ignite Docker & Self-Hosted Deployment

# Multi-stage build
# Dockerfile.hvd-apex

ARG PTH_VERSION

# 1/Building apex with pytorch:*-devel
FROM pytorch/pytorch:${PTH_VERSION}-devel AS apex-hvd-builder

ENV CUDA_HOME=/usr/local/cuda
# Set env vars to allow pip install
ENV PIP_BREAK_SYSTEM_PACKAGES=1
ENV PIP_ROOT_USER_ACTION=ignore

# Temporary NCCL library setup (missing in pytorch 2.13 image)
RUN export NCCL_ROOT=$(python -c "import nvidia; print(nvidia.__path__[0])")/nccl && \
        libnccl_filename=$(ls $NCCL_ROOT/lib) && \
        libnccl_path=$NCCL_ROOT/lib/$libnccl_filename && \
        ln -s $libnccl_path /lib/x86_64-linux-gnu/$libnccl_filename && \
        ln -s /lib/x86_64-linux-gnu/$libnccl_filename /lib/x86_64-linux-gnu/libnccl.so && \
        cp -R $NCCL_ROOT/include/* /usr/include && \
        # quick check:
        ls /lib/x86_64-linux-gnu/libnccl.so && \
        ls /usr/include/nccl.h

# Install git
RUN apt-get update && apt-get install -y --no-install-recommends git && \
     rm -rf /var/lib/apt/lists/*

# Build apex
RUN echo "Setup NVIDIA Apex" && \
    tmp_apex_path="/tmp/apex" && \
    rm -rf $tmp_apex_path && \
    git clone https://github.com/NVIDIA/apex $tmp_apex_path && \
    cd $tmp_apex_path && \
    pip install packaging && \
    pip wheel -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" .

ARG HVD_VERSION

# Build Horovod
RUN apt-get update && apt-get install -y git && \
        pip install -U "cmake<4" && \
        git clone --depth 1 https://github.com/horovod/horovod.git /horovod && \
        cd /horovod && \
        git fetch --tags && \
        git checkout ${HVD_VERSION} && \
        git submodule update --init --recursive && \
        # temporary -std=c++17 fix
        sed -i "s/CMAKE_CXX_STANDARD 14/CMAKE_CXX_STANDARD 17/g" CMakeLists.txt && \
        sed -i "s/CMAKE_CXX_STANDARD 14/CMAKE_CXX_STANDARD 17/g" horovod/torch/CMakeLists.txt && \
        sed -i "s/VERSION 2.8/VERSION 3.5/g" third_party/gloo/CMakeLists.txt && \
        sed -i '1i #include <string>' horovod/common/group_table.h && \
        sed -i '1i #include <string>' horovod/common/group_table.cc && \
        sed -i '1i #include <stdexcept>' horovod/common/common.h && \
        HOROVOD_GPU_OPERATIONS=NCCL HOROVOD_NCCL_LINK=SHARED HOROVOD_WITHOUT_MPI=1 HOROVOD_WITH_PYTORCH=1 pip wheel --no-cache-dir --no-build-isolation . && \
        rm -rf /var/lib/apt/lists/*

# Build runtime image
FROM pytorch/pytorch:${PTH_VERSION}-runtime

# Set env vars to allow pip install
ENV PIP_BREAK_SYSTEM_PACKAGES=1
ENV PIP_ROOT_USER_ACTION=ignore


# Apex
COPY --from=apex-hvd-builder /tmp/apex/apex-*.whl /tmp/apex/
RUN pip install --no-cache-dir /tmp/apex/apex-*.whl && \
    rm -fr /tmp/apex

# Install tzdata / git
RUN apt-get update && \
    ln -fs /usr/share/zoneinfo/Europe/Paris /etc/localtime && \
    apt-get -y install --no-install-recommends tzdata git && \
    dpkg-reconfigure --frontend noninteractive tzdata && \
    apt-get autoremove -y && \
    apt-get clean -y && \
    rm -rf /var/lib/apt/lists/*

# Ignite main dependencies
RUN pip install --upgrade --no-cache-dir pytorch-ignite \
                                         tensorboard \
                                         tqdm \
                                         fire

# Replace pillow with pillow-simd
RUN apt-get update && apt-get -y install --no-install-recommends g++ && \
    pip uninstall -y pillow && \
    CC="cc -mavx2" pip install --upgrade --no-cache-dir --force-reinstall pillow-simd && \
    apt-get remove -y g++ && \
    apt-get autoremove -y && \
    rm -rf /var/lib/apt/lists/*

# Checkout Ignite examples only
RUN mkdir -p pytorch-ignite-examples && \
    cd pytorch-ignite-examples && \
    git init && \
    git config core.sparsecheckout true && \
    echo examples >> .git/info/sparse-checkout && \
    git remote add -f origin https://github.com/pytorch/ignite.git && \
    git pull origin master && \
    # rm very large .git folder
    rm -rf .git

COPY --from=apex-hvd-builder /horovod/horovod-*.whl /horovod/

RUN cd /horovod && \
    pip install --no-cache-dir horovod-*.whl && \
    rm -fr /horovod