.h>
 
 #include <linalg/init.h>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
 
 namespace MLCommon {
 namespace Label {
diff --git a/cpp/src_prims/linalg/batched/matrix.cuh b/cpp/src_prims/linalg/batched/matrix.cuh
index f3d5627a12..b6a108c109 100644
--- a/cpp/src_prims/linalg/batched/matrix.cuh
+++ b/cpp/src_prims/linalg/batched/matrix.cuh
@@ -20,12 +20,12 @@
 
 #include <common/fast_int_div.cuh>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
+#include <raft/linalg/add.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/execution_policy.h>
diff --git a/cpp/src_prims/linalg/block.cuh b/cpp/src_prims/linalg/block.cuh
index ef23f506dc..d401b989e5 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/cudart_utils.h>
 
 // Anonymous namespace for internal auxiliary functions
 namespace {
diff --git a/cpp/src_prims/linalg/init.h b/cpp/src_prims/linalg/init.h
index 4dfb930880..c3dadd2895 100644
--- a/cpp/src_prims/linalg/init.h
+++ b/cpp/src_prims/linalg/init.h
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
diff --git a/cpp/src_prims/matrix/grammatrix.cuh b/cpp/src_prims/matrix/grammatrix.cuh
index bd66f815e2..80397307e6 100644
--- a/cpp/src_prims/matrix/grammatrix.cuh
+++ b/cpp/src_prims/matrix/grammatrix.cuh
@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cublas_wrappers.hpp>
-#include <raft/linalg/gemm.hpp>
+#include <raft/linalg/gemm.cuh>
 
 namespace MLCommon {
 namespace Matrix {
diff --git a/cpp/src_prims/matrix/kernelfactory.cuh b/cpp/src_prims/matrix/kernelfactory.cuh
index 78e5b9e4d8..beaeae7a84 100644
--- a/cpp/src_prims/matrix/kernelfactory.cuh
+++ b/cpp/src_prims/matrix/kernelfactory.cuh
@@ -19,7 +19,7 @@
 #include "grammatrix.cuh"
 #include "kernelmatrices.cuh"
 #include <cuml/matrix/kernelparams.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 namespace MLCommon {
 namespace Matrix {
diff --git a/cpp/src_prims/matrix/kernelmatrices.cuh b/cpp/src_prims/matrix/kernelmatrices.cuh
index 6e40dc243a..010b6ccf50 100644
--- a/cpp/src_prims/matrix/kernelmatrices.cuh
+++ b/cpp/src_prims/matrix/kernelmatrices.cuh
@@ -18,8 +18,8 @@
 
 #include "grammatrix.cuh"
 #include <raft/cuda_utils.cuh>
-#include <raft/distance/distance.hpp>
-#include <raft/linalg/gemm.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/linalg/gemm.cuh>
 
 namespace MLCommon {
 namespace Matrix {
diff --git a/cpp/src_prims/metrics/adjusted_rand_index.cuh b/cpp/src_prims/metrics/adjusted_rand_index.cuh
index edc0943dde..6e4e976632 100644
--- a/cpp/src_prims/metrics/adjusted_rand_index.cuh
+++ b/cpp/src_prims/metrics/adjusted_rand_index.cuh
@@ -25,11 +25,11 @@
 #include "contingencyMatrix.cuh"
 #include <cub/cub.cuh>
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/reduce.hpp>
-#include <raft/stats/histogram.hpp>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/reduce.cuh>
+#include <raft/stats/histogram.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/metrics/batched/information_criterion.cuh b/cpp/src_prims/metrics/batched/information_criterion.cuh
index 8770246f07..d1907538eb 100644
--- a/cpp/src_prims/metrics/batched/information_criterion.cuh
+++ b/cpp/src_prims/metrics/batched/information_criterion.cuh
@@ -25,7 +25,7 @@
  *  - BIC: https://en.wikipedia.org/wiki/Bayesian_information_criterion
  */
 
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 #include <cmath>
 
diff --git a/cpp/src_prims/metrics/contingencyMatrix.cuh b/cpp/src_prims/metrics/contingencyMatrix.cuh
index ca621c8ca8..dbb9ba5594 100644
--- a/cpp/src_prims/metrics/contingencyMatrix.cuh
+++ b/cpp/src_prims/metrics/contingencyMatrix.cuh
@@ -16,8 +16,8 @@
 
 #pragma once
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 
 #include <thrust/device_ptr.h>
 #include <thrust/execution_policy.h>
diff --git a/cpp/src_prims/metrics/dispersion.cuh b/cpp/src_prims/metrics/dispersion.cuh
index b2d3c007fb..701329db6f 100644
--- a/cpp/src_prims/metrics/dispersion.cuh
+++ b/cpp/src_prims/metrics/dispersion.cuh
@@ -18,10 +18,10 @@
 
 #include <cub/cub.cuh>
 #include <memory>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/linalg/eltwise.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/metrics/entropy.cuh b/cpp/src_prims/metrics/entropy.cuh
index 55650a3345..171cf42967 100644
--- a/cpp/src_prims/metrics/entropy.cuh
+++ b/cpp/src_prims/metrics/entropy.cuh
@@ -21,10 +21,10 @@
 
 #include <cub/cub.cuh>
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/divide.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
+#include <raft/linalg/divide.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/metrics/kl_divergence.cuh b/cpp/src_prims/metrics/kl_divergence.cuh
index bce3bf7283..8dee48d8fd 100644
--- a/cpp/src_prims/metrics/kl_divergence.cuh
+++ b/cpp/src_prims/metrics/kl_divergence.cuh
@@ -22,9 +22,9 @@
 #pragma once
 
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/map_then_reduce.hpp>
+#include <raft/linalg/map_then_reduce.cuh>
 #include <rmm/device_scalar.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/metrics/mutual_info_score.cuh b/cpp/src_prims/metrics/mutual_info_score.cuh
index f20de778e4..849797bee4 100644
--- a/cpp/src_prims/metrics/mutual_info_score.cuh
+++ b/cpp/src_prims/metrics/mutual_info_score.cuh
@@ -27,9 +27,9 @@
 #include "contingencyMatrix.cuh"
 #include <cub/cub.cuh>
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/reduce.hpp>
+#include <raft/linalg/reduce.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/metrics/rand_index.cuh b/cpp/src_prims/metrics/rand_index.cuh
index f1acf30ac5..7a337fe2b0 100644
--- a/cpp/src_prims/metrics/rand_index.cuh
+++ b/cpp/src_prims/metrics/rand_index.cuh
@@ -54,8 +54,8 @@
 
 #include <cub/cub.cuh>
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/metrics/scores.cuh b/cpp/src_prims/metrics/scores.cuh
index 782e413726..75790debbf 100644
--- a/cpp/src_prims/metrics/scores.cuh
+++ b/cpp/src_prims/metrics/scores.cuh
@@ -16,13 +16,13 @@
 
 #pragma once
 
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
-#include <raft/linalg/eltwise.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/linalg/eltwise.cuh>
 #include <raft/linalg/power.cuh>
-#include <raft/linalg/subtract.hpp>
-#include <raft/spatial/knn/knn.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/linalg/subtract.cuh>
+#include <raft/spatial/knn/knn.cuh>
+#include <raft/stats/mean.cuh>
 
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
diff --git a/cpp/src_prims/metrics/silhouette_score.cuh b/cpp/src_prims/metrics/silhouette_score.cuh
index fa23e85722..3f47746de8 100644
--- a/cpp/src_prims/metrics/silhouette_score.cuh
+++ b/cpp/src_prims/metrics/silhouette_score.cuh
@@ -23,13 +23,13 @@
 #include <math.h>
 #include <numeric>
 #include <raft/cuda_utils.cuh>
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/eltwise.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/reduce.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/eltwise.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/reduce.cuh>
 #include <raft/linalg/reduce_cols_by_key.cuh>
 #include <rmm/device_scalar.hpp>
 
diff --git a/cpp/src_prims/metrics/trustworthiness_score.cuh b/cpp/src_prims/metrics/trustworthiness_score.cuh
index 8a0a08213f..4d7f3d2747 100644
--- a/cpp/src_prims/metrics/trustworthiness_score.cuh
+++ b/cpp/src_prims/metrics/trustworthiness_score.cuh
@@ -15,9 +15,9 @@
  */
 
 #include <cuml/metrics/metrics.hpp>
-#include <raft/distance/specializations.hpp>
-#include <raft/spatial/knn/knn.hpp>
-#include <raft/spatial/knn/specializations.hpp>
+#include <raft/distance/specializations.cuh>
+#include <raft/spatial/knn/knn.cuh>
+#include <raft/spatial/knn/specializations.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <selection/columnWiseSort.cuh>
diff --git a/cpp/src_prims/random/make_arima.cuh b/cpp/src_prims/random/make_arima.cuh
index ba961df158..bec189cfec 100644
--- a/cpp/src_prims/random/make_arima.cuh
+++ b/cpp/src_prims/random/make_arima.cuh
@@ -23,7 +23,7 @@
 #include <thrust/iterator/counting_iterator.h>
 
 #include <cuml/tsa/arima_common.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <timeSeries/arima_helpers.cuh>
 
 namespace MLCommon {
diff --git a/cpp/src_prims/selection/knn.cuh b/cpp/src_prims/selection/knn.cuh
index 4568c3a781..bfc25bb68d 100644
--- a/cpp/src_prims/selection/knn.cuh
+++ b/cpp/src_prims/selection/knn.cuh
@@ -23,9 +23,9 @@
 
 #include <cuml/neighbors/knn.hpp>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 #include <raft/distance/distance_type.hpp>
 
 #include <faiss/gpu/GpuDistance.h>
diff --git a/cpp/src_prims/selection/processing.cuh b/cpp/src_prims/selection/processing.cuh
index b559efda45..203f44666c 100644
--- a/cpp/src_prims/selection/processing.cuh
+++ b/cpp/src_prims/selection/processing.cuh
@@ -17,12 +17,12 @@
 
 #include <cuml/neighbors/knn.hpp>
 
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/norm.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/norm.cuh>
+#include <raft/linalg/unary_op.cuh>
 
-#include <raft/stats/mean.hpp>
-#include <raft/stats/mean_center.hpp>
+#include <raft/stats/mean.cuh>
+#include <raft/stats/mean_center.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/src_prims/sparse/batched/csr.cuh b/cpp/src_prims/sparse/batched/csr.cuh
index 1cdcf40e9f..c299c35884 100644
--- a/cpp/src_prims/sparse/batched/csr.cuh
+++ b/cpp/src_prims/sparse/batched/csr.cuh
@@ -29,8 +29,8 @@
 #include <cuml/common/utils.hpp>
 
 #include <linalg/batched/matrix.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/matrix/matrix.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/execution_policy.h>
diff --git a/cpp/src_prims/timeSeries/arima_helpers.cuh b/cpp/src_prims/timeSeries/arima_helpers.cuh
index ecdf32ca74..b454db9b11 100644
--- a/cpp/src_prims/timeSeries/arima_helpers.cuh
+++ b/cpp/src_prims/timeSeries/arima_helpers.cuh
@@ -21,10 +21,10 @@
 #include "jones_transform.cuh"
 #include <cuml/tsa/arima_common.h>
 #include <linalg/batched/matrix.cuh>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/unary_op.cuh>
 
 // Private helper functions and kernels in the anonymous namespace
 namespace {
diff --git a/cpp/src_prims/timeSeries/fillna.cuh b/cpp/src_prims/timeSeries/fillna.cuh
index 89df6f355a..6b6fd61c2d 100644
--- a/cpp/src_prims/timeSeries/fillna.cuh
+++ b/cpp/src_prims/timeSeries/fillna.cuh
@@ -22,10 +22,10 @@
 #include "jones_transform.cuh"
 #include <cuml/tsa/arima_common.h>
 #include <linalg/batched/matrix.cuh>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/unary_op.cuh>
 #include <rmm/device_uvector.hpp>
 
 // Auxiliary functions in anonymous namespace
diff --git a/cpp/src_prims/timeSeries/jones_transform.cuh b/cpp/src_prims/timeSeries/jones_transform.cuh
index 2b50852d40..ccfd4b97b5 100644
--- a/cpp/src_prims/timeSeries/jones_transform.cuh
+++ b/cpp/src_prims/timeSeries/jones_transform.cuh
@@ -22,9 +22,9 @@
 #pragma once
 
 #include <math.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/unary_op.hpp>
+#include <raft/linalg/unary_op.cuh>
 
 namespace MLCommon {
 
diff --git a/cpp/src_prims/timeSeries/stationarity.cuh b/cpp/src_prims/timeSeries/stationarity.cuh
index d06069c099..b90bd675c8 100644
--- a/cpp/src_prims/timeSeries/stationarity.cuh
+++ b/cpp/src_prims/timeSeries/stationarity.cuh
@@ -27,10 +27,10 @@
 
 #include "arima_helpers.cuh"
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/matrix_vector_op.hpp>
-#include <raft/linalg/reduce.hpp>
-#include <raft/stats/mean.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/matrix_vector_op.cuh>
+#include <raft/linalg/reduce.cuh>
+#include <raft/stats/mean.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/test/mg/knn.cu b/cpp/test/mg/knn.cu
index a5a6a9ea27..30b947e7e9 100644
--- a/cpp/test/mg/knn.cu
+++ b/cpp/test/mg/knn.cu
@@ -19,7 +19,7 @@
 #include <cuml/neighbors/knn_mg.hpp>
 #include <gtest/gtest.h>
 #include <memory>
-#include <raft/random/make_blobs.hpp>
+#include <raft/random/make_blobs.cuh>
 
 #include <raft/comms/mpi_comms.hpp>
 
diff --git a/cpp/test/mg/knn_test_helper.cuh b/cpp/test/mg/knn_test_helper.cuh
index 9beab8b03d..26fddf4d80 100644
--- a/cpp/test/mg/knn_test_helper.cuh
+++ b/cpp/test/mg/knn_test_helper.cuh
@@ -19,7 +19,7 @@
 #include <cuml/neighbors/knn_mg.hpp>
 #include <gtest/gtest.h>
 #include <memory>
-#include <raft/random/make_blobs.hpp>
+#include <raft/random/make_blobs.cuh>
 
 #include <raft/comms/mpi_comms.hpp>
 
diff --git a/cpp/test/mg/pca.cu b/cpp/test/mg/pca.cu
index 2a17cdc0f8..1d6e8ba35c 100644
--- a/cpp/test/mg/pca.cu
+++ b/cpp/test/mg/pca.cu
@@ -20,9 +20,9 @@
 #include <cumlprims/opg/linalg/gemm.hpp>
 #include <cumlprims/opg/matrix/matrix_utils.hpp>
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/matrix/matrix.hpp>
+#include <raft/matrix/matrix.cuh>
 #include <test_utils.h>
 
 #include <raft/comms/mpi_comms.hpp>
diff --git a/cpp/test/prims/add_sub_dev_scalar.cu b/cpp/test/prims/add_sub_dev_scalar.cu
index c2fd3b7730..953bb6604b 100644
--- a/cpp/test/prims/add_sub_dev_scalar.cu
+++ b/cpp/test/prims/add_sub_dev_scalar.cu
@@ -16,11 +16,11 @@
 
 #include "test_utils.h"
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/subtract.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/subtract.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/random/rng.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/test/prims/adjusted_rand_index.cu b/cpp/test/prims/adjusted_rand_index.cu
index 4958ce5d00..66d6cdbb98 100644
--- a/cpp/test/prims/adjusted_rand_index.cu
+++ b/cpp/test/prims/adjusted_rand_index.cu
@@ -20,7 +20,7 @@
 #include <iostream>
 #include <metrics/adjusted_rand_index.cuh>
 #include <metrics/contingencyMatrix.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/batched/csr.cu b/cpp/test/prims/batched/csr.cu
index 8968d9c4d6..57da86a7fa 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <gtest/gtest.h>
 
diff --git a/cpp/test/prims/batched/gemv.cu b/cpp/test/prims/batched/gemv.cu
index edb83c328e..8a78f59ebe 100644
--- a/cpp/test/prims/batched/gemv.cu
+++ b/cpp/test/prims/batched/gemv.cu
@@ -17,8 +17,8 @@
 #include "../test_utils.h"
 #include <gtest/gtest.h>
 #include <linalg/batched/gemv.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 #include <test_utils.h>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/batched/information_criterion.cu b/cpp/test/prims/batched/information_criterion.cu
index 177693889a..1bf1b8ade6 100644
--- a/cpp/test/prims/batched/information_criterion.cu
+++ b/cpp/test/prims/batched/information_criterion.cu
@@ -18,7 +18,7 @@
 
 #include <metrics/batched/information_criterion.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <gtest/gtest.h>
diff --git a/cpp/test/prims/batched/make_symm.cu b/cpp/test/prims/batched/make_symm.cu
index 677977d06e..7c00db3a62 100644
--- a/cpp/test/prims/batched/make_symm.cu
+++ b/cpp/test/prims/batched/make_symm.cu
@@ -17,8 +17,8 @@
 #include "../test_utils.h"
 #include <gtest/gtest.h>
 #include <linalg/batched/make_symm.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 
diff --git a/cpp/test/prims/batched/matrix.cu b/cpp/test/prims/batched/matrix.cu
index b4e7dec57e..2ab5514531 100644
--- a/cpp/test/prims/batched/matrix.cu
+++ b/cpp/test/prims/batched/matrix.cu
@@ -19,8 +19,8 @@
 
 #include <linalg/batched/matrix.cuh>
 
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/linalg/add.cuh>
 
 #include <gtest/gtest.h>
 
diff --git a/cpp/test/prims/cache.cu b/cpp/test/prims/cache.cu
index 6c41469fc5..31b1603c73 100644
--- a/cpp/test/prims/cache.cu
+++ b/cpp/test/prims/cache.cu
@@ -18,9 +18,9 @@
 #include <cache/cache.cuh>
 #include <gtest/gtest.h>
 #include <iostream>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/columnSort.cu b/cpp/test/prims/columnSort.cu
index 8c7ba8aa97..56dc0355b9 100644
--- a/cpp/test/prims/columnSort.cu
+++ b/cpp/test/prims/columnSort.cu
@@ -18,7 +18,7 @@
 #include <algorithm>
 #include <gtest/gtest.h>
 #include <numeric>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 #include <selection/columnWiseSort.cuh>
 
diff --git a/cpp/test/prims/completeness_score.cu b/cpp/test/prims/completeness_score.cu
index aa32114082..e3942ac4fc 100644
--- a/cpp/test/prims/completeness_score.cu
+++ b/cpp/test/prims/completeness_score.cu
@@ -18,7 +18,7 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/completeness_score.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/contingencyMatrix.cu b/cpp/test/prims/contingencyMatrix.cu
index bbdb309682..0e4c8ff56c 100644
--- a/cpp/test/prims/contingencyMatrix.cu
+++ b/cpp/test/prims/contingencyMatrix.cu
@@ -19,7 +19,7 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/contingencyMatrix.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/test/prims/decoupled_lookback.cu b/cpp/test/prims/decoupled_lookback.cu
index 1da43a3f4b..7a7911ffc4 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/cudart_utils.h>
-#include <raft/interruptible.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.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 abecf4c36f..25369171a4 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/cudart_utils.h>
-#include <raft/interruptible.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/dispersion.cu b/cpp/test/prims/dispersion.cu
index 78c0746ee4..6a3ed52564 100644
--- a/cpp/test/prims/dispersion.cu
+++ b/cpp/test/prims/dispersion.cu
@@ -17,9 +17,9 @@
 #include "test_utils.h"
 #include <gtest/gtest.h>
 #include <metrics/dispersion.cuh>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/interruptible.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <stdio.h>
 #include <stdlib.h>
diff --git a/cpp/test/prims/dist_adj.cu b/cpp/test/prims/dist_adj.cu
index b423f1568b..5dc05bb6be 100644
--- a/cpp/test/prims/dist_adj.cu
+++ b/cpp/test/prims/dist_adj.cu
@@ -17,9 +17,9 @@
 #include "test_utils.h"
 #include <distance/distance.cuh>
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Distance {
diff --git a/cpp/test/prims/distance_base.cuh b/cpp/test/prims/distance_base.cuh
index c76671a5f6..81482e6567 100644
--- a/cpp/test/prims/distance_base.cuh
+++ b/cpp/test/prims/distance_base.cuh
@@ -17,9 +17,9 @@
 #include "test_utils.h"
 #include <distance/distance.cuh>
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Distance {
diff --git a/cpp/test/prims/eltwise2d.cu b/cpp/test/prims/eltwise2d.cu
index bc5e094503..1e34940355 100644
--- a/cpp/test/prims/eltwise2d.cu
+++ b/cpp/test/prims/eltwise2d.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <gtest/gtest.h>
 #include <linalg/eltwise2d.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace LinAlg {
diff --git a/cpp/test/prims/entropy.cu b/cpp/test/prims/entropy.cu
index 8afcfa60ef..466f0fc903 100644
--- a/cpp/test/prims/entropy.cu
+++ b/cpp/test/prims/entropy.cu
@@ -18,8 +18,8 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/entropy.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <random>
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/test/prims/fast_int_div.cu b/cpp/test/prims/fast_int_div.cu
index 19fb3f52b1..f4e742018a 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/cudart_utils.h>
+#include <raft/core/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 f711d34328..ed791c8650 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/cudart_utils.h>
-#include <raft/handle.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include "test_utils.h"
diff --git a/cpp/test/prims/gather.cu b/cpp/test/prims/gather.cu
index 45884cc8c7..6c0e223b9c 100644
--- a/cpp/test/prims/gather.cu
+++ b/cpp/test/prims/gather.cu
@@ -17,10 +17,10 @@
 #include "test_utils.h"
 #include <gtest/gtest.h>
 #include <matrix/gather.cuh>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/gram.cu b/cpp/test/prims/gram.cu
index 7f97b2fe8d..f192285335 100644
--- a/cpp/test/prims/gram.cu
+++ b/cpp/test/prims/gram.cu
@@ -21,9 +21,9 @@
 #include <matrix/grammatrix.cuh>
 #include <matrix/kernelfactory.cuh>
 #include <memory>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/grid_sync.cu b/cpp/test/prims/grid_sync.cu
index f0b03cb405..c7f6fdb9da 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/hinge.cu b/cpp/test/prims/hinge.cu
index 89a061487a..fb2ae26663 100644
--- a/cpp/test/prims/hinge.cu
+++ b/cpp/test/prims/hinge.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <functions/hinge.cuh>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/test/prims/homogeneity_score.cu b/cpp/test/prims/homogeneity_score.cu
index e9c7cd5fbb..5a44fa3fe8 100644
--- a/cpp/test/prims/homogeneity_score.cu
+++ b/cpp/test/prims/homogeneity_score.cu
@@ -18,7 +18,7 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/homogeneity_score.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/jones_transform.cu b/cpp/test/prims/jones_transform.cu
index a2ec674bf1..7a992fd32f 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <random>
 #include <rmm/device_uvector.hpp>
 #include <timeSeries/jones_transform.cuh>
diff --git a/cpp/test/prims/kl_divergence.cu b/cpp/test/prims/kl_divergence.cu
index 9e7991e4e0..93c5f9a2aa 100644
--- a/cpp/test/prims/kl_divergence.cu
+++ b/cpp/test/prims/kl_divergence.cu
@@ -18,7 +18,7 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/kl_divergence.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/knn_classify.cu b/cpp/test/prims/knn_classify.cu
index 666ee8a712..124c82bdb0 100644
--- a/cpp/test/prims/knn_classify.cu
+++ b/cpp/test/prims/knn_classify.cu
@@ -17,11 +17,11 @@
 #include "test_utils.h"
 #include <gtest/gtest.h>
 #include <iostream>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/label/classlabels.hpp>
-#include <raft/random/make_blobs.hpp>
-#include <raft/spatial/knn/knn.hpp>
+#include <raft/label/classlabels.cuh>
+#include <raft/random/make_blobs.cuh>
+#include <raft/spatial/knn/knn.cuh>
 #include <rmm/device_uvector.hpp>
 #include <selection/knn.cuh>
 #include <vector>
diff --git a/cpp/test/prims/knn_regression.cu b/cpp/test/prims/knn_regression.cu
index 34090db674..aaabfd3a39 100644
--- a/cpp/test/prims/knn_regression.cu
+++ b/cpp/test/prims/knn_regression.cu
@@ -20,11 +20,11 @@
 
 #include <label/classlabels.cuh>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/reduce.hpp>
-#include <raft/random/rng.hpp>
-#include <raft/spatial/knn/knn.hpp>
+#include <raft/linalg/reduce.cuh>
+#include <raft/random/rng.cuh>
+#include <raft/spatial/knn/knn.cuh>
 
 #include <rmm/device_uvector.hpp>
 
diff --git a/cpp/test/prims/kselection.cu b/cpp/test/prims/kselection.cu
index dd2e641b74..b3e38c4030 100644
--- a/cpp/test/prims/kselection.cu
+++ b/cpp/test/prims/kselection.cu
@@ -17,8 +17,8 @@
 #include <algorithm>
 #include <gtest/gtest.h>
 #include <limits>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 #include <selection/kselection.cuh>
 #include <stdlib.h>
 
diff --git a/cpp/test/prims/label.cu b/cpp/test/prims/label.cu
index 0146406542..fd1ef4fd5b 100644
--- a/cpp/test/prims/label.cu
+++ b/cpp/test/prims/label.cu
@@ -19,9 +19,9 @@
 #include <label/classlabels.cuh>
 
 #include "test_utils.h"
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <iostream>
diff --git a/cpp/test/prims/linalg_block.cu b/cpp/test/prims/linalg_block.cu
index 75839c3b2c..562a849317 100644
--- a/cpp/test/prims/linalg_block.cu
+++ b/cpp/test/prims/linalg_block.cu
@@ -19,10 +19,10 @@
 #include <random>
 #include <vector>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 #include "test_utils.h"
 
diff --git a/cpp/test/prims/linearReg.cu b/cpp/test/prims/linearReg.cu
index aabe5a413b..3a1c7ad532 100644
--- a/cpp/test/prims/linearReg.cu
+++ b/cpp/test/prims/linearReg.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <functions/linearReg.cuh>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/test/prims/log.cu b/cpp/test/prims/log.cu
index 443ead2571..3ef4f23007 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/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/logisticReg.cu b/cpp/test/prims/logisticReg.cu
index 8f0fc24c3d..0ec0d8a4ed 100644
--- a/cpp/test/prims/logisticReg.cu
+++ b/cpp/test/prims/logisticReg.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <functions/logisticReg.cuh>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/test/prims/make_arima.cu b/cpp/test/prims/make_arima.cu
index 3eae72be2a..1bb566d799 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/cudart_utils.h>
-#include <raft/interruptible.hpp>
 #include <random/make_arima.cuh>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/merge_labels.cu b/cpp/test/prims/merge_labels.cu
index 04ce43d3d1..76e040d20d 100644
--- a/cpp/test/prims/merge_labels.cu
+++ b/cpp/test/prims/merge_labels.cu
@@ -18,8 +18,8 @@
 #include <label/merge_labels.cuh>
 
 #include "test_utils.h"
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <rmm/device_uvector.hpp>
 #include <thrust/device_ptr.h>
 
diff --git a/cpp/test/prims/mutual_info_score.cu b/cpp/test/prims/mutual_info_score.cu
index 1087256fe3..fe37afb0ba 100644
--- a/cpp/test/prims/mutual_info_score.cu
+++ b/cpp/test/prims/mutual_info_score.cu
@@ -19,7 +19,7 @@
 #include <iostream>
 #include <metrics/contingencyMatrix.cuh>
 #include <metrics/mutual_info_score.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/penalty.cu b/cpp/test/prims/penalty.cu
index c17d41f5a1..0c8674f079 100644
--- a/cpp/test/prims/penalty.cu
+++ b/cpp/test/prims/penalty.cu
@@ -17,8 +17,8 @@
 #include "test_utils.h"
 #include <functions/penalty.cuh>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 
 namespace MLCommon {
 namespace Functions {
diff --git a/cpp/test/prims/rand_index.cu b/cpp/test/prims/rand_index.cu
index 9deab5f17a..ee4db450e4 100644
--- a/cpp/test/prims/rand_index.cu
+++ b/cpp/test/prims/rand_index.cu
@@ -18,7 +18,7 @@
 
 #include <metrics/rand_index.cuh>
 
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 
 #include <gtest/gtest.h>
 
diff --git a/cpp/test/prims/reverse.cu b/cpp/test/prims/reverse.cu
index edba2ac839..208c14e6a2 100644
--- a/cpp/test/prims/reverse.cu
+++ b/cpp/test/prims/reverse.cu
@@ -18,8 +18,8 @@
 #include <gtest/gtest.h>
 #include <matrix/reverse.cuh>
 #include <memory>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/score.cu b/cpp/test/prims/score.cu
index e0a0102c39..4c4cd3773a 100644
--- a/cpp/test/prims/score.cu
+++ b/cpp/test/prims/score.cu
@@ -18,9 +18,9 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/scores.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/interruptible.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/interruptible.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <vector>
 
diff --git a/cpp/test/prims/sigmoid.cu b/cpp/test/prims/sigmoid.cu
index 716f24b61d..7b10a31bce 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/cudart_utils.h>
 #include <rmm/device_uvector.hpp>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/silhouette_score.cu b/cpp/test/prims/silhouette_score.cu
index f42e038e4e..81a6049600 100644
--- a/cpp/test/prims/silhouette_score.cu
+++ b/cpp/test/prims/silhouette_score.cu
@@ -20,7 +20,7 @@
 #include <iostream>
 #include <metrics/batched/silhouette_score.cuh>
 #include <metrics/silhouette_score.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/distance/distance_type.hpp>
 #include <random>
 #include <rmm/device_uvector.hpp>
diff --git a/cpp/test/prims/test_utils.h b/cpp/test/prims/test_utils.h
index b756c4b6ca..a0fa3a5fc3 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/cudart_utils.h>
-#include <raft/interruptible.hpp>
 
 namespace raft {
 
diff --git a/cpp/test/prims/trustworthiness.cu b/cpp/test/prims/trustworthiness.cu
index f51f42ffa4..a6cdaab4ea 100644
--- a/cpp/test/prims/trustworthiness.cu
+++ b/cpp/test/prims/trustworthiness.cu
@@ -18,8 +18,8 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/trustworthiness_score.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/distance/distance.cuh>
 #include <vector>
 
 namespace MLCommon {
diff --git a/cpp/test/prims/v_measure.cu b/cpp/test/prims/v_measure.cu
index 750fc0ff07..0a8b16cffd 100644
--- a/cpp/test/prims/v_measure.cu
+++ b/cpp/test/prims/v_measure.cu
@@ -18,7 +18,7 @@
 #include <gtest/gtest.h>
 #include <iostream>
 #include <metrics/v_measure.cuh>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <random>
 
 namespace MLCommon {
diff --git a/cpp/test/sg/dbscan_test.cu b/cpp/test/sg/dbscan_test.cu
index 958ab959c0..2b2adf0655 100644
--- a/cpp/test/sg/dbscan_test.cu
+++ b/cpp/test/sg/dbscan_test.cu
@@ -15,18 +15,18 @@
  */
 
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <vector>
 
 #include <cuml/cluster/dbscan.hpp>
 #include <cuml/datasets/make_blobs.hpp>
 #include <cuml/metrics/metrics.hpp>
-#include <raft/distance/distance.hpp>
+#include <raft/distance/distance.cuh>
 #include <raft/distance/distance_type.hpp>
 
-#include <raft/handle.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/linalg/transpose.cuh>
 
 #include <test_utils.h>
 
diff --git a/cpp/test/sg/fil_test.cu b/cpp/test/sg/fil_test.cu
index 7ab0896055..72d0c99061 100644
--- a/cpp/test/sg/fil_test.cu
+++ b/cpp/test/sg/fil_test.cu
@@ -20,9 +20,9 @@
 
 #include <cuml/fil/fil.h>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <test_utils.h>
 #include <thrust/execution_policy.h>
 #include <thrust/functional.h>
diff --git a/cpp/test/sg/fnv_hash_test.cpp b/cpp/test/sg/fnv_hash_test.cpp
index 67bdb2c9ac..b2fcc85c01 100644
--- a/cpp/test/sg/fnv_hash_test.cpp
+++ b/cpp/test/sg/fnv_hash_test.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #include <cuml/fil/fnv_hash.h>
 #include <gtest/gtest.h>
-#include <raft/error.hpp>
+#include <raft/core/error.hpp>
 
 struct fnv_vec_t {
   std::vector<char> input;
diff --git a/cpp/test/sg/genetic/evolution_test.cu b/cpp/test/sg/genetic/evolution_test.cu
index 87af227121..6ce80d2402 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.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 3925d402dc..5ea426a65d 100644
--- a/cpp/test/sg/genetic/node_test.cpp
+++ b/cpp/test/sg/genetic/node_test.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #include <cuml/genetic/node.h>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/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 5ce753f205..d07b629c51 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/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.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 b546818112..f5d4b45760 100644
--- a/cpp/test/sg/hdbscan_test.cu
+++ b/cpp/test/sg/hdbscan_test.cu
@@ -17,8 +17,8 @@
 #include "hdbscan_inputs.hpp"
 
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <vector>
 
 #include <cuml/cluster/hdbscan.hpp>
@@ -34,9 +34,9 @@
 #include <raft/sparse/hierarchy/detail/agglomerative.cuh>
 
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/linalg/transpose.cuh>
 #include <raft/sparse/coo.hpp>
-#include <raft/sparse/op/sort.hpp>
+#include <raft/sparse/op/sort.cuh>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/execution_policy.h>
diff --git a/cpp/test/sg/holtwinters_test.cu b/cpp/test/sg/holtwinters_test.cu
index 29bca7bae6..c16b09e08b 100644
--- a/cpp/test/sg/holtwinters_test.cu
+++ b/cpp/test/sg/holtwinters_test.cu
@@ -19,9 +19,9 @@
 #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/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 
diff --git a/cpp/test/sg/kmeans_test.cu b/cpp/test/sg/kmeans_test.cu
index 7c43f1526e..ecd3cf1c50 100644
--- a/cpp/test/sg/kmeans_test.cu
+++ b/cpp/test/sg/kmeans_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/cudart_utils.h>
-#include <raft/handle.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/knn_test.cu b/cpp/test/sg/knn_test.cu
index 9c56d47fdd..28bf9c411a 100644
--- a/cpp/test/sg/knn_test.cu
+++ b/cpp/test/sg/knn_test.cu
@@ -16,9 +16,9 @@
 
 #include <gtest/gtest.h>
 #include <iostream>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/lars_test.cu b/cpp/test/sg/lars_test.cu
index 6baa991c79..7d852d9354 100644
--- a/cpp/test/sg/lars_test.cu
+++ b/cpp/test/sg/lars_test.cu
@@ -16,11 +16,11 @@
 
 #include <gtest/gtest.h>
 #include <iomanip>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 // #TODO: Replace with public header when ready
 #include <raft/linalg/detail/cusolver_wrappers.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <solver/lars_impl.cuh>
 #include <sstream>
diff --git a/cpp/test/sg/linear_svm_test.cu b/cpp/test/sg/linear_svm_test.cu
index 9da7926c71..5ce201b83d 100644
--- a/cpp/test/sg/linear_svm_test.cu
+++ b/cpp/test/sg/linear_svm_test.cu
@@ -18,11 +18,11 @@
 #include <cuml/datasets/make_regression.hpp>
 #include <cuml/svm/linear.hpp>
 #include <gtest/gtest.h>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/reduce.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/linalg/unary_op.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/reduce.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/linalg/unary_op.cuh>
+#include <raft/random/rng.cuh>
 #include <rmm/device_scalar.hpp>
 #include <rmm/device_uvector.hpp>
 #include <test_utils.h>
diff --git a/cpp/test/sg/linkage_test.cu b/cpp/test/sg/linkage_test.cu
index 300ac2b1ae..0d1600c723 100644
--- a/cpp/test/sg/linkage_test.cu
+++ b/cpp/test/sg/linkage_test.cu
@@ -15,8 +15,8 @@
  */
 
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <vector>
 
 #include <cuml/cluster/linkage.hpp>
@@ -24,7 +24,7 @@
 #include <hierarchy/pw_dist_graph.cuh>
 
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/linalg/transpose.cuh>
 #include <raft/sparse/coo.hpp>
 
 #include <cuml/common/logger.hpp>
diff --git a/cpp/test/sg/multi_sum_test.cu b/cpp/test/sg/multi_sum_test.cu
index dbc2c8e42e..d7d14a0638 100644
--- a/cpp/test/sg/multi_sum_test.cu
+++ b/cpp/test/sg/multi_sum_test.cu
@@ -18,9 +18,9 @@
 
 #include <cuml/fil/multi_sum.cuh>
 
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
 #include <thrust/device_vector.h>
 #include <thrust/fill.h>
diff --git a/cpp/test/sg/ols.cu b/cpp/test/sg/ols.cu
index 150fcb8d3e..1496f3be18 100644
--- a/cpp/test/sg/ols.cu
+++ b/cpp/test/sg/ols.cu
@@ -16,8 +16,8 @@
 
 #include <glm/ols.cuh>
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <rmm/cuda_stream_pool.hpp>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/pca_test.cu b/cpp/test/sg/pca_test.cu
index 72f644dd24..7c707a6f7b 100644
--- a/cpp/test/sg/pca_test.cu
+++ b/cpp/test/sg/pca_test.cu
@@ -17,9 +17,9 @@
 #include <cuml/decomposition/params.hpp>
 #include <gtest/gtest.h>
 #include <pca/pca.cuh>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 #include <test_utils.h>
 #include <vector>
 
diff --git a/cpp/test/sg/quasi_newton.cu b/cpp/test/sg/quasi_newton.cu
index 50f8eec8f6..f76b2015cb 100644
--- a/cpp/test/sg/quasi_newton.cu
+++ b/cpp/test/sg/quasi_newton.cu
@@ -20,9 +20,9 @@
 #include <glm/qn/glm_softmax.cuh>
 #include <glm/qn/qn.cuh>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
+#include <raft/linalg/transpose.cuh>
 #include <test_utils.h>
 #include <vector>
 
diff --git a/cpp/test/sg/rf_test.cu b/cpp/test/sg/rf_test.cu
index b6cf8fcfd1..1001a6753d 100644
--- a/cpp/test/sg/rf_test.cu
+++ b/cpp/test/sg/rf_test.cu
@@ -23,12 +23,12 @@
 #include <cuml/ensemble/randomforest.hpp>
 #include <cuml/fil/fil.h>
 #include <cuml/tree/algo_helper.h>
-#include <raft/random/rng.hpp>
+#include <raft/random/rng.cuh>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/handle.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/linalg/transpose.cuh>
 
 #include <thrust/binary_search.h>
 #include <thrust/copy.h>
diff --git a/cpp/test/sg/ridge.cu b/cpp/test/sg/ridge.cu
index d5406b43f1..3cebc612fd 100644
--- a/cpp/test/sg/ridge.cu
+++ b/cpp/test/sg/ridge.cu
@@ -17,8 +17,8 @@
 #include <glm/ols.cuh>
 #include <glm/ridge.cuh>
 #include <gtest/gtest.h>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
 #include <test_utils.h>
 
 namespace ML {
diff --git a/cpp/test/sg/rproj_test.cu b/cpp/test/sg/rproj_test.cu
index 34d5526fd3..35f74c8755 100644
--- a/cpp/test/sg/rproj_test.cu
+++ b/cpp/test/sg/rproj_test.cu
@@ -18,10 +18,10 @@
 #include <cuml/random_projection/rproj_c.h>
 #include <gtest/gtest.h>
 #include <iostream>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
-#include <raft/linalg/transpose.hpp>
+#include <raft/distance/distance.cuh>
+#include <raft/linalg/transpose.cuh>
 #include <random>
 #include <test_utils.h>
 #include <vector>
diff --git a/cpp/test/sg/sgd.cu b/cpp/test/sg/sgd.cu
index 4255f89242..22fca1385a 100644
--- a/cpp/test/sg/sgd.cu
+++ b/cpp/test/sg/sgd.cu
@@ -15,8 +15,8 @@
  */
 
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/matrix/matrix.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/matrix/matrix.cuh>
 #include <rmm/device_uvector.hpp>
 #include <solver/sgd.cuh>
 #include <test_utils.h>
diff --git a/cpp/test/sg/shap_kernel.cu b/cpp/test/sg/shap_kernel.cu
index 327afe6cdf..40d328d7a6 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/cudart_utils.h>
-#include <raft/handle.hpp>
 #include <rmm/device_uvector.hpp>
 
 #include <thrust/count.h>
diff --git a/cpp/test/sg/svc_test.cu b/cpp/test/sg/svc_test.cu
index 3c987c452d..04a0d04200 100644
--- a/cpp/test/sg/svc_test.cu
+++ b/cpp/test/sg/svc_test.cu
@@ -25,12 +25,12 @@
 #include <iostream>
 #include <matrix/grammatrix.cuh>
 #include <matrix/kernelmatrices.cuh>
+#include <raft/core/cudart_utils.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/linalg/add.hpp>
-#include <raft/linalg/map_then_reduce.hpp>
-#include <raft/linalg/transpose.hpp>
-#include <raft/random/rng.hpp>
+#include <raft/linalg/add.cuh>
+#include <raft/linalg/map_then_reduce.cuh>
+#include <raft/linalg/transpose.cuh>
+#include <raft/random/rng.cuh>
 #include <rmm/device_uvector.hpp>
 #include <string>
 #include <svm/smoblocksolve.cuh>
diff --git a/cpp/test/sg/trustworthiness_test.cu b/cpp/test/sg/trustworthiness_test.cu
index 109668e810..b015772597 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/cudart_utils.h>
-#include <raft/handle.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 8c310d4400..d789ea5b8a 100644
--- a/cpp/test/sg/tsne_test.cu
+++ b/cpp/test/sg/tsne_test.cu
@@ -17,7 +17,7 @@
 #include <cuml/manifold/tsne.h>
 #include <cuml/metrics/metrics.hpp>
 #include <raft/distance/distance_type.hpp>
-#include <raft/linalg/map.hpp>
+#include <raft/linalg/map.cuh>
 
 #include <cuml/common/logger.hpp>
 #include <datasets/boston.h>
@@ -26,7 +26,7 @@
 #include <datasets/digits.h>
 #include <gtest/gtest.h>
 #include <iostream>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <stdio.h>
 #include <stdlib.h>
 #include <thrust/reduce.h>
diff --git a/cpp/test/sg/tsvd_test.cu b/cpp/test/sg/tsvd_test.cu
index 34ee3e8ec0..b72a45df03 100644
--- a/cpp/test/sg/tsvd_test.cu
+++ b/cpp/test/sg/tsvd_test.cu
@@ -16,8 +16,8 @@
 
 #include <cuml/decomposition/params.hpp>
 #include <gtest/gtest.h>
-#include <raft/cudart_utils.h>
-#include <raft/random/rng.hpp>
+#include <raft/core/cudart_utils.hpp>
+#include <raft/random/rng.cuh>
 #include <test_utils.h>
 #include <tsvd/tsvd.cuh>
 #include <vector>
diff --git a/cpp/test/sg/umap_parametrizable_test.cu b/cpp/test/sg/umap_parametrizable_test.cu
index cf90cba1b4..66f7085309 100644
--- a/cpp/test/sg/umap_parametrizable_test.cu
+++ b/cpp/test/sg/umap_parametrizable_test.cu
@@ -24,17 +24,17 @@
 #include <cuml/metrics/metrics.hpp>
 #include <cuml/neighbors/knn.hpp>
 #include <datasets/digits.h>
-#include <raft/cudart_utils.h>
+#include <raft/core/cudart_utils.hpp>
 #include <test_utils.h>
 
 #include <datasets/digits.h>
 #include <raft/linalg/reduce_rows_by_key.cuh>
 #include <selection/knn.cuh>
 
+#include <raft/core/cudart_utils.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/cuda_utils.cuh>
-#include <raft/cudart_utils.h>
-#include <raft/distance/distance.hpp>
-#include <raft/handle.hpp>
+#include <raft/distance/distance.cuh>
 #include <selection/knn.cuh>
 #include <umap/runner.cuh>
 
diff --git a/wiki/cpp/DEVELOPER_GUIDE.md b/wiki/cpp/DEVELOPER_GUIDE.md
index 0792f7e2b4..adf3954836 100644
--- a/wiki/cpp/DEVELOPER_GUIDE.md
+++ b/wiki/cpp/DEVELOPER_GUIDE.md
@@ -368,7 +368,7 @@ void foo(const double* const srcdata, double* const result)
 ```
 No work in any stream should start in `ML::algo` before the `cudaMemcpyAsync` in `stream` launched before the call to `ML::algo` is done. And all work in all streams used in `ML::algo` should be done before the `cudaMemcpyAsync` in `stream` launched after the call to `ML::algo` starts.
 
-This can be ensured by introducing interstream dependencies with CUDA events and `cudaStreamWaitEvent`. For convenience, the header `raft/handle.hpp` provides the class `raft::stream_syncer` which lets all `raft::handle_t` internal CUDA streams wait on `raft::handle_t:get_stream()` in its constructor and in its destructor and lets `raft::handle_t::get_stream()` wait on all work enqueued in the `raft::handle_t` internal CUDA streams. The intended use would be to create a `raft::stream_syncer` object as the first thing in a entry function of the public cuML API:
+This can be ensured by introducing interstream dependencies with CUDA events and `cudaStreamWaitEvent`. For convenience, the header `raft/core/handle.hpp` provides the class `raft::stream_syncer` which lets all `raft::handle_t` internal CUDA streams wait on `raft::handle_t::get_stream()` in its constructor and in its destructor and lets `raft::handle_t::get_stream()` wait on all work enqueued in the `raft::handle_t` internal CUDA streams. The intended use would be to create a `raft::stream_syncer` object as the first thing in a entry function of the public cuML API:
 
 ```cpp
 void cumlAlgo(const raft::handle_t& handle, ...)
@@ -408,13 +408,13 @@ int main(int argc, char** argv)
 
 ## Multi-GPU
 
-The multi GPU paradigm of cuML is **O**ne **P**rocess per **G**PU (OPG). Each algorithm should be implemented in a way that it can run with a single GPU without any specific dependencies to a particular communication library. A multi-GPU implementation should use the methods offered by the class `raft::comms::comms_t` from [raft/comms/comms.hpp] for inter-rank/GPU communication. It is the responsibility of the user of cuML to create an initialized instance of `raft::comms::comms_t`.
+The multi GPU paradigm of cuML is **O**ne **P**rocess per **G**PU (OPG). Each algorithm should be implemented in a way that it can run with a single GPU without any specific dependencies to a particular communication library. A multi-GPU implementation should use the methods offered by the class `raft::comms::comms_t` from [raft/core/comms.hpp] for inter-rank/GPU communication. It is the responsibility of the user of cuML to create an initialized instance of `raft::comms::comms_t`.
 
 E.g. with a CUDA-aware MPI, a cuML user could use code like this to inject an initialized instance of `raft::comms::mpi_comms` into a `raft::handle_t`:
 
 ```cpp
 #include <mpi.h>
-#include <raft/handle.hpp>
+#include <raft/core/handle.hpp>
 #include <raft/comms/mpi_comms.hpp>
 #include <mlalgo/mlalgo.hpp>
 ...