From 31ff34776854a8ac0fb3805c8b454488229ebeba Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jonatan=20K=C5=82osko?= Date: Tue, 17 Jun 2025 21:01:09 +0200 Subject: [PATCH] Bump Docker images to CUDA 12.8 and include NCCL (#3018) --- .github/workflows/release.yml | 3 ++- Dockerfile | 6 +++--- 2 files changed, 5 insertions(+), 4 deletions(-) diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml index d55836792..f0cc88fe2 100644 --- a/.github/workflows/release.yml +++ b/.github/workflows/release.yml @@ -75,7 +75,8 @@ jobs: tag_suffix: "-cuda12" build_args: | VARIANT=cuda - CUDA_VERSION=12-6 + CUDA_VERSION_MAJOR=12 + CUDA_VERSION_MINOR=8 steps: - uses: actions/checkout@v4 - run: | diff --git a/Dockerfile b/Dockerfile index 522e74656..a7520bf92 100644 --- a/Dockerfile +++ b/Dockerfile @@ -7,7 +7,8 @@ FROM ${BASE_IMAGE} AS base-default FROM ${BASE_IMAGE} AS base-cuda -ARG CUDA_VERSION +ARG CUDA_VERSION_MAJOR +ARG CUDA_VERSION_MINOR RUN distro="ubuntu$(. /etc/lsb-release; echo "$DISTRIB_RELEASE" | tr -d '.')" && \ # Official Docker images use the sbsa packages when targetting arm64. @@ -20,8 +21,7 @@ RUN distro="ubuntu$(. /etc/lsb-release; echo "$DISTRIB_RELEASE" | tr -d '.')" && # the CUDA toolkit that is required by Elixir numerical packages # (nvcc and runtime libraries). Note that we do not need to install # the driver, it is already provided by NVIDIA Container Toolkit. - cuda_version="${CUDA_VERSION}" && cuda_major="${cuda_version%-*}" && \ - apt-get install -y git cuda-nvcc-${CUDA_VERSION} cuda-libraries-${CUDA_VERSION} libcudnn9-cuda-$cuda_major && \ + apt-get install -y git cuda-nvcc-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR} cuda-libraries-${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR} libcudnn9-cuda-${CUDA_VERSION_MAJOR} libnccl2=*+cuda${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} && \ apt-get clean -y && rm -rf /var/lib/apt/lists/* ENV PATH="/usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH"