🤝｜tech-support

[cosmic furnace]

Successfully uninstalled torchvision-0.23.0+cu128
WARNING: Skipping torchaudio as it is not installed.

(venv) B:\SD-FORGE-NEO\sd-webui-forge-neo>python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision
Looking in indexes: https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch

(venv) B:\SD-FORGE-NEO\sd-webui-forge-neo>

[ornate elk]

Let me retest that tomorrow. Maybe therock changed the distribution link

Then I'll update the guide

Yea worked for me yesterday too and today just black xD

Dev will probably fix it in a day

[sick merlin]

I'm installing swarm as a backup anyways

[eternal cave]

hello - i accidentally logged into stablediffusionweb using my google account. does this mean my account is now hacked?

[fair oxide]

hello everyone i am SynthWaveDreams or Derek, if you prefer, i am new here trying to learn how to use stable defusion to generate videos, currently am trying to learn how to even word what i want via text to video using wan 2.2 but all my videos come out blank or fail alltogether and ive watched over 50 youtube videos so far and i just dont understand what im doing so i wanted to join here to maybe find someone who could screenshare teach me how to use and set it up.

[lucid yoke]

If you used the same password i recommend changing it and if it's a single sign on button, you* should* be fine but still activate 2fa if you haven't already

[vocal burrow]

Probably not. They're using predatory techniques but as far as I can tell they're not hacking. But yes disconnect, change your password if that worries you and voilà

[tawny saddle]

.-.

msfs2020 makes me frangry

updates and installs of updates should happen on steam, not the game itself

gfingfnw

[tawny saddle]

xd

i would never ever ever ever ever evr evre rv, login to stabledeiffweb, bc, 1, u can do it on ur PC urself, and 2, xd, huggingface, stablityAI websites

that shi sus asf

xd

[rich wasp]

Im trying to work witth t2iadapter on comfy but i have been trying to findout why t2iadapter nodes doesnt appear on my node list for last 5 hours.I have tried to dowloand dozens of different t2i-xl adapter but eventho i put them in correct folder and they are safe tensor file and their names start with t2iadapter somehow comfyui doesnt see them.Since i just dowloand the comfy 2 days ago as much as i understand in order to connect the t2iadapter node i need a helper node.I've tried canny,openpose,depth,facerecognition and more of it but somehow i cant see their names on neither node libary neither at the t2iadapter custom weights node.If anyone knows something about this tips are greatly appericated .I guess thats as long as i can go with chatgpt

[tawny saddle]

is there a way to install comfy, without needing to redownload and install stable

like, is there an, "hacky" way to do it

cuz it is just the Ui init

[onyx solar]

If you are referring to your models/checkpoints, just move the model folder somewhere, reinstall comfy, then move the model folder back.

[tawny saddle]

nu, i just dont feel like it downloading and installing dependencies

if i can just, swawp the basic Ui with the comfy ui

[onyx solar]

That doesn't make sense. Are you trying to move from Automatic1111 to ComfyUI?

[lucid yoke]

If you want both comfy and a easy to use UI i recommend swarmUI

Though you have to know what your doing a little more compared to forge etc

[tawny saddle]

im assuming that bc u dunt understand wut im asking, that must mean that the file structure and intneral stuff is not at all the same as basic auto1111

[onyx solar]

So you are trying to change from A1111 to ComfyUI like I said, right?
You can just install ComfyUI and point it at your existing models folder or you can move your models over after you install ComfyUI if you don't plan on using A1111 anymore.

[lucid yoke]

In short, no you cant escape downloading different dependencies since it uses a venv

To avoid conflict with other programs

[tawny saddle]

i understand cloning and double clicking webui-user.bat and waiting 30mins for it finish downloading a bunch of stuff as usual, but is there not just simple things i can swap out in auto1111, to then have the comfyUI interface, instead of going through the entire install process

[lucid yoke]

No

[tawny saddle]

darn

[onyx solar]

They are two different applications with two different sets of dependencies.

[tawny saddle]

xd

[tawny saddle]

ah oki

i was thinking it was just the interface

welpppppppp

agh

i might try comfy, i like the stretchy noodles and the colours

looks like fun

also heard, or rather read that u can do more with it

than u can with reg

[lucid yoke]

you can do a whole lot more with it yes but considering your 500mb vram

[onyx solar]

It's a node system...it can do what you make of it.

[tawny saddle]

xd

[onyx solar]

500mb vram...yikes

[tawny saddle]

yas :>

gud ole apu

xd, ill get a card at sum point

[onyx solar]

I think your focus should be hardware, not software.

[lucid yoke]

i mean with comfy they could use a API for generating

sfw only tho but still

[tawny saddle]

i mean, i can make 620x750 things pretty gud

only take like 8mins using apu

[lucid yoke]

erm

[onyx solar]

😬

[tawny saddle]

xdd

[lucid yoke]

||5 seconds here for double that, 15s for 2k image||

hardware should be your focus

[tawny saddle]

yas ik

my reasoning for not having one atm is that all the games i play run great at the res my nuggety monitor has, and i just never had a real need for a card, but having now started using MSFS2024, and BeamNG's later updates and features, i am starting to see more reason to begin card hunting

plus with Stables new models becoming larger

6GBs seeming to be about the average ;o

[lucid yoke]

for cards? yeah, for models? more like 16gb+

[tawny saddle]

might get a 3090 to mess wit tbh, bc it cheap and has about 8GB vram, which is better than 512mbs

xd

[lucid yoke]

3090 has 24gb

[tawny saddle]

ah

even better aye

[lucid yoke]

but dont forget you also need a powesupply, space in your machine etc

[tawny saddle]

yas :V, tru, xD, i did indeed completly forget about that, tinx for reminder, atm i have a 450watt, but, p sure i need like at least 700

yikes, 850watts

[lucid yoke]

at some point its easier to design a new pc lmao

[tawny saddle]

xD yas

i might do that too tbh, i wan addressable RGB, and GA-a320-m-h doesnt have any fun feature

other than the name a320, like the airliner

[rich wasp]

is there any discord i can reach people who knows about controlnet or t2iadapters?

[ornate elk]

What questions do you have?

And for which webui?

[lucid yoke]

this message:
#🤝｜tech-support message

[ornate elk]

Ah comfyui related

[rich wasp]

yes

[ornate elk]

you can try ask in #🧣｜comfy-ui

[cosmic furnace]

Wassup, sooo. Have u looked into Therock ?

[ornate elk]

worked for me

link didnt changed

[cosmic furnace]

it worked now for me too o___o . but yesterday didnt ...wtf?

@echo off
set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1

:: set PYTHON=
:: set GIT=
:: set VENV_DIR=

set COMMANDLINE_ARGS= --skip-python-version-check --cuda-stream --attention-pytorch

:: --xformers --sage --uv
:: --pin-shared-memory --cuda-malloc --cuda-stream
:: --skip-python-version-check --skip-torch-cuda-test --skip-version-check --skip-prepare-environment --skip-install

call webui.bat

this is correct webui-user y ?

btw whats the difference if i run it with basic model 3.5gb size or this one v1-5-pruned of 7.17gb size? 🤔

[ornate elk]

more vram usage

[cosmic furnace]

is it good or bad? 🤔

[ornate elk]

1.5 is bad

[cosmic furnace]

I mean does it affect anything while using pony-illustrous or flux-qwen ? 🤨

[ornate elk]

I dont understand the question

1.5 has nothing to do with pony or flux

[cosmic furnace]

Why they even release 1.5 then ? Or its early & unoptimized? So ull advice to use usual 3.5 model? & not this 7.17 pruned 1.5 ?)

[ornate elk]

1.5 is 3 years old

The size difference is just float point precision

[cosmic furnace]

and in the end it says this again

• Installing collected packages: torch, torchvision
  Successfully installed torch-2.8.0+cu128 torchvision-0.23.0+cu128
  Traceback (most recent call last):
  File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 52, in <module>
  main()
  File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 41, in main
  prepare_environment()
  File "B:\SD-FORGE-NEO\sd-webui-forge-neo\modules\launch_utils.py", line 315, in prepare_environment
  raise RuntimeError("PyTorch is not able to access CUDA")
  RuntimeError: PyTorch is not able to access CUDA
  What i might be missing?🤔

[ornate elk]

You need to install the rocm torch

In this error message it shows the nvidia one

[cosmic furnace]

RX 7900XTX (GFX110x): python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision - u mean this one? urgh i did & removed torch before. pretty sure

[ornate elk]

Probably you weren't inside the activated venv with cmd

Try it again

[ornate elk]

@cosmic furnace make sure you copy and paste only the command and not the GPU type before

[cosmic furnace]

installed all over again from scratch and have this error -
venv "B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Scripts\Python.exe"
Python 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
Version: neo
Traceback (most recent call last):
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 52, in <module>
main()
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 41, in main
prepare_environment()
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\modules\launch_utils.py", line 315, in prepare_environment
raise RuntimeError("PyTorch is not able to access CUDA")
RuntimeError: PyTorch is not able to access CUDA

Y i used this to install and it said all is fine -
python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision

---

Mb i messed up installing hip sdk? but how, its just setup installer 🤨

So u advice to use the default stable model & avoid this 1.5 pruned?)

[timber gate]

something i need to change for this?

[ornate elk]

oh mybad its not models-dir its model-ref

[cosmic furnace]

if i try run this command again it says everything is okay & installed -
B:\SD-FORGE-NEO\sd-webui-forge-neo>python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision
Looking in indexes: https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/
Requirement already satisfied: torch in g:\sd\python_3.12.10\lib\site-packages (2.9.0+rocm7.10.0a20251019)
Requirement already satisfied: torchvision in g:\sd\python_3.12.10\lib\site-packages (0.24.0+rocm7.10.0a20251019)
Requirement already satisfied: filelock in g:\sd\python_3.12.10\lib\site-packages (from torch) (3.19.1)
Requirement already satisfied: typing-extensions>=4.10.0 in g:\sd\python_3.12.10\lib\site-packages (from torch) (4.14.1)
Requirement already satisfied: setuptools in g:\sd\python_3.12.10\lib\site-packages (from torch) (80.9.0)
Requirement already satisfied: sympy>=1.13.3 in g:\sd\python_3.12.10\lib\site-packages (from torch) (1.14.0)
Requirement already satisfied: networkx>=2.5.1 in g:\sd\python_3.12.10\lib\site-packages (from torch) (3.5)
Requirement already satisfied: jinja2 in g:\sd\python_3.12.10\lib\site-packages (from torch) (3.1.6)
Requirement already satisfied: fsspec>=0.8.5 in g:\sd\python_3.12.10\lib\site-packages (from torch) (2025.7.0)
Requirement already satisfied: rocm==7.10.0a20251019 in g:\sd\python_3.12.10\lib\site-packages (from rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)

Requirement already satisfied: rocm-sdk-core==7.10.0a20251019 in g:\sd\python_3.12.10\lib\site-packages (from rocm==7.10.0a20251019->rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)
Requirement already satisfied: rocm-sdk-libraries-gfx110X-dgpu==7.10.0a20251019 in g:\sd\python_3.12.10\lib\site-packages (from rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)
Requirement already satisfied: numpy in g:\sd\python_3.12.10\lib\site-packages (from torchvision) (2.3.2)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in g:\sd\python_3.12.10\lib\site-packages (from torchvision) (11.3.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in g:\sd\python_3.12.10\lib\site-packages (from sympy>=1.13.3->torch) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in g:\sd\python_3.12.10\lib\site-packages (from jinja2->torch) (3.0.2)

[ornate elk]

its not installed inside the venv

cmd was opened the wrong way

[timber gate]

so its model-ref for neo and models-dir for basic forge?

gotcha

thank you

[cosmic furnace]

U mean like this ? Kinda says same? -
B:\SD-FORGE-NEO\sd-webui-forge-neo>venv\Scripts\activate.bat

(venv) B:\SD-FORGE-NEO\sd-webui-forge-neo>python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision
Looking in indexes: https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/
Requirement already satisfied: torch in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (2.8.0)
Requirement already satisfied: torchvision in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (0.24.0+rocm7.10.0a20251019)
Requirement already satisfied: filelock in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.20.0)
Requirement already satisfied: typing-extensions>=4.10.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (4.15.0)
Requirement already satisfied: sympy>=1.13.3 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (1.14.0)
Requirement already satisfied: networkx in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.5)
Requirement already satisfied: jinja2 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.1.6)

Requirement already satisfied: fsspec in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (2025.9.0)
Requirement already satisfied: setuptools in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (69.5.1)
Requirement already satisfied: numpy in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torchvision) (1.26.4)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torchvision) (11.3.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from sympy>=1.13.3->torch) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from jinja2->torch) (2.1.5)

[ornate elk]

that looks good

[cosmic furnace]

y o.0 , But when i launch webui user i get error -
venv "B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Scripts\Python.exe"
Python 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
Version: neo
Traceback (most recent call last):
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 52, in <module>
main()
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\launch.py", line 41, in main
prepare_environment()
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\modules\launch_utils.py", line 315, in prepare_environment
raise RuntimeError("PyTorch is not able to access CUDA")
RuntimeError: PyTorch is not able to access CUDA

[ornate elk]

check the toch folder in the venv/lib/sitepackages folder

[cosmic furnace]

just torch? 1.22gb files seems fresh. what to check exactly ?

[ornate elk]

still the wrong torch

make sure you launch and install the stuff in the same webui

something is clearly wrong where you install the rocm torch or which webui you launch after that

my torch after using the commands

[cosmic furnace]

(venv) B:\SD-FORGE-NEO\sd-webui-forge-neo>python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision
Looking in indexes: https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/
Collecting torch
Using cached https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/torch-2.9.0%2Brocm7.10.0a20251019-cp312-cp312-win_amd64.whl (547.1 MB)
Collecting torchvision
Using cached https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/torchvision-0.24.0%2Brocm7.10.0a20251019-cp312-cp312-win_amd64.whl (1.4 MB)
Requirement already satisfied: filelock in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.20.0)
Requirement already satisfied: typing-extensions>=4.10.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (4.15.0)
Requirement already satisfied: setuptools in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (69.5.1)
Requirement already satisfied: sympy>=1.13.3 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (1.14.0)
Requirement already satisfied: networkx>=2.5.1 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.5)
Requirement already satisfied: jinja2 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (3.1.6)
Requirement already satisfied: fsspec>=0.8.5 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torch) (2025.9.0)
Requirement already satisfied: rocm==7.10.0a20251019 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)
hmgm

Requirement already satisfied: rocm-sdk-core==7.10.0a20251019 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from rocm==7.10.0a20251019->rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)
Requirement already satisfied: rocm-sdk-libraries-gfx110X-dgpu==7.10.0a20251019 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from rocm[libraries]==7.10.0a20251019->torch) (7.10.0a20251019)
Requirement already satisfied: numpy in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torchvision) (1.26.4)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from torchvision) (11.3.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from sympy>=1.13.3->torch) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in b:\sd-forge-neo\sd-webui-forge-neo\venv\lib\site-packages (from jinja2->torch) (2.1.5)
Installing collected packages: torch, torchvision
Successfully installed torch-2.9.0+rocm7.10.0a20251019 torchvision-0.24.0+rocm7.10.0a20251019

Yeah now launched fine. weird. Pretty sure i did all the same 🤨 . w/e

[ornate elk]

now that looks correct

good that it works now

[cosmic furnace]

Installing nunchaku
Failed to install nunchaku; Please manually install it
Launching Web UI with arguments: --skip-python-version-check --cuda-stream --attention-pytorch
ERROR:bitsandbytes.cuda_specs:Could not detect ROCm GPU architecture: [WinError 2] Не удается найти указанный файл
WARNING:bitsandbytes.cuda_specs:
ROCm GPU architecture detection failed despite ROCm being available.

ERROR:bitsandbytes.cextension:bitsandbytes library load error: Configured ROCm binary not found at B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll
Traceback (most recent call last):
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 313, in <module>
lib = get_native_library()
^^^^^^^^^^^^^^^^^^^^
File "B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 282, in get_native_library
raise RuntimeError(f"Configured {BNB_BACKEND} binary not found at {cuda_binary_path}")
RuntimeError: Configured ROCm binary not found at B:\SD-FORGE-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll
--- Shud i be bothered by this? 🤔

[ornate elk]

Nope, bitsandbytes is still not ready for rocm on windows. So thats why it complaints

Can be ignored

[cosmic furnace]

Why stable diffusion folder is empty? It usually auto downloads basic safetensor 3.5gb model there 🤨

Is this some kind of new forge neo adapatation ?

[ornate elk]

yea forge neo doesnt download any model

[cosmic furnace]

Btw u dont have guides on running flux-qwen-wan ? It sometimes hard to tell is model need anything else to run & where to get it

[cosmic furnace]

https://civitai.com/models/1864281/qwen-image - this for example . Just an 19gb model. No explaining what esle (files) is needed & where 😑

[timber gate]

do i make a t5 folder in models and drag my wan checkpoint there?

[cosmic furnace]

Y see, kinda similar problem there 🤔

[ornate elk]

Nope you need a t5 clip model

[timber gate]

how do i do that

[ornate elk]

I'm AFK, finishing BF6 Campaign

[timber gate]

alright

[ornate elk]

nope, no guides, but flux need t5 and clip l and ae.vae, and qwen needs a clip too and wan also

best is to use the smaller gguf variants

[timber gate]

will this work? pip install --upgrade transformers sentencepiece safetensors open_clip_torch

then make a t5 and clip folder and put the correct files in it

[ornate elk]

No please don't use chatgpt

You just need to download a clipt5 model for flux or whatever model you try to use and put it into the models/clip folder

You also probably need a custom VAE for it too

Goes in models/vae

[cosmic furnace]

Yeah with flux i use mostly 3 same files and it works, But qwen o.0 , and whuht is wan(video?)? And what about chroma?

[ornate elk]

qwen needs a t5 and a vae

chroma needs flan t5 and flux vae

wan, havent tried yet

[timber gate]

u got any model sugggestions?

[jade dew]

@ornate elk I'm in cmd and following a youtube tutorial to install forge neo webui

[cosmic furnace]

yeah i can see now it has tabs with vae and text encoder >_< , Btw ive seen an grease monkey script that makes tabs of civitai loras not in 1 row, but in 3-4 rows, so u dont have to scroll-click them, and see all at once, but cant remember now how it called & where to get it 🙁

[jade dew]

but the dude don't tell me in the video how to get my hands on the .safetensors so I can put them into forge neo directory's Stable Diffusion folder for checkpoint

[ornate elk]

You get checkpoints (models) from Civitai.com and then put them in the models/stable-diffusion folder if they are larger than 2gb

[jade dew]

???

https://github.com/nunchaku-tech/nunchaku/tree/main#

I'm trying to get the .safetensor for nunchaku

[cosmic furnace]

what this nunchaku btw?

[jade dew]

I cloned the repository via cmd, but which file is the safetensor there if it has any?

[ornate elk]

Then search for 4bit quant models to use nunchaku with

Nunchaku is just a technique, not a model itself

It adds support for 4bit quant models

And they link them on their github page, under News

[jade dew]

yeah that's why I'll need

I know what model I need for Civitai but since I have a completely blank stable diffusion being opened up.

this is what I got by fully install Forge Neo

and FYI

I'm a boomer

so please explain it as easy for me as possible

[ornate elk]

What's your graphics card?

[jade dew]

NVidia Gefore GTX 1650 Super

[ornate elk]

Ah okay

You need to download the model you want to use forge neo with and put it into models/stable-diffusion folder

Then you can select it in the checkpoint dropdown

[jade dew]

from the civitai image generator site?

ok

[jade dew]

once I place it inside the dropdown menu makes it available?

[ornate elk]

Yep once its selected in the dropdown it can be used

[jade dew]

it downloads now

[ornate elk]

But there are different type of models and they have to get used differently

[jade dew]

I wait for it before I can grab it

that's a huge ass checkpoint size wow

6.5 GB

good thing I installed the whole damn thing on my external

[ornate elk]

Hopefully not a HDD xD

[jade dew]

ok done

ehm...

I cannot select the thing

[ornate elk]

Now click on the refresh button

On the right of the dropdowns

[jade dew]

arrow unclickable

[ornate elk]

Read my message please

[jade dew]

oh refresh

it won't disappear

ok

oh it's there now

[ornate elk]

Now select XL on the top left

[jade dew]

selected

[ornate elk]

And then your ready to prompt

[jade dew]

okay here's the thing

so far I used google colab via an API key and notebook installation before I got an interface like this

generated various images before

have experience with Regional Prompter of how it works and such etc...

I'm just unfamiliar with the stuff working on a locally installed version.

[cosmic furnace]

and btw what can you say about qwen compared to flux? Gemini sayts qwen slower but just more accurate about what you get from promt 🤔 , and chroma?

And what about that Grease monkey script that shows all tabs in more than 1 rows instead of scrolling?

[ornate elk]

Qwen is nice but slow, chroma is the slowest and I didnt got any useful image out of it.
Dont know about the script

[ornate elk]

No problem, test around a bit and if you have the files you want, its pretty much the same

[cosmic furnace]

Slower than flux ? 😄 😄 🤦‍♂️

[jade dew]

and character loras also go the same folder?

[ornate elk]

Chroma for sure. Its insane slow. Qwen maybe on par with flux

[ornate elk]

Loras go into models/lora

[jade dew]

oh here right?

[ornate elk]

Best is to make subfolders in it for the different types

[jade dew]

oh thank you man

[ornate elk]

No problem 🙂

Upscalers go into models/ESRGAN

VAE in models/vae

[jade dew]

VAE models are stuff like Regional Prompter?

or Forge Couple?

I'll need at least either of these because in my Google Colab sessions I used those to be able to depict 2 or more characters on the same image.

[ornate elk]

Vae are something else.
Forge couple and regional prompter are called Extensions

[cosmic furnace]

If i want Forge Neo to use my checkpoints-lora from old zluda version without moving them around i shud add at the end of webui-user this ? -
@REM --ckpt-dir G:/SD/SD - Forge-Webui-ZLUD/stable-diffusion-webui-amdgpu-forge/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir G:/SD/SD - Forge-Webui-ZLUD/stable-diffusion-webui-amdgpu-forge/models/Lora

Like this Y?

[jade dew]

oh yeah I remember now that you mention the name Extensions

there has to be a panel for extension installs within the SD interface

I think I found something

[ornate elk]

Check my forge setup guide there you see at the bottom how its done correctly

[jade dew]

my friend has pinned me Regional Prompter, but I wonder if I have Forge Couple somewhere around here

as the latter is more advanced (i think)

[ornate elk]

Forge couple has to be installed via the "install from URL"

[jade dew]

https://github.com/Haoming02/sd-forge-couple

from this one specifically

grab the link there?

I gotta download via cmd

bingo

[ornate elk]

That doesnt look correct

[jade dew]

that's just the collective download directory

nothing else

[ornate elk]

Forge couple needs to go into the extensions folder

[jade dew]

like...

the entire thing?

is it correct if I place there

or should I install it via cmd there to start with?

[ornate elk]

You can either git clone it in there directly or install it via the install from URL in the extensions tab

[jade dew]

I don't know what the latter means so I just type "git clone" and paste the copied link there in cmd.

is this correct?

what a relief

[ornate elk]

Then relaunch forge neo so the extension will show up

But I'm off now as its very late here

[jade dew]

okay

thank you for your time it was a huge help!

❤️

[ornate elk]

No problem, feel free to ask here again if you have any questions

[jade dew]

I think I can go on like this on my own for now.

[jade dew]

Or just not.

tomorrow I gather some more help

[jade dew]

Oh my F god!

@ornate elk Oh boy, save my life tomorrow.

The thing unable to recognize neither civitai models nor gguf ones.

[tawny saddle]

u could try the ole reinstall stable, :3

[tawny saddle]

oh, i was gonna get comfy

ehh

do that later

maybe

cuts to 10 days later and i still havent

xd

[jade dew]

I just installed it and now trying to fill the damn thing with loras and checkpoint.

I'm waiting for CS1o for an answer, he must know what to do in such case.

[tawny saddle]

i bet cs1o do

cs1o is like, idek, just knows stable like, back of hand

almost as if cs1o is secrelty the creator

👀

xd

[jade dew]

So far he could help me bunch, so why he wouldn't in such a case?

[jade dew]

anyway it's really a good news

[vocal burrow]

what s the log ? and screenshot the whole webui plz

this kind of cropped screenshot is not telling much.

[queen burrow]

Hi! I have been testing the text-to-image generation models provided by Stability AI and they work great. Now, I am looking to use an image-to-image generation model where I can input both an image and some text as a prompt to generate a new image. Is this possible with any of the models provided by Stability AI?

[jade dew]

@vocal burrow My entire interface.

My cmd which runs this engine.

@vocal burrow I don't know where the log can be found.

[tawny saddle]

yas xd

[jade dew]

I try a different installation method into a different directory maybe that helps.

In the currently installed directory by the meantime I'm looking for an answer from CS1o.

[near dome]

How to reduce GPU load?

Can I distribute it for a long time?

[tawny saddle]

seems scamilicious

[vocal burrow]

the log is pretty much what s in the console you screenshot (abeit there is more stuff above)

[jade dew]

there is no way further above, but I realized things are missing

so I'll do the same method I used before in the former installation method but flash 2.8.2 and that sage thingy will be installed now

and try that way

[vocal burrow]

I m gonna be honest, I m at work and I m not comfortable searching "miaomiaoHAREM" lol. What kind of model is that ? what its weight ?

[jade dew]

6.5 GB, but there is a gguf version exist with much lighter space requirement. 1.49 GB to be more specific with Q4_K_S.gguf ending. That one is only NSFW to use, but SFW to search for because it can be found on a Hugging Face page.

really depends

yep Miao Miao Harem is an explicit checkpoint

uncertain it will work if I try to make an SFW checkpoint here and bring it into working order since I don't know for certain if I solve this problem it gonna refer to the rest of the IL checkpoints I'm planning to use or not

@ornate elk Okay here's the thing:

Stuff installed there:

Entire cmd from top to bottom:

[ornate elk]

Hey, can you try using a normal safetensor instead of a gguf file?

[jade dew]

I've been doing that same results.

[ornate elk]

Can you try it again? And share the cmd log again? Just mark everything, copy and paste it in here as txt. File

[jade dew]

ok

I'll see what I can do

[ornate elk]

My second guess is that nunchaku and the other optimisations are not compatible with your gtx1650.

[jade dew]

that would be a hellish pit news

I don't get why it would spit the models but run for 33 seconds

[ornate elk]

The good thing is that the normal Forge version supports your GPU

[jade dew]

here we go using Civitai checkpoint

now for the txt

here it is

@ornate elk

[jade dew]

So I should install simple Forge webui instead of Forge Neo?

[near dome]

@ornate elk

[ornate elk]

Yes that should work without issues

[jade dew]

Okay because I did install that.

Can you tell please

that I can use IL loras or checkpoints with simple Forge?

[ornate elk]

Yes that works

[jade dew]

thank god finally

[ornate elk]

Yes you can. But you can't recude the vram usage

[jade dew]

Can I enforce my Forge to use lowvram launch mode?

Forge Neo kinda spitted it out with an error without the interface booting up.

[near dome]

How

I need max quality, large prompts and 5 Lora’s

But only 8 gb GPU..

[jade dew]

Then I guess CS1o is about to break it to you there's no hope.

[near dome]

I asked about distribute load

[jade dew]

okay

[near dome]

Is not possible?

[jade dew]

How I supposed to know as a green eared shit?

I'm just a tryhard having lucky guesses

Someone help me I cannot even send the engine log in txt it doesn't appear!

@ornate elk I got an error doing the xl thing

[ornate elk]

Have you installed forge the correct way?

Like stated in the linked guide of this channel?

[ornate elk]

Idk exactly what you mean with "distribute for a long time"
To reduce vram you need to use lower resolution and then upscale after that

8GB can do sdxl images with 5 loras

Go for 832x1216 for Illustrious models

[jade dew]

yes I did

[ornate elk]

Okay good

Forge doesnt use --lowvram or --medvram as it does it automatically

[jade dew]

I just got a blue death screen for less

[ornate elk]

How much RAM do you have?

[jade dew]

8 GB RAM

why it's not enough?

[ornate elk]

Yep

Either get more RAM or increase the Windows Pagefile for a temporary fix

[jade dew]

oh the one thing that maximizes RAM usage in control panel?

I already done that

[ornate elk]

The pagefile setting for the C drive

[jade dew]

Can you tell how to do that at the case of Windows 10?

[ornate elk]

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set it to 16000 Min and 24000 Max for C
Then apply and reboot the PC.

[jade dew]

@ornate elk I already remember I made the settings to it's allowed maximum before in Megabytes.

[ornate elk]

Thats wrong

Its set to 8GB fixed which is to less

[jade dew]

for my background space it's not enabled

but that's what being used as storage for everything else

[ornate elk]

Set it to 16000 min and 24000 max for C

Then reboot

[jade dew]

isn't it gonna kill my PC?

[ornate elk]

No it makes everything faster

[jade dew]

okay

if the system says free space for C: to use is 37575 MB

[ornate elk]

37gb

[jade dew]

does that mean this border being set will be okay?

[ornate elk]

Yes you just need to have at least 24GB free space on C

[jade dew]

I don't know what way I can free up more space at the case of not trying to format C drive as it's a 128 GB SDD.

but I'll just boot up the system that way

[ornate elk]

Its okay

[jade dew]

alright

I pressed okay but jumped back to 8192 MB

oh okay

I just have to press that setting button to apply what I changed.

I reboot

@ornate elk Okay

Is everything alright with this?

should I give it a shot?

[jade dew]

@ornate elk Considering the long time struggle, the resolution minimalism, the prompt processing speed and struggle and her hairpods ended up being black after all this I think I'll go back to Google Collab. Because this feels dangerous.

[ornate elk]

okay, but it worked 😄
the 1650 is just to old/weak for ai stuff

[jade dew]

it worked but I definitely heard my GPU in it's unusual sounding

and I pressed the generate button and like almost 2-3 minutes passed before the generation finally ended

[ornate elk]

yea thats normal, the gpu is used to the max and probably makes a coil whine noise

[jade dew]

soo...

if I want to generate more daring pictures my GPU might be fried

because this is not what I generate in Google Collab

but much higher stuff

since I can connect to a virtual T4 GPU there

[ornate elk]

nope, the gpu should handle that well, i used a gtx1080 a year long and my new gpu freqently too , no issues

[jade dew]

my only problem with that is that is limited

so I have like...2 hours total to make a rushed work or something

and I still have to inpaint

[jade dew]

I see it works

but will my PC even handle if I go for "masterpiece/best quality" prompts with 35 sampling steps, with 6 CFG Scale, 896x1152 image size Euler a sampling method with Automatic Schedule type, crazy NSFW prompts

all this in XL ui

a technically IL compatible checkpoint lora

and technically IL compatible character lora

without I have to see blue death screen ever again?

[ornate elk]

yes it should

[jade dew]

why it went blue death screen before?

[ornate elk]

hard to guess, they mostly show an error code what it was related too

i guess out of RAM xD

[cosmic furnace]

Btw what video-gif related things forge neo can & cant run ?

[jade dew]

should I park control before every single generate run?

[ornate elk]

Wan 2.2

[ornate elk]

park control?

[jade dew]

a tool to clean my RAM before doing anything

[ornate elk]

nope, you should be fine now with the adjusted pagefile

[jade dew]

okay

[jade dew]

@ornate elk Thank you everything works fine. It's slow as FUCK, but it's working and I just have to sit and watch my favourite series with a sack of popcorn because it's that slow.

At least it's limitless unlike Google Colab.

[ornate elk]

no problem 🙂

[cosmic furnace]

Btw, Gemini suggests that the need for such a huge page file (and more than 32 GB of RAM) is likely due to a memory leak. I mean, 20 GB of VRAM plus 32 GB of RAM really should be enough for almost everything, right? Also, is it safe to use an SSD for the page file? I mean, won't it kill the SSD much faster?

So, I guess it's better to close and restart SD from time to time? Or is it normal for it to require more than 32 GB of RAM?

[lucid yoke]

depending on the model

theres huge ones (20gb+)

[cosmic furnace]

hmmm. So if it happens when using flux its okay? But im pretty sure it happened when using Illustrous for very long. not remember exactly but size of pagefile i saw was about 30gb-ish . Too much id say -_-

[ornate elk]

sometiumes it takes 30gb ram when loading stuff, but not when generating

only maybe with flux or chroma etc

[ornate elk]

yes its safe to use ssd pagefile

[cosmic furnace]

hm. this is weird. Civitai says u put embedding into root folder of sd. but i can see this folder in models folder of forge neo. So, where do i put em to work? 🤨

i guess where the folder is alrdy? models?

[ornate elk]

Models/embeddings

[cosmic furnace]

Id say indeed impressive generation speed with therock + forge neo. 1320-1712 image in 12 sec or mb even 11 🤔 , probab even 10 to 11 👀

https://www.youtube.com/watch?v=X5S0OjTq9kY

[cosmic furnace]

forgot to enable tiled vae & crashed 🙁 how do i do it? like this ?

[cosmic furnace]

Nah. Not sure but zluda felt more stable. NO freeze-stuttering. And no long finishing times to times,this or im doing smthing wrong

[cosmic furnace]

Does this neo forge behaves same as zluda ? Because as i said. Feels like heavy overload GPU when starting-ending,also finishing takes too long, Smthing is fishy or wrong

[cosmic furnace]

#🤝｜tech-support message ah i see ...

sooo, it makes image in 6-7 seconds, and finishes something (mb vae) for veeeeery long, in the end making it 80 seconds generation + this heavy stutter freezes of system during generation. Ive never had on zluda+ usual forge -_-

1280-1640 image generation taking ages =\ . mb i shudnt use sdxl.vae16 ? Dunno if it runs smoothly & fast for ya. Mb i shud set it up somehow in settings or smthing 🙁

img2img shud work fine y? 6 minutes for 1280-1640 😄 . Everytime i switch between txt2img into img2img & then into inpaint it takes long for it to start 🙁

[sick merlin]

new error with Forge Neo 😞

and the prompt is all messed up on the ones I successfully generate, is anyone else having this issue?

[sick merlin]

I wanna know if I'm the only one before attempting to reinstall

[sick merlin]

it shows me the image generating then does this

[cosmic furnace]

It generates image very quickly but finishing last 7-8% left always takes ages & stutters. Reminds me of problem with vae. Not sure

[sick merlin]

completely different issues then

is it successfully generating the image for you?

[cosmic furnace]

yeah =\

btw i can see it does memory clean up every time. taking approx 2.5sec. while kinda useful. It also in theory takes more time to generate an image

w/e i think well have to wait until CS1o comes, or whoever also helps here,besides him 😑 , or post in reddit, but dont feel like it yet

again stutters, & 3 mins to create an 1280-1640 img, and all this in one session. Not once like u said in guide -_-

[cosmic furnace]

It goes up until 92% of finishing image and it takes very fast , and it seems when it finalizes it with vae is the final time. -

and its like this every image,

if i uncheck this than yeah. speed & generation time is good

but the problem is, that without it it crashes, mb settings of tiled is bad or need some another version, dunno

and yet again very long launching when switching over to img2img. Zluda didnt had any of this issues, or mb this is forge neo. Dunno w/e . Seems like without tiled vae kinda same happened, but driver crashed on top of it -_-

[ornate elk]

Don't enable multi diffusion. Just enable tiled VAE at the "never oom"

Like stated in my guide TheRock has a bug at the VAE step. So you need to always enable Tiled VAE and sdxl fp16 VAE helps too.

[cosmic furnace]

Yeah when i found how where to enable it in history i did, and this is problems im having . doesnt seem like any difference 16vae or not, weird , First laucnh more was more success , kinda

[ornate elk]

Also its slow at the first time if you enabled new stuff like adetailer or img2img

[cosmic furnace]

any idea what i might be doing wrong?

but i mostly do it in this order txt2img

and then upscale -adjust it into img2img

so if i follow it in this order, ill have this issues? 🤨

[ornate elk]

AFK, driving rn

[cosmic furnace]

oh, ok)

All this feels like. Well zluda was working alot better & stable? 😄 While zluda was made by an fan enthusiast -_-

[cosmic furnace]

even if i simply create image after image in txt2img . without switching anywhere else, it stutters at start & lil bit after & every time it takes alot of time to create this 896-1152 res image

roughly about 80 sec everytime and even when i press interrupt it still takes as much time to finish

well almost , cuz ive insta pressed interrupt

[ornate elk]

Your it/s seem low too

Its strange because I have no such issues where it hangs

Which model have you tried? I see clip l and g mentioned?

[cosmic furnace]

Different kinda of illustrous. Also illustmix(kinda mix between pony & lustrous) unusual one with 6.91gb size

[ornate elk]

Hmm okay but they should work

[cosmic furnace]

y. but something somewhere broke it seems🤔

[ornate elk]

You should share the full txt2img settings for the image where it starts to hang

Then I can retest

[royal briar]

what could i do with a really great GPU system, came across this server to find some use cases

[ornate elk]

You can do image and video generation localy

Also chatbots with llm work good too

[royal briar]

its an rtx 6000 pro ada

96GB

[cosmic furnace]

U mean like this? erm

[royal briar]

are there any profitable avenues (not interested in mining)

[ornate elk]

Maybe offer image/video gen or model/lora training for people with lower vram

[royal briar]

where can i find them? i intend to help out (only covering electricity costs) but dont want to go through the hassle of hiveOS or vast setup.

[ornate elk]

Hmm idk :/

[cosmic furnace]

all same...mb some extensions broke smthing? they kinda shudnt

[ornate elk]

Looks okay

[ornate elk]

Make sure your forge neo is updated

By adding git pull to the webui-user.bat

[cosmic furnace]

indeed it wasnt updated o.0

and yet it still sttuters & stuck at 96% of generation and taking this long time to create 896-1152 image

very same 80 seconds -_-

hmmm -
Launching Web UI with arguments: --skip-python-version-check --cuda-stream --attention-pytorch --git pull
ERROR:bitsandbytes.cuda_specs:Could not detect ROCm GPU architecture: [WinError 2] The requested file could not be found
WARNING:bitsandbytes.cuda_specs:
ROCm GPU architecture detection failed despite ROCm being available.

ERROR:bitsandbytes.cextension:bitsandbytes library load error: Configured ROCm binary not found at G:\SD\SD-FORG-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll
Traceback (most recent call last):
File "G:\SD\SD-FORG-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 313, in <module>
lib = get_native_library()
^^^^^^^^^^^^^^^^^^^^
File "G:\SD\SD-FORG-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 282, in get_native_library
raise RuntimeError(f"Configured {BNB_BACKEND} binary not found at {cuda_binary_path}")
RuntimeError: Configured ROCm binary not found at G:\SD\SD-FORG-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll

Does this matter? It still launches it but ... if it matters this is a copy i moved for illust runs, after installing by guide. But i dont think it shud matter 🤨

Yeah probab shudnt adress this problem to AI but. This is what gemini says -
Bitsandbytes Error (Main Problem)
This error is critical and likely prevents the system from running optimally, which may be the reason for slow VAE finalization and "freezing/hangs."

Error Log:
RuntimeError: Configured ROCm binary not found at G:\SD\SD-FORG-NEO\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll

Cause:
The bitsandbytes library (used for memory and speed optimization, especially with quantization) is attempting to use its version compiled for ROCm 7.1 (or a similar version) but cannot find the corresponding DLL file. On Windows, bitsandbytes support with ROCm/PyTorch is very unstable and frequently causes issues.

Solution:
Disable bitsandbytes: In your startup file (webui-user.bat or another), find and remove any flags that might activate bitsandbytes (e.g., --bf16 or specific quantization flags).

Uninstall the library: Temporarily remove bitsandbytes from your virtual environment to prevent it from attempting to load:

Bash

cd G:\SD\SD-FORG-NEO\sd-webui-forge-neo
.\venv\Scripts\activate
pip uninstall bitsandbytes

[ornate elk]

Bitsandbytes is not supported by AMD on windows right now, so can be ignored or uninstalled

[floral snow]

Quick heads-up: ComfyUI just released v0.3.66. Key updates include faster workflow cancellation and a fix for NVIDIA VAE memory issues.

Question for the community: How do you typically handle these updates? We built a one-click launcher that handles dependencies automatically to avoid the hassle, but curious about other approaches.

Release Notes: https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.66
Our Suite: https://github.com/NexusAI-Lab/ComfyUI-Professional-Workflows

[ornate elk]

Hey, Comfyui has an update-comfyui-stable.bat, at least the portable version

[lime adder]

Hello , i just bought new video card 5070 TI and i installed stable diffusion web ui but when i try to generate give me this error "RuntimeError: CUDA error: no kernel image is available for execution on the device Compile with TORCH_USE_CUDA_DSA to enable device-side assertions." .I consultate with chat gpt and says pytorch hasnt compiled cuda cores

[ornate elk]

hey, the basic webui is outdated and doesnt support rtx50 cards nativly.
you need to setup Forge Neo instead.
In the pinned messages here, you find an setup Guide for it. First Link

[pure glade]

guys which are the best semi realism anime models?

[cosmic furnace]

So u didnt had any of this weird issues with time generation & stutters when rock + forge neo ? & got no idea what it cud be ? -_- . mb shud do clean reisntall again or smthing? Or there might be some logs? Or steps can be done to understand it . Dunno mb shud post this to reddit 🤔 . Yet reinstalling again from scratch might work? or nah ? hm

[jade dew]

@ornate elk I'm running out of memory and getting crash errors in mid generating images. Why? Where does my C: drive store additional stuff? I remember I had 37GB when I launched the damn thing. Now I have 28,8GB.

[viscid badge]

Hi, I need help
I recently changed my CG from a RTX4070 to an RTX 4080
I redownloaded Stable Diffusion and got this message when I tried :

RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
Time taken: 1.4 sec.

[ornate elk]

Open up a cmd and run
Pip cache purge
That will get you 5-10GB back

[ornate elk]

Hey, you should setup Forge Neo instead of Auto1111. Its more updated and has a better performance.
A guide for that is in the first link of the pinned messages.

[viscid badge]

I'll do that so

[viscid badge]

It work !

[ornate elk]

Perfect

[jade dew]

thank you

It looks like there are a lot of files been removed with this.

[ornate elk]

Np, yep they are temporary files left over from the install

[cosmic furnace]

where do i download this 16 vae fixed?

if no luck after reinstall again , dunno might just try comfyui with the rock or go to reddit 🙁

[ornate elk]

Here:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

TheRock is still WIP and unstable. But its much faster than zluda.

I generated with illustrious models + hires fix by 1.5 without issues (when tiled VAE is enabled)

Changing the resolution often will cause again a longer gen time because therock will do its first time VAE stuff again.

[cosmic furnace]

hmmm. probab better to wait until its somewhat updated & stick to zluda . i was hoping that hip sdk rocm installing & copying zluda files into zluda folder of usual forge is enough, But seems like no 🙁

Trying to run zluda (faster-simplier steps) on new windows >_> . seems like have to do zluda folder & path - somewhere 😑

[ornate elk]

For forge with Zluda you have to follow my setup guides

[cosmic furnace]

yay. for now if i dont change resolution & stay within txt2img all is fine

first generation i cancelled mb shudnt had. but this might be just temporary dunno...

[jade dew]

So these files are generated each flipping time I use Stable Diffusion locally like that? After each image generation I paid attention to my C: drive's state of space usage and I noticed it decreased overtime.

[cosmic furnace]

yeah when made res higher seems like again it needs some time. this is pretty sad comparing to zluda tho =\

[jade dew]

Well in the control panel's system settings CS1o recommended me to change pagefile on my C: drive to 16K MB minimum and 24K MB maximum and that solved the problem of my workflow crashes unless I don't pay attention to my C: drives space frequently. Don't necessarily listen to me in this, just wait for CS1o's advice at your case.

[cosmic furnace]

896-1156 res in 16 seconds isnt fast id say 🙁

[jade dew]

well it takes me 3-4 minutes lol xD

yep this same resolution

in your case this looks much better

[cosmic furnace]

y had it alot faster with zluda if i remember correct 🤔

like twice faster =\

something seems a bit off with the rock for now 😑

every res changing it needs to pause& adapt

Can i run Forge Neo on zluda ? 🤔

[cosmic furnace]

official cuda adaptation works worse than the one made by an enthusiast =\ & it came out a lot later on top of it. nice. nothing new tho ...

[cosmic furnace]

i didnt get used to go hires fix way. i kinda upscale in img2img -_- & go into inpaint after. this kind of workflow feels better on zluda =\

[cosmic furnace]

Had lots of hopes about this new combo 🙁 . Probab , have to wait for updates =\ . Btw qwen didnt even launched for some reason . Or rock also unoficial enthusiast develop ? & early probab 🤨

[cosmic furnace]

I can see ppl have kinda similar issues

[ornate elk]

No these files are from the first installation of a webui.

[ornate elk]

Its still work in progress, I expect that it will work much better in a few months.
But I'm glad we have now Multiple options. Also zluda can't be used anymore to patch other webuis. I tried to get forge neo and reforge work with Zluda again, no chance. Because it needs some updates.

[cosmic furnace]

896-1152 zluda forge is making it in 9 seconds =\ . almost twice faster, i guess its just because zluda is SD oriented or whuht? While theRock is more global one? It was pretty much better even on earliest stages. y ?

So what are good-best options with newest UI's now on zluda 🤔 ? Just comfy? 🤨

thats because zluda wasnt updated in a while now? or what? And yeah, ive read TheRock is actively developed-updated, so i guess yur right 🙂 . Its a good thing)

[cosmic furnace]

Id ewen cud consider 2x times slower speed, but problem is that every change of resolution (espec to higher) or type of generation(txt-img-inp , etc), kinda resets "optimization" it did, or dunno. Well compiling or w/e hes doing, when taking lots of time for that first gen

[ornate elk]

very strange, for me therock is much faster.
these settings on boath with an illustrious model and sdxl fp16 vae, no adetailer.
Tiled Vae enabled.
Forge with Zluda: 33.2 seconds
Forge Neo with TheRock: 17.2 seconds

[cosmic furnace]

im not using hires.fix -_-

its usually slower & i wanna see what generated first. yeah i cud lock seed & upscale with hires later but , blah, hires.fix usually slower 😒

i get this kinda of res ur showing in about 16-18 sec with forge zluda

that can be true even with long promt and quite a few loras on top of it

what makes me wonder, is . i added paths . installed latest rocm. and forge on zluda doesnt run on new windows. It wont run on newest 3.12 python ive installed for Therock or whuht?
yandere upscaler ? 🤨

[cosmic furnace]

So it seems like uve got Forge neo as of now working with same speed on upscaler than Forge zluda without ? U can say it works faster then 🤔 Weird

[ornate elk]

yea without hiresfix i have this speeds:
zluda: 10 seconds and 3.35 it/s
rocm: 6.8 seconds and 4.71 it/s
for hires fix steps:
zluda: 1.10 s/it
rocm: 1.81 it/s

[ornate elk]

open up a cmd (not in any webui) and run this:
pip uninstall rocm rocm-sdk-core rocm-sdk-libraries-gfx110X-dgpu -y
then try again to launch forge with zluda

[cosmic furnace]

Mb ive yet installed smthing wrong again? Is it me or feels like wer having kinda too much different "speeds" ?

Although id probab better wait for next update. Probab gonna be soon. like in a week or two . Well i hope so)

[ornate elk]

retest and let me know your speeds of the 2 image, because the first one loads the model into ram, that ads time.
also share your it/s

[cosmic furnace]

Is this okay that highest line when launching is this - TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL' is not recognized as an internal or external command, operable program or batch file. ?

[ornate elk]

nope thats a problem

[ornate elk]

the webui-user.bat should look like this:

ignore model-ref, im just pointing to my models folder

[cosmic furnace]

it started installing some s*it and laucnhed lil bit different now o.0 . hmmm)

hmmmm. if it wasnt finishing for vae (96%) that takes long & also makes stuttering. Mb cuz 1st time(hope so). its done 896-1152 in about 8-9 sec

Zluda wasnt making this heavy stutters 🙁

[ornate elk]

yea its the first time for vae, make sure you always enabled tiled vae, if not then the gpu driver might crash and pc freezes + black screen xD

[cosmic furnace]

yah ive had to reset button few times 🙂 . gosh

Yeah for 2nd time same res speeds are kinda same as zluda

but im not allowed to change res or switch types yeah?

[ornate elk]

you can

can you show me your full cmd log?

[cosmic furnace]

Like this? or like text copy & more full ?

u do have higher version of 7900 . But can speeds be that much different between them ?

dang swtiched to higher res in img2img & forgot to turn on tiled

[ornate elk]

You should get at least 3.8 - 4 it/s if torch aotriton works.
Make sure your webui-user.bat has the same launch args (except model-red) and that set TORCH_ROCM... is there too

[cosmic furnace]

I switched to higher res in img2img & its taking ages to start o.0 . is that okay ?And again pauses on 93%

from txt2 into 2img but higher res

[ornate elk]

Yea sounds normal rn

They are still working on the VAE issue. Hopefully they fix in soon

[cosmic furnace]

yeah. its doing its thing like its first time lauching (optimizing vae or smthing) if i stay in this tab (img2img) & gen on same high res speeds are good

well at least kinda same as on zluda

res 1304x1680

u think when they fix vae its gonna be good same-as zluda ? 🤔

[ornate elk]

Yea if they fix the VAE issue there isnt any more stuff that won't work. Maybe they also need to add more specific performance improvements like sage attention or xformers support.

[cosmic furnace]

And btw is it me. Or Forge Neo simplified UI , or even removed something ? 🤨

So i cant run same SD forge from same folder ive had on previous windows? WIth lower Python version & probab lower HIpSDK-ROcm ? I need to do clean install or copy it & adjust changes? WIll zluda run on 3.12 python . And this wud be enough if i say copy thiz zluda forge for this windows? - #🤝｜tech-support message

[ornate elk]

Forge Neo removed a lot of unnecessary stuff from Forge yes.
You can also setup normal forge with my TheRock guide.
Also forge with Zluda should work fine with python 3.12

If you had an old forge install, just delete the venv and .zluda folder and relaunch

[cosmic furnace]

U wanna say in theory i shud be able to lauch this zluda forge from very same folder ive had it running on previous windows ? 🤨

[ornate elk]

Make sure you have git pull in the webui-user.bat, then delete the venv folder and .zluda folder and relaunch the webui-user.bat and it should work

If you get an error when launching forge, I had to install onnxruntime and joblib inside the venv.

[cosmic furnace]

Qwen doesnt even want to run on forge neo - AssertionError: You do not have CLIP state dict! What clip ?

[ornate elk]

For me it works with the gguf clip

[cosmic furnace]

is that the one from flux ? or where do i get it?

[ornate elk]

Nope from here:
https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/tree/main

I'm off now

[cosmic furnace]

Mmkay ). Cya. i guess ill be foine ) . Well hope so >_>

Wait . theres too many files here. Dang xD

[cosmic furnace]

Both installing forge from scratch or deleting venv-zluda result in same error 🙁

[ornate elk]

Which hip SDK version is installed?
Also open up a cmd and run
hipcc --version
And then
where hipcc

Screenshot the results

[cosmic furnace]

latest rocm 🤨 like by the -Therock guide

6.4 smthing (6.4 for sure) , sec - G:\SD\SD - Forge-Webui-ZLUD\333\stable-diffusion-webui-amdgpu-forge>HIP version: 7.1.25410-dc2a92a337
AMD clang version 20.0.0git (https://github.com/ROCm/llvm-project.git a7d47b26ca0ec0b3e9e4da83825cace5d761f4bc+PATCHED:dedaea5f8cda00ff94102bce954ff85b1e8a7b5b)
Target: x86_64-pc-windows-msvc
Thread model: posix

while on prev windows its 6.2 it seems 🤔

[ornate elk]

I have rocm 7.10

Not 6.4

[vast crest]

Hi, sorry to bother you, but I've had problems these days. I changed 3 components on my PC, including the MOBO. And of course I have problems. SD Forge Neo no longer starts, saying that the webui-usare.bat file cannot be found and the path has changed, etc. What do I do?

[lucid yoke]

did you move the folder where it was installed in?

[vast crest]

I changed my motherboard and then installed my M2 SSD on the new motherboard. Now SD Forge Neo won't boot and gives me an error.

[vast crest]

At this point I have to format everything from scratch and reinstall

[rose apex]

How did you install 7.10? Also, is Forgeneo faster than Forge and comfyui rn? ( RX 9060 )

[lucid yoke]

So in short, you moved the install folder?

You can just reinstall forge neo in a new folder and move over the models/loras etc

[ornate elk]

I installed it via theatest rocm nightly command

Forge neo with rocm is faster than any zluda based webui.

You can install rocm into comfyui too

[rose apex]

I did a clean windows install recently and rn trying to install Forge neo but I get this error. I haven't installed HIP 6.4.2 because it wasn't mentioned in the Therock forge neo guide. AMD driver version is 25.9.1

[ornate ruin]

Hi 🙂 what local ui do people run in late 2025? I'm still using a1111

The loads of 'a1111 is outdated' posts is kinda getting to me lol.

[rose apex]

Forge and Comfyui ig. Forge is a really good upgrade from A1111

[vocal burrow]

forge dev is also on break for an undefined period.

[ornate ruin]

Still those two? Argh. I always figured something else would pop up in the future and hesitated on jumping to one of those 😂

Do you know if I should whipe my pc before installing or can I just delete the entire stable diffusion folder and install forge?

Oh

What even would be an advantage of jumping to , I guess then from a1111 to comfy? Is it handling generations faster for example? Is it just interface stuff?

[vocal burrow]

If you have to move to a different UI either go for forge neo if you want to stick to something similar to a1111, go comfyUI their team have been very consistent with their updates, or maybe even sd-next.

[vocal burrow]

comfyui is a very very different workflow.

it is node based, so it gives you more freedom to do whatever you want, there s tons of custom nodes. But you do have to understand what you re doing and it can be slower to iterate through different extensions, models, etc...

In A1111 based UI you basically have everything already installed or can import extension with one click and it will be ready to use.
In ComfyUI however if you want to try an extension you will have to import the node into your workflow, find where and how to connect the nodes input/outputs. Or throw your graph out and get a different one done by someone else.

Both have pros and cons.

speed wise they re about on par if configured correctly (on average, in reality it depends of the setup and hardware, you can screw one or the other to make it slower)

[ornate ruin]

I see, yea I'll probably don't need comfy ui - though it has the ability to create videos too no? Which a1111 unfortunately does not

[vocal burrow]

you can do videos with Forge Neo

But yes ComfyUI tend to have new tech faster than A1111 based UI.

[ornate ruin]

Thank you!

If I want to try something a bit more up to date like forge neo then, do I need to fully delete a1111 or can I just install it on top?

[vocal burrow]

I d say, definitely don t "install it on top". You don t want to create a mess by mixing the two different UI.
But you can install Forge Neo on the side and use the command launch args to make it use the A111 model folder, the A1111 lora folder, the A1111 etc etc

--models-dir "c:\SomePathToA111\models"

time for me to commute back home, see ya

[ornate elk]

In forge neo its --model-ref

[ornate elk]

Oh you need hip SDK 6.4, I forgot to add it to the guide.

Will do when im at home later

[ornate ruin]

I'll have to look into that then, alright. Sounds simple at least - kinda 😄

[cosmic furnace]

But official setup installators can install only 6.4 on windows? no 🤨

[ornate elk]

Hip SDK 6.4 is not rocm 6.4

Hip SDK 6.4 is needed to use any rocm on windows

[cosmic furnace]

Wait but question was about hip sdk not rocm ? 🤔

yeah but thought thats kinda same )

time to time >_>

[tawny trail]

I get the error "Given normalized_shape=[3584], expected input with shape [*, 3584], but got input of size[1, 77, 768]" from KSampler; this is the setup I'm using, which is the recommended preset for Qwen-Image

Does anyone know what's actually happening here?

[rich garden]

I’m trying to use a checkpoint in stable diffusion but it takes forever to load and after it’s done loading, it doesn’t even load in properly in the top left corner of the window. I also got an error saying something like “NotImplementedError: Cannot copy out of meta tensor; no data!” It’s been awhile since I’ve used SD so I’m not sure how to fix it

[high scaffold]

Hi guys, I’m using A1111. Is there a way to merge one element of a picture onto a specific region of another picture?

[ornate elk]

Hey, what's your GPU and how much RAM do you have?

[violet dagger]

I installed SD 3.5 Large but run in error, I think my method is wrong, could some point to the correct way to download and install?

I am using tydiffusion

[rose apex]

Even with that, I am unable to install it. always get this error when i run webui-user.bat

[ornate elk]

It shows the wrong torch version

The cuda one

You need to make sure to enable the venv and then uninstall the cuda torch and install the rocm one

[rose apex]

I did and now it shows this.

[ornate elk]

check the venv/lib/site-packages folder for the torch version

[rose apex]

There are lots of folders inside. How do i check what version i have?

[ornate elk]

it shows torch 2.8.0, still the old cuda version

While in the webuis folder click in the File Explorer bar (not searchbar) and type cmd then press enter.
Now run these commands one by one:
venv\Scripts\activate.bat
pip uninstall torch torchvision torchaudio -y
python -m pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision

[rose apex]

I did this already but let me do it again ig.

Thanks, It worked this time and the webui launched. Really did the installing torch part twice though. There's also this error I got idk if it should be ignored:
Launching Web UI with arguments: --skip-python-version-check --cuda-stream --attention-pytorch
ERROR:bitsandbytes.cuda_specs:Could not detect ROCm GPU architecture: [WinError 2] The system cannot find the file specified
WARNING:bitsandbytes.cuda_specs:
ROCm GPU architecture detection failed despite ROCm being available.

ERROR:bitsandbytes.cextension:bitsandbytes library load error: Configured ROCm binary not found at C:\TheRock\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll
Traceback (most recent call last):
File "C:\TheRock\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 313, in <module>
lib = get_native_library()
^^^^^^^^^^^^^^^^^^^^
File "C:\TheRock\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\cextension.py", line 282, in get_native_library
raise RuntimeError(f"Configured {BNB_BACKEND} binary not found at {cuda_binary_path}")
RuntimeError: Configured ROCm binary not found at C:\TheRock\sd-webui-forge-neo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll

[ornate elk]

bitsandbytes is currently not availble for AMD, so this error can be ignored

[ornate elk]

what does this 5gb contains?

[lucid yoke]

in #💬｜general-chat message he explains it a little more but i dont trust it tbh

[ornate elk]

wow wtf a portable comfyui + workflow... just why

and your right mega and the account seems sketchy

[lucid yoke]

also looking up his post history its only posting about his service/workflow

[ornate elk]

fixed it

[jade dew]

@ornate elk Hello! Is this slider gonna help me with anything?

[ornate elk]

with 4gb vram, not that much

[jade dew]

so should I just leave it the way it is

[cosmic furnace]

#🤝｜tech-support message - so what do i do with this ?

[ornate elk]

Failed to install ZLUDA: HTTP Error 404: Not Found

try delete the .zluda folder if you have one and the venv and relaunch

[cosmic furnace]

Same errors 🙁

[ornate elk]

What's in your webui-user.bat?

Also open up a cmd and run pip cache purge

[cosmic furnace]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --zluda --cuda-stream
git pull

@REM Uncomment following code to reference an existing A1111 checkout.
set A1111_HOME=G:/SD/SD-Zluda/stable-diffusion-webui-directml

@REM set A1111_HOME=Your A1111 checkout dir

@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir G:/SD/SD-Zluda/stable-diffusion-webui-directml/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat

G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK>pip cache purge
Files removed: 1420 (10827.2 MB) 😮

[ornate elk]

you have to change --zluda to --use-zluda

thats your issue why it didnt worked

[cosmic furnace]

Same error -

mb i dont have some visual studio ? o.0

..\meson.build:1:0: ERROR: Unable to detect linker for compiler clang -Wl,--version
stdout:
stderr: clang: warning: unable to find a Visual Studio installation; try running Clang from a developer command prompt [-Wmsvc-not-found]
clang: error: unable to execute command: program not executable
clang: error: linker command failed with exit code 1 (use -v to see invocation)

[ornate elk]

oh, i have it installed

[cosmic furnace]

ugh, what kind of ? where do i get it ? Weird i thought i have it too but...

[ornate elk]

here:
https://visualstudio.microsoft.com/de/downloads/?q=build+tools
visual studio 2022 build tools is the only package needed

[cosmic furnace]

This " meson " error stays & seems like doesnt let me launch or install 🤨

\meson.build:1:0: ERROR: Unable to detect linker for compiler clang -Wl,--version
stdout:

[ornate elk]

very strange

[cosmic furnace]

or i might be missing smthing else needed to launch ? SOme netframework install ?Or smthing dunno -_-

[ornate elk]

i dont think so

[cosmic furnace]

G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK>clang -Wl,--version
clang: warning: unable to find a Visual Studio installation; try running Clang from a developer command prompt [-Wmsvc-not-found]
clang: error: unable to execute command: program not executable
clang: error: linker command failed with exit code 1 (use -v to see invocation)

G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK>

o___o

but i just installed it o.0

[ornate elk]

have you restarted the PC?

[cosmic furnace]

no) do i have to ? 🤨

[ornate elk]

and this in the installed programs list:

[cosmic furnace]

G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK>clang -Wl,--version
clang: warning: unable to find a Visual Studio installation; try running Clang from a developer command prompt [-Wmsvc-not-found]
clang: error: unable to execute command: program not executable
clang: error: linker command failed with exit code 1 (use -v to see invocation)

restarted & its same

well yah ive installed this

[ornate elk]

can you check under installed programms?

[cosmic furnace]

Commit hash: 96ad3a736bd616460a86dc43cd5672acd9017e6f
ROCm agent enumerator failed: [WinError 267] Invalid folder name
ROCm: version=PYTHON_3.12.10
ZLUDA support: experimental
Failed to install ZLUDA: HTTP Error 404: Not Found
Using CPU-only torch
Is this 2 errors okay ? 🤔

[ornate elk]

nope, it installed torch with cpu, you need to delete the venv folder again

also do you have a .zluda folder in your forge folder?

ah this issues here is bad too:
ROCm agent enumerator failed: [WinError 267] Invalid folder name

[cosmic furnace]

yah its tehre in installed

[ornate elk]

open up a normal cmd again and just run this:
pip uninstall rocm rocm-sdk-core rocm-sdk-libraries-gfx110X-dgpu -y
then delete the venv and relaunch the webui-user.bat

[cosmic furnace]

nope no .zluda folder. and ive already deleted venv doesnt seem to help hm

welp it was doing alot of "installing" & ended up same error 🙁

[ornate elk]

can you show the full cmd log?

[cosmic furnace]

U mean this one?

[ornate elk]

your python install path has . in its folder name...

that will cause issues

never use any special characters other than _ or - in folder names

[cosmic furnace]

what do i do now? Can i just rename it or need to reinstall ?

[ornate elk]

reinstall

[ornate elk]

also you have to do something in your visual studio build tools too:
launch the visual studio build tools and make sure
Desktop development with C++ is selectet (checked/installed)
Also make sure that MSVC v143 (or newer) and Windows 10 SDK is checked

After doing this delete the venv folder again and relaunch the webui-user.bat

also please install python in its default path on C, it doesnt take that much space

[cosmic furnace]

hmgm. well previous windows it worked. and seems like made same mistake i have empty folder with dots there

yeah i guess didnt checked what needed when installing -_-. thx for help)

[ornate elk]

np, did it launched now?

and yea better dont use dots, while it can work, it also can cause issues that can easily be avoided

[cosmic furnace]

like this y ?

Installation Details

▸ MSBuild Tools

▾ Desktop development...

▾ Included

• Core build tools components...
• Redistributable updates...
• Basic components for desktop...

▾ Optional (or Additional)

• MSVC v143 — VS 2022 C++ x64/x8...
• Windows 11 SDK (10.0.26100.6...)
• CMake C++ tools for Windows
• Core components of testing tools...
• C++ AddressSanitizer
• vcpkg package manager
  All this ? 6.7 gb space o.o

[ornate elk]

Yep its needed for building python packages that aren't available for python directly

Like insightface for faceswap for example

[cosmic furnace]

Is that okay that ive installed windows 11 sdk while having 10 ? Its what it offered automaticly 😊

[ornate elk]

I guess thats fine. But you should upgrade to Windows 11 soon as the official Win10 security update support ended last week.

[cosmic furnace]

Didnt they promised to keep updating it until like... 28 or smthing ? 🤨

[lucid yoke]

Win10 has a year of additional security updates but loses all other support

Businesses backtracking on promises arent something new sadly

[lucid yoke]

Wait thats europe only iirc

[tawny saddle]

cheers to the ones that we got, toasts to the ones that we lost 'long the way as the dreams bring back all the memories, and the memories bring back memories bring back you.. 🎵

[ornate elk]

And only if you login with an MS account

[ornate elk]

Nope it ended on 14 October 2025.
Same goes for Office 2016 and 2019

But they let you pay if you want for 1-3 additional years of updates...

[cosmic furnace]

And now it gives different kind of error o.0

[ornate elk]

Ah the onnxruntime error

I got that too a few days ago. You have to open up the cmd and run the venv activate command again.
After that run pip install onnxruntime

[ornate elk]

There Is also a strange error caused by an extension that seems to be not compatible with forge.

[cosmic furnace]

Y i have some of those. Is that bad or can be ignored ?

[ornate elk]

I would remove it because it tries to install stuff at each launch and then fails

[cosmic furnace]

Btw have u tried inpaint anything ? Is it working for ya ? Well paint over masked image, Scratch painting or sketch ? Forgot it called

[ornate elk]

I have only tried Inpainting, that worked

[cosmic furnace]

like this y ?

and i still get this o.0

[ornate elk]

Hmm then try to add --skip-ort to the launch args of the webui-user.bat and then delete the venv again and relaunch

I need to go off now.
If you still get the same error then try use pip install optimum-onnx
But make sure you run the venv activate bat command before again.

[cosmic furnace]

main()
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\launch.py", line 50, in main
start()
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\modules\launch_utils.py", line 675, in start
import webui
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\webui.py", line 21, in <module>
initialize_forge()
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\modules_forge\initialization.py", line 68, in initialize_forge
from modules_forge.bnb_installer import try_install_bnb
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\modules_forge\bnb_installer.py", line 1, in <module>
import pkg_resources
File "G:\SD\SD - Forge-Webui-ZLUD\NEW WINDOWS-RMK\venv\Lib\site-packages\pkg_resources_init_.py", line 2172, in <module>
register_finder(pkgutil.ImpImporter, find_on_path)
^^^^^^^^^^^^^^^^^^^
AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?

this modules and their errors. whats weird is that somehow this "clean" version somehow also remembers extensions ive had installed on prev win version, or ive tried launching that different folder & old windows sd forge on zluda & it wrote that file somewhere hmgm

[cosmic furnace]

Root Cause:
Python 3.12 Changes: The ImpImporter class, part of the pkgutil module, was removed in Python 3.12. Older packages or libraries that relied on pkgutil.ImpImporter (often indirectly through setuptools' pkg_resources module) will encounter this AttributeError when run in Python 3.12 or later.
Solutions:
Update setuptools: The most common and recommended solution is to update setuptools to a recent version. Modern versions of setuptools have been updated to use importlib instead of pkgutil.ImpImporter, resolving the compatibility issue.
Is what gemini says. so i do need to downgrade to python 3.11 from 3.12 or whuht ? 🤨

https://stackoverflow.com/questions/77364550/attributeerror-module-pkgutil-has-no-attribute-impimporter-did-you-mean -not sure this fits my case hm

[cosmic furnace]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/14865

[ornate elk]

Okay, yea seems like forge only supports up to python 3.11.9 on windows

[tawny saddle]

why they call it python anyway

wut about it is snakey, unless their implying noodle code, in which case thatd mean spaghtti code, and ive seen ghetti code 👀

[maiden coral]

I think it's still better to stick with Python 3.11.9, the newest version has tons of incompatibilities.
The solutions are:
create a new environment, use Miniconda if you want or just the basic env command and create environment which you'll activate and start Automatic111 through

[cosmic furnace]

Can i have 2 versions of python installed? 😄 Guess no ? 🙃

[lucid yoke]

its not recommended no

you can use a venv or anaconda/docker but that means having some knowhow on how to set it up

[ornate elk]

you can but then you need to link the correct python path inside the webui-user.bat

but i would downgrade and uninstall 3.12 as you have no advantage of it currently

[summer ember]

Can someone please help me with vram issues on forge webui?

I'm using a 2060 with 6 GB VRAM but I had no prior issues to generating last week

Now it's crashing whenever I try to generate

I've done a reinstall, downgraded my windows and nvidia drivers, added command launch arguments and no luck

[ornate elk]

Can you share the complete cmd log?

And how much RAM do you have?

[viscid badge]

hello,
I tried to install stable diffusion (sd-webui-forge-classic) to a friend but I keep get this message :

[Memory Management] Current Free GPU Memory (MB) = 14709.159549713135
[Memory Management] Model Memory (MB) = 1639.4137649536133
[Memory Management] Minimal Inference Memory (MB) = 1024.0
[Memory Management] Estimated Remaining GPU Memory (MB) = 12045.745784759521
Moving model(s) has taken 0.17 seconds
0%| | 0/20 [00:00<?, ?it/s]CUDA error (C:/a/xformers/xformers/third_party/flash-attention/hopper\flash_fwd_launch_template.h:188): invalid argument
Appuyez sur une touche pour continuer...

[ornate elk]

Hey, what's your GPU?

[summer ember]

hi i just woke up and was told this was the issue:

i will post my CMD log soon but I tried what was in the image by sliding the GPU weights but it gave me the same error

[summer ember]

i have 16 gb ram

[ornate elk]

Make sure your Nvidia driver is updated.
Then reinstall forge like shown in the guide from the pinned messages. You could also install Forge Neo as its more optimised

If you then still get errors share the full cmd log here again.

With 16GB ram you also need to increase the Windows Pagefile for C.

[maiden coral]

Hello. I'd like to ask about Wan 2.2, I've got this error (there is a whole thread for it):"Given groups=1, weight of size [1536, 32, 1, 2, 2], expected input[1, 16, 13, 60, 90] to have 32 channels, but got 16 channels instead"
https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/discussions/3 or here https://github.com/comfyanonymous/ComfyUI/issues/9596

I confirmed that I don't have the flow2 nodde mentioned in the topic. The dependencies had to be edited since ComfyUI was screamin at me with incompatible CUDA (since I use legacy GPUGTX 1060).
In the embedded environment I have Python 3.13 (I guess that's for ComfyUI), I use the newest ComfyUI version that was released 5 days ago (portable Nvidia)
Pytorch 2.7.1+cu118 using CUDA 11.8

I'm trying out Wan 2.1 VAE to see if it fixes the problem. I use Wan 2.2. UNET model 2-bit variant image to video model (the lowest version that has around 5GB)

Thank you.

[cosmic furnace]

Sup, so whats best to run on zluda for now? With simple & pleasant UI . Still forge? or ?

Also have u tried running inpaint(segment?) anything extension ?

[maiden coral]

The problem was solved using Wan 2.1 VAE, it's also written here (had to dig it up, really confusing why 14B model needs 2.1 VAE). Now I have a different problem, I see lines in my output a bit, it's like looking through some sort of filter, but that may be configuration. https://comfyanonymous.github.io/ComfyUI_examples/wan22

[ornate elk]

Forge then

[viscid badge]

RTX 5070 Ti

[ornate elk]

Then you should install forge neo instead of forge classic. Because neo supports rtx50 nativly and has a much better performance

A guide for that can be found in the pinned messages here

[prisma moss]

neo is slower in my case compared to normal forge (non classic)

for sdxl stuff at least

[ornate elk]

hmm thats strange

[prisma moss]

like, by 8 seconds

per image (hrfix)

[ornate elk]

which launch args in boath?

[prisma moss]

--cuda-malloc

[ornate elk]

only this?

[prisma moss]

yes

[ornate elk]

yea then it makes sense that its slow

[prisma moss]

it doesnt lol

unless neo fights it

[ornate elk]

normal forge needs only --cuda-stream
and Forge Neo needs: --cuda-stream --flash --fast-fp16 --xformers

for the best performance

[prisma moss]

cuda stream on normal forge is slower for me

[ornate elk]

wouldnt recommend --cuda-malloc as it doesnt help in speed much

[prisma moss]

for comparison:

1mpx+hrifx on normal forge: 15s

lemme launch neo

[ornate elk]

use the launch args from above please with neo

[prisma moss]

not yet

Neo 1mpx+hrfix 22sec

ill try ur args

[ornate elk]

have you also compared the second image gen time, because model to Ram loading doesnt count to the image process

[prisma moss]

yes

thats the 2nd gen

after the first one has loaded n shit

u syain neo doesnt come with built in memory managment like normal forge and has to rely on xformers?

[ornate elk]

it does come with built in stuff, but not with xformers or sage or flash enabled.
xformers is only applied to the vae loading time as flash and sage are faster for the initial image

[prisma moss]

yea so neo is still slower by 5-7s

with ur args

it simply hangs more when going to hrfix mode

[ornate elk]

very strange

which settings do you compare? (resolution, steps, sampler, hires steps, upscale by etc=

[prisma moss]

all r same

[ornate elk]

i know...

but which values

[prisma moss]

1mpx, 25 steps, 3cfg

10 hrfix steps, remacri

0.35 denoise

[ornate elk]

okay

and your forge neo is updated? git pullinside the webui-user.bat?

[prisma moss]

sec

uh what

are these already enabled if here

or is that a comment

[ornate elk]

nope

[prisma moss]

how do i update

[ornate elk]

add git pull to the second line of the bat

below @echo off

[prisma moss]

i dont think it updated anything so im up tp date

[ornate elk]

does it shows already up to date in the first line of the cmd after launching?

[ornate elk]

can you screenshot the bottom of your webui? where it shows the version number

[ornate elk]

okay looks good

[prisma moss]

besides the fact thats its slower

[ornate elk]

have you tried now with the launch args added in the correct line?

[prisma moss]

since i have coilwhine, i can physically hear my gpu sounds different using neo lol

[prisma moss]

^

[ornate elk]

can you share the full cmd log?

[ornate elk]

thats not the complete one

[prisma moss]

well thats what it generated

lol

[ornate elk]

you can scroll upwards while masking everything

then ctrl+c

[prisma moss]

(malloc is enabled cuz im going back n forth trying different args)

[ornate elk]

ohh your not using an rtx50

[prisma moss]

no

[ornate elk]

then change --flash to --sage and remove --cuda-malloc

then save and relaunch

[prisma moss]

sec

[ornate elk]

that should make a difference now

[prisma moss]

lets see

[ornate elk]

also your python is outdated, would suggest updating python to 3.11.9 as its the latest available and fixed version of 3.11 for windows

[prisma moss]

still as slow as the rest

[prisma moss]

u mean system wide python?

thats sorted at least

[ornate elk]

you have to delete the venv folders of the webuis then so they get their updated packages on relaunch

[ornate elk]

hmm well, my guess is then the new performance stuff cant to much to sdxl, but they should improve the speed alot on flux, qwen, wan etc

[prisma moss]

do i just run the webui-user again to regenerate it

for fucks sake how do i stop the cmd from scrolling down

i want to screenshot

so what does this mean

i clearly just installed 3.11.9

[ornate elk]

do you have 3.10 also installed on the system?

[prisma moss]

i think

[ornate elk]

then its using the wrong one

best is to just uninstall 3.10 as everything now works with 3.11.9 anyway

[prisma moss]

whats the corrrect way o unsinstall it

[ornate elk]

system settings -> apps then sarch the python and hit uninstall

or via system control panel, program and features

[prisma moss]

aight

aight i guess its using 3.11.9 now

im going off now

ttyl

[ornate elk]

alright np and cya

[fair oxide]

hey, installed sb, everything works but it doesnt generate images, Im getting this message, anyone knows the solution?

[ornate elk]

hey, whats your gpu?

and pls share the full cmd log too

[fair oxide]

hold on, I wanted to give you screemshot cmd and now it says of lack of GPU memory

[ornate elk]

okay, and whats your gpu model?

[fair oxide]

Its rx 6600 XT

[ornate elk]

oh then you should not use directml

you need to follow the Automatic with Zluda or the Forge with Zluda setup Guide from the Pinned messages first link

[fair oxide]

ok will do, just to know, do I need to uninstall sb and webui for now?

[ornate elk]

yep, as you only need the new correct version for AMD

you find the setup Guides here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
Feel free to ask if you have any question

[fair oxide]

sure I will reach out tomorrow eventually after work, thank you ❤️

[ornate elk]

alright, no problem 🙂

[dusty sand]

Hello. I have a weird Issue. I don't know where to ask so I hope here is okay. I have been using Forge since May. I had no issues whatsoever. Until a few days/weeks ago. Suddenly Adetailer with BBox models is not as detailed anymore for faces/eyes (iris) and segmented face detailer don't work at all anymore (produce image as if no face detailer was used at all). My friend had the same issue and I found a reddit thread from 3 days ago with multiple people facing the same issue as well. No one has found a solution so far. My system: 9950x3D 5080 64GB Ram. My friend has a 2060 and now has the same issue.

Were there any reports here? is this known? Or is this a freak accident for some people?

I tried reforge etc and same thing.

[prisma moss]

try removing the config file

(back it up)

[summer ember]

I bought a 5060 TI and installed it and trying forge neo but trying to generate anything crashes it immediately

[untold flame]

Hey guys, has anybody else noticed SDXL being slower than ever lately? I remember I used to get like 4.1it/s on my 3090, now I can't even hit 3it/s anymore. I thought comfy recently got an update that was supposed to speed up image gen

Any ideas? I use comfy, and it never used to be this slow for me

even SD1.5 seems slower. I get like 13it/s at 512x when I used to get like 20

still been using SDXL cause SD3 is still really bad at a lot of stuff, but I feel like its just gotten slower for some reason

oh shit, its my monitors

Went down from my 4 monitors to 1 and went from 2.9it/s to 4.31it/s instantly

[untold flame]

ok, with some changes to my power limit, the removal of the problem monitor, a change in samplers, steps, and schedulers, I am down from 11.2 seconds per image to 4.6, so a huge improvement 😅

[dusty sand]

I actually completely reinstalled already with windows full reinstall and Stability Matrix reinstall and Forge solo reinstall. Same issue. My friend didn't touch his Forge install for 2 months. When I asked to check his gens suddenly were messed up like mine even though he changed nothing. Then all the other people with a similiar problem on reddit? It's so odd. I appreciate the help.

[maiden coral]

something might be wrong with you PC, I use the newest ComfyUI and it's fine

[untold flame]

I ended up figuring it out thankfully

now my speed is much better

[cosmic furnace]

@ornate elk Weird . when im generating for an hour or 2. iget this error at some point. Well its mb useful to restart SD at that point. TO clean out cash-memory or smthing. BUt still o.0 . If it can be avoided , probab shud 🤔 . I dont rememb if i installed correct torch or deleted it. after changing python. Only venv folder deleted

[maiden coral]

Have you tried ConfyUI? They have a specific instructions for Nvidia GPUs. Check if the series of your GPU match on the ComfyUI github

I rember I had a problem with Automatic111 which I couldn't find a solution to, I ended up going for ComfyUI.

Forge is just a newer fork of Automatic111.

Also what resolution are you using (+ image or a video?)and what's your GPU?

[ornate elk]

No clue about that, haven't let it run for an hour lately

[ornate elk]

I also noticed this behaviour lately, but only with some adetailer models.

[storm knoll]

Did all the node colors in comfyui break with the recent manager update for anyone else?

it doesn't matter but like damn...

need my organization man

[maiden coral]

Just zoom in

I see nothing wrong with it. It's a normal behaviour.

[storm knoll]

nah like they're broken broken

i can't color nodes they are all just black

everything works its just the color

[maiden coral]

Did you use update bat or update stable? I heard you shouldn't use update_stable.bat file but idk, I just redownloaded comfyui and moved nodes etc

[storm knoll]

no i updated comfyui manager in the node manager

[maiden coral]

Do not use that next time.

The right way should probably ve through the bat file you see in the root folder of ComfyUI portable version. I never dealt with the installer.

But the portable version is the way to go if you don't know which one to choose. Also the embedded python folder contains the environment xomfyui uses, so you can just reinstall things that don't fit your GPU.

Check what the console tells you in ComfyUI cmd that opens when you run it.

The current version of ComfyUI uses CUDA 12+ and newest pytorch which is incompatible with my Legacy GPU so I just went into the environment and used the python.exe from there to reinstall pytorch, pytorch vision and audio libraries.

[dusty sand]

So another friend of mine has the same issue. All segmented models are not working. One other friend has it working. There is no reason or rhyme why it's like this. It came out of nowhere.

https://www.reddit.com/r/StableDiffusion/comments/1oc3410/i_reinstalled_forge_and_reforge_and_now/

here are more people

with the same issue

[maiden coral]

Check the version of the requirements and compare with the ones you have installed in your python environment, they list out pretty good troubleshooting ideas

And I had lots of problems before in order to run some nodes they needed specific version range and so on.

Idk if forge uses its own embedded environment or the one you install or if you use your own environment you created

[dusty sand]

I mean its multiple people popping out of nowhere. Some haven't touched their installs until I asked them to try only for it to now not work anymore.

Forge wasn't updated in quite some time no? Neither was Adetailer? Or am I missing something.

I tried various versions.

One guy in the thread went above and beyond.

testing out different versions.

Python etc.

Checking requirements.

[maiden coral]

Doesn't matter. Multiple people had the same issue with Wan and it turned out Wan 2.2 14B needs 2.1 VAE which was listed as an instruction on obscure comfyUI tutorial webpage

Does forge use embedded python environment or it uses yours?

[dusty sand]

I would agree witht his if it wasnt the literal same install that worked 4 weeks prior.

without touching anything.

[maiden coral]

If you didn't input the exact same commands then it might not work the way you expect

[dusty sand]

Forge uses its own. The Ilyasviel one.

Just like Stabilty Matrix

I tried all of them.

Forge

Reforge

A1111