README.md

Setup : AUTOMATIC1111 / Stable Diffusion WebUI ver 1.6.0 for vast.ai

• GPU sharing cloud service Vast.ai : https://cloud.vast.ai/
• Setup memo : 2023/09/01 ver for AUTOMATIC1111/stable-diffusion-webui 1.6.0 (for SDXL 1.0)
• Video explaining in Japanese (Youtube) : https://www.youtube.com/watch?v=U4HrpzkinP4

Preparation

1. Setup your local PC(Windows, macOS, Linux,...)'s SSH client setting. Generate a SSH keypair(private key & public key) for remote access.
2. Create a new account for vast.ai Client Account Type.
3. Regist your SSH public key into vast.ai Account menu.
4. Add credit 10 USD with your credit card on Billing menu. I DON'T recommend Auto charge setting.

GPU Instance configuration

• Base docker image : nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04
• Launch Type: ssh
• On-start script: Not set
• Disk Space To Allocate: over ~60GB (recommend)
• Launch mode :
  • Run interactive shell server, SSH. This will allow you to connect and run commands using an SSH client.
  • Checked : Use direct SSH connection - faster than proxy, but limited to machines with open ports. Proxy ssh still available as backup.
• GPU Type :
  • Interruptible and On-Demand : both OK
  • #GPUs : 1X (if the instance has multi GPUs, the current version of Stable Diffusion uses only 1X GPU.)

Setup procedures

1. Start a instance based on the above Instance configuration.
2. Wait until the instance loads the docker image and starts up. (5 minutes at the fastest, 20 minutes at the longest)
3. Push the >_ CONNECT button and copy the Direct ssh connect command.

• ex)

ssh -p XXXXX root@AAA.BBB.CCC.DDD -L 8080:localhost:8080

4. Paste the command into your terminal/command prompt and add -L 7860:localhost:7860 for browser access (SSH local port fowarding)

• ex)

ssh -p XXXXX root@AAA.BBB.CCC.DDD -L 7860:localhost:7860

5. Connect the instance via SSH with the above 4. ssh command.
6. Install the AUTOMATIC1111 / stable-diffusion-webui (v1.1.0)

6.1 step1 (as ROOT)

apt-get install vim unzip libgl1-mesa-dev libcairo2-dev wget git -y
apt-get install python3 python3-venv python3-dev build-essential -y
adduser user1 --disabled-password --gecos ""
su user1

6.2 step2 (as user1 = not root user)

cd ~
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui
# Install and launch AUTOMATIC1111 WebUI (with xformers)
./webui.sh --xformers

7.Access the stable diffusion by your local PC's web browser

• http://localhost:7860
• via SSH port forward by -L 7860:localhost:7860

how to terminate stable diffusion and restart

• terminate by Ctrl + C
• restart :

./webui.sh --xformers

restart from SSH connection :

su user1
cd ~
cd ~/stable-diffusion-webui/
./webui.sh --xformers

how to install checkpoint model

Execute as user1

wget -P /home/user1/stable-diffusion-webui/models/Stable-diffusion/ http://example.com/HogeHogeModel.safetensors

• SDXL 1.0 (base + refiner)

wget -P /home/user1/stable-diffusion-webui/models/Stable-diffusion/ https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors
wget -P /home/user1/stable-diffusion-webui/models/Stable-diffusion/ https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/resolve/main/sd_xl_refiner_1.0.safetensors

how to install VAE

Execute as user1

wget -P /home/user1/stable-diffusion-webui/models/VAE/ http://example.com/vae-ft-mse-840000-ema-pruned.safetensors

• SDXL 1.0 VAE

wget -P /home/user1/stable-diffusion-webui/models/VAE/ https://huggingface.co/stabilityai/sdxl-vae/resolve/main/sdxl_vae.safetensors

how to install LoRA

Execute as user1

wget -P /home/user1/stable-diffusion-webui/models/Lora/ http://example.com/HogeHogeLora.safetensors

how to install embeddings

Execute as user1

wget -P /home/user1/stable-diffusion-webui/embeddings/ http://example.com/EasyNegative.safetensors

Download the outputs from the instance to local PC.

Execute from your Local PC's terminal

scp -r -P XXXXX root@AAA.BBB.CCC.DDD:/home/user1/stable-diffusion-webui/outputs ./outputs/

(Option) Apache2 for output directory

• as a root user

apt install apache2 -y
a2enmod userdir
/etc/init.d/apache2 restart
su user1

• as a user1 user

cd
mkdir ~/public_html
chmod 711 $HOME
chmod 755 ~/public_html
ln -s ~/stable-diffusion-webui/outputs/ ~/public_html/outputs

• recoonect with port forwarding setting for web server

ssh -p XXXXX root@AAA.BBB.CCC.DDD -L 8080:localhost:80 -L 7860:localhost:7860

• access from local PC with web browser (via SSH port forwarding)
• http://localhost:8080/~user1/

(Additional info) Multi GPU settings

• [1] start a instance which has 2 multi GPUs in vast.ai
• [2] connect the instance with SSH port forward

ssh -p XXXXX root@AAA.BBB.CCC.DDD -L 7860:localhost:7860 -L 7861:localhost:7861

• [3] start AUTOMATIC1111 webui. (1st time) and stop it.
• [4] clone configfile

cd ~/stable-diffusion-webui/
cp ui-config.json ui-config2.json
cp webui.sh webui2.sh

• [5] edit ui-config2.json
• [6] start two webuis with &

cd ~/stable-diffusion-webui/
./webui.sh --device-id 0 --listen --port 7860 --xformers --ui-settings-file ui-config.json &
./webui2.sh --device-id 1 --listen --port 7861 --xformers --ui-settings-file ui-config2.json &

• [7] access the two webui with 2 browser
  • http://127.0.0.1:7860
  • http://127.0.0.1:7861
• [8] change save directory name
  • Settings -> Saving to a directory -> change the name of Directory name pattern from [data] to 0_[date] for GPU0 / 1_[date] for GPU1

Ref. ( &Special thanks)

• https://cloud.vast.ai
• https://github.com/AUTOMATIC1111/stable-diffusion-webui
• https://hub.docker.com/r/nvidia/cuda/tags
• https://pytorch.org/get-started/pytorch-2.0/
• https://github.com/CompVis/stable-diffusion