ARG PYTHON=python3
ARG PYTHON_VERSION=3.10.8
ARG PYTHON_SHORT_VERSION=3.10
ARG MAMBA_VERSION=23.1.0-1
ARG PYTORCH_VERSION_EC2=2.0.1
ARG PYTORCH_VERSION_SM=2.0.0

# SMD binaries
ARG SMD_DATA_PARALLEL_URL=https://smdataparallel.s3.amazonaws.com/binary/pytorch/2.0.0/cu118/2023-03-20/smdistributed_dataparallel-1.8.0-cp310-cp310-linux_x86_64.whl
ARG SMD_MODEL_PARALLEL_URL=https://sagemaker-distributed-model-parallel.s3.us-west-2.amazonaws.com/pytorch-2.0.0/build-artifacts/2023-04-14-20-14/smdistributed_modelparallel-1.15.0-cp310-cp310-linux_x86_64.whl

# ZeRO-2D Binary
ARG ZERO_2D_URL=https://aws-deepspeed-zero-2d-binaries.s3.us-west-2.amazonaws.com/r2.0.0/20230407-184728/1ea3d4b6aa41fe66277daacbb78b3743a310d85a/deepspeed-0.6.1%2B1ea3d4b-py3-none-any.whl

FROM nvidia/cuda:11.8.0-base-ubuntu20.04 AS base_image

ENV DEBIAN_FRONTEND=noninteractive \
    LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/local/lib"

RUN apt-get update \
 && apt-get upgrade -y \
 && apt-get autoremove -y \
 && apt-get clean \
 && rm -rf /var/lib/apt/lists/*

FROM base_image AS common

LABEL maintainer="Amazon AI"
LABEL dlc_major_version="1"

ARG PYTHON
ARG PYTHON_VERSION
ARG PYTHON_SHORT_VERSION
ARG MAMBA_VERSION

ARG CUBLAS_VERSION=11.11.3.6
ARG EFA_PATH=/opt/amazon/efa

# This arg required to stop docker build waiting for region configuration while installing tz data from ubuntu 20
ARG DEBIAN_FRONTEND=noninteractive

# Python won’t try to write .pyc or .pyo files on the import of source modules
# Force stdin, stdout and stderr to be totally unbuffered. Good for logging
ENV CUDA_HOME=/opt/conda/
ENV PYTHONDONTWRITEBYTECODE=1
ENV PYTHONUNBUFFERED=1
ENV LD_LIBRARY_PATH="/usr/local/lib:${LD_LIBRARY_PATH}"
ENV LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"
ENV PYTHONIOENCODING=UTF-8
ENV LANG=C.UTF-8
ENV LC_ALL=C.UTF-8
ENV PATH /opt/conda/bin:$PATH
ENV TORCH_CUDA_ARCH_LIST="3.7 5.0 7.0+PTX 8.0"
ENV TORCH_NVCC_FLAGS="-Xfatbin -compress-all"

ENV CUDNN_VERSION=8.7.0.84
ENV NCCL_VERSION=2.16.2
ENV HOROVOD_VERSION=0.26.1
ENV EFA_VERSION=1.21.0
ENV OMPI_VERSION=4.1.5
ENV BRANCH_OFI=1.5.0-aws
ENV GDRCOPY_VERSION=2.3.1

ENV CMAKE_PREFIX_PATH="$(dirname $(which conda))/../"
ENV OPEN_MPI_PATH=/opt/amazon/openmpi

ENV DGLBACKEND=pytorch
ENV MANUAL_BUILD=0
ENV RDMAV_FORK_SAFE=1
ENV DLC_CONTAINER_TYPE=training

RUN apt-get update \
 # TODO: Remove systemd upgrade once it is updated in base image
 && apt-get -y upgrade --only-upgrade systemd \
 && apt-get install -y --allow-change-held-packages --no-install-recommends \
    build-essential \
    ca-certificates \
    cmake \
    # cuda-toolkit-11-8 \
    # cuda-command-line-tools-11-8 \
    # cuda-nvcc-11-8 \
    # libcudnn8=$CUDNN_VERSION-1+cuda11.8 \
    curl \
    emacs \
    git \
    hwloc \
    jq \
    libcurl4-openssl-dev \
    libglib2.0-0 \
    libgl1-mesa-glx \
    libsm6 \
    libxext6 \
    libxrender-dev \
    libgomp1 \
    libibverbs-dev \
    libhwloc-dev \
    libnuma1 \
    libnuma-dev \
    libssl1.1 \
    libssl-dev \
    libtool \
    openssl \
    python3-dev \
    unzip \
    vim \
    wget \
    zlib1g-dev \
    autoconf \
    pkg-config \
    check \
    libsubunit0 \
    libsubunit-dev \
 && rm -rf /var/lib/apt/lists/* \
 && apt-get clean

# for conda ssl verification
ENV REQUESTS_CA_BUNDLE=/etc/ssl/certs/ca-certificates.crt
RUN curl -L -o ~/mambaforge.sh https://github.com/conda-forge/miniforge/releases/download/${MAMBA_VERSION}/Mambaforge-${MAMBA_VERSION}-Linux-x86_64.sh \
 && chmod +x ~/mambaforge.sh \
 && ~/mambaforge.sh -b -p /opt/conda \
 && rm ~/mambaforge.sh

# Install common conda packages
RUN /opt/conda/bin/mamba install -y --override-channels \
  -c https://aws-ml-conda-ec2.s3.us-west-2.amazonaws.com \
  -c nvidia/label/cuda-11.8.0 \
  pytorch-cuda=11.8 cuda-toolkit=11.8 \
  && /opt/conda/bin/mamba install -y -c conda-forge \
    pyopenssl \
    requests \
    libgcc \
    pybind11 \
    cmake \
    python=$PYTHON_VERSION \
    cython \
    mkl \
    mkl-include \
    parso \
    typing \
    conda-content-trust \
    charset-normalizer \
    packaging \
    opencv \
    awscli \
    boto3 \
    pyyaml \
    scipy \
    click \
    psutil \
    ipython \
    # Adding package for studio kernels
    ipykernel \
    # patch CVE
    "cryptography>39.0.0" \
    # patch CVE
    "pillow>=9.4" \
    # force conda to install h5py with openmpi build, not mpich (rubik error)
    # https://anaconda.org/conda-forge/h5py/files?page=2
    "mpi4py>=3.1.4,<3.2" \
    h5py \
  && /opt/conda/bin/mamba install -y -c dglteam/label/cu118 dgl \
  && /opt/conda/bin/mamba clean -afy \
  && rm -rf /etc/apt/sources.list.d/*

# Install NCCL
RUN cd /tmp \
 && git clone https://github.com/NVIDIA/nccl.git -b v${NCCL_VERSION}-1 \
 && cd nccl \
 && make -j64 src.build BUILDDIR=/usr/local CUDA_LIB=${CUDA_HOME}/lib \
 && rm -rf /tmp/nccl

# Install EFA alone without AWS OPEN_MPI
RUN mkdir /tmp/efa \
 && cd /tmp/efa \
 && curl -O https://s3-us-west-2.amazonaws.com/aws-efa-installer/aws-efa-installer-${EFA_VERSION}.tar.gz \
 && tar -xf aws-efa-installer-${EFA_VERSION}.tar.gz \
 && cd aws-efa-installer \
 && apt-get update \
 && ./efa_installer.sh -y --skip-kmod -g \
 && rm -rf $OPEN_MPI_PATH \
 && rm -rf /tmp/efa \
 && rm -rf /tmp/aws-efa-installer-${EFA_VERSION}.tar.gz \
 && rm -rf /var/lib/apt/lists/* \
 && apt-get clean

# Install OpenMPI without libfabric support
RUN mkdir /tmp/openmpi \
 && cd /tmp/openmpi \
 && wget --quiet https://download.open-mpi.org/release/open-mpi/v4.1/openmpi-${OMPI_VERSION}.tar.gz \
 && tar zxf openmpi-${OMPI_VERSION}.tar.gz \
 && cd openmpi-${OMPI_VERSION} \
 && ./configure --enable-orterun-prefix-by-default --prefix=$OPEN_MPI_PATH --with-cuda=${CUDA_HOME} \
 && make -j $(nproc) all \
 && make install \
 && ldconfig \
 && cd / \
 && rm -rf /tmp/openmpi \
 && ompi_info --parsable --all | grep mpi_built_with_cuda_support:value

ENV PATH="$OPEN_MPI_PATH/bin:$EFA_PATH/bin:$PATH"
ENV LD_LIBRARY_PATH=$OPEN_MPI_PATH/lib/:$EFA_PATH/lib/:$LD_LIBRARY_PATH

# Install GDRCopy which is a dependency of SM Distributed DataParallel binary
# The test binaries requires cuda driver library which could be found in conda
# So update the linker path to point to it to avoid -Lcuda not found
# Note that starting with 11.8, cuda relate lib are moved to conda instead /usr/local/cuda
RUN cd /tmp \
 && git clone https://github.com/NVIDIA/gdrcopy.git -b v${GDRCOPY_VERSION} \
 && cd gdrcopy \
 && sed -ie '12s@$@ -L $(CUDA)/lib/stubs@' tests/Makefile \
 && CUDA=/opt/conda make install \
 && rm -rf /tmp/gdrcopy

RUN pip install --no-cache-dir --upgrade pip --no-cache-dir --trusted-host pypi.org --trusted-host files.pythonhosted.org \
 && ln -s /opt/conda/bin/pip /usr/local/bin/pip3

WORKDIR /root

# Configure Open MPI and configure NCCL parameters
RUN mv $OPEN_MPI_PATH/bin/mpirun $OPEN_MPI_PATH/bin/mpirun.real \
 && echo '#!/bin/bash' > $OPEN_MPI_PATH/bin/mpirun \
 && echo "${OPEN_MPI_PATH}/bin/mpirun.real --allow-run-as-root \"\$@\"" >> $OPEN_MPI_PATH/bin/mpirun \
 && chmod a+x $OPEN_MPI_PATH/bin/mpirun \
 && echo "hwloc_base_binding_policy = none" >> $OPEN_MPI_PATH/etc/openmpi-mca-params.conf \
 && echo "rmaps_base_mapping_policy = slot" >> $OPEN_MPI_PATH/etc/openmpi-mca-params.conf \
 && echo NCCL_DEBUG=INFO >> /etc/nccl.conf \
 && echo NCCL_SOCKET_IFNAME=^docker0 >> /etc/nccl.conf

# Install OpenSSH for MPI to communicate between containers, allow OpenSSH to talk to containers without asking for confirmation
RUN apt-get update \
 && apt-get install -y  --allow-downgrades --allow-change-held-packages --no-install-recommends \
 && apt-get install -y --no-install-recommends openssh-client openssh-server \
 && mkdir -p /var/run/sshd \
 && cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new \
 && echo "    StrictHostKeyChecking no" >> /etc/ssh/ssh_config.new \
 && mv /etc/ssh/ssh_config.new /etc/ssh/ssh_config \
 && rm -rf /var/lib/apt/lists/* \
 && apt-get clean

# Configure OpenSSH so that nodes can communicate with each other
RUN mkdir -p /var/run/sshd && \
 sed 's@session\s*required\s*pam_loginuid.so@session optional pam_loginuid.so@g' -i /etc/pam.d/sshd

RUN rm -rf /root/.ssh/ && \
 mkdir -p /root/.ssh/ && \
 ssh-keygen -q -t rsa -N '' -f /root/.ssh/id_rsa && \
 cp /root/.ssh/id_rsa.pub /root/.ssh/authorized_keys \
 && printf "Host *\n StrictHostKeyChecking no\n" >> /root/.ssh/config

RUN mkdir -p /etc/pki/tls/certs && cp /etc/ssl/certs/ca-certificates.crt /etc/pki/tls/certs/ca-bundle.crt

RUN curl -o /license.txt https://aws-dlc-licenses.s3.amazonaws.com/pytorch-2.0/license.txt

# Removing the cache as it is needed for security verification
RUN rm -rf /root/.cache | true

########################################################
#  _____ ____ ____    ___
# | ____/ ___|___ \  |_ _|_ __ ___   __ _  __ _  ___
# |  _|| |     __) |  | || '_ ` _ \ / _` |/ _` |/ _ \
# | |__| |___ / __/   | || | | | | | (_| | (_| |  __/
# |_____\____|_____| |___|_| |_| |_|\__,_|\__, |\___|
#                                         |___/
#  ____           _
# |  _ \ ___  ___(_)_ __   ___
# | |_) / _ \/ __| | '_ \ / _ \
# |  _ <  __/ (__| | |_) |  __/
# |_| \_\___|\___|_| .__/ \___|
#                  |_|
########################################################

FROM common AS ec2

ARG PYTHON
ARG PYTHON_VERSION
ARG PYTHON_SHORT_VERSION

# PyTorch Binaries
ARG PT_EC2_TRAINING_URL
ARG PT_TORCHVISION_URL
ARG PT_TORCHAUDIO_URL
ARG PT_TORCHDATA_URL

WORKDIR /

# Install ec2 PyTorch
RUN /opt/conda/bin/mamba install -y pytorch=${PYTORCH_VERSION_EC2} \
  torchvision torchaudio torchtext aws-ofi-nccl-dlc \
  --override-channels \
  -c https://aws-ml-conda-ec2.s3.us-west-2.amazonaws.com \
  -c conda-forge

# install PyTorch related python packages (depends on PyTorch)
RUN /opt/conda/bin/mamba install -y -c conda-forge \
  # needed by fastai.distributed
  accelerate \
  && /opt/conda/bin/mamba install -y -c fastai fastai \
  # fastai conda strictly depends on dataclasses
  && pip uninstall -y dataclasses \
  && pip install torchtnt \
# explicitly install triton build-time dependencies, which won't get installed by conda for some reason
  cmake \
  lit

# Patches
RUN pip install "pillow>=9.5"
RUN /opt/conda/bin/mamba install -y -c conda-forge \
  "requests>=2.31.0"

# Install Nvidia Apex (needs pytorch)
## Pin apex commit requested by sm-model-parallel team
RUN git clone https://github.com/NVIDIA/apex \
  && cd apex \
  # recent commit, torch._six fix is included https://github.com/NVIDIA/apex/commit/a78ccf0b3e3f7130b3f157732dc8e8e651389922
  && git checkout 27de66c \
  && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./ \
  && cd .. \
  && rm -rf apex

# Install Horovod (needs pytorch)
RUN pip uninstall -y horovod \
#  && ldconfig /usr/local/cuda-11.8/targets/x86_64-linux/lib/stubs \
 && HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_CUDA_HOME=${CUDA_HOME} HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir horovod==${HOROVOD_VERSION} \
 && ldconfig

RUN HOME_DIR=/root \
 && curl -o ${HOME_DIR}/oss_compliance.zip https://aws-dlinfra-utilities.s3.amazonaws.com/oss_compliance.zip \
 && unzip ${HOME_DIR}/oss_compliance.zip -d ${HOME_DIR}/ \
 && cp ${HOME_DIR}/oss_compliance/test/testOSSCompliance /usr/local/bin/testOSSCompliance \
 && chmod +x /usr/local/bin/testOSSCompliance \
 && chmod +x ${HOME_DIR}/oss_compliance/generate_oss_compliance.sh \
 && ${HOME_DIR}/oss_compliance/generate_oss_compliance.sh ${HOME_DIR} ${PYTHON} \
 && rm -rf ${HOME_DIR}/oss_compliance* \
 && rm -rf /tmp/tmp*

# Removing the cache as it is needed for security verification
RUN rm -rf /root/.cache | true

# Starts framework
CMD ["/bin/bash"]

#################################################################
#  ____                   __  __       _
# / ___|  __ _  __ _  ___|  \/  | __ _| | _____ _ __
# \___ \ / _` |/ _` |/ _ \ |\/| |/ _` | |/ / _ \ '__|
#  ___) | (_| | (_| |  __/ |  | | (_| |   <  __/ |
# |____/ \__,_|\__, |\___|_|  |_|\__,_|_|\_\___|_|
#              |___/
#  ___                              ____           _
# |_ _|_ __ ___   __ _  __ _  ___  |  _ \ ___  ___(_)_ __   ___
#  | || '_ ` _ \ / _` |/ _` |/ _ \ | |_) / _ \/ __| | '_ \ / _ \
#  | || | | | | | (_| | (_| |  __/ |  _ <  __/ (__| | |_) |  __/
# |___|_| |_| |_|\__,_|\__, |\___| |_| \_\___|\___|_| .__/ \___|
#                      |___/                        |_|
#################################################################

FROM common AS sagemaker

LABEL maintainer="Amazon AI"
LABEL dlc_major_version="1"

ARG PYTHON
ARG PYTHON_VERSION
ARG PYTHON_SHORT_VERSION

ARG RMM_VERSION=0.15.0

# The smdebug pipeline relies for following format to perform string replace and trigger DLC pipeline for validating
# the nightly builds. Therefore, while updating the smdebug version, please ensure that the format is not disturbed.
ARG SMDEBUG_VERSION=1.0.34

ENV SAGEMAKER_TRAINING_MODULE=sagemaker_pytorch_container.training:main

# SMD model parallel and data parallel binaries
ARG SMD_DATA_PARALLEL_URL
ARG SMD_MODEL_PARALLEL_URL

# ZeRO-2D Binary
ARG ZERO_2D_URL

# PyTorch Binaries
ARG PT_SM_TRAINING_URL
ARG PT_TORCHVISION_URL
ARG PT_TORCHAUDIO_URL
ARG PT_TORCHDATA_URL

# Install sm PyTorch (force remove will increase docker image size drastically)
RUN /opt/conda/bin/mamba install -y pytorch=${PYTORCH_VERSION_SM} \
  torchvision torchaudio torchtext aws-ofi-nccl-dlc \
  --override-channels \
  -c https://aws-ml-conda-pre-prod-ec2.s3.us-west-2.amazonaws.com \
  -c conda-forge

# install PyTorch related python packages (depends on PyTorch)
RUN /opt/conda/bin/mamba install -y -c conda-forge \
  # needed by fastai.distributed
  accelerate \
  && /opt/conda/bin/mamba install -y -c fastai fastai \
  # fastai conda strictly depends on dataclasses
  && pip uninstall -y dataclasses \
  && pip install torchnet \
# explicitly install triton build-time dependencies, which won't get installed by conda for some reason
  cmake \
  lit

# Patches
RUN pip install "pillow>=9.5"
RUN /opt/conda/bin/mamba install -y -c conda-forge \
  "requests>=2.31.0"

WORKDIR /

# Install Nvidia Apex (needs pytorch)
## Pin apex commit requested by sm-model-parallel team
RUN git clone https://github.com/NVIDIA/apex \
  && cd apex \
  # requested by Rubik. Will wait for resolution on the illegal mem access error
  && git checkout aa756ce \
  && sed -i 's/from torch._six import string_classes//g' apex/amp/_initialize.py \
  && sed -i 's/string_classes/str/g' apex/amp/_initialize.py \
  && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./ \
  && cd .. \
  && rm -rf apex

# Install Horovod (needs pytorch)
RUN pip uninstall -y horovod \
#  && ldconfig /usr/local/cuda-11.8/targets/x86_64-linux/lib/stubs \
 && HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_CUDA_HOME=${CUDA_HOME} HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir horovod==${HOROVOD_VERSION} \
 && ldconfig

# Install libboost from source. This package is needed for smdataparallel functionality [for networking asynchronous IO].
RUN wget https://sourceforge.net/projects/boost/files/boost/1.73.0/boost_1_73_0.tar.gz/download -O boost_1_73_0.tar.gz \
 && tar -xzf boost_1_73_0.tar.gz \
 && cd boost_1_73_0 \
 && ./bootstrap.sh \
 && ./b2 threading=multi --prefix=/opt/conda -j 64 cxxflags=-fPIC cflags=-fPIC install || true \
 && cd .. \
 && rm -rf boost_1_73_0.tar.gz \
 && rm -rf boost_1_73_0 \
 && cd /opt/conda/include/boost

WORKDIR /opt/pytorch

# Copy workaround script for incorrect hostname
COPY changehostname.c /
COPY start_with_right_hostname.sh /usr/local/bin/start_with_right_hostname.sh

RUN chmod +x /usr/local/bin/start_with_right_hostname.sh

WORKDIR /root

RUN pip install --no-cache-dir -U \
    smclarify \
    "sagemaker>=2,<3" \
    "sagemaker-experiments<1" \
    sagemaker-pytorch-training \
    flash-attn==0.2.8 \
    triton==2.0.0.dev20221202

# Install smdebug from source
RUN cd /tmp \
  && git clone https://github.com/awslabs/sagemaker-debugger --depth 1 --single-branch --branch ${SMDEBUG_VERSION} \
  && cd sagemaker-debugger \
  && pip install . \
  && rm -rf /tmp/*

# Install extra packages
RUN /opt/conda/bin/mamba install -y -c conda-forge \
  bokeh \
  imageio \
  numba \
  pandas \
  plotly \
  shap \
  scikit-learn \
  seaborn \
 && /opt/conda/bin/mamba clean -afy

# install metis
RUN rm -rf /etc/apt/sources.list.d/* \
 && git clone https://github.com/KarypisLab/GKlib \
 && cd GKlib \
 && make config \
 && make \
 && make install \
 && cd .. \
 && git clone https://github.com/KarypisLab/METIS.git \
 && cd METIS \
 && make config shared=1 cc=gcc prefix=/root/local \
 && make install \
 && cd .. \
 && rm -rf METIS GKlib \
 && rm -rf /var/lib/apt/lists/* \
 && apt-get clean

# Install RAPIDSMemoryManager.
# Requires cmake>=3.14.
RUN  wget -nv https://github.com/rapidsai/rmm/archive/v${RMM_VERSION}.tar.gz \
 && tar -xvf v${RMM_VERSION}.tar.gz \
 && cd rmm-${RMM_VERSION} \
 && INSTALL_PREFIX=/usr/local ./build.sh librmm \
 && cd .. \
 && rm -rf v${RMM_VERSION}.tar* \
 && rm -rf rmm-${RMM_VERSION}

# Install ZeRO-2D Binary
RUN pip install --no-cache-dir -U ${ZERO_2D_URL}

# Install SM Distributed Modelparallel binary
RUN pip install --no-cache-dir -U ${SMD_MODEL_PARALLEL_URL}

# Install SM Distributed DataParallel binary
RUN SMDATAPARALLEL_PT=1 pip install --no-cache-dir ${SMD_DATA_PARALLEL_URL}

ENV LD_LIBRARY_PATH="/opt/conda/lib/python${PYTHON_SHORT_VERSION}/site-packages/smdistributed/dataparallel/lib:$LD_LIBRARY_PATH"

WORKDIR /

RUN HOME_DIR=/root \
 && curl -o ${HOME_DIR}/oss_compliance.zip https://aws-dlinfra-utilities.s3.amazonaws.com/oss_compliance.zip \
 && unzip ${HOME_DIR}/oss_compliance.zip -d ${HOME_DIR}/ \
 && cp ${HOME_DIR}/oss_compliance/test/testOSSCompliance /usr/local/bin/testOSSCompliance \
 && chmod +x /usr/local/bin/testOSSCompliance \
 && chmod +x ${HOME_DIR}/oss_compliance/generate_oss_compliance.sh \
 && ${HOME_DIR}/oss_compliance/generate_oss_compliance.sh ${HOME_DIR} ${PYTHON} \
 && rm -rf ${HOME_DIR}/oss_compliance* \
 && rm -rf /tmp/tmp*

# Removing the cache as it is needed for security verification
RUN rm -rf /root/.cache | true

ENTRYPOINT ["bash", "-m", "start_with_right_hostname.sh"]
CMD ["/bin/bash"]