How to Run Forge-neo on RunPod: Complete Guide to the DCAI Forge Neo Minimal Template [WAN2.2 Support]

This article explains how to install Forge-neo, which was previously introduced, on a RunPod Pod. The Pod template is basically the same as the Forge template introduced earlier. First, please note that this template runs on CUDA version 13.0. It can also be used with GPUs like the RTX3090 and RTX4090, but when selecting a GPU Pod, you need to set the CUDA versions in Additional filters to 13.0 and then select your GPU. If you are unfamiliar with network volumes, I recommend reading the following article first.

Build Stable Diffusion WebUI Forge on RunPod｜Fast Startup & Low‑Cost with NV

🔗Read later

As a prerequisite, you need to have a RunPod account already created. If you have not registered yet, please refer to the following guide first.

Installation guide for Kohya ss_GUI using GPU Cloud RunPod

🔗Read later

What You’ll Learn in This Article🔗

• Features and overview of the “DCAI Forge Neo Minimal” template for running Forge-neo on RunPod.
• Steps to generate a RunPod API key and how to securely register it as a Secret.
• Steps to deploy the template and how to configure Environment Variables.
• How to download models using the JupyterLab terminal.
• How to use the auto shutdown, cleaning, and backup/restore features.
• Verifying WAN2.2 operation and comparing generation speeds by GPU. (💎Members only)

“DCAI Forge Neo Minimal” Template Overview🔗

DCAI Forge Neo Minimal / RunPod

https://console.runpod.io/deploy?template=tnl5ts1b7f&ref=8x35frmy

🔗External Link

No official Forge-neo template was found on RunPod. Using the lightweight DCAI template “DCAI Forge Neo Minimal” introduced here, you can set up a Forge-neo environment quickly and simply. The key features of the template are summarized below.

• Fast startup: Minimal configuration with unnecessary software removed as much as possible.
• PV/NV support: Compatible with both standard pod volumes and network volumes.
• Forge-neo (port 7860): Python 3.13 / Torch 2.11.0+cu130, Sage Attention 2.0 + CivBrowser included by default.
• JupyterLab (port 8888): Convenient for file management and terminal operations.
• Auto shutdown: Automatically stops if there is no access for a set period of time (stops when PV is unused / deletes when NV is in use).
• SageAttention Auto-rebuild: Automatically reinstalls Sage Attention based on the GPU architecture.
• Output folder auto cleaning: Deletes generated images/videos older than the specified time at startup.
• Trash auto cleaning: Empties the trash at startup.
• Open/closed mode: Password authentication can be configured for Forge-neo/JupyterLab.
• Backup/restore: Save and restore settings with forge-backup.

✅This template is not officially supported. Installation and operation are done at your own risk. If errors or issues occur, please follow the official documentation of Stable Diffusion WebUI Forge-neo itself, or the Issue guidelines of the respective extension developers.

RunPod API Setup🔗

The RunPod API is an interface for controlling RunPod’s GPU resources from an external source. It is required for the auto shutdown feature of the “DCAI Forge Neo Minimal” template used in this article.

How to Generate an API Key🔗

　

• Open RunPod’s “Settings” → “API Keys”.
• “+ Create API Key” → Enter an easy-to-understand name.
• Select Restricted for “API Key Options”.
• Set api.runpod.io/graphql to Read / Write, and api.runpod.ai to None.
• Issue by clicking “Create”. Be sure to save the displayed rpa_xxxx... temporarily (it can only be viewed at this time).

  　

How to Use RunPod Secrets🔗

It is not recommended to write sensitive information (API keys and passwords) directly into Environment Variables. Register them in RunPod’s Secret feature and reference them from the template to handle them securely.

How to Create a Secret🔗

　

• Open the “Secret” page on RunPod.
• Click “+ Create Secret”.
• Secret Name: Enter API_AutoStop.
• Secret Value: Enter the API key you generated above.
• Save by clicking “Create Secret”.

How to Deploy the “DCAI Forge Neo Minimal” Template🔗

DCAI Forge Neo Minimal / RunPod

https://console.runpod.io/deploy?template=tnl5ts1b7f&ref=8x35frmy

🔗External Link

Select the template and deploy it. The options are as follows.

• Select template: Click the link above to go to the deployment screen.
• (Optional) Select Network Volume: At the top of the “Deploy a Pod” screen, select a previously created NV.
• ⚠️Important: Set GPU filter: Click Additional filters and select 13.0 from CUDA versions.

  　

• Select GPU: RTX5090 and RTX4090 are recommended, but if they are busy, the RTX PRO 4500 also offers great cost performance.
• Environment Variables: Configure the required ENV via “Edit Template”. ✅For the first startup, it is recommended to change IDLE_ENABLE to false.

The following items should be checked first.

　

• Make sure RUNPOD_API_KEY is set to {{ RUNPOD_SECRET_API_AutoStop }}.
• For the first startup, set IDLE_ENABLE=false to disable auto shutdown at startup.
• For the timezone setting format, enter the TZ identifier from 🔗List of tz database time zones.
• For closed mode operation, add the following at the bottom in the Raw Editor and click “Update Variables” to apply (✅also apply RunPod Secrets).

  　

WEBUI_USERNAME={{ RUNPOD_SECRET_forge_username }}
WEBUI_PASSWORD={{ RUNPOD_SECRET_forge_password }}
JUPYTER_AUTH_ENABLE=true
JUPYTER_PASSWORD={{ RUNPOD_SECRET_Jupyter_password }}

✅Always use RunPod Secrets to reference sensitive information such as usernames and passwords.

For other features, switch between false/true as needed. The main ENV variables are as follows.

Once the template is configured, launch it via “Deploy On-Demand”. The first time will take approximately 17 minutes to build the environment. When Startup time: ... (prepare environment: ...) appears in the log, the startup is complete. ✅Even if Forge-neo shows as Ready on the Pod’s Connect screen, wait until you see this in the log.

Directory Structure🔗

The workspace has the following structure. Models are placed in /workspace/forge/models/. Generated output is written to /workspace/outputs/.

/workspace
├─ forge/
│  ├─ embeddings/
│  ├─ models/
│  │  ├─ ControlNet/
│  │  ├─ ControlNetPreprocessor/
│  │  ├─ diffusers/
│  │  ├─ ESRGAN/
│  │  ├─ Lora/
│  │  ├─ Stable-diffusion/
│  │  ├─ text_encoder/
│  │  └─ VAE/
│  ├─ sd-webui-forge-neo/   (FORGE_DIR)
│  ├─ venv/                 (Python 3.13)
│  └─ pip-constraints.txt
└─ outputs/

How to Use the “DCAI Forge Neo Minimal” Template🔗

How to Install and Download Models🔗

There are two main ways to install models. (1) Using CivBrowser, and (2) downloading directly from the JupyterLab terminal.

Method 1: Download with CivBrowser

This template includes “sd-webui-civbrowser” as a standard feature, so you can easily download models from Civitai using it.

sd-civitai-browser (CivBrowser) · GitHub

https://github.com/SignalFlagZ/sd-webui-civbrowser

🔗External Link

Method 2: Download via the JupyterLab Terminal

Download directly using wget in the JupyterLab terminal. For Civitai models, going through the extension is faster, but this method is convenient when downloading from Hugging Face.

This template creates symbolic links for the major folders in convenient locations. Models are located under /workspace/forge/models. ⚠️embeddings is one level up.

Uploading from local is also possible, but transferring several GBs takes time. Downloading directly from Hugging Face or Civitai is recommended.

Here, we will use wget. ⚠️For models that require login (e.g., Flux.1[Dev]), please consider using curl for downloading.

As an example, to save a checkpoint model, first open the JupyterLab “Terminal”.

Here is an example of downloading the following model.

QuantStack/Wan2.2-I2V-A14B-GGUF · Hugging Face

https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF

🔗External Link

Copy the link using “copy download link” on the model page, then replace LINK, FOLDER, and FILE_NAME in the command below and run it. Switch the FOLDER depending on the model.

wget --wait=10 --tries=10 "LINK" -O ~/workspace/forge/models/FOLDER/FILE_NAME.safetensors

For this example, it would look like this.

Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf

wget --wait=10 --tries=10 "https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/resolve/main/HighNoise/Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf" -O /workspace/forge/models/Stable-diffusion/Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf

Running the above command will start the download.

SageAttention Auto-rebuild🔗

This feature is only relevant when using NV. With pod volume operation, you cannot switch GPUs midway, but with a network volume, you can switch between GPUs with different architectures (Compute Capability).

For example, if you build an NV with an RTX4090 (Ada Lovelace) and then start it with an RTX5090 (Blackwell), it will automatically install Sage Attention for Blackwell. This feature allows you to switch GPUs depending on your needs when operating with a network volume.

Auto Shutdown🔗

“DCAI Forge Neo Minimal” has auto shutdown set to true by default. If there is no external access for the time specified by IDLE_TIMEOUT_MINUTES, it will automatically stop. When using PV (standard pod volume), it stops; when using NV, it is deleted.

Note: If Forge is left open in the browser (including in the background), access may continue depending on the situation, which may prevent the shutdown from triggering.

Output Folder Auto Cleanup🔗

This is false by default. It automatically deletes generated images older than MAX_AGE_HOURS at startup. This is a useful option for workflows where NV storage tends to be tight.

⚠️When enabled, download and save any necessary generated images before stopping the Pod. Also, data deleted by this operation cannot be recovered with the mv command.

Trash Auto Cleanup🔗

✅Files deleted in JupyterLab have been changed to be deleted directly without going through the trash. (Cannot be recovered)

This is false by default. Setting EMPTY_TRASH_ON_START=true empties the NV trash at startup. If you want to check the contents, run the following code in the terminal from /workspace.

ls -al /workspace/.Trash-0/{files,info} 2>/dev/null

Closed Mode (Password Authentication)🔗

With open mode, RunPod URLs are accessible by anyone. Considering the risk of URLs being seen in public places, closed mode is strongly recommended when using from multiple devices or locations. Please configure the ENV variables mentioned above (WEBUI_USERNAME / WEBUI_PASSWORD / JUPYTER_AUTH_ENABLE / JUPYTER_PASSWORD) using RunPod Secrets.

Backup🔗

“DCAI Forge Neo Minimal” can back up Forge settings on NV (⚠️models are not included). This is useful when you want to stop storage charges while not using NV for a while.

Backup Steps

Before deleting the NV, run forge-backup in JupyterLab or the RunPod standard Web Terminal. A file like forge_cfg_extlist_YYYYMMDD_HHMMSS.tar will be created in /workspace/outputs, so download and store it locally.

Note: After running forge-backup, config-backup/ and extensions.txt will be created in /workspace/forge. If these remain, they will be loaded at the next startup and overwrite subsequent configuration changes, so please delete them after taking the backup.

Restore🔗

Restore Forge settings and extensions from the backed-up archive (⚠️models are not restored).

Restore Steps

• Create a new NV and deploy with “DCAI Forge Neo Minimal”.
• Launch JupyterLab and upload the backup to /workspace/forge (config-backup/ and extensions.txt).
• Restart the Pod, and settings and extensions will be automatically restored on the next startup.
• ⚠️After restarting, delete the uploaded archive from forge (to prevent the restore from running every time).

Troubleshooting🔗

Cannot Access Forge (Port 7860)🔗

Please connect via the HTTP port from RunPod’s “Connect”. Check whether 7860 is included in the template’s HTTP ports, and whether Startup time or Running on is displayed in the Pod log. To avoid browser cache issues, trying again in an incognito window is also effective.

Suddenly Stopped Working During Use🔗

If there is a RunPod volume link error or the assigned GPU is unstable, models may suddenly become unavailable. Please restart the Pod. If restarting does not resolve the issue, RunPod itself may be experiencing problems. Try starting again after a while.

First Boot Takes Too Long or Appears Frozen🔗

The first time takes approximately 17 minutes to build the Forge core and venv on the volume. Check the progress using prepare environment and Startup time in the log as a guide. From the second startup onward, it starts quickly. ⚠️If the log stops midway after waiting more than 30 minutes, it is possible that a RunPod issue has stopped the process.

Auto Shutdown Not Working or Shutting Down Too Soon🔗

If it is not working, check whether IDLE_ENABLE=true and IDLE_TIMEOUT_MINUTES are set correctly. Also, if Forge is left open in the browser (including background tabs), it may not shut down. Conversely, if it shuts down too quickly, the timeout may be too short. Extend the value as needed and check whether there is regular access to the monitored ports (Forge/JupyterLab).

NV Storage Full / Write Error (No space left on device)🔗

Clean up unnecessary generated content from Output and delete unused models and duplicate files. Setting IMAGE_CLEAN_ON_START=true and MAX_AGE_HOURS as needed to enable auto cleanup is convenient. If storage is consistently insufficient even after this, expand the NV capacity.

Failed to Download Extensions or Models🔗

There may be rate limiting or URL changes on the Civitai side. Retry after some time. If you get a 403 from Hugging Face, the model requires login (token required). Switch to an authenticated download method.

Authentication Not Working / Cannot Log In🔗

Check the values and Secret names of WEBUI_USERNAME/WEBUI_PASSWORD and JUPYTER_AUTH_ENABLE/JUPYTER_PASSWORD, and make sure there are no typos. After changing settings, restart the Pod to apply the changes.

Cannot Connect to JupyterLab (Port 8888)🔗

Check whether JUPYTER_ENABLE=true is set and whether 8888 is included in the HTTP ports. Select the target port in RunPod’s “Connect” to access it. If authentication is enabled, also re-check the password Secret settings.

✅Even if RunPod itself is unstable, it may become unreachable. (e.g., browser goes blank) Try starting again after a while.

Backup / Restore Runs Every Time🔗

There may be a restore archive remaining in /workspace/outputs, or config-backup/ or extensions.txt may still exist in /workspace/forge. After recovery, delete these to prevent the restore from running every time.

Accidentally Deleted pip-constraints.txt🔗

pip-constraints.txt is a file that pins the versions of torch/torchvision to ensure environment reproducibility. If you accidentally deleted it, run the following code in the terminal.

echo -e 'torch==2.11.0+cu130\ntorchvision==0.26.0+cu130' > /workspace/forge/pip-constraints.txt

Running WAN2.2 with the “DCAI Forge Neo Minimal” Template🔗

Here, we will try running WAN2.2. Usage of WAN2.2 is also limited to paid Patreon supporters, but please refer to the following article.

How to Generate WAN 2.2 Videos with Forge Neo: Easier Video Generation Than ComfyUI#Running WAN 2.2 with Stable Diffusion WebUI Forge - Neo

🔗Read later

The following paid article covers how to download models in bulk and compares generation speeds across different GPUs.

Summary🔗

This article covered everything from setting up to operating the lightweight “DCAI Forge Neo Minimal” template for running Forge-neo on RunPod. By simply selecting a GPU compatible with CUDA version 13.0 and configuring the API key and Secrets correctly, you can quickly build a Forge-neo environment with SageAttention and CivBrowser included as standard. By making use of the auto shutdown and auto cleaning features, you can operate comfortably while keeping cloud costs down.