This is a tutorial on how to install LLaMa on your Windows machine using WSL (Windows Subsystem for Linux). Although with some tweaks you may get this to work properly on another hardware or on multi-GPU setups, this tutorial is specifically designed to work with Nvidia graphics cards - and I only cover a Single-GPU configuration. I'll be using my RTX2060 with 6GB to load Llama model 7B with 4-bit quantization, but you can try this tutorial with other models that fit your Nvidia hardware.
These instructions are delivered for educational purposes only, and it assumes that you have a fresh install of windows 10/11 latest version (and it is obvious that you can try it without having to reinstall Windows).
While we will need to restart the Ubuntu terminal a few times, I've done my best to simplify the instructions to minimize the number of steps. This will help you explore these AI models as fast as possible.
I used some lines "git reset..." to maximize compatibility throughout time.
Open windows update (right-click on start menu; settings; update & security), click on advanced options and enable “receive updates for other Microsoft products when you update Windows”, then check for updates and restart your PC if required.
Go to nvidia website (or through your Geforce Experience app if u have installed) and install latest version of your nvidia driver (make sure it is the “Game ready” version), restart again if required.
Click on start menu; type cmd and click on “run as administrator”
wsl --install
It will ask to restart PC if this is the first WSL installation on this machine
After reboot WSL will continue installation...
While it does, click on start menu; type cmd and open
notepad .wslconfig
Click on yes to create a new file; copy and paste the content below:
[wsl2]
memory=6GB
processors=2
swap=20GB
localhostforwarding=true
Close notepad, click on save; close prompt;
When WSL installation ends, enter a new username and pwd
clear && git clone https://github.com/Highlyhotgames/fast_txtgen_7B.git > /dev/null 2>&1 && cd fast_txtgen_7B && chmod +x requirements && ./requirements
Wait for message to close terminal
To ensure WSL2 is activated and to update:
Click on Start menu; type cmd and run as Administrator
wsl --set-version Ubuntu 2 & wsl --update
Without closing prompt, click on start menu; right-click on Ubuntu; more -> pin to taskbar;
Open windows update and check for updates - it will get an update for WSL;
After update, go back to Admin prompt:
cd lxss\lib & del libcuda.so & del libcuda.so.1 & wsl -e /bin/bash -c "ln -s libcuda.so.1.1 libcuda.so.1 && ln -s libcuda.so.1.1 libcuda.so"
Close prompt and restart your PC;
Open Ubuntu:
./install
When "checking CUDA Installation" it will display a message "release 11.7" in cyan color.
After model completes loading, if last line says
"To create a public link, set share=True in launch().",
open your browser and enter this URL:
To start server, open Ubuntu:
conda activate webui && cd webui/text-generation-webui
Run:
python server.py --wbits 4 --model llama-7b-4bit-128g --groupsize 128 --chat
To remove completely (including WSL), open cmd as Admin:
wsl --shutdown & wsl --unregister Ubuntu