diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
new file mode 100644
index 0000000000..af7e9806aa
--- /dev/null
+++ b/.github/workflows/wheels.yml
@@ -0,0 +1,62 @@
+name: cuML wheels
+
+on:
+  workflow_call:
+    inputs:
+      versioneer-override:
+        type: string
+        default: ''
+      build-tag:
+        type: string
+        default: ''
+      branch:
+        required: true
+        type: string
+      date:
+        required: true
+        type: string
+      sha:
+        required: true
+        type: string
+      build-type:
+        type: string
+        default: nightly
+
+concurrency:
+  group: "cuml-${{ github.workflow }}-${{ github.ref }}"
+  cancel-in-progress: true
+
+jobs:
+  cuml-wheel:
+    uses: rapidsai/shared-action-workflows/.github/workflows/wheels-manylinux.yml@main
+    with:
+      repo: rapidsai/cuml
+
+      build-type: ${{ inputs.build-type }}
+      branch: ${{ inputs.branch }}
+      sha: ${{ inputs.sha }}
+      date: ${{ inputs.date }}
+
+      package-dir: python
+      package-name: cuml
+
+      # Note that this approach to cloning repos obviates any modification to
+      # the CMake variables in get_cumlprims_mg.cmake since CMake will just use
+      # the clone as is.
+      extra-repo: rapidsai/cumlprims_mg
+      extra-repo-sha: branch-22.12
+      extra-repo-deploy-key: CUMLPRIMS_SSH_PRIVATE_DEPLOY_KEY
+
+      python-package-versioneer-override: ${{ inputs.versioneer-override }}
+      python-package-build-tag: ${{ inputs.build-tag }}
+
+      skbuild-configure-options: "-DCUML_BUILD_WHEELS=ON -DDETECT_CONDA_ENV=OFF -DCPM_cumlprims_mg_SOURCE=/project/cumlprims_mg/"
+
+      # Always want to test against latest dask/distributed.
+      test-before-amd64: "pip install git+https://github.com/dask/dask.git@main git+https://github.com/dask/distributed.git@main git+https://github.com/rapidsai/dask-cuda.git@branch-22.12"
+      # On arm also need to install cupy from the specific webpage and CMake
+      # because treelite needs to be compiled (no wheels available for arm).
+      test-before-arm64: "pip install cupy-cuda11x -f https://pip.cupy.dev/aarch64 && pip install cmake && pip install git+https://github.com/dask/dask.git@main git+https://github.com/dask/distributed.git@main git+https://github.com/rapidsai/dask-cuda.git@branch-22.12"
+      test-extras: test
+      test-unittest: "pytest -v ./python/cuml/tests -k 'not test_silhouette_score_batched'"
+    secrets: inherit
diff --git a/.gitignore b/.gitignore
index 6a85de8349..b1bc1283dd 100644
--- a/.gitignore
+++ b/.gitignore
@@ -31,6 +31,7 @@ dask-worker-space/
 tmp/
 .hypothesis
 wheels/
+wheelhouse/
 _skbuild/
 
 ## files pickled in notebook when ran during python docstring generation
diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index d42f28e1b8..9e9e040a56 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -28,11 +28,6 @@ rapids_cuda_init_architectures(CUML)
 
 project(CUML VERSION 22.12.00 LANGUAGES CXX CUDA)
 
-# Needed because GoogleBenchmark changes the state of FindThreads.cmake, causing subsequent runs to
-# have different values for the `Threads::Threads` target. Setting this flag ensures
-# `Threads::Threads` is the same value in first run and subsequent runs.
-set(THREADS_PREFER_PTHREAD_FLAG ON)
-
 # Write the version header
 rapids_cmake_write_version_file(include/cuml/version_config.hpp)
 
@@ -71,6 +66,15 @@ option(CUDA_STATIC_RUNTIME "Statically link the CUDA toolkit runtime and librari
 option(CUML_USE_RAFT_STATIC "Build and statically link the RAFT libraries" OFF)
 option(CUML_USE_FAISS_STATIC "Build and statically link the FAISS library for nearest neighbors search on GPU" OFF)
 option(CUML_USE_TREELITE_STATIC "Build and statically link the treelite library" OFF)
+option(CUML_EXPORT_TREELITE_LINKAGE "Whether to publicly or privately link treelite to libcuml++" OFF)
+option(CUML_USE_CUMLPRIMS_MG_STATIC "Build and statically link the cumlprims_mg library" OFF)
+# The options below allow incorporating libcuml into another build process
+# without installing all its components. This is useful if total file size is
+# at a premium and we do not expect other consumers to use any APIs of the
+# dependency except those that are directly linked to by the dependent library.
+option(CUML_EXCLUDE_RAFT_FROM_ALL "Exclude RAFT targets from cuML's 'all' target" OFF)
+option(CUML_EXCLUDE_TREELITE_FROM_ALL "Exclude Treelite targets from cuML's 'all' target" OFF)
+option(CUML_EXCLUDE_CUMLPRIMS_MG_FROM_ALL "Exclude cumlprims_mg targets from cuML's 'all' target" OFF)
 option(CUML_RAFT_CLONE_ON_PIN "Explicitly clone RAFT branch when pinned to non-feature branch" ON)
 
 message(VERBOSE "CUML_CPP: Building libcuml_c shared library. Contains the cuML C API: ${BUILD_CUML_C_LIBRARY}")
@@ -104,6 +108,13 @@ set(RMM_LOGGING_LEVEL "INFO" CACHE STRING "Choose the logging level.")
 set_property(CACHE RMM_LOGGING_LEVEL PROPERTY STRINGS "TRACE" "DEBUG" "INFO" "WARN" "ERROR" "CRITICAL" "OFF")
 message(VERBOSE "CUML_CPP: RMM_LOGGING_LEVEL = '${RMM_LOGGING_LEVEL}'.")
 
+if(BUILD_CUML_TESTS OR BUILD_PRIMS_TESTS)
+  # Needed because GoogleBenchmark changes the state of FindThreads.cmake, causing subsequent runs to
+  # have different values for the `Threads::Threads` target. Setting this flag ensures
+  # `Threads::Threads` is the same value in first run and subsequent runs.
+  set(THREADS_PREFER_PTHREAD_FLAG ON)
+endif()
+
 ##############################################################################
 # - Target names -------------------------------------------------------------
 
@@ -132,15 +143,7 @@ endif()
 
 set(_ctk_static_suffix "")
 if(CUDA_STATIC_RUNTIME)
-  # If we're statically linking CTK cuBLAS,
-  # we also want to statically link BLAS
-  set(BLA_STATIC ON)
-  set(_ctk_static_suffix "_static")
-  set(_ctk_static_suffix_cufft "_static_nocallback")
-  # Control legacy FindCUDA.cmake behavior too
-  # Remove this after we push it into rapids-cmake:
-  # https://github.com/rapidsai/rapids-cmake/pull/259
-  set(CUDA_USE_STATIC_CUDA_RUNTIME ON)
+  set(_ctk_static_suffix "_static_nocallback")
 endif()
 
 if (NOT DISABLE_OPENMP)
@@ -212,8 +215,11 @@ endif()
 
 # add third party dependencies using CPM
 rapids_cpm_init()
+rapids_cmake_install_lib_dir(lib_dir)
 
-find_package(Threads)
+if(BUILD_CUML_TESTS OR BUILD_PRIMS_TESTS)
+  find_package(Threads)
+endif()
 
 include(cmake/thirdparty/get_raft.cmake)
 
@@ -262,6 +268,20 @@ SECTIONS
 ]=])
 endif()
 
+# Copy the interface include directories from INCLUDED_TARGET to TARGET.
+# This is necessary when INCLUDED_TARGET was compiled statically but includes
+# public APIs that may still require consumers to have the same interface
+# headers available.
+function(copy_interface_excludes)
+  set(_options "")
+  set(_one_value TARGET INCLUDED_TARGET)
+  set(_multi_value "")
+  cmake_parse_arguments(_CUML_INCLUDES "${_options}" "${_one_value}"
+                        "${_multi_value}" ${ARGN})
+  get_target_property(_includes ${_CUML_INCLUDES_INCLUDED_TARGET} INTERFACE_INCLUDE_DIRECTORIES)
+  target_include_directories(${_CUML_INCLUDES_TARGET} PUBLIC ${_includes})
+endfunction()
+
 if(BUILD_CUML_CPP_LIBRARY)
 
   # single GPU components
@@ -534,19 +554,67 @@ if(BUILD_CUML_CPP_LIBRARY)
       $<INSTALL_INTERFACE:include>
   )
 
+  set(_cuml_cpp_public_libs)
+  set(_cuml_cpp_private_libs)
+
+  if(CUML_USE_RAFT_STATIC AND (TARGET raft::raft))
+    copy_interface_excludes(INCLUDED_TARGET raft::raft TARGET ${CUML_CPP_TARGET})
+
+    if(CUML_USE_RAFT_NN AND (TARGET faiss::faiss))
+      copy_interface_excludes(INCLUDED_TARGET faiss::faiss TARGET ${CUML_CPP_TARGET})
+    endif()
+
+    if(CUML_USE_RAFT_DIST AND (TARGET cuco::cuco))
+      list(APPEND _cuml_cpp_private_libs cuco::cuco)
+    endif()
+  endif()
+
+  if(CUML_USE_TREELITE_STATIC AND (TARGET treelite::treelite_static))
+    # By default, TREELITE_LIBS will contain both treelite::treelite_static and
+    # treelite::treelite_runtime_static if we are linking statically, but these
+    # two targets have duplicate symbols so we can only link to one of them.
+    set(TREELITE_LIBS treelite::treelite_static)
+
+    copy_interface_excludes(INCLUDED_TARGET treelite::treelite_static TARGET ${CUML_CPP_TARGET})
+  elseif(CUML_EXPORT_TREELITE_LINKAGE)
+    list(APPEND _cuml_cpp_public_libs ${TREELITE_LIBS})
+  endif()
+
+  if(CUML_USE_CUMLPRIMS_MG_STATIC AND (TARGET cumlprims_mg::cumlprims_mg))
+    copy_interface_excludes(INCLUDED_TARGET cumlprims_mg::cumlprims_mg TARGET ${CUML_CPP_TARGET})
+  endif()
+
+  # These are always private:
+  list(APPEND _cuml_cpp_private_libs
+    $<TARGET_NAME_IF_EXISTS:GPUTreeShap::GPUTreeShap>
+    $<$<BOOL:${LINK_CUFFT}>:CUDA::cufft${_ctk_static_suffix}>
+    ${TREELITE_LIBS}
+    ${OpenMP_CXX_LIB_NAMES}
+    $<$<OR:$<BOOL:${BUILD_CUML_STD_COMMS}>,$<BOOL:${BUILD_CUML_MPI_COMMS}>>:NCCL::NCCL>
+    $<$<BOOL:${BUILD_CUML_MPI_COMMS}>:${MPI_CXX_LIBRARIES}>
+  )
+
+  set(_cuml_cpp_libs_var_name "_cuml_cpp_public_libs")
+  if(CUDA_STATIC_RUNTIME)
+    set(_cuml_cpp_libs_var_name "_cuml_cpp_private_libs")
+    # Add CTK include paths because we're going to make our CTK library links private below
+    target_include_directories(${CUML_CPP_TARGET} SYSTEM PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
+  endif()
+
+  # The visibility of these depend on whether we're linking the CTK statically,
+  # because cumlprims_mg and cuML inherit their CUDA libs from the raft::raft
+  # INTERFACE target.
+  list(APPEND ${_cuml_cpp_libs_var_name}
+    raft::raft
+    $<$<BOOL:${CUML_USE_RAFT_NN}>:raft::nn>
+    $<$<BOOL:${CUML_USE_RAFT_DIST}>:raft::distance>
+    $<TARGET_NAME_IF_EXISTS:cumlprims_mg::cumlprims_mg>
+  )
+
   target_link_libraries(${CUML_CPP_TARGET}
-    PUBLIC
-      raft::raft
-      $<$<BOOL:${CUML_USE_RAFT_NN}>:raft::nn>
-      $<$<BOOL:${CUML_USE_RAFT_DIST}>:raft::distance>
-    PRIVATE
-      $<$<BOOL:${LINK_CUFFT}>:CUDA::cufft${_ctk_static_suffix_cufft}>
-      ${TREELITE_LIBS}
-      $<$<BOOL:${treeshap_algo}>:GPUTreeShap::GPUTreeShap>
-      ${OpenMP_CXX_LIB_NAMES}
-      $<$<OR:$<BOOL:${BUILD_CUML_STD_COMMS}>,$<BOOL:${BUILD_CUML_MPI_COMMS}>>:NCCL::NCCL>
-      $<$<BOOL:${BUILD_CUML_MPI_COMMS}>:${MPI_CXX_LIBRARIES}>
-      $<$<BOOL:${ENABLE_CUMLPRIMS_MG}>:cumlprims_mg::cumlprims_mg>
+    PUBLIC  rmm::rmm
+            ${_cuml_cpp_public_libs}
+    PRIVATE ${_cuml_cpp_private_libs}
   )
 
   # If we export the libdmlc symbols, they can lead to weird crashes with other
@@ -556,7 +624,6 @@ if(BUILD_CUML_CPP_LIBRARY)
   target_link_options(${CUML_CPP_TARGET} PRIVATE "-Wl,--exclude-libs,libprotobuf.a")
   # ensure CUDA symbols aren't relocated to the middle of the debug build binaries
   target_link_options(${CUML_CPP_TARGET} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/fatbin.ld")
-
 endif()
 
 #############################################################################
@@ -611,7 +678,6 @@ endif()
 
 # ###################################################################################################
 # # - install targets -------------------------------------------------------------------------------
-rapids_cmake_install_lib_dir( lib_dir )
 include(CPack)
 
 set(CUML_TARGETS ${CUML_CPP_TARGET})
diff --git a/cpp/cmake/thirdparty/get_cumlprims_mg.cmake b/cpp/cmake/thirdparty/get_cumlprims_mg.cmake
index 16137ed340..0bbc91fc17 100644
--- a/cpp/cmake/thirdparty/get_cumlprims_mg.cmake
+++ b/cpp/cmake/thirdparty/get_cumlprims_mg.cmake
@@ -19,12 +19,20 @@ set(CUML_BRANCH_VERSION_cumlprims_mg "${CUML_VERSION_MAJOR}.${CUML_VERSION_MINOR
 
 function(find_and_configure_cumlprims_mg)
 
-    set(oneValueArgs VERSION FORK PINNED_TAG CLONE_ON_PIN)
+    set(oneValueArgs VERSION FORK PINNED_TAG BUILD_STATIC EXCLUDE_FROM_ALL CLONE_ON_PIN)
     cmake_parse_arguments(PKG "" "${oneValueArgs}" "" ${ARGN} )
 
     if(PKG_CLONE_ON_PIN AND NOT PKG_PINNED_TAG STREQUAL "branch-${CUML_BRANCH_VERSION_cumlprims_mg}")
         message("Pinned tag found: ${PKG_PINNED_TAG}. Cloning cumlprims locally.")
         set(CPM_DOWNLOAD_cumlprims_mg ON)
+    elseif(PKG_BUILD_STATIC AND (NOT CPM_cumlprims_mg_SOURCE))
+      message(STATUS "CUML: Cloning cumlprims_mg locally to build static libraries.")
+      set(CPM_DOWNLOAD_cumlprims_mg ON)
+    endif()
+
+    set(CUMLPRIMS_MG_BUILD_SHARED_LIBS ON)
+    if(PKG_BUILD_STATIC)
+      set(CUMLPRIMS_MG_BUILD_SHARED_LIBS OFF)
     endif()
 
     rapids_cpm_find(cumlprims_mg ${PKG_VERSION}
@@ -32,8 +40,14 @@ function(find_and_configure_cumlprims_mg)
       BUILD_EXPORT_SET    cuml-exports
       INSTALL_EXPORT_SET  cuml-exports
         CPM_ARGS
-            GIT_REPOSITORY git@github.com:${PKG_FORK}/cumlprims_mg.git
-            GIT_TAG        ${PKG_PINNED_TAG}
+          GIT_REPOSITORY git@github.com:${PKG_FORK}/cumlprims_mg.git
+          GIT_TAG        ${PKG_PINNED_TAG}
+          EXCLUDE_FROM_ALL ${PKG_EXCLUDE_FROM_ALL}
+          SOURCE_SUBDIR    cpp
+          OPTIONS
+            "BUILD_TESTS OFF"
+            "BUILD_BENCHMARKS OFF"
+            "BUILD_SHARED_LIBS ${CUMLPRIMS_MG_BUILD_SHARED_LIBS}"
     )
 
 endfunction()
@@ -48,9 +62,11 @@ endfunction()
 # cumlprims_mg as part of building cuml itself, set the CMake variable
 # `-D CPM_cumlprims_mg_SOURCE=/path/to/cumlprims_mg`
 ###
-find_and_configure_cumlprims_mg(VERSION    ${CUML_MIN_VERSION_cumlprims_mg}
+find_and_configure_cumlprims_mg(VERSION          ${CUML_MIN_VERSION_cumlprims_mg}
                                 FORK       rapidsai
                                 PINNED_TAG branch-${CUML_BRANCH_VERSION_cumlprims_mg}
+                                BUILD_STATIC     ${CUML_USE_CUMLPRIMS_MG_STATIC}
+                                EXCLUDE_FROM_ALL ${CUML_EXCLUDE_CUMLPRIMS_MG_FROM_ALL}
                                 # When PINNED_TAG above doesn't match cuml,
                                 # force local cumlprims_mg clone in build directory
                                 # even if it's already installed.
diff --git a/cpp/cmake/thirdparty/get_raft.cmake b/cpp/cmake/thirdparty/get_raft.cmake
index 9fe2e222b5..feba32e1fa 100644
--- a/cpp/cmake/thirdparty/get_raft.cmake
+++ b/cpp/cmake/thirdparty/get_raft.cmake
@@ -18,54 +18,60 @@ set(CUML_MIN_VERSION_raft "${CUML_VERSION_MAJOR}.${CUML_VERSION_MINOR}.00")
 set(CUML_BRANCH_VERSION_raft "${CUML_VERSION_MAJOR}.${CUML_VERSION_MINOR}")
 
 function(find_and_configure_raft)
-    set(oneValueArgs VERSION FORK PINNED_TAG USE_RAFT_DIST USE_RAFT_NN USE_RAFT_STATIC USE_FAISS_STATIC CLONE_ON_PIN NVTX)
+    set(oneValueArgs VERSION FORK PINNED_TAG EXCLUDE_FROM_ALL USE_RAFT_DIST USE_RAFT_NN USE_RAFT_STATIC USE_FAISS_STATIC CLONE_ON_PIN NVTX)
     cmake_parse_arguments(PKG "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN} )
 
     if(PKG_CLONE_ON_PIN AND NOT PKG_PINNED_TAG STREQUAL "branch-${CUML_BRANCH_VERSION_raft}")
-        message(STATUS "CUML: RAFT pinned tag found: ${PKG_PINNED_TAG}. Cloning raft locally.")
-        set(CPM_DOWNLOAD_raft ON)
-    endif()
-
-    if(PKG_USE_RAFT_STATIC)
-        message(STATUS "CUML: Cloning raft locally to build static libraries.")
-        set(CPM_DOWNLOAD_raft ON)
+      message(STATUS "CUML: RAFT pinned tag found: ${PKG_PINNED_TAG}. Cloning raft locally.")
+      set(CPM_DOWNLOAD_raft ON)
+    elseif(PKG_USE_RAFT_STATIC AND (NOT CPM_raft_SOURCE))
+      message(STATUS "CUML: Cloning raft locally to build static libraries.")
+      set(CPM_DOWNLOAD_raft ON)
     endif()
 
     if(PKG_USE_RAFT_DIST)
-        string(APPEND RAFT_COMPONENTS "distance")
+      string(APPEND RAFT_COMPONENTS "distance")
     endif()
+
     if(PKG_USE_RAFT_NN)
-        string(APPEND RAFT_COMPONENTS " nn")
+      string(APPEND RAFT_COMPONENTS " nn")
     endif()
 
     if(PKG_USE_RAFT_DIST AND PKG_USE_RAFT_NN)
-        set(RAFT_COMPILE_LIBRARIES ON)
+      set(RAFT_COMPILE_LIBRARIES ON)
     else()
-        set(RAFT_COMPILE_LIBRARIES OFF)
+      set(RAFT_COMPILE_LIBRARIES OFF)
     endif()
 
     # We need to set this each time so that on subsequent calls to cmake
     # the raft-config.cmake re-evaluates the RAFT_NVTX value
     set(RAFT_NVTX ${PKG_NVTX})
 
+    set(RAFT_BUILD_SHARED_LIBS ON)
+    if(${PKG_USE_RAFT_STATIC})
+      set(RAFT_BUILD_SHARED_LIBS OFF)
+    endif()
+
     message(VERBOSE "CUML: raft FIND_PACKAGE_ARGUMENTS COMPONENTS ${RAFT_COMPONENTS}")
 
     rapids_cpm_find(raft ${PKG_VERSION}
-            GLOBAL_TARGETS      raft::raft
-            BUILD_EXPORT_SET    cuml-exports
-            INSTALL_EXPORT_SET  cuml-exports
-            COMPONENTS          ${RAFT_COMPONENTS}
-            CPM_ARGS
-            GIT_REPOSITORY https://github.com/${PKG_FORK}/raft.git
-            GIT_TAG        ${PKG_PINNED_TAG}
-            SOURCE_SUBDIR  cpp
-            OPTIONS
-              "BUILD_TESTS OFF"
-              "RAFT_COMPILE_LIBRARIES ${RAFT_COMPILE_LIBRARIES}"
-              "RAFT_COMPILE_NN_LIBRARY ${PKG_USE_RAFT_NN}"
-              "RAFT_COMPILE_DIST_LIBRARY ${PKG_USE_RAFT_DIST}"
-              "RAFT_USE_FAISS_STATIC ${PKG_USE_FAISS_STATIC}"
+      GLOBAL_TARGETS      raft::raft
+      BUILD_EXPORT_SET    cuml-exports
+      INSTALL_EXPORT_SET  cuml-exports
+      COMPONENTS          ${RAFT_COMPONENTS}
+      CPM_ARGS
+        GIT_REPOSITORY         https://github.com/${PKG_FORK}/raft.git
+        GIT_TAG                ${PKG_PINNED_TAG}
+        SOURCE_SUBDIR          cpp
+        EXCLUDE_FROM_ALL       ${PKG_EXCLUDE_FROM_ALL}
+        OPTIONS
+          "BUILD_TESTS OFF"
+          "BUILD_SHARED_LIBS ${RAFT_BUILD_SHARED_LIBS}"
+          "RAFT_COMPILE_LIBRARIES ${RAFT_COMPILE_LIBRARIES}"
+          "RAFT_COMPILE_NN_LIBRARY ${PKG_USE_RAFT_NN}"
+          "RAFT_COMPILE_DIST_LIBRARY ${PKG_USE_RAFT_DIST}"
+          "RAFT_USE_FAISS_STATIC ${PKG_USE_FAISS_STATIC}"
     )
 
     if(raft_ADDED)
@@ -82,8 +88,8 @@ endfunction()
 # CPM_raft_SOURCE=/path/to/local/raft
 find_and_configure_raft(VERSION          ${CUML_MIN_VERSION_raft}
                         FORK             rapidsai
-			PINNED_TAG       branch-${CUML_BRANCH_VERSION_raft}
-
+                        PINNED_TAG       branch-${CUML_BRANCH_VERSION_raft}
+                        EXCLUDE_FROM_ALL ${CUML_EXCLUDE_RAFT_FROM_ALL}
                         # When PINNED_TAG above doesn't match cuml,
                         # force local raft clone in build directory
                         # even if it's already installed.
diff --git a/cpp/cmake/thirdparty/get_treelite.cmake b/cpp/cmake/thirdparty/get_treelite.cmake
index 199b2a5fdb..ec6929be9e 100644
--- a/cpp/cmake/thirdparty/get_treelite.cmake
+++ b/cpp/cmake/thirdparty/get_treelite.cmake
@@ -14,16 +14,17 @@
 # limitations under the License.
 #=============================================================================
 
+
 function(find_and_configure_treelite)
 
-    set(oneValueArgs VERSION PINNED_TAG BUILD_STATIC_LIBS)
+    set(oneValueArgs VERSION PINNED_TAG EXCLUDE_FROM_ALL BUILD_STATIC_LIBS)
     cmake_parse_arguments(PKG "${options}" "${oneValueArgs}"
                           "${multiValueArgs}" ${ARGN} )
 
     if(NOT PKG_BUILD_STATIC_LIBS)
-        list(APPEND TREELITE_LIBS treelite::treelite treelite::treelite_runtime)
+      list(APPEND TREELITE_LIBS treelite::treelite treelite::treelite_runtime)
     else()
-        list(APPEND TREELITE_LIBS treelite::treelite_static treelite::treelite_runtime_static)
+      list(APPEND TREELITE_LIBS treelite::treelite_static treelite::treelite_runtime_static)
     endif()
 
     rapids_cpm_find(Treelite ${PKG_VERSION}
@@ -32,6 +33,7 @@ function(find_and_configure_treelite)
         CPM_ARGS
             GIT_REPOSITORY   https://github.com/dmlc/treelite.git
             GIT_TAG          ${PKG_PINNED_TAG}
+            EXCLUDE_FROM_ALL ${PKG_EXCLUDE_FROM_ALL}
             OPTIONS
               "USE_OPENMP ON"
               "BUILD_STATIC_LIBS ${PKG_BUILD_STATIC_LIBS}"
@@ -90,4 +92,5 @@ endfunction()
 
 find_and_configure_treelite(VERSION     3.0.1
                         PINNED_TAG  cb09d539dce7e67f60cc33e4fac6b95b7185847a
+                        EXCLUDE_FROM_ALL  ${CUML_EXCLUDE_TREELITE_FROM_ALL}
                         BUILD_STATIC_LIBS ${CUML_USE_TREELITE_STATIC})
diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
index 0a2131cdde..6f2f59189d 100644
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -18,6 +18,11 @@ include(../fetch_rapids.cmake)
 
 set(CUML_VERSION 22.12.00)
 
+# We always need CUDA for cuML because the raft dependency brings in a
+# header-only cuco dependency that enables CUDA unconditionally.
+include(rapids-cuda)
+rapids_cuda_init_architectures(cuml-python)
+
 project(
   cuml-python
   VERSION ${CUML_VERSION}
@@ -25,12 +30,13 @@ project(
             # language to be enabled here. The test project that is built in scikit-build to verify
             # various linking options for the python library is hardcoded to build with C, so until
             # that is fixed we need to keep C.
-            C CXX
+            C CXX CUDA
 )
 
 ################################################################################
 # - User Options  --------------------------------------------------------------
 option(FIND_CUML_CPP "Search for existing CUML C++ installations before defaulting to local files" OFF)
+option(CUML_BUILD_WHEELS "Whether this build is generating a Python wheel." OFF)
 option(SINGLEGPU "Disable all mnmg components and comms libraries" OFF)
 set(CUML_RAFT_CLONE_ON_PIN OFF)
 
@@ -44,64 +50,62 @@ set(CUML_CPP_TARGET "cuml++")
 ################################################################################
 # - Process User Options  ------------------------------------------------------
 
-
-# If the user requested it,  we attempt to find cuml.
+# If the user requested it, we attempt to find cuml.
 if(FIND_CUML_CPP)
+  # We need to call get_treelite explicitly because we need the correct
+  # ${TREELITE_LIBS} definition for RF
   include(rapids-cpm)
   include(rapids-export)
   rapids_cpm_init()
-  include(rapids-cuda)
-  rapids_cuda_init_architectures(cuml-python)
-  enable_language(CUDA)
-
-  # variables used by get_raft.cmake, we can get rid of them when this is solved:
-  # https://github.com/rapidsai/rapids-cmake/issues/228
-  set(CUML_VERSION_MAJOR "${cuml-python_VERSION_MAJOR}")
-  set(CUML_VERSION_MINOR "${cuml-python_VERSION_MINOR}")
-  set(CUML_USE_RAFT_NN ON)
-  set(CUML_USE_RAFT_DIST ON)
-
-  # We need to call get_raft due to cuML asking for raft::nn and
-  # raft::distance targets
-  # see issue https://github.com/rapidsai/cuml/issues/4843
-  include(../cpp/cmake/thirdparty/get_raft.cmake)
-
-  # We need to call get_treelite explicitly because we need the correct
-  # ${TREELITE_LIBS} definition for RF
   include(../cpp/cmake/thirdparty/get_treelite.cmake)
   find_package(cuml ${CUML_VERSION} REQUIRED)
-
 else()
   set(cuml_FOUND OFF)
 endif()
 
 include(rapids-cython)
 
-if(NOT cuml_FOUND)
-  # TODO: This will not be necessary once we upgrade to CMake 3.22, which will pull in the required
-  # languages for the C++ project even if this project does not require those languges.
-  include(rapids-cuda)
-  rapids_cuda_init_architectures(cuml-python)
-  enable_language(CUDA)
-
-  # Since cuml only enables CUDA optionally, we need to manually include the file that
-  # rapids_cuda_init_architectures relies on `project` including.
-  include("${CMAKE_PROJECT_cuml-python_INCLUDE}")
+if(CUML_BUILD_WHEELS)
+  set(CUML_PYTHON_TREELITE_TARGET treelite::treelite_static)
+else()
+  set(CUML_PYTHON_TREELITE_TARGET treelite::treelite)
+endif()
 
+if(NOT cuml_FOUND)
   set(BUILD_CUML_TESTS OFF)
   set(BUILD_PRIMS_TESTS OFF)
   set(BUILD_CUML_C_LIBRARY OFF)
   set(BUILD_CUML_EXAMPLES OFF)
   set(BUILD_CUML_BENCH OFF)
   set(BUILD_CUML_PRIMS_BENCH OFF)
-  message(STATUS "installing packages")
-  add_subdirectory(../cpp cuml-cpp)
+  set(CUML_EXPORT_TREELITE_LINKAGE ON)
+
+  set(_exclude_from_all "")
+  if(CUML_BUILD_WHEELS)
+    # Statically link dependencies if building wheels
+    set(CUDA_STATIC_RUNTIME ON)
+    set(CUML_USE_RAFT_STATIC ON)
+    set(CUML_USE_FAISS_STATIC ON)
+    set(CUML_USE_TREELITE_STATIC ON)
+    set(CUML_USE_CUMLPRIMS_MG_STATIC ON)
+    # Don't install the static libs into wheels
+    set(CUML_EXCLUDE_RAFT_FROM_ALL ON)
+    set(RAFT_EXCLUDE_FAISS_FROM_ALL ON)
+    set(CUML_EXCLUDE_TREELITE_FROM_ALL ON)
+    set(CUML_EXCLUDE_CUMLPRIMS_MG_FROM_ALL ON)
+
+    # Don't install the cuML C++ targets into wheels
+    set(_exclude_from_all EXCLUDE_FROM_ALL)
+  endif()
+
+  add_subdirectory(../cpp cuml-cpp ${_exclude_from_all})
 
   set(cython_lib_dir cuml)
   install(TARGETS ${CUML_CPP_TARGET} DESTINATION ${cython_lib_dir})
 endif()
 
 set(cuml_sg_libraries cuml::${CUML_CPP_TARGET})
+set(cuml_mg_libraries cuml::${CUML_CPP_TARGET})
 
 if(NOT SINGLEGPU)
   include(../cpp/cmake/thirdparty/get_cumlprims_mg.cmake)
diff --git a/python/LICENSE b/python/LICENSE
new file mode 120000
index 0000000000..ea5b60640b
--- /dev/null
+++ b/python/LICENSE
@@ -0,0 +1 @@
+../LICENSE
\ No newline at end of file
diff --git a/python/_custom_build/backend.py b/python/_custom_build/backend.py
new file mode 100644
index 0000000000..46a7d79855
--- /dev/null
+++ b/python/_custom_build/backend.py
@@ -0,0 +1,40 @@
+# Copyright (c) 2022, NVIDIA CORPORATION.
+
+"""Custom build backend for cuml to get versioned requirements.
+
+Based on https://setuptools.pypa.io/en/latest/build_meta.html
+"""
+import os
+from functools import wraps
+
+from setuptools import build_meta as _orig
+
+# Alias the required bits
+prepare_metadata_for_build_wheel = _orig.prepare_metadata_for_build_wheel
+build_wheel = _orig.build_wheel
+build_sdist = _orig.build_sdist
+
+
+def replace_requirements(func):
+    @wraps(func)
+    def wrapper(config_settings=None):
+        orig_list = getattr(_orig, func.__name__)(config_settings)
+        suffix = os.getenv('RAPIDS_PY_WHEEL_CUDA_SUFFIX', default='')
+        append_list = [
+            f"rmm{suffix}",
+            f"pylibraft{suffix}",
+        ]
+        return orig_list + append_list
+
+    return wrapper
+
+
+get_requires_for_build_wheel = replace_requirements(
+    _orig.get_requires_for_build_wheel
+)
+get_requires_for_build_sdist = replace_requirements(
+    _orig.get_requires_for_build_sdist
+)
+get_requires_for_build_editable = replace_requirements(
+    _orig.get_requires_for_build_editable
+)
diff --git a/python/cuml/cluster/CMakeLists.txt b/python/cuml/cluster/CMakeLists.txt
index 53e0d4986f..de5a23b655 100644
--- a/python/cuml/cluster/CMakeLists.txt
+++ b/python/cuml/cluster/CMakeLists.txt
@@ -30,7 +30,7 @@ add_subdirectory(hdbscan)
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX cluster_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/cluster/hdbscan/CMakeLists.txt b/python/cuml/cluster/hdbscan/CMakeLists.txt
index 0a242a5768..2b7722eae6 100644
--- a/python/cuml/cluster/hdbscan/CMakeLists.txt
+++ b/python/cuml/cluster/hdbscan/CMakeLists.txt
@@ -19,7 +19,7 @@ set(cython_sources
 rapids_cython_create_modules(
         CXX
         SOURCE_FILES "${cython_sources}"
-        LINKED_LIBRARIES ${cuml_sg_libraries}
+        LINKED_LIBRARIES "${cuml_sg_libraries}"
         MODULE_PREFIX cluster_hdbscan_
         ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/common/CMakeLists.txt b/python/cuml/common/CMakeLists.txt
index 5e49f65a8a..ea81f3c623 100644
--- a/python/cuml/common/CMakeLists.txt
+++ b/python/cuml/common/CMakeLists.txt
@@ -29,7 +29,7 @@ endif()
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX common_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/common/input_utils.py b/python/cuml/common/input_utils.py
index 971dac4aa9..d450d3eb5b 100644
--- a/python/cuml/common/input_utils.py
+++ b/python/cuml/common/input_utils.py
@@ -314,7 +314,10 @@ def check_order(arr_order):
 
     # format conversion
 
-    if isinstance(X, (dask_cudf.core.Series, dask_cudf.core.DataFrame)):
+    if (
+        has_dask_cudf()
+        and isinstance(X, (dask_cudf.core.Series, dask_cudf.core.DataFrame))
+    ):
         # TODO: Warn, but not when using dask_sql
         X = X.compute()
 
@@ -573,7 +576,10 @@ def convert_dtype(X,
     if the conversion would lose information.
     """
 
-    if isinstance(X, (dask_cudf.core.Series, dask_cudf.core.DataFrame)):
+    if (
+        has_dask_cudf()
+        and isinstance(X, (dask_cudf.core.Series, dask_cudf.core.DataFrame))
+    ):
         # TODO: Warn, but not when using dask_sql
         X = X.compute()
 
diff --git a/python/cuml/decomposition/CMakeLists.txt b/python/cuml/decomposition/CMakeLists.txt
index 668952a178..70bcdb0de5 100644
--- a/python/cuml/decomposition/CMakeLists.txt
+++ b/python/cuml/decomposition/CMakeLists.txt
@@ -29,7 +29,7 @@ endif()
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX decomposition_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/ensemble/CMakeLists.txt b/python/cuml/ensemble/CMakeLists.txt
index 34c8456bc0..957440a4e0 100644
--- a/python/cuml/ensemble/CMakeLists.txt
+++ b/python/cuml/ensemble/CMakeLists.txt
@@ -12,8 +12,6 @@
 # the License.
 # =============================================================================
 
-
-
 set(cython_sources
     randomforest_common.pyx
     randomforest_shared.pyx
@@ -21,7 +19,7 @@ set(cython_sources
     randomforestregressor.pyx
 )
 
-set(linked_libraries 
+set(linked_libraries
     ${cuml_sg_libraries}
     ${TREELITE_LIBS})
 
diff --git a/python/cuml/explainer/CMakeLists.txt b/python/cuml/explainer/CMakeLists.txt
index 17fb835f49..70f41bd14f 100644
--- a/python/cuml/explainer/CMakeLists.txt
+++ b/python/cuml/explainer/CMakeLists.txt
@@ -12,19 +12,21 @@
 # the License.
 # =============================================================================
 
-
-
 set(cython_sources
     base.pyx
     kernel_shap.pyx
     permutation_shap.pyx
     tree_shap.pyx
 )
+set(linked_libraries
+  "${cuml_sg_libraries}"
+  "${CUML_PYTHON_TREELITE_TARGET}"
+)
 
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES "${cuml_sg_libraries}"
+  LINKED_LIBRARIES "${linked_libraries}"
   MODULE_PREFIX explainer_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/fil/CMakeLists.txt b/python/cuml/fil/CMakeLists.txt
index 69eac4d2ea..b720caac78 100644
--- a/python/cuml/fil/CMakeLists.txt
+++ b/python/cuml/fil/CMakeLists.txt
@@ -17,11 +17,15 @@
 set(cython_sources
     fil.pyx
 )
+set(linked_libraries
+  "${cuml_sg_libraries}"
+  "${CUML_PYTHON_TREELITE_TARGET}"
+)
 
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES "${cuml_sg_libraries}"
+  LINKED_LIBRARIES "${linked_libraries}"
   MODULE_PREFIX fil_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/linear_model/CMakeLists.txt b/python/cuml/linear_model/CMakeLists.txt
index f9e6981d73..6bf5b9dc95 100644
--- a/python/cuml/linear_model/CMakeLists.txt
+++ b/python/cuml/linear_model/CMakeLists.txt
@@ -35,7 +35,7 @@ endif()
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX linear_model_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/metrics/cluster/CMakeLists.txt b/python/cuml/metrics/cluster/CMakeLists.txt
index 179f6c70f3..d678de583e 100644
--- a/python/cuml/metrics/cluster/CMakeLists.txt
+++ b/python/cuml/metrics/cluster/CMakeLists.txt
@@ -31,4 +31,3 @@ rapids_cython_create_modules(
   LINKED_LIBRARIES "${cuml_sg_libraries}"
   ASSOCIATED_TARGETS cuml
 )
-
diff --git a/python/cuml/neighbors/CMakeLists.txt b/python/cuml/neighbors/CMakeLists.txt
index 3af0257f5e..cd3e671318 100644
--- a/python/cuml/neighbors/CMakeLists.txt
+++ b/python/cuml/neighbors/CMakeLists.txt
@@ -31,7 +31,7 @@ endif()
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX neighbors_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/solvers/CMakeLists.txt b/python/cuml/solvers/CMakeLists.txt
index e8b2137cd8..15404c372d 100644
--- a/python/cuml/solvers/CMakeLists.txt
+++ b/python/cuml/solvers/CMakeLists.txt
@@ -29,7 +29,7 @@ endif()
 rapids_cython_create_modules(
   CXX
   SOURCE_FILES "${cython_sources}"
-  LINKED_LIBRARIES $<IF:$<BOOL:${SINGLE_GPU}>,${cuml_sg_libraries},${cuml_mg_libraries}>
+  LINKED_LIBRARIES "${cuml_mg_libraries}"
   MODULE_PREFIX solvers_
   ASSOCIATED_TARGETS cuml
 )
diff --git a/python/cuml/tests/conftest.py b/python/cuml/tests/conftest.py
index 0515c952fa..686acd4d06 100644
--- a/python/cuml/tests/conftest.py
+++ b/python/cuml/tests/conftest.py
@@ -100,7 +100,7 @@ def pytest_collection_modifyitems(config, items):
     # Mark the tests as "skip" if needed
     if not should_run_unit:
         skip_unit = pytest.mark.skip(
-            reason="Stress tests run with --run_unit flag.")
+            reason="Unit tests run with --run_unit flag.")
         for item in items:
             if "unit" in item.keywords:
                 item.add_marker(skip_unit)
diff --git a/python/cuml/tests/test_doctest.py b/python/cuml/tests/test_doctest.py
index 1ae20238db..6ff940af09 100644
--- a/python/cuml/tests/test_doctest.py
+++ b/python/cuml/tests/test_doctest.py
@@ -104,7 +104,20 @@ def test_docstring(docstring):
     with contextlib.redirect_stdout(doctest_stdout):
         runner.run(docstring)
         results = runner.summarize()
-    assert not results.failed, (
-        f"{results.failed} of {results.attempted} doctests failed for "
-        f"{docstring.name}:\n{doctest_stdout.getvalue()}"
-    )
+    try:
+        assert not results.failed, (
+            f"{results.failed} of {results.attempted} doctests failed for "
+            f"{docstring.name}:\n{doctest_stdout.getvalue()}"
+        )
+    except AssertionError:
+        # If some failed but all the failures were due to lack of multiGPU
+        # support, we can skip. This code assumes that any MG-related failures
+        # mean that all doctests failed for that reason, which is heavy-handed
+        # and could miss a few things but is much easier than trying to
+        # identify every line corresponding to any Exception raised.
+        if (
+            "cuML has not been built with multiGPU support"
+            in doctest_stdout.getvalue()
+        ):
+            pytest.skip("Doctest requires MG support.")
+        raise
diff --git a/python/pyproject.toml b/python/pyproject.toml
index 0261c0b09d..2afcab578c 100644
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -21,4 +21,9 @@ requires = [
     "scikit-build>=0.13.1",
     "cmake>=3.23.1,!=3.25.0",
     "ninja",
+    "treelite==3.0.1",
+    "treelite_runtime==3.0.1",
+    "versioneer",
 ]
+build-backend = "backend"
+backend-path = ["_custom_build"]
diff --git a/python/setup.py b/python/setup.py
index e17a513218..65cb110322 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -14,6 +14,7 @@
 # limitations under the License.
 #
 
+import glob
 import os
 import shutil
 import sys
@@ -21,36 +22,50 @@
 
 from setuptools import find_packages
 
-from setuputils import clean_folder
-from setuputils import get_environment_option
-from setuputils import get_cli_option
-
 import versioneer
 from skbuild import setup
 
-install_requires = ['numba', 'cython']
 
 ##############################################################################
-# - Print of build options used by setup.py  --------------------------------
+# - Helper functions
+def get_cli_option(name):
+    if name in sys.argv:
+        print("-- Detected " + str(name) + " build option.")
+        return True
 
-cuda_home = get_environment_option("CUDA_HOME")
-libcuml_path = get_environment_option('CUML_BUILD_PATH')
+    else:
+        return False
 
-clean_artifacts = get_cli_option('clean')
 
-##############################################################################
-# - Dependencies include and lib folder setup --------------------------------
+def clean_folder(path):
+    """
+    Function to clean all Cython and Python artifacts and cache folders. It
+    cleans the folder as well as its direct children recursively.
+
+    Parameters
+    ----------
+    path : String
+        Path to the folder to be cleaned.
+    """
+    shutil.rmtree(path + '/__pycache__', ignore_errors=True)
+
+    folders = glob.glob(path + '/*/')
+    for folder in folders:
+        shutil.rmtree(folder + '/__pycache__', ignore_errors=True)
+
+        clean_folder(folder)
 
-if not cuda_home:
-    nvcc_path = shutil.which('nvcc')
-    if (not nvcc_path):
-        raise FileNotFoundError("nvcc not found.")
+        cython_exts = glob.glob(folder + '/*.cpp')
+        cython_exts.extend(glob.glob(folder + '/*.cpython*'))
+        for file in cython_exts:
+            os.remove(file)
 
-    cuda_home = str(Path(nvcc_path).parent.parent)
-    print("-- Using nvcc to detect CUDA, found at " + str(cuda_home))
 
-cuda_include_dir = os.path.join(cuda_home, "include")
-cuda_lib_dir = os.path.join(cuda_home, "lib64")
+##############################################################################
+# - Print of build options used by setup.py  --------------------------------
+
+clean_artifacts = get_cli_option('clean')
+
 
 ##############################################################################
 # - Clean target -------------------------------------------------------------
@@ -60,7 +75,6 @@
 
     # Reset these paths since they may be deleted below
     treelite_path = False
-    libcuml_path = False
 
     try:
         setup_file_path = str(Path(__file__).parent.absolute())
@@ -89,31 +103,69 @@
         sys.exit(0)
 
 
-if not libcuml_path:
-    libcuml_path = '../cpp/build/'
-
-cmdclass = versioneer.get_cmdclass()
-
 ##############################################################################
 # - Python package generation ------------------------------------------------
 
-setup(name='cuml',
+# Make versioneer produce PyPI-compatible nightly versions for wheels.
+if "RAPIDS_PY_WHEEL_VERSIONEER_OVERRIDE" in os.environ:
+    orig_get_versions = versioneer.get_versions
+
+    version_override = os.environ["RAPIDS_PY_WHEEL_VERSIONEER_OVERRIDE"]
+
+    def get_versions():
+        data = orig_get_versions()
+        data["version"] = version_override
+        return data
+
+    versioneer.get_versions = get_versions
+
+
+cuda_suffix = os.getenv("RAPIDS_PY_WHEEL_CUDA_SUFFIX", default="")
+setup(name=f'cuml{cuda_suffix}',
+      version=os.getenv("RAPIDS_PY_WHEEL_VERSIONEER_OVERRIDE",
+                        default=versioneer.get_version()),
       description="cuML - RAPIDS ML Algorithms",
-      version=versioneer.get_version(),
+      url="https://github.com/rapidsai/cuml",
+      author="NVIDIA Corporation",
+      license="Apache",
       classifiers=[
           "Intended Audience :: Developers",
           "Programming Language :: Python",
           "Programming Language :: Python :: 3.8",
           "Programming Language :: Python :: 3.9"
       ],
-      author="NVIDIA Corporation",
-      url="https://github.com/rapidsai/cuml",
-      setup_requires=['Cython>=0.29,<0.30'],
+      cmdclass=versioneer.get_cmdclass(),
+      include_package_data=True,
       packages=find_packages(include=['cuml', 'cuml.*']),
       package_data={
           key: ["*.pxd"] for key in find_packages(include=['cuml', 'cuml.*'])
       },
-      install_requires=install_requires,
-      license="Apache",
-      cmdclass=cmdclass,
+      install_requires=[
+        "numba",
+        "scipy",
+        "seaborn",
+        "treelite==3.0.1",
+        "treelite_runtime==3.0.1",
+        f"cudf{cuda_suffix}",
+        f"dask-cudf{cuda_suffix}",
+        f"pylibraft{cuda_suffix}",
+        f"raft-dask{cuda_suffix}",
+      ],
+      extras_require={
+          "test": [
+              "pytest",
+              "hypothesis",
+              "pytest-xdist",
+              "pytest-benchmark",
+              "nltk",
+              "dask-ml",
+              "numpydoc",
+              "umap-learn",
+              "statsmodels",
+              "scikit-learn==0.24",
+              "hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master",  # noqa:E501
+              "dask-glm @ git+https://github.com/dask/dask-glm@main",
+              "dask-cuda"
+          ]
+      },
       zip_safe=False)
diff --git a/python/setuputils.py b/python/setuputils.py
deleted file mode 100644
index a8085802be..0000000000
--- a/python/setuputils.py
+++ /dev/null
@@ -1,86 +0,0 @@
-#
-# 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.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-
-import glob
-import os
-import re
-import shutil
-import sys
-
-
-def get_environment_option(name):
-    env_variable = os.environ.get(name, False)
-
-    if not env_variable:
-        print("-- " + name + " environment variable not set.")
-
-    else:
-        print("-- " + name + " detected with value: " + str(env_variable))
-
-    return env_variable
-
-
-def get_cli_option(name):
-    if name in sys.argv:
-        print("-- Detected " + str(name) + " build option.")
-        return True
-
-    else:
-        return False
-
-
-def clean_folder(path):
-    """
-    Function to clean all Cython and Python artifacts and cache folders. It
-    clean the folder as well as its direct children recursively.
-
-    Parameters
-    ----------
-    path : String
-        Path to the folder to be cleaned.
-    """
-    shutil.rmtree(path + '/__pycache__', ignore_errors=True)
-
-    folders = glob.glob(path + '/*/')
-    for folder in folders:
-        shutil.rmtree(folder + '/__pycache__', ignore_errors=True)
-
-        clean_folder(folder)
-
-        cython_exts = glob.glob(folder + '/*.cpp')
-        cython_exts.extend(glob.glob(folder + '/*.cpython*'))
-        for file in cython_exts:
-            os.remove(file)
-
-
-def get_cuda_version_from_header(cuda_include_dir, delimiter=""):
-
-    cuda_version = None
-
-    with open(os.path.join(cuda_include_dir, "cuda.h"), encoding="utf-8") as f:
-        for line in f.readlines():
-            if re.search(r"#define CUDA_VERSION ", line) is not None:
-                cuda_version = line
-                break
-
-    if cuda_version is None:
-        raise TypeError("CUDA_VERSION not found in cuda.h")
-    cuda_version = int(cuda_version.split()[2])
-    return "%d%s%d" % (
-        cuda_version // 1000,
-        delimiter,
-        (cuda_version % 1000) // 10,
-    )