先日、Flairを使ったモデルを構築し、SageMakerのトレーニングジョブに投げたところモデルの保存で躓いた。原因を調べたところ、pickleでダンプしようとしていたオブジェクトの中に、Python 3.6ではダンプできないオブジェクトがあるようだった。そこで、SageMakerのトレーニングで使われているPythonのバージョンを3.6から3.7に上げたところモデルの保存をできるようになった。
以下に、作成したDockerfileを貼っておく。これをCodeBuildでビルドし、ECRに登録後、SageMakerのEstimatorで登録したイメージを指定すれば使うことができる。
FROM nvidia/cuda@sha256:4979db047661dc0003594fb20d37cce6d6c7e989252f4e3fb0beb39874a078e2
LABEL maintainer="Amazon AI"
ARG PYTHON_VERSION=3.7.7
ARG OPEN_MPI_VERSION=4.0.1
ARG CUBLAS_VERSION=10.2.1.243-1_amd64
# 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 PYTHONDONTWRITEBYTECODE=1
ENV PYTHONUNBUFFERED=1
ENV LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/local/lib"
ENV LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/opt/conda/lib"
ENV PYTHONIOENCODING=UTF-8
ENV LANG=C.UTF-8
ENV LC_ALL=C.UTF-8
ENV PATH /opt/conda/bin:$PATH
ENV TORCH_CUDA_ARCH_LIST="3.5 5.2 6.0 6.1 7.0+PTX"
ENV TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
ENV HOROVOD_VERSION=0.19.1
ENV DGLBACKEND=pytorch
ENV CMAKE_PREFIX_PATH="$(dirname $(which conda))/../"
ENV SAGEMAKER_TRAINING_MODULE=sagemaker_pytorch_container.training:main
RUN apt-get update \
&& apt-get install -y --allow-downgrades --allow-change-held-packages --no-install-recommends \
build-essential \
ca-certificates \
cmake \
cuda-command-line-tools-10-1 \
cuda-cufft-10-1 \
cuda-curand-10-1 \
cuda-cusolver-10-1 \
cuda-cusparse-10-1 \
curl \
git \
jq \
libglib2.0-0 \
libgl1-mesa-glx \
libsm6 \
libxext6 \
libxrender-dev \
libgomp1 \
libibverbs-dev \
libhwloc-dev \
libnuma1 \
libnuma-dev \
vim \
wget \
zlib1g-dev \
&& apt-get remove -y cuda-cufft-dev-10-1 \
cuda-curand-dev-10-1 \
cuda-cusolver-dev-10-1 \
cuda-npp-dev-10-1 \
cuda-nvgraph-dev-10-1 \
cuda-nvjpeg-dev-10-1 \
cuda-nvrtc-dev-10-1 \
&& apt-get clean \
&& rm -rf /var/lib/apt/lists/*
RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/libcublas10_${CUBLAS_VERSION}.deb \
&& dpkg -i libcublas10_${CUBLAS_VERSION}.deb \
&& apt-get install -f -y \
&& rm libcublas10_${CUBLAS_VERSION}.deb
RUN wget https://www.open-mpi.org/software/ompi/v4.0/downloads/openmpi-$OPEN_MPI_VERSION.tar.gz \
&& gunzip -c openmpi-$OPEN_MPI_VERSION.tar.gz | tar xf - \
&& cd openmpi-$OPEN_MPI_VERSION \
&& ./configure --prefix=/home/.openmpi \
&& make all install \
&& cd .. \
&& rm openmpi-$OPEN_MPI_VERSION.tar.gz \
&& rm -rf openmpi-$OPEN_MPI_VERSION
ENV PATH="$PATH:/home/.openmpi/bin"
ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/home/.openmpi/lib/"
RUN ompi_info --parsable --all | grep mpi_built_with_cuda_support:value \
&& curl -L -o ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh \
&& chmod +x ~/miniconda.sh \
&& ~/miniconda.sh -b -p /opt/conda \
&& rm ~/miniconda.sh \
&& /opt/conda/bin/conda install -y -c anaconda \
python=$PYTHON_VERSION \
numpy==1.16.4 \
ipython==7.10.1 \
mkl==2019.4 \
mkl-include==2019.4 \
cython==0.29.12 \
future==0.17.1 \
"pyopenssl>=17.5.0" \
&& conda install -c dglteam -y dgl-cuda10.1==0.4.3 \
&& /opt/conda/bin/conda clean -ya
RUN conda install -c pytorch magma-cuda101==2.5.1 \
&& conda install -c conda-forge \
opencv==4.0.1 \
&& conda install -y scikit-learn==0.21.2 \
pandas==0.25.0 \
h5py==2.9.0 \
requests==2.22.0 \
libgcc \
&& conda clean -ya
RUN pip install psutil==5.6.7 \
Pillow==7.1.0
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
WORKDIR /root
RUN /opt/conda/bin/conda config --set ssl_verify False \
&& pip install --upgrade pip --trusted-host pypi.org --trusted-host files.pythonhosted.org \
&& ln -s /opt/conda/bin/pip /usr/local/bin/pip3
# Uninstall torch and torchvision before installing the custom versions from an S3 bucket
RUN pip install \
--no-cache-dir smdebug==0.7.2 \
sagemaker==1.50.17 \
sagemaker-experiments==0.1.7 \
--no-cache-dir fastai==1.0.59 \
awscli \
scipy==1.2.2 \
&& pip install --no-cache-dir -U torch==1.5.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html\
&& pip uninstall -y torchvision \
&& pip install --no-deps --no-cache-dir -U torchvision==0.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
# Install Horovod
RUN pip uninstall -y horovod \
&& ldconfig /usr/local/cuda-10.1/targets/x86_64-linux/lib/stubs \
&& HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_CUDA_HOME=/usr/local/cuda-10.1 HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir horovod==${HOROVOD_VERSION} \
&& ldconfig
# Install Nvidia Apex
RUN git clone https://github.com/NVIDIA/apex.git \
&& cd apex \
&& git checkout f3a960f \
&& pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
# Configure Open MPI and configure NCCL parameters
RUN mv /home/.openmpi/bin/mpirun /home/.openmpi/bin/mpirun.real \
&& echo '#!/bin/bash' > /home/.openmpi/bin/mpirun \
&& echo 'mpirun.real --allow-run-as-root "$@"' >> /home/.openmpi/bin/mpirun \
&& chmod a+x /home/.openmpi/bin/mpirun \
&& echo "hwloc_base_binding_policy = none" >> /home/.openmpi/etc/openmpi-mca-params.conf \
&& echo "rmaps_base_mapping_policy = slot" >> /home/.openmpi/etc/openmpi-mca-params.conf \
&& echo "btl_tcp_if_exclude = lo,docker0" >> /home/.openmpi/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 --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 \
&& apt-get clean \
&& rm -rf /var/lib/apt/lists/*
WORKDIR /
RUN pip install --no-cache-dir "sagemaker-pytorch-training<2"
RUN chmod +x /usr/local/bin/start_with_right_hostname.sh
RUN curl -o /license.txt https://aws-dlc-licenses.s3.amazonaws.com/pytorch-1.5.0/license.txt
# Starts framework
ENTRYPOINT ["bash", "-m", "start_with_right_hostname.sh"]
CMD ["/bin/bash"]
考えてみれば、ARGが定義されているので、Pythonのバージョンに関してはビルド時に引数渡して指定すればよかった。その他、typingモジュールをインストールしないようにしたりしているが、基本的には元のDockerfileと大きく変わるところはない。
