This script automates the setup of an Amazon EC2 Graviton ARM-based instances for deep learning tasks. It takes care of installing essential utilities, setting up latest Nvidia drivers and CUDA 12.2 toolkit and cuDNN library, and build PyTorch from source. The step-by-step guided can be checked here. https://jumpshare.com/blog/deep-learning-on-a…

ec2_graviton_dl_bootstrap.sh

#!/bin/bash
set -e # Exit on any error

# Check if required arguments are provided
if [ -z "$REGION" ] || [ -z "$SECURITY_GROUPS" ] || [ -z "$KEY_PAIR" ] || [ -z "$SUBNET" ]; then
    echo "Error: You must provide REGION, SECURITY_GROUPS, KEY_PAIR, and SUBNET as environment variables."
    echo "Example:"
    echo "  export REGION=us-east-1"
    echo "  export SECURITY_GROUPS=sg-12345678,sg-87654321"
    echo "  export KEY_PAIR=my-key-pair"
    echo "  export SUBNET=subnet-12345678"
    exit 1
fi

USER_DATA=$(cat <<'EOF'
#!/bin/bash
set -e # Exit on any error

CUDA_HOME=/usr/local/cuda
HOME_DIR=/home/ec2-user
SRC_DIR=$HOME_DIR/sources
USER_EXEC="sudo -u ec2-user"
CPUS=$(nproc)
LOG_FILE="$HOME_DIR/install.log"

if [[ "$1" != "--stdout" ]]; then
exec >>"$LOG_FILE" 2>&1
fi

# Create source directory
mkdir -p $SRC_DIR
pushd $SRC_DIR

# Install system utilities and updates
install_utils() {
    echo "Installing utilities..."
    dnf -y update
    dnf -y groupinstall "Development Tools"
    dnf install -y openssl-devel cmake3 rust cargo amazon-efs-utils htop iotop yasm nasm jq python3-pip python-devel cronie cronie-anacron
    echo "Success : Updates and developer tools installed."
    echo "PATH=$CUDA_HOME/bin:\$PATH" | sudo tee -a $HOME_DIR/.bashrc
    echo "LD_LIBRARY_PATH=$CUDA_HOME/lib64:\$LD_LIBRARY_PATH" | sudo tee -a $HOME_DIR/.bashrc

    echo "/usr/local/lib" | sudo tee /etc/ld.so.conf.d/usr-local-lib.conf
    echo "/usr/local/lib64" | sudo tee -a /etc/ld.so.conf.d/usr-local-lib.conf    
}

# Setup GPU, CUDA and CUDNN
setup_gpu() {
    echo "Setting up GPU..."
    wget https://us.download.nvidia.com/tesla/535.104.05/NVIDIA-Linux-aarch64-535.104.05.run
    sh NVIDIA-Linux-aarch64-535.104.05.run --disable-nouveau --silent

    wget https://developer.download.nvidia.com/compute/cuda/12.2.2/local_installers/cuda_12.2.2_535.104.05_linux_sbsa.run
    sh cuda_12.2.2_535.104.05_linux_sbsa.run --silent --override --toolkit --samples --toolkitpath=/usr/local/cuda-12.2 --samplespath=$CUDA_HOME --no-opengl-libs

    wget https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-sbsa/cudnn-linux-sbsa-8.9.4.25_cuda12-archive.tar.xz
    tar -xf cudnn-linux-sbsa-8.9.4.25_cuda12-archive.tar.xz
    cp -P cudnn-linux-sbsa-8.9.4.25_cuda12-archive/include/* $CUDA_HOME/include/
    cp -P cudnn-linux-sbsa-8.9.4.25_cuda12-archive/lib/* $CUDA_HOME/lib64/
    chmod a+r $CUDA_HOME/lib64/*
    ldconfig
    rm -fr cu* NVIDIA*
}

# Install PyTorch from source
install_pytorch() {
    echo "Installing PyTorch..."
    wget https://github.com/ccache/ccache/releases/download/v4.8.3/ccache-4.8.3.tar.xz
    tar -xf ccache-4.8.3.tar.xz
    pushd ccache-4.8.3
    cmake .
    make -j $CPUS
    popd
    dnf install -y numpy
    pip3 install typing-extensions
    git clone --recursive https://github.com/pytorch/pytorch.git
    pushd pytorch
    python3 setup.py install
    popd
    ldconfig
    $USER_EXEC pip3 install sympy filelock fsspec networkx
}

mount -o remount,size=15G /tmp/

# Execute Functions
install_utils
setup_gpu
source $HOME_DIR/.bashrc
install_pytorch

# Cleanup
popd
rm -fr $SRC_DIR

# Test the installation
$USER_EXEC python3 -c "import torch; print('Congratulations, you are all set to go.' if torch.cuda.is_available() else 'Something went wrong. Please check if you missed any steps.')"

EOF
)

AMI_ID="ami-0b9ce70cf1bc24fc3"

aws ec2 run-instances \
    --image-id $AMI_ID \
    --instance-type g5g.4xlarge \
    --key-name $KEY_PAIR \
    --subnet-id $SUBNET \
    --security-group-ids $SECURITY_GROUPS \
    --region $REGION \
    --block-device-mappings '[{"DeviceName":"/dev/xvda","Ebs":{"VolumeSize":20,"VolumeType":"gp3"}}]' \
    --tag-specifications 'ResourceType=instance,Tags=[{Key=Name,Value=AMI-Builder}]' \
    --user-data "$USER_DATA"

Just kidding, all i needed to was make sure libpng/jpg development packages were installed and

python3 setup.py install from the cloned torchvision repo

# Install torchvision from source
install_torchvision() {
    echo "Installing torchvision..."
    dnf install -y libjpeg-turbo-devel libpng-devel
    wget https://github.com/pytorch/vision/archive/refs/tags/v0.17.0.tar.gz
    tar -xf v0.17.0.tar.gz
   
    pushd vision-0.17.0
    python3 setup.py install
    popd
    ldconfig
}

From interactive terminal post launch

[ec2-user /home/ec2-user]$ sudo su
[root /home/ec2-user]# dnf install -y libjpeg-turbo-devel libpng-devel
[root /home/ec2-user]# source /home/ec2-user/.bashrc
[root /home/ec2-user]# mkdir -p /home/ec2-user/sources && pushd /home/ec2-user/sources
[root /home/ec2-user/sources]# wget https://github.com/pytorch/vision/archive/refs/tags/v0.17.0.tar.gz
[root /home/ec2-user/sources]# tar -xf v0.17.0.tar.gz && pushd vision-0.17.0
[root /home/ec2-user/sources/vision-0.17.0]# python3 setup.py install
[root /home/ec2-user/sources/vision-0.17.0]# ldconfig
[root /home/ec2-user/sources/vision-0.17.0]# exit
[ec2-user /home/ec2-user]$  python3 -c "import torch; print('Congratulations, you are all set to go.' if torch.cuda.is_available() else 'Something went wrong. Please check if you missed any steps.')"
[ec2-user /home/ec2-user]$  python3 -c "import torchvision; print('Imported torchvision, you are probably all set to go.')"