Forem: gladevise

Building a Stable Diffusion WebUI local environment on Ubuntu 22.04

gladevise — Mon, 29 May 2023 09:51:30 +0000

This article guides you on how to set up a Stable Diffusion environment on Ubuntu 22.04.

Installation of Python, wget, git

First, install the necessary applications such as python, wget, and git.



sudo apt install wget git python3 python3-venv

After installation, check the Python version.



python --version
Python 3.10.6

CUDA Installation

Next, install CUDA.

Confirming the version of CUDA compatible with PyTorch

Before installing CUDA, check the version of CUDA that is compatible with PyTorch. By accessing the PyTorch official site and setting the PyTorch Build to Stable, the OS to Linux, the Package to Pip, and the Language to Python, you can find out which version of CUDA is available.

In this case, it was confirmed that CUDA 11.7 or 11.8 could be used. Since this article might be outdated, I recommend you to confirm the available CUDA versions using this method in advance.

Installing CUDA

To download an older version of CUDA, first access the cuda-toolkit-archive.

Since we are downloading version 11.7 this time, choose either 11.7.0 or 11.7.1. As corrections at the patch level might be included, it is recommended to choose 11.7.1 unless you have a specific reason.

Next, by entering your environment on the CUDA Toolkit 11.7 Update 1 Downloads site, the necessary installation commands will be generated. This time we are using an Ubuntu 22.04 environment with an Intel CPU, so make the selections as follows. Please choose according to your PC environment.

As a result of the selection, the following commands are displayed.



wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-ubuntu2204.pin
sudo mv cuda-ubuntu2204.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/11.7.1/local_installers/cuda-repo-ubuntu2204-11-7-local_11.7.1-515.65.01-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2204-11-7-local_11.7.1-515.65.01-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2204-11-7-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt update
sudo apt install cuda-11-7

Before starting the installation process, there is one thing to note. The last apt install cuda command needs to include a version, like sudo apt install cuda-11-7. Otherwise, the latest version of CUDA will be installed. PyTorch may not be compatible with the latest version of CUDA. Therefore, always specify the version when installing.

After installation is complete, you can check the version of CUDA by running the following command.



/usr/local/cuda-11.7/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Jun__8_16:49:14_PDT_2022
Cuda compilation tools, release 11.7, V11.7.99
Build cuda_11.7.r11.7/compiler.31442593_0

If the path is not set correctly, you may run into issues later. Therefore, I recommend adding the following description to your ~/.bashrc file to set the path to CUDA.



# Setting CUDA PATH
export PATH="/usr/local/cuda-11.7/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-11.7/lib64:$LD_LIBRARY_PATH"

After saving, open a new terminal and execute which nvcc to confirm that the path is set correctly.



❯ which nvcc
/usr/local/cuda-11.7/bin/nvcc

Checking the PyTorch Installation Command

Visit the PyTorch official page and select the command to install.

Currently, as the latest version of CUDA that PyTorch supports is 11.7, you should use pip3 install torch torchvision torchaudio. However, it is recommended to note down the command specifying the version of CUDA for future reference when versions like 11.8 or 12.0 are added.



pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117

This command is to be noted in the place where you set the command to install PyTorch in the Stable Diffusion WebUI.

Installing Stable Diffusion WebUI

First, clone the repository.



git clone --depth=1 --branch v1.2.1 https://github.com/AUTOMATIC1111/stable-diffusion-webui

The webui-user.sh file is used for detailed configuration.

For instance, the Stable Diffusion WebUI is designed to be cloned under ~/ by default. But if you've cloned it somewhere else, you need to change install_dir as follows.



install_dir="/home/$(whoami)/Apps"

The PyTorch installation command is also set here.



export TORCH_COMMAND="pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117"

xformers, which slightly speeds up image generation, can also be enabled by adding COMMANDLINE_ARGS as follows.



export COMMANDLINE_ARGS="--autolaunch --xformers"

The script is granted execution rights and run with the following command.



chmod 755 webui-user.sh
bash ./webui-user.sh

Then run webui.sh in the same terminal.



bash ./webui.sh

On the first run, it may take 1-3 hours as it will need to download the model and dependencies.

If the installation is successful, the WebUI will be displayed at http://127.0.0.1:7860/.

Troubleshooting

Here is a list of common errors and their solutions.

The program 'nvcc' is currently not installed. You can install it by typing: sudo apt install nvidia-cuda-toolkit

This error occurs when CUDA is not installed, or when the path is not set correctly. Please make sure that an executable file exists at /usr/local/cuda-11.7/bin/nvcc.

No module 'xformers'. Proceeding without it.

This error occurs when xformers is not installed. Add --xformers to COMMANDLINE_ARGS in webui-user.sh.



export COMMANDLINE_ARGS="--autolaunch --xformers"

RuntimeError: Couldn't install torch.

This error occurs when the installation of PyTorch fails for some reason.

In my case, the cause of this error was that the version of CUDA supported by the PyTorch I was trying to install did not match the version of CUDA installed on my system.

By default, the Stable Diffusion WebUI tries to install PyTorch with the following command:



pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113

Here, it is trying to install PyTorch that is compatible with CUDA 11.3. However, as a result of running sudo apt install cuda when I was installing CUDA, CUDA 12.1 was installed instead.

As explained at the beginning of the article, I was able to successfully install by matching the versions of PyTorch and CUDA.

However, this solution may not work in all environments. I hope that those who encounter a similar error will find this information helpful in resolving the issue.

Building a Stable Diffusion WebUI environment with RunPod

gladevise — Sun, 07 May 2023 09:01:42 +0000

As Google Colab has restricted the execution of Stable Diffusion in the free tier, we introduce RunPod as an alternative cloud resource.

What you need:

$10
RunPod account
Backblaze account

First, visit RunPod and Backblaze to create accounts.

Paying the fee

RunPod requires a prepayment, so you will need to pay a minimum of $10 upfront.

You will be redirected to the Stripe page, where you can enter your credit card information.

If your payment is successful, the following screen will be displayed.

Building a Stable Diffusion environment

We will build a Stable Diffusion environment with RunPod.

Go to the Secure Cloud and select the resources you want to use. In this case, we will choose the cheapest option, the RTX A4000.

From the existing templates, select RunPod Fast Stable Diffusion.

Once the confirmation screen is displayed, click Deploy.

Click My Pods and check the deployed Pod.

After confirming that the CPU usage has decreased, click Connect and then Connect to Jupyter Lab.

Open the RNPD-A1111.ipynb file.

Once the notebook is open, execute the cells in order from Dependencies, Install/Update AUTOMATIC1111 repo, Model Download/Load, ControlNet, to Start Stable-Diffusion.

When you run Start Stable-Diffusion, a URL like Running on local URL: https://foobar-3000.proxy.runpod.net will be displayed, so access it.

Enjoy generating images within the limits of your credit with the familiar interface.

To stop, click Stop in My Pods.

A confirmation screen will appear, so click Stop Pod.

As mentioned on the confirmation screen, RunPod will continue to charge you until the storage you are using is released, even after stopping the Pod. In the example above, $0.014 will be deducted per hour.

The only way to stop this is to click Terminate to completely delete the Pod.

Saving RunPod data with Backblaze

As it stands, the generated images will be lost, so we will save them to Backblaze.

Go to Backblaze and click Create a Bucket from Buckets.

Enter the Bucket Unique Name and click Create a Bucket.

Next, go to Application Keys.

Click Add a New Application Key to issue a new key.

Enter the key name and specify the bucket you set earlier.

Take note of the keyID and keyName displayed here.

Return to RunPod and click CloudSync.

When the cloud storage selection screen appears, click Backblaze B2.

Click Copy to Backblaze B2.

Enter the keyID and keyName you noted earlier as follows:



Account ID: keyID displayed when generating the key in Backblaze
Application Key: keyName displayed when generating the key in Backblaze
Bucket Path: {bucket} is the bucketName, {folder} is an appropriate name. ex: RunPodCloudSyncSDWebUI/fast-stable-diffusion
Pod Path: /workspace

This will save the output results and properties of Stable Diffusion to Backblaze B2.

Once the upload is complete, you can check the specified folder in Backblaze's Browse Files to see that the files have been saved.

Now, even if you terminate the Pod, the generated images will not be lost. When launching the next Pod, click Copy from Backblaze B2 to copy files from Backblaze B2, allowing you to continue generating images from where you left off.

Conclusion

If you frequently generate images with Stable Diffusion, Google Colab Pro is more cost-effective.
If you only generate images occasionally with Stable Diffusion, RunPod's pay-as-you-go flexibility may be better.
If you temporarily need more specs than those provided by Google Colab, such as for model training, RunPod is more suitable.