MAINT: Bump rapids to 25.08 and some other versions

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

@@ -44,13 +44,13 @@ jobs:
         ARCH:
           - "amd64"
         CUDA_VER:
-          - "12.5.1"
+          - "12.9.1"
         PY_VER:
           - "3.11"
           - "3.12"
     runs-on: linux-${{ matrix.ARCH }}-cpu4
     container:
-      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"
+      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu24.04-py${{ matrix.PY_VER }}"
     steps:
       - uses: aws-actions/configure-aws-credentials@v4
         with:

.github/workflows/docs-build.yaml

@@ -29,12 +29,11 @@ jobs:
       matrix:
         include:
           - ARCH: amd64
-            CUDA_VER: "12.5.1"
+            CUDA_VER: "12.9.1"
             PY_VER: "3.12"
-    # NOTE: Try going back to "latest" runners once we bump RAPIDS 25.04
-    runs-on: linux-${{ matrix.ARCH }}-gpu-v100-earliest-1
+    runs-on: linux-${{ matrix.ARCH }}-gpu-l4-earliest-1
     container:
-      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"
+      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu24.04-py${{ matrix.PY_VER }}"
       env:
         NVIDIA_VISIBLE_DEVICES: ${{ env.NVIDIA_VISIBLE_DEVICES }}
     steps:

.github/workflows/pr.yaml

@@ -23,7 +23,7 @@ jobs:
       - conda-python-build
       - conda-python-cpu-tests
       - conda-python-gpu-tests
-    uses: rapidsai/shared-workflows/.github/workflows/pr-builder.yaml@branch-25.04
+    uses: rapidsai/shared-workflows/.github/workflows/pr-builder.yaml@branch-25.08
 
   pre-commit:
     runs-on: ubuntu-latest
@@ -49,15 +49,14 @@ jobs:
       matrix:
         include:
           - ARCH: "amd64"
-            CUDA_VER: "12.5.1"
+            CUDA_VER: "12.9.1"
             PY_VER: "3.11"
           - ARCH: "amd64"
-            CUDA_VER: "12.5.1"
+            CUDA_VER: "12.9.1"
             PY_VER: "3.12"
-    # NOTE: Try going back to "latest" runners once we bump RAPIDS 25.04
     runs-on: linux-${{ matrix.ARCH }}-cpu16
     container:
-      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"
+      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu24.04-py${{ matrix.PY_VER }}"
     steps:
       - uses: actions/checkout@v4
         with:
@@ -90,15 +89,14 @@ jobs:
       matrix:
         include:
           - ARCH: "amd64"
-            CUDA_VER: "12.5.1"
+            CUDA_VER: "12.9.1"
             PY_VER: "3.11"
           - ARCH: "amd64"
-            CUDA_VER: "12.5.1"
+            CUDA_VER: "12.9.1"
             PY_VER: "3.12"
-    # NOTE: Try going back to "latest" runners once we bump RAPIDS 25.04
-    runs-on: linux-${{ matrix.ARCH }}-gpu-v100-earliest-1
+    runs-on: linux-${{ matrix.ARCH }}-gpu-l4-latest-1
     container:
-      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu22.04-py${{ matrix.PY_VER }}"
+      image: "rapidsai/ci-conda:cuda${{ matrix.CUDA_VER }}-ubuntu24.04-py${{ matrix.PY_VER }}"
       env:
         NVIDIA_VISIBLE_DEVICES: ${{ env.NVIDIA_VISIBLE_DEVICES }}
     steps:

Dockerfile

@@ -13,21 +13,21 @@
 # limitations under the License.
 #
 
-ARG CUDA_VERSION="12.5.1"
+ARG CUDA_VERSION="12.9.1"
 ARG PYTHON_VERSION="3.11"
 
-ARG BASE_IMAGE="rapidsai/miniforge-cuda:cuda${CUDA_VERSION}-base-ubuntu22.04-py${PYTHON_VERSION}"
+ARG BASE_IMAGE="rapidsai/miniforge-cuda:cuda${CUDA_VERSION}-base-ubuntu24.04-py${PYTHON_VERSION}"
 FROM ${BASE_IMAGE}
 RUN apt-get update && apt-get install -y --no-install-recommends git && rm -rf /var/lib/apt/lists/*
 
 # Legate-dataframe: create conda environment, build, and install
 #
-# Update conda the environment (we use `cut` to convert the format of CUDA_VERSION from "12.5.1" to "125")
+# Update conda the environment (we use `cut` to convert the format of CUDA_VERSION from "12.9.1" to "129")
 RUN mkdir -p /opt/legate-dataframe/conda-env-file
 COPY ./conda/environments/*.yaml /opt/legate-dataframe/conda-env-file/
 
 # To ensure we find the GPU version of legate in the docker build.
-ARG CONDA_OVERRIDE_CUDA=12.4
+ARG CONDA_OVERRIDE_CUDA=12.9
 RUN /bin/bash -c '/opt/conda/bin/mamba env create --name legate-dev --file \
   /opt/legate-dataframe/conda-env-file/all_cuda-$(cut --output-delimiter="" -d "." -f 1,2 <<< ${CONDA_OVERRIDE_CUDA})_arch-x86_64.yaml'
 

conda/environments/all_cuda-124_arch-x86_64.yaml → conda/environments/all_cuda-129_arch-x86_64.yaml

@@ -16,19 +16,19 @@ dependencies:
 - cuda-nvtx-dev
 - cuda-profiler-api
 - cuda-sanitizer-api
-- cuda-version=12.4
-- cudf==25.04.*,>=0.0.0a0
+- cuda-version=12.9
+- cudf==25.08.*,>=0.0.0a0
 - cupy>=12.0.0
 - cupynumeric==25.10.*,>=0.0.0.dev0
 - cxx-compiler
 - cython>=3.0.3
-- dask-cuda==25.04.*
-- dask-cudf==25.04.*
+- dask-cuda==25.08.*
+- dask-cudf==25.08.*
 - gcc_linux-64=11.*
 - legate==25.10.*,>=0.0.0.dev0
 - libarrow-acero
-- libcudf==25.04.*,>=0.0.0a0
-- librmm==25.04.*,>=0.0.0a0
+- libcudf==25.08.*,>=0.0.0a0
+- librmm==25.08.*,>=0.0.0a0
 - make
 - myst-parser>=4.0
 - nccl>=2.19
@@ -37,7 +37,7 @@ dependencies:
 - openssh
 - polars>=1.25,<1.32
 - pydata-sphinx-theme>=0.16.0
-- pylibcudf==25.04.*,>=0.0.0a0
+- pylibcudf==25.08.*,>=0.0.0a0
 - pynvjitlink<=0.6
 - pytest>=7.0
 - python>=3.11,<3.13
@@ -46,4 +46,4 @@ dependencies:
 - sphinx>=8.0,<8.2.0
 - sysroot_linux-64==2.17
 - valgrind
-name: all_cuda-124_arch-x86_64
+name: all_cuda-129_arch-x86_64

conda/recipes/legate-dataframe/conda_build_config.yaml

@@ -24,7 +24,7 @@ gpu_enabled:
   - false
 
 legate_version:
-  - "=25.10.*,>=0.0.0.dev0"
+  - "=25.10.*,>=0.0.0.dev0,!=25.10.00.rc1"
 
 cupynumeric_version:
   - "=25.10.*,>=0.0.0.dev0"
@@ -38,4 +38,4 @@ arrow_version:
   - "19.0.*"
 
 rapids_version:
-  - =25.04.*
+  - =25.08.*

cpp/CMakeLists.txt

@@ -14,6 +14,8 @@
 
 cmake_minimum_required(VERSION 3.26.4 FATAL_ERROR)
 
+set(rapids-cmake-version 25.08)
+
 include(cmake/fetch_rapids.cmake)
 include(rapids-cmake)
 include(rapids-cpm)
@@ -144,22 +146,14 @@ set_target_properties(
   PROPERTIES BUILD_RPATH "\$ORIGIN"
              INSTALL_RPATH "\$ORIGIN"
              # set target compile options
-             CXX_STANDARD 17
+             CXX_STANDARD 20
              CXX_STANDARD_REQUIRED ON
              # For std:: support of __int128_t. Can be removed once using cuda::std
              CXX_EXTENSIONS ON
              POSITION_INDEPENDENT_CODE ON
              INTERFACE_POSITION_INDEPENDENT_CODE ON
 )
 
-if(Legion_USE_CUDA)
-  set_target_properties(LegateDataframe PROPERTIES CUDA_STANDARD 17 CUDA_STANDARD_REQUIRED ON)
-  # Need to add this define, as is done in legate CMakeLists.txt as well for __half support.  If
-  # __half support fails elsewhere it may be needed there. This may be a CCCL 2.7.0 issue and become
-  # unnecessary in the future.
-  target_compile_definitions("LegateDataframe" PUBLIC _LIBCUDACXX_HAS_NVFP16=1)
-endif()
-
 list(APPEND LDF_CUDA_FLAGS --expt-extended-lambda)
 list(APPEND LDF_CUDA_FLAGS --expt-relaxed-constexpr)
 
@@ -173,6 +167,14 @@ target_include_directories(
   PUBLIC "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>"
   INTERFACE "$<INSTALL_INTERFACE:include>"
 )
+# Ridiculous hack, remove as soon as possible (i.e. after legate 25.10 for sure)! Legion headers
+# barf on a cudart issue with cuda::std::complex<__half> when using C++20.  So we inject our own
+# `legion_defines.h` before legate does it. This should fix it:
+# https://github.com/nv-legate/legate.internal/pull/2965
+target_include_directories(
+  LegateDataframe SYSTEM
+  PUBLIC "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include/legate_dataframe/legion_hack>"
+)
 
 target_link_libraries(
   LegateDataframe

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-25.04/RAPIDS.cmake
+  file(DOWNLOAD https://raw.githubusercontent.com/rapidsai/rapids-cmake/branch-25.08/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 25.04 GIT_REPO https://github.com/rapidsai/cudf.git GIT_TAG branch-25.04
+  VERSION 25.08 GIT_REPO https://github.com/rapidsai/cudf.git GIT_TAG branch-25.08
 )

cpp/examples/CMakeLists.txt

@@ -16,7 +16,7 @@ set(TEST_INSTALL_PATH bin/tests/hello_world)
 set(TEST_NAME hello_world)
 
 add_executable(hello_world hello.cpp)
-set_target_properties(hello_world PROPERTIES INSTALL_RPATH "\$ORIGIN/..")
+set_target_properties(hello_world PROPERTIES INSTALL_RPATH "\$ORIGIN/.." CXX_STANDARD 20)
 target_link_libraries(hello_world PRIVATE LegateDataframe::LegateDataframe)
 
 if(CMAKE_COMPILER_IS_GNUCXX)

cpp/include/legate_dataframe/legion_hack/legion_defines.h

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 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.
+ * 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.
+ */
+
+#include "legate/deps/legion_defines.h"  // if this fails, try removing the hack!
+
+/*
+ * As of early 25.10, legion headers are broken with C++20 which we need.
+ * We can hack around that by removing the legion redop half definition
+ * so let's just *assume* we can find the correct header here (where it should
+ * be).
+ * The problem is that legion wants to use cuda::std::complex<__half> inside
+ * an atomic and that is broken (an issue with cuda::std::complex<__half> or
+ * rather `__half2` used by it.
+ * Legate should work around this soon, until then undef redop half and hope
+ * we do not run into this.
+ *
+ * On the CPU undefinig this means __half isn't defined, so don't do it there.
+ */
+#ifdef LEGION_USE_CUDA
+#undef LEGION_REDOP_HALF
+#endif

cpp/src/join.cu

@@ -16,7 +16,8 @@
 
 #include <cudf/column/column_factories.hpp>
 #include <cudf/copying.hpp>
-#include <cudf/join.hpp>
+#include <cudf/join/hash_join.hpp>
+#include <cudf/join/join.hpp>
 #include <cudf/table/table.hpp>
 
 #include <legate_dataframe/core/repartition_by_hash.hpp>

cpp/src/parquet.cpp

@@ -471,8 +471,8 @@ ParquetReadInfo get_parquet_info(const std::vector<std::string>& file_paths,
       // but it doesn't work either way (As of legate 25.07).  For table reading
       // 1-D rects can be used (and are for eager mode).
       row_group_ranges.emplace_back(
-        legate::Rect<1>({static_cast<int64_t>(nrows_total)},
-                        {static_cast<int64_t>(nrows_total + nrows_in_group - 1)}));
+        legate::Rect<1>({static_cast<legate::coord_t>(nrows_total)},
+                        {static_cast<legate::coord_t>(nrows_total + nrows_in_group - 1)}));
       nrows_total += nrows_in_group;
     }
   }

cpp/src/unaryop.cu

@@ -46,7 +46,7 @@ namespace legate::dataframe::task {
   TaskContext ctx{context};
 
   const auto input      = argument::get_next_input<PhysicalColumn>(ctx);
-  auto digits           = argument::get_next_scalar<int32_t>(ctx);
+  auto decimal_places   = argument::get_next_scalar<int32_t>(ctx);
   auto mode             = argument::get_next_scalar<std::string>(ctx);
   auto output           = argument::get_next_output<PhysicalColumn>(ctx);
   cudf::column_view col = input.column_view();
@@ -58,8 +58,10 @@ namespace legate::dataframe::task {
   } else {
     throw std::invalid_argument("Unsupported rounding method: " + mode);
   }
+  // TODO(seberg): Need to switch to round_decimal, but it failed tests due to
+  // some input types in our tests and I have not yet checked why or what to use.
   std::unique_ptr<cudf::column> ret =
-    cudf::round(col, digits, rounding_method, ctx.stream(), ctx.mr());
+    cudf::round(col, decimal_places, rounding_method, ctx.stream(), ctx.mr());
   if (get_prefer_eager_allocations()) {
     output.copy_into(std::move(ret));
   } else {

cpp/tests/CMakeLists.txt

@@ -31,11 +31,11 @@ set_target_properties(
   cpp_tests
   PROPERTIES RUNTIME_OUTPUT_DIRECTORY "$<BUILD_INTERFACE:${LegateDataframe_BINARY_DIR}/gtests>"
              # INSTALL_RPATH "\$ORIGIN/../../../lib"
-             CXX_STANDARD 17
+             CXX_STANDARD 20
              CXX_STANDARD_REQUIRED ON
              # For std:: support of __int128_t. Can be removed once using cuda::std
              CXX_EXTENSIONS ON
-             CUDA_STANDARD 17
+             CUDA_STANDARD 20
              CUDA_STANDARD_REQUIRED ON
 )