Upgrade to CUDA 12.9 (#77)

* Upgrade to CUDA 12.9

* Use latest NCCL

* fix

Author: hcho3

containers/ci_container.yml

@@ -7,8 +7,8 @@ x-rapids_versions:
   stable: &rapids_version "25.12"
 
 x-cuda_versions:
-  cuda: &cuda_version "12.8.0"
-  nccl: &nccl_version "2.27.7-1"
+  cuda: &cuda_version "12.9.0"
+  nccl: &nccl_version "2.29.2-1"
 
 xgb-ci.gpu_build_rockylinux8:
   container_def: gpu_build_rockylinux8

containers/dockerfile/Dockerfile.gpu

@@ -18,13 +18,14 @@ ENV JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/
 # Install all basic requirements
 RUN \
     { [ $ARCH = "aarch64" ] && export CUDA_REPO_ARCH="sbsa" || export CUDA_REPO_ARCH="x86_64"; } && \
+    export CUDA_SHORT=`echo $CUDA_VERSION | grep -o -E '[0-9]+\.[0-9]'` && \
     export NCCL_VERSION=$NCCL_VERSION && \
     sed -i 's/ports.ubuntu.com/mirrors.ocf.berkeley.edu/g' /etc/apt/sources.list && \
     apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/${CUDA_REPO_ARCH}/3bf863cc.pub && \
     apt-get update && \
     apt-get install -y wget unzip bzip2 libgomp1 build-essential openjdk-8-jdk-headless && \
-    apt-get install "libnccl2=${NCCL_VERSION}+cuda12.9" \
-      "libnccl-dev=${NCCL_VERSION}+cuda12.9" -y --allow-change-held-packages && \
+    apt-get install "libnccl2=${NCCL_VERSION}+cuda${CUDA_SHORT}" \
+      "libnccl-dev=${NCCL_VERSION}+cuda${CUDA_SHORT}" -y --allow-change-held-packages && \
     # Miniforge
     wget -nv -O conda.sh https://github.com/conda-forge/miniforge/releases/download/$MINIFORGE_VERSION/Miniforge3-$MINIFORGE_VERSION-Linux-${ARCH}.sh && \
     bash conda.sh -b -p /opt/miniforge

containers/dockerfile/Dockerfile.gpu_build_cuda13_rockylinux8

@@ -35,10 +35,11 @@ RUN \
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
     { [ $ARCH = "aarch64" ] && export CUDA_REPO_ARCH="sbsa" || export CUDA_REPO_ARCH="x86_64"; } && \
+    export CUDA_SHORT=`echo $CUDA_VERSION | grep -o -E '[0-9]+\.[0-9]'` && \
     export NCCL_VERSION=$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}+cuda13.0 libnccl-devel-${NCCL_VERSION}+cuda13.0 libnccl-static-${NCCL_VERSION}+cuda13.0
+    dnf install -y libnccl-${NCCL_VERSION}+cuda13.1 libnccl-devel-${NCCL_VERSION}+cuda13.1 libnccl-static-${NCCL_VERSION}+cuda13.1
 
 # Install lightweight sudo (not bound to TTY)
 RUN set -ex; \

containers/dockerfile/Dockerfile.gpu_build_rockylinux8

@@ -36,10 +36,11 @@ RUN \
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
     { [ $ARCH = "aarch64" ] && export CUDA_REPO_ARCH="sbsa" || export CUDA_REPO_ARCH="x86_64"; } && \
+    export CUDA_SHORT=`echo $CUDA_VERSION | grep -o -E '[0-9]+\.[0-9]'` && \
     export NCCL_VERSION=$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}+cuda12.9 libnccl-devel-${NCCL_VERSION}+cuda12.9
+    dnf install -y libnccl-${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

containers/dockerfile/Dockerfile.jvm_gpu_build

@@ -35,10 +35,11 @@ RUN \
 
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
+    export CUDA_SHORT=`echo $CUDA_VERSION | grep -o -E '[0-9]+\.[0-9]'` && \
     export NCCL_VERSION=$NCCL_VERSION && \
     dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo && \
     dnf -y update && \
-    dnf install -y libnccl-${NCCL_VERSION}+cuda12.9 libnccl-devel-${NCCL_VERSION}+cuda12.9 libnccl-static-${NCCL_VERSION}+cuda12.9
+    dnf install -y libnccl-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-devel-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-static-${NCCL_VERSION}+cuda${CUDA_SHORT}
 
 # Install Python packages
 RUN pip install numpy pytest scipy scikit-learn wheel kubernetes awscli