0715ab3c10
This PR adds optional support for loading nccl with `dlopen` as an alternative of compile time linking. This is to address the size bloat issue with the PyPI binary release. - Add CMake option to load `nccl` at runtime. - Add an NCCL stub. After this, `nccl` will be fetched from PyPI when using pip to install XGBoost, either by a user or by `pyproject.toml`. Others who want to link the nccl at compile time can continue to do so without any change. At the moment, this is Linux only since we only support MNMG on Linux.
50 lines
1.9 KiB
Docker
50 lines
1.9 KiB
Docker
ARG CUDA_VERSION_ARG
|
|
FROM nvidia/cuda:$CUDA_VERSION_ARG-runtime-ubuntu22.04
|
|
ARG CUDA_VERSION_ARG
|
|
ARG RAPIDS_VERSION_ARG
|
|
ARG NCCL_VERSION_ARG
|
|
|
|
# Environment
|
|
ENV DEBIAN_FRONTEND noninteractive
|
|
SHELL ["/bin/bash", "-c"] # Use Bash as shell
|
|
|
|
# Install all basic requirements
|
|
RUN \
|
|
apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub && \
|
|
apt-get update && \
|
|
apt-get install -y wget unzip bzip2 libgomp1 build-essential openjdk-8-jdk-headless && \
|
|
# Python
|
|
wget -nv -O conda.sh https://github.com/conda-forge/miniforge/releases/download/22.11.1-2/Mambaforge-22.11.1-2-Linux-x86_64.sh && \
|
|
bash conda.sh -b -p /opt/mambaforge
|
|
|
|
ENV PATH=/opt/mambaforge/bin:$PATH
|
|
|
|
# Create new Conda environment with cuDF, Dask, and cuPy
|
|
RUN \
|
|
conda install -c conda-forge mamba && \
|
|
mamba create -n gpu_test -c rapidsai-nightly -c rapidsai -c nvidia -c conda-forge -c defaults \
|
|
python=3.10 cudf=$RAPIDS_VERSION_ARG* rmm=$RAPIDS_VERSION_ARG* cudatoolkit=$CUDA_VERSION_ARG \
|
|
nccl>=$(cut -d "-" -f 1 << $NCCL_VERSION_ARG) \
|
|
dask \
|
|
dask-cuda=$RAPIDS_VERSION_ARG* dask-cudf=$RAPIDS_VERSION_ARG* cupy \
|
|
numpy pytest pytest-timeout scipy scikit-learn pandas matplotlib wheel python-kubernetes urllib3 graphviz hypothesis \
|
|
pyspark>=3.4.0 cloudpickle cuda-python && \
|
|
mamba clean --all && \
|
|
conda run --no-capture-output -n gpu_test pip install buildkite-test-collector
|
|
|
|
ENV GOSU_VERSION 1.10
|
|
ENV JAVA_HOME /usr/lib/jvm/java-8-openjdk-amd64/
|
|
|
|
# Install lightweight sudo (not bound to TTY)
|
|
RUN set -ex; \
|
|
wget -nv -O /usr/local/bin/gosu "https://github.com/tianon/gosu/releases/download/$GOSU_VERSION/gosu-amd64" && \
|
|
chmod +x /usr/local/bin/gosu && \
|
|
gosu nobody true
|
|
|
|
# Default entry-point to use if running locally
|
|
# It will preserve attributes of created files
|
|
COPY entrypoint.sh /scripts/
|
|
|
|
WORKDIR /workspace
|
|
ENTRYPOINT ["/scripts/entrypoint.sh"]
|