Tensordock Server for AI Generation/Training

• Tutorial on how to setup a server on Tensordock for AI Image generation or Model training
• First wrote it on rlim, but github has is easier with markdown and I can edit it in the future so, here we are

---

This is a guide on how to setup a server for AI (That I use)

• Navigation
  • Preparation
  • Server Preparation
  • UI Preparation
    • Commands used (For Non-Linux Users)
    • A1111 & Forge
    • ComfyUI
    • Kohya_SS
    • Ease of Use for UIs
• General Information
  • Hosting
    • Finding a good hosting for AI can be quite an issue generally. You can have something like GCS (Google Cloud Services) with many features, but it will be expensive. Or you can go for something like runpod which is mostly specified for AI.
    • Since I like having control of the operating system and I'm quite cheap, I went the option of Tensordock where you don't have that much customizability as on GCS, but you have full control of the OS unlike runpod. (I am not sponsored by Tensordock, I use it because I like it)
  • Operating System
    • I'm going to show steps on how to setup Ubuntu 22.04 for AI, since that's the easiest and the one that Tensordock provides. It also provides Windows Server, but I don't have good experience with that one.
    • I will go through all you need to install and most likely want to install for simplicity, the commands will be shown alongside the steps.

Okay that's it for the introduction, let's get into it.

1. Preparation

   1. Create an Account on Tensordock and create your Organization (It is nothing paid, It's just for users who want their account to be used by multiple people)
   2. Go to Your Organization section and deposit some funds, even the minimum 5$ is enough for start.
   3. Now go into your console (On Windows use Terminal or CMD)
      1. We're going to generate a ssh key file for your machine, so you don't need to login every time you use it.
      2. type ssh-keygen -f <location-where-you-want-to-save-it>/<name-of-the-key> -t rsa -b 4096
         1. The location of the file should be generally at ~/.ssh for Linux users and C:\Users\<Your-Name>\.ssh for Windows users.
         2. I like to keep the key name simple, mostly just tensordiffusion, but you can name it whatever you want
         3. Windows specific info: The dividers between directories in a console is different between the OS's, on linux we use a frontslash, on windows it is a backslash
         4. Press enter and your file will be generated where you specified
   4. Go back to your browser and open New Cloud GPU
      1. Now, select a GPU based on your financial status (how much do you want to spend on a server)
         1. I usually go for A4000 for Generation and A5000 or higher for Training
      2. Specify your system resources
         1. One GPU is enough
         2. Use at least 32GB of RAM
         3. At least 8 CPU vCores
         4. And use at minimum like 150GB of storage (you can go wild, it is quite cheap)
      3. Select a location where you want your server
         1. Pricier location have better transfer speeds and more GPUs (If you turn off a server and disassociate a GPU, it can be used up by someone else)
         2. Cheaper ones have less GPUs and sometimes worse transfer speeds (I mean internet)
      4. Select Ubuntu 22.04 LTS
      5. Now let's use your ssh key

         1. Open the key file with .pub extension in a text editor and Ctrl+A copy everything inside (on linux just use cat and copy the output [.pub files are the public files which are supposed to be on a server, the file with no extension is a private file which should be on your computer. Do NOT share the private key file.]

         2. Click on the text show in the image and a popup will emerge, in there you will enter your public key, press update and close the popup

            

      6. Type in the name of your server
      7. Specify a port you want to open
         1. The 5 digit number is the outside port (you will be typing these in your browser later), The 4 digit and less are the internal ports (7860 is used by SD Webui as a port).
         2. Open at least one more port of your choosing, if you choose random number we will fix that later.
      8. Click Deploy Server and you're done. You've successfully deployed an Ubuntu server.

2. Server Preparation

   1. Prepare a file from which you can run the server easily.
      1. The file will look like this: ssh -i <path-to-the-file> -p <ssh-port> user@<server-ip>
         1. The path to the file will be ~/ssh/<filename> on linux and C:\Users\<username>\.ssh\<filename> on Windows (Please use the private key for this
         2. The server ip can be found under your server's settings
         3. The port can be found under the ip address (the 5digit one that assigned to internal port 22)
      2. Save the file with extension .bat on windows or .sh on linux
         1. On linux you will also need to chmod +x <file> to make it executable
      3. Windows users can just double-click the file to log-in, linux users run it through the console (as per usual)
   2. Once connected to the server you will run these commands (On linux you paste into a console by Ctrl+Shift+V, because Ctrl by itself is used in different ways ex. Ctrl+C stops the command you have run)
      1. sudo apt update && sudo apt upgrade -y | Check for packages that can be updated and update them without asking
      2. sudo apt install -y python3 python3-pip python3-venv python3-tk git wget google-perftools zip unzip | Install python and all the dependencies for webui's
   3. Let's open up the firewall
      1. sudo ufw allow ssh | ufw (Universal Firewall) allow running ssh port (port 22)
      2. sudo ufw allow <port> | run this for the ports you specified and the port 8888 (as a backup)
      3. sudo ufw enable | It will ask you if you are sure, select yes unless you haven't allowed ssh. If you run this and you haven't allowed ssh, your ssh connection will break and you will not be able to connect to the server.
   4. (Optional) Install software for helping you keep track of the system and manage files
      1. btop (also know as bashtop)
         1. It is from the family of top software/cli utilities, which are essentially system monitors
         2. git clone https://github.com/aristocratos/btop.git | We're gonna build it from source to get GPU support
         3. cd btop | learn this command, it is a Change Directory command, to go back from a directory use cd ..
         4. make | this will build all the files from the source
         5. sudo make install | this will install the software to the system
         6. Then you can run it just by typing btop into the console. Please run it alongside your AI session in a New Window/Tab/Split.
      2. dust (file tree and file size cli utility)
         1. Just install it through snap (a package manager for Linux, which is generally bad but for this it is enough) sudo snap install dust
         2. Run it with dust in a directory
         3. Advanced users may install it through Homebrew or build it from source, but that complication is not needed in this guide
   5. On your HOME PC install Filezilla Client, for easier access to files on the server through SFTP (Secure/SSH File Transfer Protocol)
      1. After installing it, open it
      2. In the top-left corner under the menu bar (under File) click on the first Button/Icon which opens the Site Manager
      3. Click New Site
         1. For protocol select SFTP
         2. For Host, it is the IP Address of your server
         3. For Port, it is the SSH port
         4. Change Logon type to Key File
         5. User is just "user"
         6. And key file is just your private ssh key file
      4. After this you can connect and browse/upload/download files from your server easily

3. UI Preparation

   (Do NOT copy the commands whole, try to write them yourself alongside this guide, I will prompt you when you can copy fully)

   1. Some commands we will be using

      1. cd ~ - change directory to home, cd .. - change directory to level above, cd <dir> - change directory to what you specify
      2. git clone - clone a github repository, git pull - update a cloned repository
      3. echo "something" - print text to console, echo "something" >> <file> - append/add text to a file
      4. You can use the tab key to autocomplete paths in the console, ex. cd stab + <TAB> -> cd stable-diffusion-webui
      5. chmod +x <file> - change file's mode/permissions, +x allow file to be executed
      6. You can stop a command you ran by pressing CTRL+C

   2. A1111 & Forge

      1. A1111

         # This can be fully copied
         cd ~
         git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
         cd stable-diffusion-webui
         echo 'SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )' >> ./start.sh
         echo '$SCRIPT_DIR/venv/bin/python3 $SCRIPT_DIR/launch.py --xformers --listen --port 7860 --enable-insecure-extension-access' >> ./start.sh # You can specify the port you want
         chmod +x start.sh
         chmod +x webui.sh
         ./webui.sh # Just for installation purposes

         Forge

         # This can be fully copied
         cd ~
         git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git
         cd stable-diffusion-webui-forge
         echo 'SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )' >> ./start.sh
         echo '$SCRIPT_DIR/venv/bin/python3 $SCRIPT_DIR/launch.py --xformers --listen --port 7860 --enable-insecure-extension-access' >> ./start.sh # You can specify the port you want
         #If you are using forge and have 16GB+ VRAM you can add "--cuda-malloc --cuda-stream --pin-shared-memory --enable-insecure-extension-access" arguments, there may be a speedup
         chmod +x start.sh
         chmod +x webui.sh
         ./webui.sh # Just for installation purposes

         After you run one of these scripts, kill the program with Ctrl+C and run it again with ./start.sh so that it will start up with your commandline arguments.
      2. If the --enable-insecure-extension-access argument causes errors for some reason, do nano start.sh (nano is commandline text editor),
         navigate with your arrow keys to the argument and just remove it. Save the file by CTRL+S and CTRL+X, in that order.
         Now you cannot install extension inside the UI as --listen blocks that and since --enable-insecure-extension-access gives errors,
         if you want to install an extension, you will need to go to the extension directory and clone the repository of the extension there.
         • example:

           cd extensions
           git clone https://github.com/adieyal/sd-dynamic-prompts.git

   3. ComfyUI

      1. # This can be fully copied
         cd ~
         git clone https://github.com/comfyanonymous/ComfyUI.git
         cd ComfyUI/
         cd custom_nodes/
         git clone https://github.com/ltdrdata/ComfyUI-Manager.git # With manager you can install custom nodes even on the server
         cd ..
         python3 -m venv venv
         source venv/bin/activate
         pip install -r requirements.txt
         pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121
         echo 'SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )' >> ./start.sh
         echo '$SCRIPT_DIR/venv/bin/python $SCRIPT_DIR/main.py --listen 0.0.0.0 --port 7860' >> ./start.sh # You can specify the port you want
         chmod +x start.sh
         deactivate
         ./start.sh

   4. Kohya_SS

      • You can copy the commands fully here
      1. Install Nvidia CUDA Toolkit
         • Go to this url https://developer.nvidia.com/cuda-downloads
         • Copy the Base Installer contents and run them in your server console
         • Example: (for CUDA Toolkit 12.5)

           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/12.5.0/local_installers/cuda-repo-ubuntu2204-12-5-local_12.5.0-555.42.02-1_amd64.deb
           sudo dpkg -i cuda-repo-ubuntu2204-12-5-local_12.5.0-555.42.02-1_amd64.deb
           sudo cp /var/cuda-repo-ubuntu2204-12-5-local/cuda-*-keyring.gpg /usr/share/keyrings/
           sudo apt-get update
           sudo apt-get -y install cuda-toolkit-12-5 nvidia-cuda-toolkit # Install both CUDA 12 and CUDA 11, bitsandbytes needs CUDA 11 for some reason 

      2. Install Nvidia cuDNN
         • Go to this url https://developer.nvidia.com/cudnn-downloads
         • Copy the Installation Instructions and run them aswell
         • Example:

           wget https://developer.download.nvidia.com/compute/cudnn/9.2.0/local_installers/cudnn-local-repo-ubuntu2204-9.2.0_1.0-1_amd64.deb
           sudo dpkg -i cudnn-local-repo-ubuntu2204-9.2.0_1.0-1_amd64.deb
           sudo cp /var/cudnn-local-repo-ubuntu2204-9.2.0/cudnn-*-keyring.gpg /usr/share/keyrings/
           sudo apt-get update
           sudo apt-get -y install cudnn

      3. Export CUDA to path permanently

         • echo 'export PATH=/usr/local/cuda/lib64:$PATH' >> ~/.bashrc
           echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64' >> ~/.bashrc
           echo 'export CUDA_PATH=/usr/local/cuda' >> ~/.bashrc
           exec bash # Reloading the console without restarting the system

      4. Install Kohya_SS

         • # This can be fully copied
           cd ~
           git clone https://github.com/bmaltais/kohya_ss.git
           cd kohya_ss
           ./setup.sh
           ./venv/bin/accelerate config
           echo 'SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )' >> ./start.sh
           echo '$SCRIPT_DIR/gui.sh --listen 0.0.0.0 --server_port 7860 --headless' >> ./start.sh # The server port can be changed to what you need
           chmod +x start.sh
           ./start.sh

         • After installing Kohya_SS, accelerate prompt will show up. Don't be scared, just follow these steps:
           • In which compute environment are you running? -> This Machine; Press Enter
           • Which type of machine are you using? -> No distributed training; Press Enter
             (Unless you got multiple GPUs when setting up the server, then you would go with Multi-GPU)
           • Do you want to run your training on CPU only? -> Type no; Press Enter
           • Do you wish to optimize your script with torch dynamo? -> Type no; Press Enter
           • Do you want to use DeepSpeed? -> Type no; Press Enter
           • What GPU(s) (by id) should be used for training on this machine as a comma-seperated list? -> Type 0; Press Enter
             (Again, exception for Multi-GPU users, type all)
           • Do you wish to use FP16 or BF16 (mixed precision)? -> no; Press Enter
           • Hooray, you live. /s

   5. Ease of Use for UIs

      • Create aliases for running the different UIs
        • As you have seen before, the export command does export to a path (environment variables on Windows)
        • We are now gonna use the alias command (This is the reason for the SCRIPT_DIR echo, without it this wouldn't work)

        •  echo 'alias webui=/home/user/stable-diffusion-webui/start.sh' >> ~/.bashrc # Example for A1111
           echo 'alias forge=/home/user/stable-diffusion-webui-forge/start.sh' >> ~/.bashrc # Example for Forge
           echo 'alias comfyui=/home/user/ComfyUI/start.sh' >> ~/.bashrc # Example for ComfyUI
           echo 'alias kohya=/home/user/kohya_ss/start.sh' >> ~/.bashrc # Example for Kohya_SS
           exec bash # Reload console/bash config

        • You can change the alias names to your liking (alias name are this alias <name>=/)
        • If you set what you want, now you can just run ex. webui as a command to start the UI