Migrate to CTK13 as the default for Linux.

.github/workflows/lint.yml

@@ -26,14 +26,14 @@ jobs:
       tag-prefix: lint
 
   clang-tidy:
-    name: Run clang-tidy
+    name: Run clang-tidy with CUDA 12.9
     needs: ci-configure
     runs-on:
       - runs-on=${{ github.run_id }}
       - runner=linux-amd64-cpu
       - tag=lint-clang-tidy-clang-cuda
     container:
-      image: ${{ needs.ci-configure.outputs.docker_registry }}/xgb-ci.gpu_build_rockylinux8:${{ needs.ci-configure.outputs.image_tag }}
+      image: ${{ needs.ci-configure.outputs.docker_registry }}/xgb-ci.gpu_build_cuda12_rockylinux8:${{ needs.ci-configure.outputs.image_tag }}
       credentials:
         username: ${{ needs.ci-configure.outputs.docker_username }}
         password: ${{ needs.ci-configure.outputs.docker_password }}

.github/workflows/main.yml

@@ -70,17 +70,16 @@ jobs:
       - uses: dmlc/xgboost-devops/actions/sccache@main
         with:
           cache-key-prefix: ${{ github.job }}-${{ matrix.build_suite }}
-      - name: Build and test
-        run: bash ops/pipeline/build-cpu.sh ${{ matrix.build_suite }}
-      - name: Test federated learning on CPU
+      - name: Build run C++ tests on CPU
         if: matrix.variant == 'default'
         shell: bash -l {0}
         run: |
           source activate linux_cpu_test
-          conda install scikit-build-core -y
-          cd python-package && pip install -e . --no-deps --no-build-isolation && cd -
-          timeout 600s pytest -v -s -rxXs --durations=0 tests/test_distributed/test_federated
+          bash ops/pipeline/build-cpu.sh ${{ matrix.build_suite }}
       - run: sccache --show-stats
+      - name: Build and test with sanitizer
+        if: matrix.variant == 'sanitizer'
+        run: bash ops/pipeline/build-cpu.sh ${{ matrix.build_suite }}
 
   build-cuda:
     name: Build CUDA ${{ matrix.cuda_version }} (${{ matrix.arch }})
@@ -94,27 +93,27 @@ jobs:
       fail-fast: false
       matrix:
         include:
-        # CUDA 12
+        # xgboost-cu12
         - cuda_version: 12
           arch: aarch64
           runner: linux-arm64-cpu
-          image_repo: xgb-ci.gpu_build_rockylinux8_aarch64
+          image_repo: xgb-ci.gpu_build_cuda12_rockylinux8_aarch64
           use_rmm: 0
         - cuda_version: 12
           arch: x86_64
           runner: linux-amd64-cpu
-          image_repo: xgb-ci.gpu_build_rockylinux8
+          image_repo: xgb-ci.gpu_build_cuda12_rockylinux8
           use_rmm: 0
-        # CUDA 13
+        # Default xgboost package (CUDA 13)
         - cuda_version: 13
           arch: aarch64
           runner: linux-arm64-cpu
-          image_repo: xgb-ci.gpu_build_cuda13_rockylinux8_aarch64
+          image_repo: xgb-ci.gpu_build_rockylinux8_aarch64
           use_rmm: 0
         - cuda_version: 13
           arch: x86_64
           runner: linux-amd64-cpu
-          image_repo: xgb-ci.gpu_build_cuda13_rockylinux8
+          image_repo: xgb-ci.gpu_build_rockylinux8
           use_rmm: 0
     container:
       image: ${{ needs.ci-configure.outputs.docker_registry }}/${{ matrix.image_repo }}:${{ needs.ci-configure.outputs.image_tag }}
@@ -142,15 +141,15 @@ jobs:
             build/testxgboost python-package/dist/*.whl
 
   build-cuda-clang:
-    name: Build CUDA with clang 21.1.8 (x86_64)
+    name: Build CUDA 12.9 with clang 22.1.8 (x86_64)
     needs: ci-configure
     runs-on:
       - runs-on=${{ github.run_id }}
       - runner=linux-amd64-cpu
       - tag=main-build-cuda-clang
       - extras=s3-cache
     container:
-      image: ${{ needs.ci-configure.outputs.docker_registry }}/xgb-ci.gpu_build_rockylinux8:${{ needs.ci-configure.outputs.image_tag }}
+      image: ${{ needs.ci-configure.outputs.docker_registry }}/xgb-ci.gpu_build_cuda12_rockylinux8:${{ needs.ci-configure.outputs.image_tag }}
       credentials:
         username: ${{ needs.ci-configure.outputs.docker_username }}
         password: ${{ needs.ci-configure.outputs.docker_password }}
@@ -161,8 +160,8 @@ jobs:
           submodules: "true"
       - uses: dmlc/xgboost-devops/actions/sccache@main
         with:
-          cache-key-prefix: ${{ github.job }}-clang21-cuda12
-      - run: bash ops/pipeline/build-cuda-clang.sh --clang-version 21.1.8
+          cache-key-prefix: ${{ github.job }}-clang22-1-8-cuda12
+      - run: bash ops/pipeline/build-cuda-clang.sh --clang-version 22.1.8
       - run: sccache --show-stats
 
   audit-cuda-wheel:
@@ -176,22 +175,22 @@ jobs:
       fail-fast: false
       matrix:
         include:
-        # CUDA 12
+        # xgboost-cu12
         - cuda_version: 12
-          cuda_variant_flag: ""
+          cuda_variant_flag: "--cuda-variant cuda12"
           arch: aarch64
           runner: linux-arm64-cpu
         - cuda_version: 12
-          cuda_variant_flag: ""
+          cuda_variant_flag: "--cuda-variant cuda12"
           arch: x86_64
           runner: linux-amd64-cpu
-        # CUDA 13
+        # Default xgboost package (CUDA 13)
         - cuda_version: 13
-          cuda_variant_flag: "--cuda-variant cuda13"
+          cuda_variant_flag: ""
           arch: aarch64
           runner: linux-arm64-cpu
         - cuda_version: 13
-          cuda_variant_flag: "--cuda-variant cuda13"
+          cuda_variant_flag: ""
           arch: x86_64
           runner: linux-amd64-cpu
     container:
@@ -225,7 +224,7 @@ jobs:
             python-package/dist/*.whl
 
   build-cuda-with-rmm:
-    name: Build CUDA with RMM
+    name: Build CUDA 13 with RMM
     needs: ci-configure
     runs-on:
       - runs-on=${{ github.run_id }}
@@ -243,9 +242,11 @@ jobs:
         with:
           submodules: "true"
       - uses: dmlc/xgboost-devops/actions/sccache@main
+        with:
+          cache-key-prefix: ${{ github.job }}-cuda13
       - run: >-
           bash ops/pipeline/build-cuda.sh
-          --cuda-version 12
+          --cuda-version 13
           --use-rmm 1
       - run: sccache --show-stats
       - name: Stash files
@@ -331,58 +332,58 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          # CUDA 12 tests
+          # CUDA 12 artifact tests using the native CUDA 12.9 runtime
           # Note: --gpus all provides GPU access; --privileged is not needed for basic GPU tests
           - cuda_version: 12
             suite: gpu
             arch: x86_64
             runner: linux-amd64-gpu
-            image_repo: xgb-ci.gpu
+            image_repo: xgb-ci.gpu_cuda12
             artifact_from: build-cuda12-x86_64
             container_options: "--gpus all --privileged"
             test_args: ""
           - cuda_version: 12
             suite: gpu
             arch: aarch64
             runner: linux-arm64-gpu
-            image_repo: xgb-ci.gpu_aarch64
+            image_repo: xgb-ci.gpu_cuda12_aarch64
             artifact_from: build-cuda12-aarch64
             container_options: "--gpus all --privileged"
             test_args: ""
-          - cuda_version: 12
-            suite: gpu-rmm
+          # Default CUDA 13 artifact tests
+          - cuda_version: 13
+            suite: gpu
             arch: x86_64
             runner: linux-amd64-gpu
             image_repo: xgb-ci.gpu
-            artifact_from: build-cuda-with-rmm
+            artifact_from: build-cuda13-x86_64
             container_options: "--gpus all --privileged"
-            test_args: "--use-rmm-pool --gtest_filter=-*DeathTest*"
-          - cuda_version: 12
+            test_args: ""
+          - cuda_version: 13
+            suite: gpu
+            arch: aarch64
+            runner: linux-arm64-gpu
+            image_repo: xgb-ci.gpu_aarch64
+            artifact_from: build-cuda13-aarch64
+            container_options: "--gpus all --privileged"
+            test_args: ""
+          - cuda_version: 13
             suite: mgpu
             arch: x86_64
             runner: linux-amd64-mgpu
             image_repo: xgb-ci.gpu
-            artifact_from: build-cuda12-x86_64
+            artifact_from: build-cuda13-x86_64
             # mgpu needs --shm-size for NCCL shared memory communication
             container_options: "--gpus all --shm-size=4g --privileged"
             test_args: "--gtest_filter=*MGPU*"
-          # CUDA 13 tests
           - cuda_version: 13
-            suite: gpu
+            suite: gpu-rmm
             arch: x86_64
             runner: linux-amd64-gpu
-            image_repo: xgb-ci.gpu_build_cuda13_rockylinux8
-            artifact_from: build-cuda13-x86_64
-            container_options: "--gpus all --privileged"
-            test_args: ""
-          - cuda_version: 13
-            suite: gpu
-            arch: aarch64
-            runner: linux-arm64-gpu
-            image_repo: xgb-ci.gpu_build_cuda13_rockylinux8_aarch64
-            artifact_from: build-cuda13-aarch64
+            image_repo: xgb-ci.gpu
+            artifact_from: build-cuda-with-rmm
             container_options: "--gpus all --privileged"
-            test_args: ""
+            test_args: "--use-rmm-pool --gtest_filter=-*DeathTest*"
     container:
       image: ${{ needs.ci-configure.outputs.docker_registry }}/${{ matrix.image_repo }}:${{ needs.ci-configure.outputs.image_tag }}
       credentials:
@@ -416,41 +417,41 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          # CUDA 12 tests
+          # xgboost-cu12 tests using the native CUDA 12.9 runtime
           - cuda_version: 12
             description: GPU-x86_64-CUDA-12
-            image_repo: xgb-ci.gpu
+            image_repo: xgb-ci.gpu_cuda12
             suite: gpu
             runner: linux-amd64-gpu
             artifact_from: audit-cuda12-wheel-x86_64
             container_options: "--gpus all --privileged"
-          - cuda_version: 12
-            description: Multi-GPU-x86_64-CUDA-12
-            image_repo: xgb-ci.gpu
-            suite: mgpu
-            runner: linux-amd64-mgpu
-            artifact_from: audit-cuda12-wheel-x86_64
-            # mgpu needs --shm-size for NCCL shared memory communication
-            container_options: "--gpus all --shm-size=4g --privileged"
-          # CUDA 12 aarch64 tests
+          # xgboost-cu12 aarch64 test using the native CUDA 12.9 runtime
           - cuda_version: 12
             description: GPU-arm64-CUDA-12
-            image_repo: xgb-ci.gpu_aarch64
+            image_repo: xgb-ci.gpu_cuda12_aarch64
             suite: gpu-arm64
             runner: linux-arm64-gpu
             artifact_from: audit-cuda12-wheel-aarch64
             container_options: "--gpus all --privileged"
-          # CUDA 13 tests
+          # Default xgboost package tests (CUDA 13)
           - cuda_version: 13
             description: GPU-x86_64-CUDA-13
-            image_repo: xgb-ci.gpu_build_cuda13_rockylinux8
+            image_repo: xgb-ci.gpu
             suite: gpu
             runner: linux-amd64-gpu
             artifact_from: audit-cuda13-wheel-x86_64
             container_options: "--gpus all --privileged"
+          - cuda_version: 13
+            description: Multi-GPU-x86_64-CUDA-13
+            image_repo: xgb-ci.gpu
+            suite: mgpu
+            runner: linux-amd64-mgpu
+            artifact_from: audit-cuda13-wheel-x86_64
+            # mgpu needs --shm-size for NCCL shared memory communication
+            container_options: "--gpus all --shm-size=4g --privileged"
           - cuda_version: 13
             description: GPU-arm64-CUDA-13
-            image_repo: xgb-ci.gpu_build_cuda13_rockylinux8_aarch64
+            image_repo: xgb-ci.gpu_aarch64
             suite: gpu-arm64
             runner: linux-arm64-gpu
             artifact_from: audit-cuda13-wheel-aarch64
@@ -476,16 +477,16 @@ jobs:
           bash ops/pipeline/test-python-wheel.sh
           --suite ${{ matrix.suite }}
           --cuda-version ${{ matrix.cuda_version }}
-      # Train a model for cross-platform testing (only for CUDA 12 x86_64)
+      # Train a model for cross-platform testing (only for default CUDA 13 x86_64)
       - name: Train cross-platform test model
-        if: matrix.cuda_version == 12 && matrix.suite == 'gpu'
+        if: matrix.cuda_version == 13 && matrix.suite == 'gpu'
         shell: bash -l {0}
         run: |
           source activate gpu_test
           python tests/cross-platform/test_cross_platform_model.py \
             --train --model-path cross_platform_model.ubj
       - name: Upload cross-platform model artifact
-        if: matrix.cuda_version == 12 && matrix.suite == 'gpu'
+        if: matrix.cuda_version == 13 && matrix.suite == 'gpu'
         uses: actions/upload-artifact@v7.0.1
         with:
           name: cross-platform-model
@@ -505,20 +506,20 @@ jobs:
     timeout-minutes: 60
     strategy:
       fail-fast: false
-      # Uses the wheel from cuda12 for tests.
+      # Uses the default CUDA 13 wheel for tests.
       matrix:
         include:
           - description: CPU-amd64
             image_repo: xgb-ci.cpu
             suite: cpu
             runner: linux-amd64-cpu
-            artifact_from: audit-cuda12-wheel-x86_64
+            artifact_from: audit-cuda13-wheel-x86_64
             container_options: "--init"
           - description: CPU-arm64
             image_repo: xgb-ci.cpu_aarch64
             suite: cpu-arm64
             runner: linux-arm64-cpu
-            artifact_from: audit-cuda12-wheel-aarch64
+            artifact_from: audit-cuda13-wheel-aarch64
             container_options: "--init"
     container:
       image: ${{ needs.ci-configure.outputs.docker_registry }}/${{ matrix.image_repo }}:${{ needs.ci-configure.outputs.image_tag }}

doc/build.rst

@@ -139,23 +139,6 @@ architectures can be found `in this page <https://developer.nvidia.com/cuda-gpus
   building XGBoost with NCCL as a shared library, while ``USE_DLOPEN_NCCL`` enables
   XGBoost to load NCCL at runtime using ``dlopen``.
 
-Federated Learning
-==================
-
-The federated learning plugin requires ``grpc`` and ``protobuf``. To install grpc, refer
-to the `installation guide from the gRPC website
-<https://grpc.io/docs/languages/cpp/quickstart/>`_. Alternatively, one can use the
-``libgrpc`` and the ``protobuf`` package from conda forge if conda is available. After
-obtaining the required dependencies, enable the flag: ``-DPLUGIN_FEDERATED=ON`` when
-running CMake. Please note that only Linux is supported for the federated plugin.
-
-
-.. code-block:: bash
-
-  cmake -B build -S . -DPLUGIN_FEDERATED=ON -GNinja
-  cd build && ninja
-
-
 .. _build_python:
 
 ***********************************
@@ -206,8 +189,7 @@ There are several ways to build and install the package from source:
         --config-settings cmake.define.USE_CUDA=ON \
         --config-settings cmake.define.USE_NCCL=ON
 
-  Common CMake options
-  (see :ref:`building-the-shared-library` for the full list):
+  Common CMake options (see :ref:`build_shared_lib` for the full list):
 
   - ``USE_CUDA`` — build with CUDA / GPU acceleration
   - ``USE_NCCL`` — build with NCCL for distributed GPU training