Ubuntu18.04, Cuda10.1, Cudnn7.0, Tensorflow2.3 singularity example on donut cluster. When building your own Docker image, you can inherit from `ryantanaka/cuda10.1:latest`

Dockerfile

# Ubuntu 18.04
from nvidia/cuda:10.1-base

# set locale
ENV LANG C.UTF-8

# install linux packages
RUN apt update && apt install -y \
        software-properties-common \
        wget \
        curl \
        sudo \
        vim3

# install python, pip, and tf2.3
RUN add-apt-repository -y ppa:deadsnakes/ppa
RUN apt update && apt install -y python3.7 python3-pip
RUN pip3 install -U pip.
RUN pip3 install tensorflow==2.3.0

# install additional cuda packages
WORKDIR /tmp
# must have cudnn7 package locally 
ADD ./libcudnn7_7.6.5.32-1+cuda10.1_amd64.deb /tmp
RUN dpkg -i libcudnn7_7.6.5.32-1+cuda10.1_amd64.deb

RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/libcublas10_10.1.0.105-1_amd64.deb
RUN dpkg -i libcublas10_10.1.0.105-1_amd64.deb

RUN wget  https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-cufft-10-1_10.1.105-1_amd64.deb
RUN dpkg -i cuda-cufft-10-1_10.1.105-1_amd64.deb

RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-curand-10-1_10.1.243-1_amd64.deb
RUN dpkg -i cuda-curand-10-1_10.1.243-1_amd64.deb

RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-cusolver-10-1_10.1.243-1_amd64.deb
RUN dpkg -i cuda-cusolver-10-1_10.1.243-1_amd64.deb

RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-cusparse-10-1_10.1.243-1_amd64.deb
RUN dpkg -i cuda-cusparse-10-1_10.1.243-1_amd64.deb

# test script
RUN echo '#!/usr/bin/env python3\n\
import sys\n\
try:\n\
    import tensorflow as tf\n\
except ModuleNotFoundError:\n\
    print("test requires tensorflow 2.3")\n\
    sys.exit(1)\n\
\n\
tf.debugging.set_log_device_placement(True)\n\
# Create some tensors\n\
a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])\n\
b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])\n\
c = tf.matmul(a, b)\n\
\n\
print(c)\n'\
>> /usr/bin/matmul.py

# test
RUN echo "Testing /usr/bin/matmul.py"
RUN python3 /usr/bin/matmul.py

test.sh

#!/bin/bash
#SBATCH --partition=donut-default
#SBATCH --nodes=1
#SBATCH --cpus-per-task=4
#SBATCH --gpus=1
#SBATCH --mem=4096

singularity exec --nv docker://ryantanaka/cuda10.1 python3 /usr/bin/matmul.py