CUDA 12.3 w/ EFA

Dockerfile

# https://nvidia.github.io/container-wiki/toolkit/container-images.html
# https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda/tags
# https://gitlab.com/nvidia/container-images/cuda
FROM nvcr.io/nvidia/cuda:12.3.2-cudnn9-runtime-centos7 as base

ARG EFA_INSTALLER_VERSION='1.30.0'
ARG AWS_OFI_NCCL_VERSION='1.8.1'

ENV PATH="$PATH:/opt/amazon/efa/bin:/opt/amazon/openmpi/bin"
ENV LD_LIBRARY_PATH="/usr/lib64/:/usr/local/cuda/lib64/:/opt/amazon/efa/lib64/:/opt/amazon/openmpi/lib64/"

# AWS EFA installer
RUN curl -sL https://efa-installer.amazonaws.com/aws-efa-installer-${EFA_INSTALLER_VERSION}.tar.gz | tar xvz \
  && pushd aws-efa-installer \
    && ./efa_installer.sh --skip-kmod --skip-limit-conf --no-verify --yes \
  && popd \
  && rm -rf aws-efa-installer* \
  && yum clean all \
  && rm -rf /var/cache/dnf/* /var/cache/yum/*

# AWS OFI NCCL
RUN yum install -y \
    automake \
    autoconf \
    libtool \
  && curl -sL https://github.com/aws/aws-ofi-nccl/releases/download/v${AWS_OFI_NCCL_VERSION}-aws/aws-ofi-nccl-${AWS_OFI_NCCL_VERSION}-aws.tar.gz | tar xvz \
  && pushd aws-ofi-nccl-${AWS_OFI_NCCL_VERSION}-aws \
    && ./autogen.sh \
    && ./configure --prefix=/opt/aws-ofi-nccl/install \
      --with-libfabric=/opt/amazon/efa/ \
      --with-mpi=/opt/amazon/openmpi/ \
      --with-cuda=/usr/local/cuda \
      --with-nccl=/usr/lib64/ \
      --enable-platform-aws \
    && make \
    && make install \
  && popd \
  && rm -rf aws-ofi-nccl-* \
  && yum clean all \
  && rm -rf /var/cache/dnf/* /var/cache/yum/*