API,MAINT: RAPIDS 24.12 and align DatetimeComponent with pylibcudf (#15)

Legate bumped to using rapids-cmake 24.12, so let's use RAPIDS 24.12
also.

Most changes are just search/replace, but:
* This enables CI for Python 3.12, I forgot about that when we went to
24.10 (I think that was compatible).
* `libcudf` doesn't ship `libcudftestutil` anymore, instead we need to
add it to our link targets (and the compile flags) to compile it here.
* **API change:** `(py)libcudf` now has `etract_component` and
`DatetimeComponent`. So align with that. This means:
* Python code needs to use the capitalizem, i.e. `DatetimeComponent.DAY`
now.
* `day_of_year` is a dedicated function in `(py)libcudf`, so it is gone
for now.
  * (On the C-side, the same change would be needed.)

---------

Signed-off-by: Sebastian Berg <[email protected]>

.github/workflows/conda-python-build.yaml

@@ -41,7 +41,7 @@ jobs:
         PY_VER:
           - "3.10"
           - "3.11"
-          # - "3.12" requires update to RAPIDS 24.10
+          - "3.12"
     runs-on: linux-${{ matrix.ARCH }}-cpu4
     container:
       image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"

.github/workflows/pr.yaml

@@ -48,14 +48,17 @@ jobs:
       #
       #  * architectures: amd64 only
       #  * CUDA: >=12.2
-      #  * Python: 3.10, 3.11, 3.12
+      #  * Python: 3.10, 3.11, 3.12  (3.11 used in doc building)
       #
       # Valid set of RAPIDS ci-conda image tags: https://hub.docker.com/r/rapidsai/ci-conda/tags
       matrix:
         include:
           - ARCH: "amd64"
             CUDA_VER: "12.5.1"
             PY_VER: "3.10"
+          - ARCH: "amd64"
+            CUDA_VER: "12.5.1"
+            PY_VER: "3.12"
     runs-on: linux-${{ matrix.ARCH }}-gpu-v100-latest-1
     container:
       image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"

ci/run_ctests.sh

@@ -11,13 +11,8 @@ else
     cd "${INSTALL_PREFIX:-${CONDA_PREFIX:-/usr}}/bin/gtests/liblegate_dataframe/"
 fi
 
-# Unless otherwise specified, use all available GPUs and set
+# Unless `LEGATE_CONFIG` is set, default to all available GPUs and set fbmem/sysmem.
 # LEGATE_TEST=1 to test broadcasting code paths (locally).
-# TODO: Set LEGATE_CONFIG instead (if undefined).  However,
-#       as of 2024-10-11 LEGATE_CONFIG seems broken:
-#       https://github.com/nv-legate/legate.core.internal/issues/1304
+LEGATE_CONFIG=${LEGATE_CONFIG:- --gpus="$(nvidia-smi -L | wc -l) --fbmem=4000 --sysmem=4000"} \
 LEGATE_TEST=${LEGATE_TEST:-1} \
-legate \
-    --gpus "$(nvidia-smi -L | wc -l)" \
-    --fbmem=4000 --sysmem=4000 \
-    ./cpp_tests --output-on-failure --no-tests=error "$@"
+legate ./cpp_tests --output-on-failure --no-tests=error "$@"

ci/run_pytests.sh

@@ -18,15 +18,11 @@ set -e -E -u -o pipefail
 # Support invoking run_cudf_pytests.sh outside the script directory
 cd "$(dirname "$(realpath "${BASH_SOURCE[0]}")")"/../python/tests/
 
-# Unless otherwise specified, use all available GPUs and set
+# Unless `LEGATE_CONFIG` is set, default to all available GPUs and set fbmem/sysmem.
 # LEGATE_TEST=1 to test broadcasting code paths (locally).
-# TODO: Set LEGATE_CONFIG instead (if undefined).  However,
-#       as of 2024-10-11 LEGATE_CONFIG seems broken:
-#       https://github.com/nv-legate/legate.core.internal/issues/1304
+LEGATE_CONFIG=${LEGATE_CONFIG:- --gpus="$(nvidia-smi -L | wc -l) --fbmem=4000 --sysmem=4000"} \
 LEGATE_TEST=${LEGATE_TEST:-1} \
 legate \
-    --gpus "$(nvidia-smi -L | wc -l)"\
-    --fbmem=4000 \
     --module pytest \
     . \
     -sv \

conda/environments/all_cuda-124_arch-x86_64.yaml

@@ -17,24 +17,24 @@ dependencies:
 - cuda-profiler-api
 - cuda-sanitizer-api
 - cuda-version=12.4
-- cudf==24.10.*,>=0.0.0a0
+- cudf==24.12.*,>=0.0.0a0
 - cupy>=12.0.0
 - cupynumeric==25.01.*,>=0.0.0.dev0
 - cxx-compiler
 - cython>=3.0.3
-- dask-cuda==24.10.*
-- dask-cudf==24.10.*
+- dask-cuda==24.12.*
+- dask-cudf==24.12.*
 - gcc_linux-64=11.*
 - legate==25.01.*,>=0.0.0.dev0
-- libcudf==24.10.*,>=0.0.0a0
-- librmm==24.10.*,>=0.0.0a0
+- libcudf==24.12.*,>=0.0.0a0
+- librmm==24.12.*,>=0.0.0a0
 - make
 - myst-parser>=4.0
 - ninja
 - numpy >=1.23,<3.0.0a0
 - openssh
 - pydata-sphinx-theme>=0.16.0
-- pylibcudf==24.10.*,>=0.0.0a0
+- pylibcudf==24.12.*,>=0.0.0a0
 - pytest>=7.0
 - python>=3.10,<3.13
 - rapids-build-backend>=0.3.2,<0.4.0.dev0

conda/recipes/legate-dataframe/conda_build_config.yaml

@@ -22,5 +22,10 @@ cuda11_compiler:
 legate_version:
   - "=25.01.*,>=0.0.0.dev0"
 
+# TODO: The != temporary blocklist cupynumeric versions
+# using cupynumeric, because it is fewer versions to block.
+cupynumeric_version:
+  - "=25.01.*,>=0.0.0.dev0,!=25.01.0.dev62,!=25.01.0.dev61,!=25.01.0.dev60,!=25.01.0.rc1"
+
 rapids_version:
-  - =24.10.*
+  - =24.12.*

conda/recipes/legate-dataframe/meta.yaml

@@ -1,4 +1,4 @@
-# Copyright (c) 2024, NVIDIA CORPORATION.
+# Copyright (c) 2024-2025, NVIDIA CORPORATION.
 
 {% set pyproject_data = load_file_data("python/pyproject.toml") %}
 {% set version = environ['LEGATEDATAFRAME_PACKAGE_VERSION'] %}
@@ -60,7 +60,7 @@ requirements:
     # Only to ensure a nightly legate version we pick up
     # is compatible with an existing cupynumeric version.
     # (may also stabilize not using debug/sanitizer builds)
-    - cupynumeric
+    - cupynumeric {{ cupynumeric_version }}
     - python
     - pip
     - cython >=3.0.3

cpp/cmake/fetch_rapids.cmake

@@ -12,7 +12,7 @@
 # the License.
 # =============================================================================
 if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/LEGATE_DATAFRAME_RAPIDS.cmake)
-  file(DOWNLOAD https://raw.githubusercontent.com/rapidsai/rapids-cmake/branch-24.10/RAPIDS.cmake
+  file(DOWNLOAD https://raw.githubusercontent.com/rapidsai/rapids-cmake/branch-24.12/RAPIDS.cmake
        ${CMAKE_CURRENT_BINARY_DIR}/LEGATE_DATAFRAME_RAPIDS.cmake
   )
 endif()

cpp/cmake/thirdparty/get_cudf.cmake

@@ -51,5 +51,5 @@ function(find_and_configure_cudf)
 endfunction()
 
 find_and_configure_cudf(
-  VERSION 24.10 GIT_REPO https://github.com/rapidsai/cudf.git GIT_TAG branch-24.10
+  VERSION 24.12 GIT_REPO https://github.com/rapidsai/cudf.git GIT_TAG branch-24.12
 )

cpp/include/legate_dataframe/timestamps.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024, NVIDIA CORPORATION.
+ * Copyright (c) 2023-2025, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,6 +18,7 @@
 
 #include <string>
 
+#include <cudf/datetime.hpp>
 #include <cudf/types.hpp>
 
 #include <legate_dataframe/core/column.hpp>
@@ -79,27 +80,14 @@ LogicalColumn to_timestamps(const LogicalColumn& input,
                             cudf::data_type timestamp_type,
                             std::string format);
 
-enum class DatetimeComponent : int32_t {
-  year,
-  month,
-  day,
-  weekday,
-  hour,
-  minute,
-  second,
-  millisecond_fraction,
-  microsecond_fraction,
-  nanosecond_fraction,
-  day_of_year
-};
-
 /**
  * @brief Extracts part of a timestamp as a int16.
  *
  * @param input Timestamp column
  * @param component The component which to extract.
  * @return New int16 column.
  */
-LogicalColumn extract_timestamp_component(const LogicalColumn& input, DatetimeComponent component);
+LogicalColumn extract_timestamp_component(const LogicalColumn& input,
+                                          cudf::datetime::datetime_component component);
 
 }  // namespace legate::dataframe

cpp/src/timestamps.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024, NVIDIA CORPORATION.
+ * Copyright (c) 2023-2025, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,10 +16,7 @@
 
 #include <string>
 
-// cudf's detail API of datetime assume that RMM has been included
-#include <rmm/mr/device/device_memory_resource.hpp>
-
-#include <cudf/detail/datetime.hpp>
+#include <cudf/datetime.hpp>
 #include <cudf/strings/convert/convert_datetime.hpp>
 
 #include <legate_dataframe/core/column.hpp>
@@ -52,52 +49,13 @@ class ExtractTimestampComponentTask
   static void gpu_variant(legate::TaskContext context)
   {
     GPUTaskContext ctx{context};
-    const auto component =
-      argument::get_next_scalar<std::underlying_type_t<DatetimeComponent>>(ctx);
-    const auto input = argument::get_next_input<PhysicalColumn>(ctx);
-    auto output      = argument::get_next_output<PhysicalColumn>(ctx);
+    const auto component = argument::get_next_scalar<cudf::datetime::datetime_component>(ctx);
+    const auto input     = argument::get_next_input<PhysicalColumn>(ctx);
+    auto output          = argument::get_next_output<PhysicalColumn>(ctx);
 
     std::unique_ptr<cudf::column> ret;
-    /* unfortunately, there seems to be no templating for this in libcudf: */
-    switch (static_cast<DatetimeComponent>(component)) {
-      case DatetimeComponent::year:
-        ret = cudf::datetime::detail::extract_year(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::month:
-        ret = cudf::datetime::detail::extract_month(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::day:
-        ret = cudf::datetime::detail::extract_day(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::weekday:
-        ret = cudf::datetime::detail::extract_weekday(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::hour:
-        ret = cudf::datetime::detail::extract_hour(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::minute:
-        ret = cudf::datetime::detail::extract_minute(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::second:
-        ret = cudf::datetime::detail::extract_second(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::millisecond_fraction:
-        ret = cudf::datetime::detail::extract_millisecond_fraction(
-          input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::microsecond_fraction:
-        ret = cudf::datetime::detail::extract_microsecond_fraction(
-          input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::nanosecond_fraction:
-        ret = cudf::datetime::detail::extract_nanosecond_fraction(
-          input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      case DatetimeComponent::day_of_year:
-        ret = cudf::datetime::detail::day_of_year(input.column_view(), ctx.stream(), ctx.mr());
-        break;
-      default: throw std::runtime_error("invalid resolution to time part extraction?");
-    }
+    ret = cudf::datetime::extract_datetime_component(
+      input.column_view(), component, ctx.stream(), ctx.mr());
 
     output.move_into(std::move(ret));
   }
@@ -119,7 +77,8 @@ LogicalColumn to_timestamps(const LogicalColumn& input,
   return ret;
 }
 
-LogicalColumn extract_timestamp_component(const LogicalColumn& input, DatetimeComponent component)
+LogicalColumn extract_timestamp_component(const LogicalColumn& input,
+                                          cudf::datetime::datetime_component component)
 {
   if (!cudf::is_timestamp(input.cudf_type())) {
     throw std::invalid_argument("extract_timestamp_component() input must be timestamp");
@@ -128,8 +87,8 @@ LogicalColumn extract_timestamp_component(const LogicalColumn& input, DatetimeCo
   auto ret     = LogicalColumn::empty_like(cudf::data_type{cudf::type_id::INT16}, input.nullable());
   legate::AutoTask task =
     runtime->create_task(get_library(), task::ExtractTimestampComponentTask::TASK_ID);
-  argument::add_next_scalar(task,
-                            static_cast<std::underlying_type_t<DatetimeComponent>>(component));
+  argument::add_next_scalar(
+    task, static_cast<std::underlying_type_t<cudf::datetime::datetime_component>>(component));
   argument::add_next_input(task, input);
   argument::add_next_output(task, ret);
   runtime->submit(std::move(task));

cpp/tests/CMakeLists.txt

@@ -34,10 +34,13 @@ set_target_properties(
              CUDA_STANDARD_REQUIRED ON
 )
 
+set(LDF_TEST_CUDA_FLAGS --expt-extended-lambda --expt-relaxed-constexpr)
+target_compile_options(cpp_tests PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${LDF_TEST_CUDA_FLAGS}>")
+
 # Note that fmt::fmt should not be required, but seems to be for debug builds.
 target_link_libraries(
-  cpp_tests PRIVATE LegateDataframe cudf cudf::cudftestutil GTest::gmock GTest::gtest fmt::fmt
-                    $<TARGET_NAME_IF_EXISTS:conda_env>
+  cpp_tests PRIVATE LegateDataframe cudf cudf::cudftestutil cudf::cudftestutil_impl
+                    GTest::gmock GTest::gtest fmt::fmt $<TARGET_NAME_IF_EXISTS:conda_env>
 )
 rapids_test_add(
   NAME cpp_tests

dependencies.yaml

@@ -175,8 +175,8 @@ dependencies:
         packages:
           - cupynumeric==25.01.*,>=0.0.0.dev0
           - pytest>=7.0
-          - dask-cuda==24.10.*
-          - dask-cudf==24.10.*
+          - dask-cuda==24.12.*
+          - dask-cudf==24.12.*
       - output_types: conda
         packages:
           - cuda-sanitizer-api
@@ -187,17 +187,17 @@ dependencies:
     common:
       - output_types: conda
         packages:
-          - &cudf_unsuffixed cudf==24.10.*,>=0.0.0a0
-          - &pylibcudf_unsuffixed pylibcudf==24.10.*,>=0.0.0a0
+          - &cudf_unsuffixed cudf==24.12.*,>=0.0.0a0
+          - &pylibcudf_unsuffixed pylibcudf==24.12.*,>=0.0.0a0
     specific:
       - output_types: [requirements, pyproject]
         matrices:
           - matrix:
               cuda: "12.*"
               cuda_suffixed: "true"
             packages:
-              - cudf-cu12==24.10.*,>=0.0.0a0
-              - pylibcudf-cu12==24.10.*,>=0.0.0a0
+              - cudf-cu12==24.12.*,>=0.0.0a0
+              - pylibcudf-cu12==24.12.*,>=0.0.0a0
           - {matrix: null, packages: [*cudf_unsuffixed, *pylibcudf_unsuffixed]}
 
   depends_on_cupy:
@@ -223,30 +223,30 @@ dependencies:
     common:
       - output_types: conda
         packages:
-          - &libcudf_unsuffixed libcudf==24.10.*,>=0.0.0a0
+          - &libcudf_unsuffixed libcudf==24.12.*,>=0.0.0a0
     specific:
       - output_types: [requirements, pyproject]
         matrices:
           - matrix:
               cuda: "12.*"
               cuda_suffixed: "true"
             packages:
-              - libcudf-cu12==24.10.*,>=0.0.0a0
+              - libcudf-cu12==24.12.*,>=0.0.0a0
           - {matrix: null, packages: [*libcudf_unsuffixed]}
 
   depends_on_librmm:
     common:
       - output_types: conda
         packages:
-          - &librmm_unsuffixed librmm==24.10.*,>=0.0.0a0
+          - &librmm_unsuffixed librmm==24.12.*,>=0.0.0a0
     specific:
       - output_types: [requirements, pyproject]
         matrices:
           - matrix:
               cuda: "12.*"
               cuda_suffixed: "true"
             packages:
-              - librmm-cu12==24.10.*,>=0.0.0a0
+              - librmm-cu12==24.12.*,>=0.0.0a0
           - matrix:
             packages:
               - *librmm_unsuffixed

python/legate_dataframe/lib/timestamps.pyi

@@ -2,15 +2,18 @@
 # SPDX-License-Identifier: Apache-2.0
 
 from numpy.typing import DTypeLike
+from pylibcudf.datetime import DatetimeComponent
 
 from legate_dataframe.lib.core.column import LogicalColumn
 
+__all__ = ["to_timestamps", "extract_timestamp_component", "DatetimeComponent"]
+
 def to_timestamps(
     col: LogicalColumn,
     timestamp_type: DTypeLike,
     format_pattern: str,
 ) -> LogicalColumn: ...
-def extract_timepart(
+def extract_timestamp_component(
     col: LogicalColumn,
-    resolution: str,
+    component: DatetimeComponent,
 ) -> LogicalColumn: ...