build wheels with CUDA 13.0.x, test wheels against mix of CTK versions, drop CUDA math libraries dependencies

.devcontainer/cuda12.9-pip/devcontainer.json

@@ -19,10 +19,10 @@
   "features": {
     "ghcr.io/rapidsai/devcontainers/features/cuda:26.4": {
       "version": "12.9",
-      "installcuBLAS": true,
-      "installcuSOLVER": true,
-      "installcuRAND": true,
-      "installcuSPARSE": true
+      "installcuBLAS": false,
+      "installcuSOLVER": false,
+      "installcuRAND": false,
+      "installcuSPARSE": false
     },
     "ghcr.io/rapidsai/devcontainers/features/rapids-build-utils:26.4": {}
   },

.devcontainer/cuda13.1-pip/devcontainer.json

@@ -19,10 +19,10 @@
   "features": {
     "ghcr.io/rapidsai/devcontainers/features/cuda:26.4": {
       "version": "13.1",
-      "installcuBLAS": true,
-      "installcuSOLVER": true,
-      "installcuRAND": true,
-      "installcuSPARSE": true
+      "installcuBLAS": false,
+      "installcuSOLVER": false,
+      "installcuRAND": false,
+      "installcuSPARSE": false
     },
     "ghcr.io/rapidsai/devcontainers/features/rapids-build-utils:26.4": {}
   },

ci/test_wheel.sh

@@ -14,14 +14,14 @@ mkdir -p "${RAPIDS_TESTS_DIR}"
 
 # generate constraints, the constraints will limit the version of the
 # dependencies that can be installed later on when installing the wheel
-rapids-generate-pip-constraints test_python ./constraints.txt
+rapids-generate-pip-constraints test_python "${PIP_CONSTRAINT}"
 
 # Install just minimal dependencies first
 rapids-pip-retry install \
+  --prefer-binary \
+  --constraint "${PIP_CONSTRAINT}" \
   "${LIBNVFOREST_WHEELHOUSE}"/libnvforest*.whl \
-  "${NVFOREST_WHEELHOUSE}"/nvforest*.whl \
-  --constraint ./constraints.txt \
-  --constraint "${PIP_CONSTRAINT}"
+  "${NVFOREST_WHEELHOUSE}"/nvforest*.whl
 
 # Try to import nvforest with just a minimal install"
 rapids-logger "Importing nvforest with minimal dependencies"
@@ -34,10 +34,10 @@ python -c "import nvforest"
 #     ignored if any other --constraint are passed via the CLI
 #
 rapids-pip-retry install \
-   "${LIBNVFOREST_WHEELHOUSE}"/libnvforest*.whl \
-  "$(echo "${NVFOREST_WHEELHOUSE}"/nvforest*.whl)[test]" \
-  --constraint ./constraints.txt \
-  --constraint "${PIP_CONSTRAINT}"
+  --prefer-binary \
+  --constraint "${PIP_CONSTRAINT}" \
+  "${LIBNVFOREST_WHEELHOUSE}"/libnvforest*.whl \
+  "$(echo "${NVFOREST_WHEELHOUSE}"/nvforest*.whl)[test]"
 
 EXITCODE=0
 trap "EXITCODE=1" ERR

conda/environments/all_cuda-129_arch-aarch64.yaml

@@ -20,11 +20,6 @@ dependencies:
 - doxygen
 - gcc_linux-aarch64=14.*
 - graphviz
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

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

@@ -20,11 +20,6 @@ dependencies:
 - doxygen
 - gcc_linux-64=14.*
 - graphviz
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

conda/environments/all_cuda-131_arch-aarch64.yaml

@@ -20,11 +20,6 @@ dependencies:
 - doxygen
 - gcc_linux-aarch64=14.*
 - graphviz
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

conda/environments/all_cuda-131_arch-x86_64.yaml

@@ -20,11 +20,6 @@ dependencies:
 - doxygen
 - gcc_linux-64=14.*
 - graphviz
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

conda/environments/clang_tidy_cuda-129_arch-x86_64.yaml

@@ -15,11 +15,6 @@ dependencies:
 - cuda-version=12.9
 - cxx-compiler
 - gcc_linux-64=14.*
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft-headers==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - llvm-openmp==15.0.7

conda/environments/clang_tidy_cuda-131_arch-x86_64.yaml

@@ -15,11 +15,6 @@ dependencies:
 - cuda-version=13.1
 - cxx-compiler
 - gcc_linux-64=14.*
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft-headers==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - llvm-openmp==15.0.7

conda/environments/cpp_all_cuda-129_arch-x86_64.yaml

@@ -13,11 +13,6 @@ dependencies:
 - cuda-version=12.9
 - cxx-compiler
 - gcc_linux-64=14.*
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft-headers==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

conda/environments/cpp_all_cuda-131_arch-x86_64.yaml

@@ -13,11 +13,6 @@ dependencies:
 - cuda-version=13.1
 - cxx-compiler
 - gcc_linux-64=14.*
-- libcublas-dev
-- libcufft-dev
-- libcurand-dev
-- libcusolver-dev
-- libcusparse-dev
 - libraft-headers==26.4.*,>=0.0.0a0
 - librmm==26.4.*,>=0.0.0a0
 - ninja

conda/recipes/libnvforest/recipe.yaml

@@ -76,11 +76,6 @@ cache:
       - rapids-logger =0.2
       - treelite ${{ treelite_version }}
       - cuda-cudart-dev
-      - libcublas-dev
-      - libcufft-dev
-      - libcurand-dev
-      - libcusolver-dev
-      - libcusparse-dev
 
 outputs:
   - package:
@@ -107,31 +102,17 @@ outputs:
         - librmm =${{ minor_version }}
         - treelite ${{ treelite_version }}
         - cuda-cudart-dev
-        - libcublas
-        - libcufft
-        - libcurand
-        - libcusolver
-        - libcusparse
       run:
         - ${{ pin_compatible("cuda-version", upper_bound="x", lower_bound="x") }}
         - cuda-cudart
-        - libcublas
-        - libcufft
-        - libcurand
-        - libcusolver
-        - libcusparse
         - librmm =${{ minor_version }}
         - rapids-logger =0.2
         - treelite ${{ treelite_version }}
       ignore_run_exports:
         by_name:
           - cuda-cudart
           - cuda-version
-          - libcurand
           - librmm
-          - libcusparse
-          - libcufft
-          - libcublas
           - libcusolver
     about:
       homepage: ${{ load_from_file("python/libnvforest/pyproject.toml").project.urls.Homepage }}
@@ -163,11 +144,6 @@ outputs:
         by_name:
           - cuda-cudart
           - cuda-version
-          - libcublas
-          - libcufft
-          - libcurand
-          - libcusolver
-          - libcusparse
           - librmm
     about:
       homepage: ${{ load_from_file("python/libnvforest/pyproject.toml").project.urls.Homepage }}

cpp/CMakeLists.txt

@@ -141,13 +141,6 @@ endif()
 # ######################################################################################################################
 # * compiler options ---------------------------------------------------------
 
-set(_ctk_static_suffix "")
-set(_ctk_fft_static_suffix "")
-if(CUDA_STATIC_MATH_LIBRARIES)
-  set(_ctk_static_suffix "_static")
-  set(_ctk_fft_static_suffix "_static_nocallback")
-endif()
-
 if(NOT DISABLE_OPENMP)
   find_package(OpenMP)
   if(OpenMP_FOUND)
@@ -273,10 +266,7 @@ set_target_properties(
              POSITION_INDEPENDENT_CODE ON
              INTERFACE_POSITION_INDEPENDENT_CODE ON)
 
-target_compile_definitions(
-  ${NVFOREST_CPP_TARGET}
-  PUBLIC DISABLE_CUSPARSE_DEPRECATED
-  PRIVATE NVFOREST_CPP_API)
+target_compile_definitions(${NVFOREST_CPP_TARGET} PRIVATE NVFOREST_CPP_API)
 
 target_compile_options(${NVFOREST_CPP_TARGET} PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${NVFOREST_CXX_FLAGS}>"
                                                       "$<$<COMPILE_LANGUAGE:CUDA>:${NVFOREST_CUDA_FLAGS}>")

dependencies.yaml

@@ -123,7 +123,6 @@ files:
     extras:
       table: project
     includes:
-      - cuda_wheels
       - depends_on_cuda_python
       - depends_on_cupy
       - depends_on_libnvforest
@@ -162,7 +161,6 @@ files:
     extras:
       table: project
     includes:
-      - cuda_wheels
       - depends_on_libraft
       - depends_on_librmm
       - depends_on_rapids_logger
@@ -243,6 +241,8 @@ dependencies:
           - packaging
           - &scikit_learn scikit-learn>=1.5
           - *treelite
+  # 'cuda_version' intentionally does not contain fallback entries... we want
+  # a loud error if an unsupported 'cuda' value is passed
   cuda_version:
     specific:
       - output_types: conda
@@ -271,41 +271,57 @@ dependencies:
               cuda: "13.1"
             packages:
               - cuda-version=13.1
-  cuda:
-    common:
-      - output_types: [conda]
-        packages:
-          - cuda-cudart-dev
-          - cuda-profiler-api
-          - libcublas-dev
-          - libcufft-dev
-          - libcurand-dev
-          - libcusolver-dev
-          - libcusparse-dev
-  cuda_wheels:
-    specific:
-      - output_types: [requirements, pyproject]
+      - output_types: requirements
         matrices:
+          # if use_cuda_wheels=false is provided, do not add dependencies on any CUDA wheels
+          # (e.g. for DLFW and pip devcontainers)
           - matrix:
-              cuda: "12.*"
+              use_cuda_wheels: "false"
+            packages:
+          - matrix:
+              arch: aarch64
+              cuda: "12.2"
               use_cuda_wheels: "true"
             packages:
-              - cuda-toolkit==12.*
+              # some components (like nvidia-cublas-cu12 and nvidia-cuda-nvcc-cu12) didn't have
+              # aarch64 wheels until CTK 12.3, so allow a slightly looser bound here
+              - cuda-toolkit>=12.2,<12.4
           - matrix:
-              cuda: "13.*"
+              cuda: "12.2"
               use_cuda_wheels: "true"
             packages:
-              - cuda-toolkit==13.*
-          # if use_cuda_wheels=false is provided, do not add dependencies on any CUDA wheels
-          # (e.g. for DLFW and pip devcontainers)
+              - cuda-toolkit==12.2.*
           - matrix:
-              use_cuda_wheels: "false"
+              cuda: "12.5"
+              use_cuda_wheels: "true"
+            packages:
+              - cuda-toolkit==12.5.*
+          - matrix:
+              cuda: "12.8"
+              use_cuda_wheels: "true"
+            packages:
+              - cuda-toolkit==12.8.*
+          - matrix:
+              cuda: "12.9"
+              use_cuda_wheels: "true"
+            packages:
+              - cuda-toolkit==12.9.*
+          - matrix:
+              cuda: "13.0"
+              use_cuda_wheels: "true"
             packages:
-          # if no matching matrix selectors passed, list the unsuffixed packages
-          # (just as a source of documentation, as this populates pyproject.toml in source control)
+              - cuda-toolkit==13.0.*
           - matrix:
+              cuda: "13.1"
+              use_cuda_wheels: "true"
             packages:
-              - cuda-toolkit>=12,<14
+              - cuda-toolkit==13.1.*
+  cuda:
+    common:
+      - output_types: [conda]
+        packages:
+          - cuda-cudart-dev
+          - cuda-profiler-api
   docs:
     common:
       - output_types: [conda, requirements]

python/libnvforest/CMakeLists.txt

@@ -44,21 +44,9 @@ set(NVFOREST_EXCLUDE_TREELITE_FROM_ALL ON)
 
 # --- CUDA --- #
 set(CUDA_STATIC_RUNTIME ON)
-set(CUDA_STATIC_MATH_LIBRARIES OFF)
 
 add_subdirectory(../../cpp nvforest-cpp)
 
-# assumes libnvforest++ is installed 2 levels deep, e.g. site-packages/nvforest/lib64/libnvforest++.so
-if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-  # starting with CTK 13 wheels, all libraries except NCCL are grouped in this 'nvidia/cu13/lib' directory
-  set(rpaths "$ORIGIN/../../nvidia/cu13/lib")
-else()
-  set(rpaths
-      "$ORIGIN/../../nvidia/cublas/lib" "$ORIGIN/../../nvidia/cufft/lib" "$ORIGIN/../../nvidia/curand/lib"
-      "$ORIGIN/../../nvidia/cusolver/lib" "$ORIGIN/../../nvidia/cusparse/lib" "$ORIGIN/../../nvidia/nvjitlink/lib")
-endif()
-list(APPEND rpaths "$ORIGIN/../../nvidia/nccl/lib")
-
 set_property(
   TARGET ${NVFOREST_CPP_TARGET}
   PROPERTY INSTALL_RPATH ${rpaths}

python/libnvforest/pyproject.toml

@@ -25,7 +25,6 @@ classifiers = [
     "Environment :: GPU :: NVIDIA CUDA",
 ]
 dependencies = [
-    "cuda-toolkit>=12,<14",
     "libraft==26.4.*,>=0.0.0a0",
     "librmm==26.4.*,>=0.0.0a0",
     "rapids-logger==0.2.*,>=0.0.0a0",

python/nvforest/pyproject.toml

@@ -28,7 +28,6 @@ license = "Apache-2.0"
 requires-python = ">=3.11"
 dependencies = [
     "cuda-python>=13.0.1,<14.0a0",
-    "cuda-toolkit>=12,<14",
     "cupy-cuda13x>=13.6.0",
     "libnvforest==26.4.*,>=0.0.0a0",
     "numpy>=1.23,<3.0a0",