Skip to content

INSTALLATION.md

Latest commit

 

History

History
80 lines (57 loc) · 1.93 KB

INSTALLATION.md

File metadata and controls

80 lines (57 loc) · 1.93 KB

Installation

Using Automated BASH Script

  • Install Nvidia Driver, CUDA Toolkit & Python Dependencies

Stable installation from requirements.txt

chmod +x install.sh stable
./install.sh

Install latest version of libraries:

chmod +x install.sh latest
./install.sh

Install CUDA 11.6 Toolkit

If installed nvidia driver version is >=450.80.02 then install CUDA 11.6 Toolkit:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-keyring_1.0-1_all.deb
sudo dpkg -i cuda-keyring_1.0-1_all.deb
sudo apt-get update
sudo apt install cuda-toolkit-11-6

If installed nvidia driver version is < 450.80.02 - Remove previously installed nvidia driver

sudo apt clean
sudo apt update
sudo apt purge nvidia-*

nvidia driver version 510 & CUDA 11.6 - download

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-keyring_1.0-1_all.deb
sudo dpkg -i cuda-keyring_1.0-1_all.deb
sudo apt-get update
sudo apt install nvidia-driver-510 cuda-toolkit-11-6

Dependency Installation

python3 -m pip install transformers tensorboard datasets nvidia-ml-py3 python-dotenv requests huggingface_hub evaluate pyext accelerate
pip install torch --force-reinstall --extra-index-url https://download.pytorch.org/whl/cu116
sudo apt install libaio-dev ninja-build
pip install deepspeed

If error "deepspeed: command not found" is visible after installation - faced this issue on ubuntu 20:

nano ~/.bashrc

Save following at end of file

export PATH="/home/user/.local/bin:$PATH"

You will then need to profile, do this by either running the command:

source ~/.bashrc