initial work on adding wheel build for libcudf

.github/workflows/pr.yaml

@@ -23,6 +23,7 @@ jobs:
       - static-configure
       - conda-notebook-tests
       - docs-build
+      - wheel-build-libcudf
       - wheel-build-cudf
       - wheel-tests-cudf
       - wheel-build-dask-cudf
@@ -118,10 +119,18 @@ jobs:
       arch: "amd64"
       container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/build_docs.sh"
-  wheel-build-cudf:
+  wheel-build-libcudf:
     needs: checks
     secrets: inherit
     uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@branch-24.06
+    with:
+      matrix_filter: group_by([.ARCH, (.CUDA_VER|split(".")|map(tonumber)|.[0])]) | map(max_by(.PY_VER|split(".")|map(tonumber)))
+      build_type: pull-request
+      script: "ci/build_wheel_libcudf.sh"
+  wheel-build-cudf:
+    needs: [checks, wheel-build-libcudf]
+    secrets: inherit
+    uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@branch-24.06
     with:
       build_type: pull-request
       script: "ci/build_wheel_cudf.sh"

build.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-# Copyright (c) 2019-2023, NVIDIA CORPORATION.
+# Copyright (c) 2019-2024, NVIDIA CORPORATION.
 
 # cuDF build script
 
@@ -17,12 +17,13 @@ ARGS=$*
 # script, and that this script resides in the repo dir!
 REPODIR=$(cd $(dirname $0); pwd)
 
-VALIDARGS="clean libcudf cudf cudfjar dask_cudf benchmarks tests libcudf_kafka cudf_kafka custreamz -v -g -n --pydevelop -l --allgpuarch --disable_nvtx --opensource_nvcomp  --show_depr_warn --ptds -h --build_metrics --incl_cache_stats"
-HELP="$0 [clean] [libcudf] [cudf] [cudfjar] [dask_cudf] [benchmarks] [tests] [libcudf_kafka] [cudf_kafka] [custreamz] [-v] [-g] [-n] [-h] [--cmake-args=\\\"<args>\\\"]
+VALIDARGS="clean libcudf cudf libcudfwheel cudfjar dask_cudf benchmarks tests libcudf_kafka cudf_kafka custreamz -v -g -n --pydevelop -l --allgpuarch --disable_nvtx --opensource_nvcomp  --show_depr_warn --ptds -h --build_metrics --incl_cache_stats"
+HELP="$0 [clean] [libcudf] [cudf] [libcudfwheel] [cudfjar] [dask_cudf] [benchmarks] [tests] [libcudf_kafka] [cudf_kafka] [custreamz] [-v] [-g] [-n] [-h] [--cmake-args=\\\"<args>\\\"]
    clean                         - remove all existing build artifacts and configuration (start
                                    over)
    libcudf                       - build the cudf C++ code only
    cudf                          - build the cudf Python package
+   libcudfwheel                  - build the cudf C++ code packaged as a python wheel package
    cudfjar                       - build cudf JAR with static libcudf using devtoolset toolchain
    dask_cudf                     - build the dask_cudf Python package
    benchmarks                    - build benchmarks
@@ -333,7 +334,14 @@ if buildAll || hasArg libcudf; then
     fi
 fi
 
-# Build and install the cudf Python package
+if buildAll || hasArg libcudfwheel; then
+
+    cd ${REPODIR}/python/libcudf
+    SKBUILD_CMAKE_ARGS="-DCMAKE_PREFIX_PATH=${INSTALL_PREFIX};-DCMAKE_LIBRARY_PATH=${LIBCUDF_BUILD_DIR};-DCMAKE_CUDA_ARCHITECTURES=${CUDF_CMAKE_CUDA_ARCHITECTURES};${EXTRA_CMAKE_ARGS}" \
+        python ${PYTHON_ARGS_FOR_INSTALL} .
+fi
+
+# Build and install the cudf Python packages
 if buildAll || hasArg cudf; then
 
     cd ${REPODIR}/python/cudf

ci/build_wheel.sh

@@ -34,6 +34,8 @@ if ! rapids-is-release-build; then
     alpha_spec=',>=0.0.0a0'
 fi
 
+sed -r -i "s/libcudf==(.*)\"/libcudf${PACKAGE_CUDA_SUFFIX}==\1${alpha_spec}\"/g" ${pyproject_file}
+
 if [[ ${package_name} == "dask-cudf" ]]; then
     sed -r -i "s/cudf==(.*)\"/cudf${PACKAGE_CUDA_SUFFIX}==\1${alpha_spec}\"/g" ${pyproject_file}
     sed -r -i "s/dask-cuda==(.*)\"/dask-cuda==\1${alpha_spec}\"/g" ${pyproject_file}

ci/build_wheel_cudf.sh

@@ -3,14 +3,20 @@
 
 set -euo pipefail
 
+package_name="cudf"
 package_dir="python/cudf"
 
 export SKBUILD_CMAKE_ARGS="-DUSE_LIBARROW_FROM_PYARROW=ON"
 
-./ci/build_wheel.sh cudf ${package_dir}
+RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
+# Downloads libcudf wheel from this current build, then points pip to it in PIP_FIND_LINKS below
+RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 cpp /tmp/libcudf_dist
 
-python -m auditwheel repair -w ${package_dir}/final_dist ${package_dir}/dist/*
+export PIP_FIND_LINKS="/tmp/libcudf_dist"
+./ci/build_wheel.sh cudf ${package_dir}
 
+cd ${package_dir}
+mkdir -p final_dist
+python -m auditwheel repair --exclude libcudf.so --exclude libarrow.so.1400 -w final_dist dist/*
 
-RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-upload-wheels-to-s3 ${package_dir}/final_dist
+RAPIDS_PY_WHEEL_NAME="${package_name}_${RAPIDS_PY_CUDA_SUFFIX}" rapids-upload-wheels-to-s3 python final_dist

ci/build_wheel_libcudf.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.
+
+set -euo pipefail
+
+package_dir="python/libcudf"
+
+./ci/build_wheel.sh libcudf ${package_dir}
+
+RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
+RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-upload-wheels-to-s3 cpp ${package_dir}/dist

ci/cudf_pandas_scripts/pandas-tests/run.sh

@@ -11,8 +11,9 @@ rapids-logger "Running Pandas tests using $PANDAS_TESTS_BRANCH branch and rapids
 rapids-logger "PR number: ${RAPIDS_REF_NAME:-"unknown"}"
 
 RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-cudf-dep
-python -m pip install $(ls ./local-cudf-dep/cudf*.whl)[test,pandas-tests]
+RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 cpp ./local-cudf-dep
+RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 python ./local-cudf-dep
+python -m pip install --find-links $(pwd)/local-cudf-dep $(ls ./local-cudf-dep/cudf*.whl)[test,pandas-tests]
 
 RESULTS_DIR=${RAPIDS_TESTS_DIR:-"$(mktemp -d)"}
 RAPIDS_TESTS_DIR=${RAPIDS_TESTS_DIR:-"${RESULTS_DIR}/test-results"}/

ci/cudf_pandas_scripts/run_tests.sh

@@ -32,8 +32,9 @@ if [ "$no_cudf" = true ]; then
     echo "Skipping cudf install"
 else
     RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-    RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-cudf-dep
-    python -m pip install $(ls ./local-cudf-dep/cudf*.whl)[test,cudf-pandas-tests]
+    RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 cpp ./local-cudf-dep
+    RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 python ./local-cudf-dep
+    python -m pip install --find-links $(pwd)/local-cudf-dep $(ls ./local-cudf-dep/cudf*.whl)[test,cudf-pandas-tests]
 fi
 
 python -m pytest -p cudf.pandas ./python/cudf/cudf_pandas_tests/

ci/test_wheel_cudf.sh

@@ -4,10 +4,11 @@
 set -eou pipefail
 
 RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./dist
+RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 cpp ./dist
+RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 python ./dist
 
 # echo to expand wildcard before adding `[extra]` requires for pip
-python -m pip install $(echo ./dist/cudf*.whl)[test]
+python -m pip install --find-links $(pwd)/dist $(echo ./dist/cudf*.whl)[test]
 
 RESULTS_DIR=${RAPIDS_TESTS_DIR:-"$(mktemp -d)"}
 RAPIDS_TESTS_DIR=${RAPIDS_TESTS_DIR:-"${RESULTS_DIR}/test-results"}/

ci/test_wheel_dask_cudf.sh

@@ -7,8 +7,9 @@ RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
 RAPIDS_PY_WHEEL_NAME="dask_cudf_${RAPIDS_PY_CUDA_SUFFIX}" RAPIDS_PY_WHEEL_PURE="1" rapids-download-wheels-from-s3 ./dist
 
 # Download the cudf built in the previous step
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-cudf-dep
-python -m pip install --no-deps ./local-cudf-dep/cudf*.whl
+RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 cpp ./local-cudf-dep
+RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 python ./local-cudf-dep
+python -m pip install --no-deps --find-links $(pwd)/local-cudf-dep ./local-cudf-dep/cudf*.whl
 
 # echo to expand wildcard before adding `[extra]` requires for pip
 python -m pip install $(echo ./dist/dask_cudf*.whl)[test]

dependencies.yaml

@@ -21,6 +21,7 @@ files:
       - py_version
       - run_common
       - run_cudf
+      - run_libcudf
       - run_dask_cudf
       - run_custreamz
       - test_cpp
@@ -75,6 +76,18 @@ files:
       - docs
       - libarrow_run
       - py_version
+  # This is the shared library, bundled as a wheel. It is meant to be consumed by the wrapper.
+  py_build_libcudf:
+    output: pyproject
+    pyproject_dir: python/libcudf
+    extras:
+      table: build-system
+    includes:
+      - build_base
+      - build_cpp
+      - build_python_common
+      - build_python_cudf
+  # This is the wrapper that gets used in Python, not the shared library wheel
   py_build_cudf:
     output: pyproject
     pyproject_dir: python/cudf
@@ -84,6 +97,7 @@ files:
       - build_base
       - build_python_common
       - build_python_cudf
+      - run_libcudf
   py_run_cudf:
     output: pyproject
     pyproject_dir: python/cudf
@@ -94,6 +108,7 @@ files:
       - run_cudf
       - pyarrow_run
       - depends_on_cupy
+      - run_libcudf
   py_test_cudf:
     output: pyproject
     pyproject_dir: python/cudf
@@ -336,6 +351,13 @@ dependencies:
           # Allow runtime version to float up to minor version
           # Disallow pyarrow 14.0.0 due to a CVE
           - pyarrow>=14.0.1,<15.0.0a0
+  run_libcudf:
+    common:
+      # TODO: Currently this is a hack for devcontainers. Need to figure out the best solution.
+      #- output_types: [requirements, pyproject]
+      - output_types: [pyproject]
+        packages:
+          - libcudf==24.6.*
   cuda_version:
     specific:
       - output_types: conda

python/cudf/CMakeLists.txt

@@ -24,72 +24,15 @@ project(
   LANGUAGES CXX CUDA
 )
 
-option(FIND_CUDF_CPP "Search for existing CUDF C++ installations before defaulting to local files"
-       OFF
-)
-option(USE_LIBARROW_FROM_PYARROW "Only use the libarrow contained in pyarrow" OFF)
-mark_as_advanced(USE_LIBARROW_FROM_PYARROW)
-
-# Find Python early so that later commands can use it
-find_package(Python 3.9 REQUIRED COMPONENTS Interpreter)
-
-# If the user requested it we attempt to find CUDF.
-if(FIND_CUDF_CPP)
-  include(rapids-cpm)
-  include(rapids-export)
-  include(rapids-find)
-  rapids_cpm_init()
-
-  if(USE_LIBARROW_FROM_PYARROW)
-    # We need to find arrow before libcudf since libcudf requires it but doesn't bundle arrow
-    # libraries. These variables have no effect because we are always searching for arrow via
-    # pyarrow, but they must be set as they are required arguments to the function in
-    # get_arrow.cmake.
-    set(CUDF_USE_ARROW_STATIC OFF)
-    set(CUDF_ENABLE_ARROW_S3 OFF)
-    set(CUDF_ENABLE_ARROW_ORC OFF)
-    set(CUDF_ENABLE_ARROW_PYTHON OFF)
-    set(CUDF_ENABLE_ARROW_PARQUET OFF)
-    include(../../cpp/cmake/thirdparty/get_arrow.cmake)
-  endif()
+find_package(cudf "${RAPIDS_VERSION}" REQUIRED)
 
-  find_package(cudf "${RAPIDS_VERSION}" REQUIRED)
-
-  # an installed version of libcudf doesn't provide the dlpack headers so we need to download dlpack
-  # for the interop.pyx
-  include(../../cpp/cmake/thirdparty/get_dlpack.cmake)
-else()
-  set(cudf_FOUND OFF)
-endif()
+# an installed version of libcudf doesn't provide the dlpack headers so we need to download dlpack
+# for the interop.pyx
+include(rapids-cpm)
+rapids_cpm_init()
+include(../../cpp/cmake/thirdparty/get_dlpack.cmake)
 
 include(rapids-cython-core)
-
-if(NOT cudf_FOUND)
-  set(BUILD_TESTS OFF)
-  set(BUILD_BENCHMARKS OFF)
-  set(CUDF_BUILD_TESTUTIL OFF)
-  set(CUDF_BUILD_STREAMS_TEST_UTIL OFF)
-  set(CUDA_STATIC_RUNTIME ON)
-
-  add_subdirectory(../../cpp cudf-cpp EXCLUDE_FROM_ALL)
-
-  # libcudf targets are excluded by default above via EXCLUDE_FROM_ALL to remove extraneous
-  # components like headers from libcudacxx, but we do need the libraries. However, we want to
-  # control where they are installed to. Since there are multiple subpackages of cudf._lib that
-  # require access to libcudf, we place the library and all its dependent artifacts in the cudf
-  # directory as a single source of truth and modify the other rpaths appropriately.
-  set(cython_lib_dir cudf)
-  include(cmake/Modules/WheelHelpers.cmake)
-  # TODO: This install is currently overzealous. We should only install the libraries that are
-  # downloaded by CPM during the build, not libraries that were found on the system.  However, in
-  # practice right this would only be a problem is if libcudf was not found but some of the
-  # dependencies were, and we have no real use cases where that happens.
-  install_aliased_imported_targets(
-    TARGETS cudf arrow_shared nvcomp::nvcomp nvcomp::nvcomp_gdeflate nvcomp::nvcomp_bitcomp
-    DESTINATION ${cython_lib_dir}
-  )
-endif()
-
 rapids_cython_init()
 
 include(cmake/Modules/LinkPyarrowHeaders.cmake)
@@ -99,3 +42,18 @@ add_subdirectory(udf_cpp)
 if(DEFINED cython_lib_dir)
   rapids_cython_add_rpath_entries(TARGET cudf PATHS "${cython_lib_dir}")
 endif()
+
+# libcudf targets are excluded by default above via EXCLUDE_FROM_ALL to remove extraneous components
+# like headers from libcudacxx, but we do need the libraries. However, we want to control where they
+# are installed to. Since there are multiple subpackages of cudf._lib that require access to
+# libcudf, we place the library and all its dependent artifacts in the cudf directory as a single
+# source of truth and modify the other rpaths appropriately.
+include(cmake/Modules/WheelHelpers.cmake)
+# TODO: This install is currently overzealous. We should only install the libraries that are
+# downloaded by CPM during the build, not libraries that were found on the system.  However, in
+# practice right this would only be a problem is if libcudf was not found but some of the
+# dependencies were, and we have no real use cases where that happens.
+install_aliased_imported_targets(
+  TARGETS cudf arrow_shared nvcomp::nvcomp nvcomp::nvcomp_gdeflate nvcomp::nvcomp_bitcomp
+  DESTINATION ${cython_lib_dir}
+)

python/cudf/cudf/__init__.py

@@ -1,5 +1,15 @@
 # Copyright (c) 2018-2024, NVIDIA CORPORATION.
 
+# If libcudf was installed as a wheel, we must request it to load the library symbols.
+# Otherwise, we assume that the library was installed in a system path that ld can find.
+try:
+    import libcudf
+except ModuleNotFoundError:
+    pass
+else:
+    libcudf.load_library()
+    del libcudf
+
 # _setup_numba _must be called before numba.cuda is imported, because
 # it sets the numba config variable responsible for enabling
 # Minor Version Compatibility. Setting it after importing numba.cuda has no effect.

python/cudf/cudf/_lib/CMakeLists.txt

@@ -68,6 +68,8 @@ target_link_libraries(strings_udf PUBLIC cudf_strings_udf)
 set(targets_using_arrow_headers interop avro csv orc json parquet)
 link_to_pyarrow_headers("${targets_using_arrow_headers}")
 
+target_include_directories(interop PUBLIC "$<BUILD_INTERFACE:${DLPACK_INCLUDE_DIR}>")
+
 add_subdirectory(cpp)
 add_subdirectory(io)
 add_subdirectory(nvtext)

python/cudf/cudf/_lib/pylibcudf/CMakeLists.txt

@@ -44,5 +44,6 @@ rapids_cython_create_modules(
   LINKED_LIBRARIES "${linked_libraries}" MODULE_PREFIX pylibcudf_ ASSOCIATED_TARGETS cudf
 )
 link_to_pyarrow_headers(pylibcudf_interop)
+target_include_directories(pylibcudf_interop PUBLIC "$<BUILD_INTERFACE:${DLPACK_INCLUDE_DIR}>")
 
 add_subdirectory(strings)

python/cudf/pyproject.toml

@@ -5,6 +5,7 @@ build-backend = "scikit_build_core.build"
 requires = [
     "cmake>=3.26.4",
     "cython>=3.0.3",
+    "libcudf==24.6.*",
     "ninja",
     "numpy==1.23.*",
     "pyarrow==14.0.2.*",
@@ -28,6 +29,7 @@ dependencies = [
     "cuda-python>=11.7.1,<12.0a0",
     "cupy-cuda11x>=12.0.0",
     "fsspec>=0.6.0",
+    "libcudf==24.6.*",
     "numba>=0.57",
     "numpy>=1.23,<2.0a0",
     "nvtx>=0.2.1",