Migrate to CTK 13 as the default for Linux.

.github/workflows/containers.yml

@@ -32,13 +32,13 @@ jobs:
       fail-fast: false
       matrix:
         image_repo:
-          - xgb-ci.clang_tidy
           - xgb-ci.cpu
           - xgb-ci.cpu_build_r_doc
           - xgb-ci.gpu
+          - xgb-ci.gpu_cuda12
           - xgb-ci.gpu_build_r_rockylinux8
           - xgb-ci.gpu_build_rockylinux8
-          - xgb-ci.gpu_build_cuda13_rockylinux8
+          - xgb-ci.gpu_build_cuda12_rockylinux8
           - xgb-ci.jvm
           - xgb-ci.jvm_gpu_build
           - xgb-ci.manylinux_2_28_x86_64
@@ -50,14 +50,16 @@ jobs:
             runner: linux-arm64-cpu
           - image_repo: xgb-ci.manylinux_2_28_aarch64
             runner: linux-arm64-cpu
-          - image_repo: xgb-ci.gpu_build_cuda13_rockylinux8_aarch64
-            runner: linux-arm64-cpu
           - image_repo: xgb-ci.gpu_build_rockylinux8_aarch64
             runner: linux-arm64-cpu
+          - image_repo: xgb-ci.gpu_build_cuda12_rockylinux8_aarch64
+            runner: linux-arm64-cpu
           - image_repo: xgb-ci.cpu_aarch64
             runner: linux-arm64-cpu
           - image_repo: xgb-ci.gpu_aarch64
             runner: linux-arm64-cpu
+          - image_repo: xgb-ci.gpu_cuda12_aarch64
+            runner: linux-arm64-cpu
     steps:
       # Restart Docker daemon so that it recognizes the ephemeral disks
       - run: sudo systemctl restart docker

containers/ci_container.yml

@@ -7,8 +7,9 @@ x-rapids_versions:
   stable: &rapids_version "26.04"
 
 x-cuda_versions:
-  cuda: &cuda_version "12.9.0"
-  nccl: &nccl_version "2.29.2-1"
+  default: &cuda_version "13.3.0"
+  cuda12: &cuda12_version "12.9.0"
+  nccl: &nccl_version "2.30.7-1"
 
 xgb-ci.gpu_build_rockylinux8:
   container_def: gpu_build_rockylinux8
@@ -26,17 +27,18 @@ xgb-ci.gpu_build_rockylinux8_aarch64:
     RAPIDS_VERSION: *rapids_version
     ARCH: aarch64
 
-xgb-ci.gpu_build_cuda13_rockylinux8:
-  container_def: gpu_build_cuda13_rockylinux8
+# CUDA 12 builders are used for the xgboost-cu12 PyPI package and clang CI jobs.
+xgb-ci.gpu_build_cuda12_rockylinux8:
+  container_def: gpu_build_cuda12_rockylinux8
   build_args:
-    CUDA_VERSION: "13.0.0"
+    CUDA_VERSION: *cuda12_version
     NCCL_VERSION: *nccl_version
     ARCH: x86_64
 
-xgb-ci.gpu_build_cuda13_rockylinux8_aarch64:
-  container_def: gpu_build_cuda13_rockylinux8
+xgb-ci.gpu_build_cuda12_rockylinux8_aarch64:
+  container_def: gpu_build_cuda12_rockylinux8
   build_args:
-    CUDA_VERSION: "13.0.0"
+    CUDA_VERSION: *cuda12_version
     NCCL_VERSION: *nccl_version
     ARCH: aarch64
 
@@ -65,10 +67,22 @@ xgb-ci.gpu_aarch64:
     RAPIDS_VERSION: *rapids_version
     ARCH: aarch64
 
-xgb-ci.clang_tidy:
-  container_def: clang_tidy
+# CUDA 12 runtime images provide native test environments for xgboost-cu12.
+xgb-ci.gpu_cuda12:
+  container_def: gpu
   build_args:
-    CUDA_VERSION: *cuda_version
+    CUDA_VERSION: *cuda12_version
+    NCCL_VERSION: *nccl_version
+    RAPIDS_VERSION: *rapids_version
+    ARCH: x86_64
+
+xgb-ci.gpu_cuda12_aarch64:
+  container_def: gpu
+  build_args:
+    CUDA_VERSION: *cuda12_version
+    NCCL_VERSION: *nccl_version
+    RAPIDS_VERSION: *rapids_version
+    ARCH: aarch64
 
 xgb-ci.cpu:
   container_def: cpu

containers/conda_env/linux_cpu_test.yml

@@ -27,19 +27,21 @@ dependencies:
 - pytest
 - pytest-timeout
 - pytest-cov
-- python-kubernetes
 - urllib3
 - jsonschema
 - loky>=3.5.1
 - pyarrow
-- protobuf
-- cloudpickle
-- modin
+- polars
 # PySpark
+- cloudpickle
 - pyspark>=4.0
 - grpcio
 - grpcio-status
 - googleapis-common-protos
 - zstandard
+# The modin on conda forge is old at the time of writing (0.32.0). Old modin constraints
+# the numpy version to <2. Pulling it from PyPI is not safe with its large dependency
+# tree. Just disable it on the CPU CI for now.
+# - modin
 - pip:
   - py-ubjson

containers/dockerfile/Dockerfile.cpu

@@ -1,6 +1,6 @@
 FROM ubuntu:22.04
 ARG ARCH=x86_64
-ARG MINIFORGE_VERSION=24.9.2-0
+ARG MINIFORGE_VERSION=26.3.2-3
 
 SHELL ["/bin/bash", "-c"]
 
@@ -26,20 +26,6 @@ RUN \
     wget -nv -O conda.sh https://github.com/conda-forge/miniforge/releases/download/$MINIFORGE_VERSION/Miniforge3-$MINIFORGE_VERSION-Linux-${ARCH}.sh && \
     bash conda.sh -b -p /opt/miniforge
 
-# Install gRPC
-# Patch Abseil to apply https://github.com/abseil/abseil-cpp/issues/1629
-RUN git clone -b v1.65.4 https://github.com/grpc/grpc.git \
-      --recurse-submodules --depth 1 && \
-    pushd grpc && \
-    pushd third_party/abseil-cpp && \
-    git fetch origin master && \
-    git cherry-pick -n cfde5f74e276049727f9556f13473a59fe77d9eb && \
-    popd && \
-    cmake -S . -B build -GNinja -DCMAKE_INSTALL_PREFIX=/opt/grpc -DCMAKE_CXX_VISIBILITY_PRESET=hidden && \
-    cmake --build build --target install && \
-    popd && \
-    rm -rf grpc
-
 RUN mamba install awscli -y	# install awscli into base
 
 # Create new Conda environment

containers/dockerfile/Dockerfile.gpu

@@ -5,7 +5,7 @@ ARG RAPIDS_VERSION
   # Should be first 4 digits (e.g. 24.06)
 ARG NCCL_VERSION
 ARG ARCH=x86_64
-ARG MINIFORGE_VERSION=24.9.2-0
+ARG MINIFORGE_VERSION=26.3.2-3
 ARG PYTHON_VERSION=3.12
 
 SHELL ["/bin/bash", "-c"]

containers/dockerfile/Dockerfile.gpu_build_cuda13_rockylinux8 → containers/dockerfile/Dockerfile.gpu_build_cuda12_rockylinux8

@@ -3,7 +3,7 @@ FROM nvcr.io/nvidia/cuda:$CUDA_VERSION-devel-rockylinux8
 ARG CUDA_VERSION
 ARG NCCL_VERSION
 ARG ARCH=x86_64
-ARG MINIFORGE_VERSION=25.11.0-1
+ARG MINIFORGE_VERSION=26.3.2-3
 ARG CMAKE_VERSION=4.1.0
 
 SHELL ["/bin/bash", "-c"]
@@ -36,7 +36,7 @@ RUN \
     export NCCL_VERSION=$NCCL_VERSION && \
     dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/${CUDA_REPO_ARCH}/cuda-rhel8.repo && \
     dnf -y update && \
-    dnf install -y libnccl-${NCCL_VERSION}+cuda13.1 libnccl-devel-${NCCL_VERSION}+cuda13.1 libnccl-static-${NCCL_VERSION}+cuda13.1
+    dnf install -y libnccl-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-devel-${NCCL_VERSION}+cuda${CUDA_SHORT}
 
 # Install lightweight sudo (not bound to TTY)
 RUN sh /scripts/install_gosu.sh

containers/dockerfile/Dockerfile.gpu_build_r_rockylinux8

@@ -2,8 +2,8 @@ ARG CUDA_VERSION=notset
 FROM nvcr.io/nvidia/cuda:$CUDA_VERSION-devel-rockylinux8
 ARG CUDA_VERSION
 ARG R_VERSION
-ARG MINIFORGE_VERSION=24.9.2-0
-ARG CMAKE_VERSION=3.31.2
+ARG MINIFORGE_VERSION=26.3.2-3
+ARG CMAKE_VERSION=4.1.0
 
 SHELL ["/bin/bash", "-c"]
 

containers/dockerfile/Dockerfile.gpu_build_rockylinux8

@@ -4,8 +4,8 @@ ARG CUDA_VERSION
 ARG NCCL_VERSION
 ARG RAPIDS_VERSION
 ARG ARCH=x86_64
-ARG MINIFORGE_VERSION=24.9.2-0
-ARG CMAKE_VERSION=3.31.2
+ARG MINIFORGE_VERSION=26.3.2-3
+ARG CMAKE_VERSION=4.1.0
 
 SHELL ["/bin/bash", "-c"]
 
@@ -39,29 +39,19 @@ RUN \
     dnf -y update && \
     dnf install -y libnccl-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-devel-${NCCL_VERSION}+cuda${CUDA_SHORT}
 
-# Install gRPC
-# Patch Abseil to apply https://github.com/abseil/abseil-cpp/issues/1629
-RUN git clone -b v1.65.4 https://github.com/grpc/grpc.git \
-      --recurse-submodules --depth 1 && \
-    pushd grpc && \
-    pushd third_party/abseil-cpp && \
-    git fetch origin master && \
-    git cherry-pick -n cfde5f74e276049727f9556f13473a59fe77d9eb && \
-    popd && \
-    cmake -S . -B build -GNinja -DCMAKE_INSTALL_PREFIX=/opt/grpc -DCMAKE_CXX_VISIBILITY_PRESET=hidden && \
-    cmake --build build --target install && \
-    popd && \
-    rm -rf grpc
-
 # Install RMM
 # Patch out -Werror
+# Force the static logger dependencies to use the container's GCC toolchain instead of
+# reusing ABI-incompatible shared libraries from Miniforge.
 RUN git clone -b v${RAPIDS_VERSION}.00 https://github.com/rapidsai/rmm.git --recurse-submodules --depth 1 && \
     pushd rmm && \
     find . -name CMakeLists.txt -print0 | xargs -0 sed -i 's/-Werror//g' && \
     pushd cpp && \
     mkdir build && \
     pushd build && \
-    cmake .. -GNinja -DCMAKE_INSTALL_PREFIX=/opt/rmm -DCUDA_STATIC_RUNTIME=ON -DBUILD_SHARED_LIBS=OFF -DBUILD_TESTS=OFF && \
+    cmake .. -GNinja -DCMAKE_INSTALL_PREFIX=/opt/rmm -DCUDA_STATIC_RUNTIME=ON \
+      -DBUILD_SHARED_LIBS=OFF -DBUILD_TESTS=OFF \
+      -DCPM_DOWNLOAD_fmt=ON -DCPM_DOWNLOAD_spdlog=ON && \
     cmake --build . --target install && \
     popd && popd && popd && \
     rm -rf rmm

containers/dockerfile/Dockerfile.jvm

@@ -1,6 +1,6 @@
 FROM rockylinux:8
-ARG MINIFORGE_VERSION=24.9.2-0
-ARG CMAKE_VERSION=3.31.2
+ARG MINIFORGE_VERSION=26.3.2-3
+ARG CMAKE_VERSION=4.1.0
 ARG MAVEN_VERSION=3.9.16
 
 SHELL ["/bin/bash", "-c"]

containers/dockerfile/Dockerfile.jvm_gpu_build

@@ -2,8 +2,8 @@ ARG CUDA_VERSION=notset
 FROM nvcr.io/nvidia/cuda:$CUDA_VERSION-devel-rockylinux8
 ARG CUDA_VERSION
 ARG NCCL_VERSION
-ARG MINIFORGE_VERSION=24.9.2-0
-ARG CMAKE_VERSION=3.31.2
+ARG MINIFORGE_VERSION=26.3.2-3
+ARG CMAKE_VERSION=4.1.0
 ARG MAVEN_VERSION=3.9.16
 
 SHELL ["/bin/bash", "-c"]

containers/dockerfile/Dockerfile.manylinux_2_28_aarch64

@@ -1,5 +1,5 @@
 FROM quay.io/pypa/manylinux_2_28_aarch64
-ARG MINIFORGE_VERSION=24.9.2-0
+ARG MINIFORGE_VERSION=26.3.2-3
 
 SHELL ["/bin/bash", "-c"]
 

containers/extract_build_args.sh

@@ -3,9 +3,9 @@
 ## given the image repo.
 ##
 ## Example input:
-##   xgb-ci.clang_tidy
+##   xgb-ci.cpu
 ## Example output:
-##   CONTAINER_DEF='clang_tidy' BUILD_ARGS='--build-arg CUDA_VERSION_ARG=12.4.1'
+##   CONTAINER_DEF='cpu' BUILD_ARGS='--build-arg ARCH=x86_64'
 
 if [ "$#" -ne 1 ]; then
     echo "Usage: $0 [image_repo]"