FROM nvidia/cuda:11.0-base-ubuntu18.04 LABEL maintainer="Amazon AI" LABEL dlc_major_version="3" # prevent stopping by user interaction ENV DEBIAN_FRONTEND noninteractive ENV DEBCONF_NONINTERACTIVE_SEEN true ENV SAGEMAKER_TRAINING_MODULE sagemaker_tensorflow_container.training:main ENV PYTHONDONTWRITEBYTECODE=1 ENV PYTHONUNBUFFERED=1 ENV PYTHONIOENCODING=UTF-8 ENV LANG=C.UTF-8 ENV LC_ALL=C.UTF-8 # Set environment variables for MKL # For more about MKL with TensorFlow see: # https://www.tensorflow.org/performance/performance_guide#tensorflow_with_intel%C2%AE_mkl_dnn ENV KMP_AFFINITY=granularity=fine,compact,1,0 ENV KMP_BLOCKTIME=1 ENV KMP_SETTINGS=0 ENV MANUAL_BUILD=0 ARG PYTHON=python3.7 ARG PYTHON_PIP=python3-pip ARG PIP=pip3 ARG PYTHON_VERSION=3.7.10 ARG OPEN_MPI_PATH=/usr/local ARG TF_URL=https://aws-tensorflow-binaries.s3-us-west-2.amazonaws.com/tensorflow/r2.3_aws/20210813_093824/gpu/cu110/py37/tensorflow_gpu-2.3.4-cp37-cp37m-manylinux2010_x86_64.whl ARG ESTIMATOR_URL=https://aws-tensorflow-binaries.s3-us-west-2.amazonaws.com/estimator/r2.3_aws/20210813_093824/tensorflow_estimator-2.3.0-py2.py3-none-any.whl # 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=1.0.9 RUN apt-get update \ && apt-get -y upgrade --only-upgrade systemd \ && apt-get install -y --no-install-recommends --allow-unauthenticated \ ca-certificates \ cuda-command-line-tools-11-0 \ cuda-cudart-dev-11-0 \ libcufft-dev-11-0 \ libcurand-dev-11-0 \ libcusolver-dev-11-0 \ libcusparse-dev-11-0 \ curl \ emacs \ libcudnn8=8.0.4.30-1+cuda11.0 \ # TensorFlow doesn't require libnccl anymore but Open MPI still depends on it libnccl2=2.7.8-1+cuda11.0 \ libgomp1 \ libnccl-dev=2.7.8-1+cuda11.0 \ libfreetype6-dev \ libhdf5-serial-dev \ liblzma-dev \ libpng-dev \ libtemplate-perl \ libzmq3-dev \ git \ wget \ unzip \ libtool \ vim \ build-essential \ libssl1.1 \ openssl \ openssh-client \ openssh-server \ zlib1g-dev \ # Install dependent library for OpenCV libgtk2.0-dev \ && apt-get update \ && apt-get install -y --no-install-recommends --allow-unauthenticated \ libcublas-11-0=11.2.0.252-1 \ libcublas-dev-11-0=11.2.0.252-1 \ # The 'apt-get install' of nvinfer-runtime-trt-repo-ubuntu1804-5.0.2-ga-cuda10.0 # adds a new list which contains libnvinfer library, so it needs another # 'apt-get update' to retrieve that list before it can actually install the # library. # We don't install libnvinfer-dev since we don't need to build against TensorRT, # and libnvinfer4 doesn't contain libnvinfer.a static library. # nvinfer-runtime-trt-repo doesn't have a 1804-cuda10.1 version yet. see: # https://developer.download.nvidia.cn/compute/machine-learning/repos/ubuntu1804/x86_64/ && apt-get update && apt-get install -y --no-install-recommends --allow-unauthenticated \ nvinfer-runtime-trt-repo-ubuntu1804-5.0.2-ga-cuda10.0 \ && apt-get update && apt-get install -y --no-install-recommends --allow-unauthenticated \ libnvinfer7=7.1.3-1+cuda11.0 \ && rm -rf /var/lib/apt/lists/* \ && mkdir -p /var/run/sshd RUN wget --quiet https://dl.bintray.com/boostorg/release/1.73.0/source/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=/usr -j 64 cxxflags=-fPIC cflags=-fPIC install || true \ && cd .. \ && rm -rf boost_1_73_0.tar.gz \ && rm -rf boost_1_73_0 ########################################################################### # Horovod & its dependencies ########################################################################### # Install Open MPI RUN mkdir /tmp/openmpi \ && cd /tmp/openmpi \ && curl -fSsL -O https://download.open-mpi.org/release/open-mpi/v4.0/openmpi-4.0.4.tar.gz \ && tar zxf openmpi-4.0.4.tar.gz \ && cd openmpi-4.0.4 \ && ./configure --enable-orterun-prefix-by-default \ && make -j $(nproc) all \ && make install \ && ldconfig \ && rm -rf /tmp/openmpi # Create a wrapper for OpenMPI to allow running as root by default RUN mv $OPEN_MPI_PATH/bin/mpirun $OPEN_MPI_PATH/bin/mpirun.real \ && echo '#!/bin/bash' > $OPEN_MPI_PATH/bin/mpirun \ && echo 'mpirun.real --allow-run-as-root "$@"' >> $OPEN_MPI_PATH/bin/mpirun \ && chmod a+x $OPEN_MPI_PATH/bin/mpirun # Configure OpenMPI to run good defaults: # --bind-to none --map-by slot --mca btl_tcp_if_exclude lo,docker0 RUN echo "hwloc_base_binding_policy = none" >> $OPEN_MPI_PATH/etc/openmpi-mca-params.conf \ && echo "rmaps_base_mapping_policy = slot" >> $OPEN_MPI_PATH/etc/openmpi-mca-params.conf # Set default NCCL parameters RUN echo NCCL_DEBUG=INFO >> /etc/nccl.conf ENV LD_LIBRARY_PATH=$OPEN_MPI_PATH/openmpi/lib:$LD_LIBRARY_PATH # /usr/local/lib/libpython* needs to be accessible for dynamic linking ENV LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH ENV PATH=$OPEN_MPI_PATH/openmpi/bin/:$PATH ENV PATH=$OPEN_MPI_PATH/nvidia/bin:$PATH # SSH login fix. Otherwise user is kicked off after login 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 # Create SSH key. RUN 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 WORKDIR / RUN apt-get update \ && apt-get install -y --no-install-recommends \ libbz2-dev \ libc6-dev \ libffi-dev \ libgdbm-dev \ libncursesw5-dev \ libreadline-gplv2-dev \ libsqlite3-dev \ libssl-dev \ tk-dev \ && rm -rf /var/lib/apt/lists/* \ && apt-get clean RUN wget https://www.python.org/ftp/python/$PYTHON_VERSION/Python-$PYTHON_VERSION.tgz \ && tar -xvf Python-$PYTHON_VERSION.tgz \ && cd Python-$PYTHON_VERSION \ && ./configure --enable-shared && make && make install \ && rm -rf ../Python-$PYTHON_VERSION* RUN ${PIP} --no-cache-dir install --upgrade \ pip \ setuptools # Some TF tools expect a "python" binary RUN ln -s $(which ${PYTHON}) /usr/local/bin/python \ && ln -s $(which ${PIP}) /usr/bin/pip # # python-dateutil==2.8.0 to satisfy botocore associated with latest awscli RUN ${PIP} install --no-cache-dir -U \ pybind11 \ cmake==3.18.2.post1 \ numpy==1.19.1 \ scipy==1.5.2 \ scikit-learn==0.23 \ pandas==1.1 \ "Pillow>=8.3,<8.4" \ python-dateutil==2.8.1 \ # install PyYAML>=5.4.1 to avoid conflict with latest awscli "pyYAML>=5.4.1,<5.5" \ requests==2.24.0 \ "awscli<2" \ mpi4py==3.0.3 \ opencv-python==4.3.0.36 \ "sagemaker>=2,<3" \ sagemaker-experiments==0.* \ "sagemaker-tensorflow>=2.3,<2.4" \ "sagemaker-tensorflow-training>=20" \ # Let's install TensorFlow separately in the end to avoid # the library version to be overwritten && ${PIP} install --no-cache-dir -U \ ${TF_URL} \ ${ESTIMATOR_URL} \ h5py==2.10.0 \ "absl-py>=0.9,<0.11" \ werkzeug==1.0.1 \ smdebug==${SMDEBUG_VERSION} \ smclarify ARG METIS=metis-5.1.0 ARG RMM_VERSION=0.15.0 # install metis RUN wget -nv http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/${METIS}.tar.gz \ && gunzip -f ${METIS}.tar.gz \ && tar -xvf ${METIS}.tar \ && cd ${METIS} \ && apt-get update \ && apt-get install -y build-essential \ && apt-get install -y cmake \ && make config shared=1 \ && make install \ && rm -rf ${METIS}.tar* \ && rm -rf ${METIS} \ && rm -rf /var/lib/apt/lists/* \ && apt-get clean # Install RAPIDSMemoryManager. # Requires cmake>=3.14. RUN wget -nv https://github.com/rapidsai/rmm/archive/v${RMM_VERSION}.tar.gz \ && tar -xvf v${RMM_VERSION}.tar.gz \ && cd rmm-${RMM_VERSION} \ && INSTALL_PREFIX=/usr/local ./build.sh librmm \ && rm -rf v${RMM_VERSION}.tar* \ && rm -rf rmm-${RMM_VERSION} ENV CPATH="/usr/local/lib/python3.7/dist-packages/pybind11/include/" # Install Horovod, temporarily using CUDA stubs RUN ldconfig /usr/local/cuda-11.0/targets/x86_64-linux/lib/stubs \ && HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_WITH_TENSORFLOW=1 ${PIP} install --no-cache-dir horovod==0.20.3 \ && ldconfig # Allow OpenSSH to talk to containers without asking for confirmation RUN 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 # Add NGC vars ENV TF_AUTOTUNE_THRESHOLD=2 # Install SMD MP binary RUN pip install --no-cache-dir -U https://sagemaker-distributed-model-parallel.s3.amazonaws.com/tensorflow-2.3/build-artifacts/2021-03-26-21-57/smdistributed_modelparallel-1.3.1-cp37-cp37m-linux_x86_64.whl # Install SM Distributed DataParallel binary ARG SMDATAPARALLEL_BINARY=https://smdataparallel.s3.amazonaws.com/binary/tensorflow/2.3.1/cu110/2021-01-14/smdistributed_dataparallel-1.0.0-cp37-cp37m-linux_x86_64.whl RUN SMDATAPARALLEL_TF=1 pip install --no-cache-dir ${SMDATAPARALLEL_BINARY} ENV LD_LIBRARY_PATH="/usr/local/lib/python3.7/site-packages/smdistributed/dataparallel/lib:$LD_LIBRARY_PATH" ADD https://raw.githubusercontent.com/aws/deep-learning-containers/master/src/deep_learning_container.py /usr/local/bin/deep_learning_container.py RUN chmod +x /usr/local/bin/deep_learning_container.py 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* RUN curl https://aws-dlc-licenses.s3.amazonaws.com/tensorflow-2.3/license.txt -o /license.txt CMD ["/bin/bash"]