RTX 5000 compatibility

[omegabrain]

I can't use Forge with a 5090 or 5080 from now on. This seems to be due to the sm_120 architecture required on RTX 5000s, which is only supported by CUDA 12.8? Is there a solution while we wait for native integration?

[licongy]

Also concerned about this issue, but first waiting for torch to support CUDA 12.8.

The good news is that the linux preview build for CUDA 12.8 has been released, the Windows build shouldn't be too long I guess

[gachapengames]

All of the other repos and SD front ends have a blackwell version out. Is there one coming for forge? I hope so as its my preferred one to use.

[Zboy18]

I ended up just swapping to reforge in the meantime but also hoping forge reaches compatability soon

[E2GO]

There's already existed Nightly Build with 12.8 support and also here's some talks how to get it work in Fooocus: lllyasviel/Fooocus#3862

So hopefully we can get it for Forge also.

[Deadstarrr]

Any update on this? I have an existing install of forgui that I used with my previous card. I want to use it with the 5090 now.

[LoksliSpb]

Is there no news about this?

[Deadstarrr]

Still waiting for it to work in forge. I don't really want to switch to something else.

[E2GO]

Just update torch stuff with nightly builds - all should work.

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --upgrade

Or you can create .bat file in run.bat dir with this to make it automatically:

@echo off
:: Load environment variables required by the system (e.g. PATH, CUDA, Python)
call environment.bat

:: Upgrade to the latest nightly builds of PyTorch, TorchVision, and Torchaudio
:: with support for CUDA 12.8
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --upgrade

:: Optional: check the installed versions
echo.
echo Installed versions:
python -c "import torch; print('Torch:', torch.__version__); print('CUDA:', torch.version.cuda)"

pause

[chreepha]

there are 100s of builds there, specify the one that works on all lenx, and windows ?

[tritant]

d:\stable-diffusion-webui-Forge\venv\Scripts\python.exe -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
Working for me, but break xformers and Fp16-fast/sageattn is not available

[Teskun]

Hi, I am just curious, did you find a solution for xformers and Fp16-fast/sageattn

[maxiarat1]

Hello everyone,
I’m relatively new to ComfyUI and WebUI Forge, but I’ve managed to put together a Docker‑based setup that runs both UIs with shared models and outputs. Preconfigured for NVIDIA RTX 50xx (CUDA 12.8 + xformers). If you’re looking for an easy, ready‑to‑use environment, feel free to take a look at the code here:
https://github.com/maxiarat1/sd-comfy-forge

[sproid-i]

I notice the script uses .ps1. Will it only work on Windows or is there a way to run it on Linux?

[Allwhey]

Hello everyone, I've managed to get the 50 series working (at least for me) using miniconda and a custom requirements file.

I am not sure if this will be able to help you, and I doubt I can give support since I probably won't be able to reproduce issues you're having (at least unless it's for a specific feature I haven't tried yet).

I can confirm it boots, has no dependency conflicts (according to pip, anyways), and can generate images quickly in batches.

You can skip installing the virtual environment as it's just unnecessary nesting.

Here's the requirements file. For these requirements to work, you MUST install CUDA 12.8.
conda_requirements.txt

Place the requirements file in the forge-webui folder before running the last command.

In the folder where you want to install forge webui:

git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git forge-webui
cd forge-webui
conda create -n forge-webui python==3.10.6
pip3 install -r condarequirements.txt --prefer-binary

Alternatively, skip the first two commands and you can attempt to bypass a pre-existing forge install.
See more on that below.

You can then make a one-click batch file (or similar bash script) to start it like this, placing it in the folder above the forge-webui folder.
If you paste this directly, it will not work. Take note of the username in condaroot and example cli args.

@ECHO OFF


set condaroot=C:/Users/<YOURUSER>/miniconda3


echo Setting command args....
set COMMANDLINE_ARGS= --example-arg1 --example-arg2
echo Activating Conda Environment...
call %condaroot%/Scripts/activate.bat forge-webui

echo Launching webui!
cd forge-webui
start "Forge WebUI" python launch.py
cd ..

We bypass the normal batch launchers to avoid unnecessary checks for venv, system location (forge's embedded python install) and other things. This way, it uses the Conda environment's files. You can probably still use them, but you'd have to configure those variables. (SKIP_VENV, PATH, PY_LIBS, PY_PIP, etc.)

Hope this helps some of you!

EDIT:
I had some spontaneous issues with black squares. I am not sure if this is a bug with the 5090 and CUDA, something
I may be experiencing, or an issue with my Forge environment. Either way, I decided to re-setup my environment and noticed some things
that will keep it far more stable.

The requirements file has been updated.
Additionally, here are some notes I made-
There is an issue of potential environment corruption if you do not modify either Forge's included requirements file or the launch_utils.py code.

Run
pip3 install -r conda_requirements.txt torch==2.7.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128

Make sure
requirements_file = os.environ.get('REQS_FILE', "conda_requirements.txt")
is set in modules/launch_utils.py

Alternatively, replace your requirements_versions.txt file with conda_requirements.txt
Otherwise, Forge's launcher has a good chance of corrupting your python environment!

[irtwangevictor]

Allwhey, super super helpful stuff. Works like a charm for me. I ran into a few issues but they were easily solved by chatgpt.

[TheActualDonKnotts]

This works flawlessly, and is much less hassle than the other two methods I was looking at. Thanks for 1, figuring this out so we didn't have to, and 2, putting this together in an easy to implement format that even a person like me that knows nothing about python, torch etc. could do it without issue.