Enable cuml-cpu nightly

build.sh

@@ -27,6 +27,7 @@ HELP="$0 [<target> ...] [<flag> ...]
    libcuml           - build the cuml C++ code only. Also builds the C-wrapper library
                        around the C++ code.
    cuml              - build the cuml Python package
+   cuml-cpu          - build the cuml CPU Python package
    cpp-mgtests       - build libcuml mnmg tests. Builds MPI communicator, adding MPI as dependency.
    prims             - build the ml-prims tests
    bench             - build the libcuml C++ benchmark
@@ -294,3 +295,9 @@ if (! hasArg --configure-only) && (completeBuild || hasArg cuml || hasArg pydocs
         make html
     fi
 fi
+
+if hasArg cuml-cpu; then
+    SKBUILD_CONFIGURE_OPTIONS="${CUML_EXTRA_CMAKE_ARGS} -DCUML_CPU=ON -DFIND_CUML_CPP=OFF" \
+        SKBUILD_BUILD_OPTIONS="-j${PARALLEL_LEVEL}" \
+        python -m pip install --no-build-isolation --no-deps ${REPODIR}/python
+fi

ci/build_python.sh

@@ -20,4 +20,13 @@ rapids-conda-retry mambabuild \
   --channel "${CPP_CHANNEL}" \
   conda/recipes/cuml
 
+# Build cuml-cpu only in CUDA 11 jobs since it only depends on python
+# version
+RAPIDS_CUDA_MAJOR="${RAPIDS_CUDA_VERSION%%.*}"
+if [[ ${RAPIDS_CUDA_MAJOR} == "11" ]]; then
+  rapids-conda-retry mambabuild \
+  --no-test \
+  conda/recipes/cuml-cpu
+fi
+
 rapids-upload-conda-to-s3 python

conda/recipes/cuml-cpu/build.sh

@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+# Copyright (c) 2023, NVIDIA CORPORATION.
+
+# This assumes the script is executed from the root of the repo directory
+./build.sh cuml-cpu -v

conda/recipes/cuml-cpu/conda_build_config.yaml

@@ -0,0 +1,11 @@
+c_compiler_version:
+  - 11
+
+cxx_compiler_version:
+  - 11
+
+cmake_version:
+  - ">=3.26.4"
+
+sysroot_version:
+  - "=2.17"

conda/recipes/cuml-cpu/meta.yaml

@@ -0,0 +1,50 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.
+
+# Usage:
+#   conda build . -c conda-forge -c numba -c rapidsai -c pytorch
+{% set version = environ.get('GIT_DESCRIBE_TAG', '0.0.0.dev').lstrip('v') + environ.get('VERSION_SUFFIX', '') %}
+{% set py_version = environ['CONDA_PY'] %}
+{% set date_string = environ['RAPIDS_DATE_STRING'] %}
+
+package:
+  name: cuml-cpu
+  version: {{ version }}
+
+source:
+  git_url: ../../..
+
+build:
+  number: {{ GIT_DESCRIBE_NUMBER }}
+  string: py{{ py_version }}_{{ date_string }}_{{ GIT_DESCRIBE_HASH }}_{{ GIT_DESCRIBE_NUMBER }}
+  script_env:
+    - VERSION_SUFFIX
+
+requirements:
+  build:
+    - cmake {{ cmake_version }}
+    - {{ compiler('c') }}
+    - {{ compiler('cxx') }}
+    - sysroot_{{ target_platform }} {{ sysroot_version }}
+    - ninja
+  host:
+    - python x.x
+    - setuptools
+    - scikit-build>=0.13.1
+    - cython>=3.0.0
+  run:
+    - python x.x
+    - numpy
+    - scikit-learn=1.2
+    - hdbscan<=0.8.30
+    - umap-learn=0.5.3
+    - nvtx
+
+tests:                                 # [linux64]
+  imports:                             # [linux64]
+    - cuml-cpu                         # [linux64]
+
+about:
+  home: http://rapids.ai/
+  license: Apache-2.0
+  # license_file: LICENSE
+  summary: cuML-CPU library

python/CMakeLists.txt

@@ -18,10 +18,16 @@ include(../fetch_rapids.cmake)
 
 set(CUML_VERSION 23.10.00)
 
-# 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)
+option(CUML_CPU "Build only cuML CPU Python components." OFF)
+set(language_list "C;CXX")
+
+if(NOT CUML_CPU)
+  # We always need CUDA for cuML GPU because the raft dependency brings in a
+  # header-only cuco dependency that enables CUDA unconditionally.
+  include(rapids-cuda)
+  rapids_cuda_init_architectures(cuml-python)
+  list(APPEND language_list "CUDA")
+endif()
 
 project(
   cuml-python
@@ -30,14 +36,12 @@ 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 CUDA
+            ${language_list}
 )
 
 ################################################################################
 # - User Options  --------------------------------------------------------------
 option(CUML_UNIVERSAL "Build all cuML Python components." ON)
-option(CUML_CPU "Build only cuML CPU Python components." OFF)
-option(CUML_GPU "Build only cuML GPU Python components. Not supported yet, defaulting to CUML_UNIVERSAL" OFF)
 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)
@@ -80,7 +84,7 @@ else()
   set(CUML_PYTHON_TREELITE_TARGET treelite::treelite)
 endif()
 
-if(NOT cuml_FOUND)
+if(NOT cuml_FOUND AND NOT ${CUML_CPU})
   set(BUILD_CUML_TESTS OFF)
   set(BUILD_PRIMS_TESTS OFF)
   set(BUILD_CUML_C_LIBRARY OFF)
@@ -118,10 +122,28 @@ if(CUML_CPU)
   set(SINGLEGPU ON)
 
   set(CUML_ALGORITHMS "linearregression")
-endif()
+  list(APPEND CUML_ALGORITHMS "pca")
+  list(APPEND CUML_ALGORITHMS "tsvd")
+  list(APPEND CUML_ALGORITHMS "elasticnet")
+  list(APPEND CUML_ALGORITHMS "logisticregression")
+  list(APPEND CUML_ALGORITHMS "ridge")
+  list(APPEND CUML_ALGORITHMS "lasso")
+  list(APPEND CUML_ALGORITHMS "umap")
+  list(APPEND CUML_ALGORITHMS "knn")
+  list(APPEND CUML_ALGORITHMS "hdbscan")
+
+  # this won't be needed when we add CPU libcuml++ (FIL)
+  set(cuml_sg_libraries "")
+
+  list(APPEND CYTHON_FLAGS
+  "--compile-time-env GPUBUILD=0")
+else()
+  set(cuml_sg_libraries cuml::${CUML_CPP_TARGET})
+  set(cuml_mg_libraries cuml::${CUML_CPP_TARGET})
 
-set(cuml_sg_libraries cuml::${CUML_CPP_TARGET})
-set(cuml_mg_libraries cuml::${CUML_CPP_TARGET})
+  list(APPEND CYTHON_FLAGS
+  "--compile-time-env GPUBUILD=1")
+endif()
 
 if(NOT SINGLEGPU)
   include("${CUML_CPP_SRC}/cmake/thirdparty/get_cumlprims_mg.cmake")
@@ -131,9 +153,20 @@ if(NOT SINGLEGPU)
   )
 endif()
 
+ ################################################################################
+ # - Build Cython artifacts -----------------------------------------------------
+
+include("${CUML_CPP_SRC}/cmake/modules/ConfigureAlgorithms.cmake")
+include(cmake/ConfigureCythonAlgorithms.cmake)
+
+if(${CUML_CPU})
+  # libcuml requires metrics built if HDSCAN is built, which is not the case
+  # for cuml-cpu
+  unset(metrics_algo)
+endif()
+
 rapids_cython_init()
 
-include(cmake/ConfigureAlgorithmsHelpers.cmake)
 
 add_subdirectory(cuml/common)
 add_subdirectory(cuml/internals)

python/cmake/ConfigureAlgorithmsHelpers.cmake → python/cmake/ConfigureCythonAlgorithms.cmake

@@ -14,19 +14,6 @@
 
 
 function(add_module_gpu_default FILENAME)
-    set (extra_args ${ARGN})
-    list(LENGTH extra_args extra_count)
-    if (${extra_count} GREATER 0 OR
-        ${CUML_UNIVERSAL} OR
-        ${CUML_GPU})
-      list(APPEND cython_sources
-           ${FILENAME})
-      set (cython_sources ${cython_sources} PARENT_SCOPE)
-    endif()
-endfunction()
-
-
-function(add_module_universal_default FILENAME)
     set (extra_args ${ARGN})
     list(LENGTH extra_args extra_count)
     if (${extra_count} GREATER 0 OR

python/cuml/__init__.py

@@ -15,96 +15,105 @@
 #
 
 from cuml.internals.base import Base, UniversalBase
+from cuml.internals.available_devices import is_cuda_available
 
 # GPU only packages
 
-import cuml.common.cuda as cuda
-from cuml.common.handle import Handle
+if is_cuda_available():
+    import cuml.common.cuda as cuda
+    from cuml.common.handle import Handle
 
-from cuml.cluster.dbscan import DBSCAN
-from cuml.cluster.kmeans import KMeans
-from cuml.cluster.agglomerative import AgglomerativeClustering
-from cuml.cluster.hdbscan import HDBSCAN
+    from cuml.cluster.dbscan import DBSCAN
+    from cuml.cluster.kmeans import KMeans
+    from cuml.cluster.agglomerative import AgglomerativeClustering
 
-from cuml.datasets.arima import make_arima
-from cuml.datasets.blobs import make_blobs
-from cuml.datasets.regression import make_regression
-from cuml.datasets.classification import make_classification
+    from cuml.datasets.arima import make_arima
+    from cuml.datasets.blobs import make_blobs
+    from cuml.datasets.regression import make_regression
+    from cuml.datasets.classification import make_classification
 
-from cuml.decomposition.pca import PCA
-from cuml.decomposition.tsvd import TruncatedSVD
-from cuml.decomposition.incremental_pca import IncrementalPCA
+    from cuml.decomposition.incremental_pca import IncrementalPCA
 
-from cuml.fil.fil import ForestInference
+    from cuml.fil.fil import ForestInference
 
-from cuml.ensemble.randomforestclassifier import RandomForestClassifier
-from cuml.ensemble.randomforestregressor import RandomForestRegressor
+    from cuml.ensemble.randomforestclassifier import RandomForestClassifier
+    from cuml.ensemble.randomforestregressor import RandomForestRegressor
 
-from cuml.explainer.kernel_shap import KernelExplainer
-from cuml.explainer.permutation_shap import PermutationExplainer
-from cuml.explainer.tree_shap import TreeExplainer
+    from cuml.explainer.kernel_shap import KernelExplainer
+    from cuml.explainer.permutation_shap import PermutationExplainer
+    from cuml.explainer.tree_shap import TreeExplainer
 
-import cuml.feature_extraction
-from cuml.fil import fil
+    import cuml.feature_extraction
+    from cuml.fil import fil
 
-from cuml.internals.global_settings import (
-    GlobalSettings,
-    _global_settings_data,
-)
+    from cuml.kernel_ridge.kernel_ridge import KernelRidge
 
-from cuml.kernel_ridge.kernel_ridge import KernelRidge
+    from cuml.linear_model.mbsgd_classifier import MBSGDClassifier
+    from cuml.linear_model.mbsgd_regressor import MBSGDRegressor
 
-from cuml.linear_model.elastic_net import ElasticNet
-from cuml.linear_model.lasso import Lasso
-from cuml.linear_model.logistic_regression import LogisticRegression
-from cuml.linear_model.mbsgd_classifier import MBSGDClassifier
-from cuml.linear_model.mbsgd_regressor import MBSGDRegressor
-from cuml.linear_model.ridge import Ridge
+    from cuml.manifold.t_sne import TSNE
+    from cuml.metrics.accuracy import accuracy_score
+    from cuml.metrics.cluster.adjusted_rand_index import adjusted_rand_score
+    from cuml.metrics.regression import r2_score
+    from cuml.model_selection import train_test_split
 
-from cuml.manifold.t_sne import TSNE
-from cuml.manifold.umap import UMAP
-from cuml.metrics.accuracy import accuracy_score
-from cuml.metrics.cluster.adjusted_rand_index import adjusted_rand_score
-from cuml.metrics.regression import r2_score
-from cuml.model_selection import train_test_split
+    from cuml.naive_bayes.naive_bayes import MultinomialNB
 
-from cuml.naive_bayes.naive_bayes import MultinomialNB
+    from cuml.neighbors.nearest_neighbors import NearestNeighbors
+    from cuml.neighbors.kernel_density import KernelDensity
+    from cuml.neighbors.kneighbors_classifier import KNeighborsClassifier
+    from cuml.neighbors.kneighbors_regressor import KNeighborsRegressor
 
-from cuml.neighbors.nearest_neighbors import NearestNeighbors
-from cuml.neighbors.kernel_density import KernelDensity
-from cuml.neighbors.kneighbors_classifier import KNeighborsClassifier
-from cuml.neighbors.kneighbors_regressor import KNeighborsRegressor
+    from cuml.preprocessing.LabelEncoder import LabelEncoder
 
-from cuml.preprocessing.LabelEncoder import LabelEncoder
+    from cuml.random_projection.random_projection import (
+        GaussianRandomProjection,
+    )
+    from cuml.random_projection.random_projection import SparseRandomProjection
+    from cuml.random_projection.random_projection import (
+        johnson_lindenstrauss_min_dim,
+    )
 
-from cuml.random_projection.random_projection import GaussianRandomProjection
-from cuml.random_projection.random_projection import SparseRandomProjection
-from cuml.random_projection.random_projection import (
-    johnson_lindenstrauss_min_dim,
-)
+    from cuml.svm import SVC
+    from cuml.svm import SVR
+    from cuml.svm import LinearSVC
+    from cuml.svm import LinearSVR
 
-from cuml.solvers.cd import CD
-from cuml.solvers.sgd import SGD
-from cuml.solvers.qn import QN
-from cuml.svm import SVC
-from cuml.svm import SVR
-from cuml.svm import LinearSVC
-from cuml.svm import LinearSVR
+    from cuml.tsa import stationarity
+    from cuml.tsa.arima import ARIMA
+    from cuml.tsa.auto_arima import AutoARIMA
+    from cuml.tsa.holtwinters import ExponentialSmoothing
+
+    from cuml.common.pointer_utils import device_of_gpu_matrix
+
+# Universal packages
 
-from cuml.tsa import stationarity
-from cuml.tsa.arima import ARIMA
-from cuml.tsa.auto_arima import AutoARIMA
-from cuml.tsa.holtwinters import ExponentialSmoothing
+from cuml.internals.global_settings import (
+    GlobalSettings,
+    _global_settings_data,
+)
 
-from cuml.common.pointer_utils import device_of_gpu_matrix
 from cuml.internals.memory_utils import (
     set_global_output_type,
     using_output_type,
 )
 
-# Universal packages
+from cuml.cluster.hdbscan import HDBSCAN
+
+from cuml.decomposition.pca import PCA
+from cuml.decomposition.tsvd import TruncatedSVD
 
 from cuml.linear_model.linear_regression import LinearRegression
+from cuml.linear_model.elastic_net import ElasticNet
+from cuml.linear_model.lasso import Lasso
+from cuml.linear_model.logistic_regression import LogisticRegression
+from cuml.linear_model.ridge import Ridge
+
+from cuml.manifold.umap import UMAP
+
+from cuml.solvers.cd import CD
+from cuml.solvers.sgd import SGD
+from cuml.solvers.qn import QN
 
 # Version configuration
 __version__ = "23.10.00"